JP2014006566A - Built-in intelligence controller, control system, control program, recording medium, and control method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a built-in intelligence controller by which position control and force control can be smoothly switched.SOLUTION: An ICS unit 50 performs position control by a position command, and force control by a force command to a control object. A position command system 51 is a fuzzy neural network that performs position control. A force command system 52 is a fuzzy neural network that performs force control. A switching unit 135 changes output of the force command system 52 according to a value of a force sensor of the control object, and continuously switches the position control and the force control by hybrid/compliance control of position and force.

Description

  The present invention relates to an embedded intelligent controller, a control system, a control program, a recording medium, and a control method, and more particularly to an embedded intelligent controller, control system, control program, recording medium, and control method for performing position control and control.

In recent years, robot manipulators are used in various work environments such as the space probe “Hayabusa” and medical sites. In addition, automation and autonomy of robots are being promoted, and high-precision control is required. However, there is a big problem in realizing such control. This is because the dynamic characteristics of the robot manipulator have non-linearities such as inertial load due to position and posture, centrifugal force due to velocity, and gravity load, as well as fluctuations caused by disturbances due to environmental conditions and external factors.
Conventionally, PID control is known as a general control method. PID control is feedback control using P (Proportional) operation, I (Integral), and D (Differential, Derivative) operation. That is, in the PID control, a proportional calculation, an integral calculation, and a differential calculation are performed on a control deviation that is a target value-control amount, and an operation amount is calculated and controlled.
PID control is classical control, but is widely used because of its ease of handling and accuracy. However, classical control such as PID control has been difficult to cope with disturbances and nonlinear elements.

Therefore, conventionally, the system design related to manipulators is generally a method of constructing an independent control system for each joint axis by considering the control target as a one-degree-of-freedom system and designing the control system using an approximate expression. Used in many products.
However, in general, a manipulator is a non-linear multi-input multi-output system that has low rigidity and receives gravity and centrifugal force. In addition, the conventional method achieves accuracy, vibration suppression, robustness, etc. using phase compensation control in PID control, optimal servo system, control using disturbance observer or H∞ control, etc., but sufficient control performance Was not obtained.

Here, as a conventional PID adjusting device, referring to Patent Document 1, PI is set so that a deviation between a control amount PVn from a controlled object and an effective target value SVn0 obtained by performing a compensation operation from the target value SVn becomes zero. Alternatively, in an adjustment device that performs PID (P: proportional, I: integration, D: differentiation) adjustment calculation and applies the obtained adjustment calculation signal as the operation signal MVn to the control target, the effective target value is calculated from the target value of the controlled variable. In order to obtain, a compensation operation means using a proportional gain 2 degrees of freedom coefficient α (a constant between 0 and 1) and an integration time 2 degrees of freedom coefficient β (a constant between β and 1 to 10) is provided, A two-degree-of-freedom PID adjusting device in which a transfer function Csv (s) between a target value and an operation signal is Kp {α + [1 / (TI · s)] · [(1 + TI · s) / (1 + βTI · s)]} However, Kp: Proportional gain, TI: During integration , S: what is the Laplace operator it is described (hereinafter, referred to as prior art 1.).
According to the prior art 1, the two-degree-of-freedom of the integration operation for the two-degree-of-freedom of the PID control is made very simple, and the two-degree-of-freedom coefficient α of the proportional gain and the two degrees of freedom of the integration time By providing the compensation calculation means that can remove the mutual interference of the quantization coefficient β, (1) the two degrees of freedom of the integration operation can be greatly simplified and can be realized in an easily understandable form. (2) It is possible to eliminate the involvement of the proportional gain 2-degree-of-freedom coefficient α in the 2-degree-of-freedom integral operation, and the adjustment becomes simple. (3) Furthermore, the involvement of the two-degree-of-freedom coefficient β in the integration operation is only the denominator, and it is simple and can be intuitively grasped when the two-degree-of-freedom coefficient β is adjusted.

Japanese Patent Application Laid-Open No. 07-072901

However, when disturbances such as centrifugal force, gravity, and Coulomb friction are conspicuous, robust control has been performed using an observer that estimates state variables from input / output signals. It wasn't possible.
In such conventional robust control, the gain, which is a parameter related to the steady state characteristics, is fixed, so that it is not possible to cope with the influence of sudden disturbance, and the occurrence of problems has been a problem. This was the same in the prior art 1.
On the other hand, the assembly work of the industrial robot involves contact between parts when the parts are assembled, and thus cannot be handled by position control, and force control is required.
However, since it is difficult to adjust force control parameters, switching between position control and force control cannot be performed smoothly.
For this reason, there has been a strong demand for a device that switches between position control and force control practically while performing intelligent (intelligent) control.

  This invention is made | formed in view of such a condition, and makes it a subject to eliminate the above-mentioned subject.

The embedded intelligent controller of the present invention is a position control system for performing position control on a control target, and in the built-in intelligent controller performing force control based on a force command, a fuzzy neural network of a position control system that performs the position control, A force control system fuzzy neural network that performs force control, and the output of the force control system is changed according to the value of the force sensor to be controlled, and the position control and force control are performed by position / force hybrid / compliance control. And a switching unit that switches continuously.
The built-in intelligence controller of the present invention comprises feedforward means for learning parameters of the displacement, speed, and acceleration of the controlled object by a neural network in the position control system, and acquiring a reverse characteristic of the controlled object characteristic, The force control system includes feedforward means for learning the parameters of the force, viscosity, and rigidity of the controlled object using a neural network and acquiring a reverse characteristic of the controlled object characteristic.
The embedded intelligent controller according to the present invention includes a fuzzy neural network of the position control system and a fuzzy neural network of the force control system each including a plurality of feedback means which are compensators for reducing nonlinear errors, and the feedback means Comprises in parallel a fuzzy PD means for performing PD (proportional, differential) compensation of the nonlinear dynamic characteristic of the controlled object and a fuzzy PI means for performing PI (proportional, integral) compensation of the nonlinear dynamic characteristic of the controlled object. It is characterized by.
The embedded intelligent controller of the present invention is configured such that the feedback means of the fuzzy neural network of the position control system and the feedback means of the fuzzy neural network of the force control system correspond to position and force errors in a fuzzy divided space. A variable feedback gain is set.
The embedded intelligent controller of the present invention is characterized in that the fuzzy neural network of the position control system and the fuzzy neural network of the force control system perform real-time processing with a predetermined sampling time.
The embedded intelligent controller of the present invention is characterized in that the position control system fuzzy neural network and the force control system fuzzy neural network perform software PWM control.
A control system according to the present invention includes the embedded intelligent controller.
The control system of the present invention includes the embedded intelligent controller and a robot including a force sensor as the control target.
The control program of the present invention is a control program for an embedded intelligent controller that performs position control with respect to a controlled object by a position command and performs force control by a force command. The embedded intelligent controller includes a fuzzy neural network of a position control system. The position control is performed, the force control is performed by a fuzzy neural network of the force control system, the output of the force control system is changed according to the value of the force sensor to be controlled, and a hybrid of position and force / The position control and the force control are continuously switched by compliance control.
The recording medium of the present invention is a recording medium in which a control program of an embedded intelligent controller that performs position control by a position command and performs force control by a force command is recorded on the control target. The position control is performed by the fuzzy neural network of the system, the force control is performed by the fuzzy neural network of the force control system, and the output of the force control system is changed by the value of the force sensor to be controlled. The computer-readable recording medium stores a program for continuously switching between the position control and the force control by position / force hybrid / compliance control.
The control method of the present invention is a control method for a built-in intelligent controller that performs position control with respect to a controlled object by a position command and performs force control with a force command, and the built-in intelligent controller includes a fuzzy neural network of a position control system. The position control is performed, the force control is performed by a fuzzy neural network of the force control system, the output of the force control system is changed according to the value of the force sensor to be controlled, and a hybrid of position and force / The position control and the force control are continuously switched by compliance control.

  According to the present invention, by using a fuzzy neural network, an intelligent controller that smoothly switches between position control and force control while realizing intelligent and robust control with practical nonlinear operation characteristics is provided. can do.

1 is an external view of an example of a control system X according to a first embodiment of the present invention. It is an external view of the ICS part 10 of the intelligent controller part 1 which concerns on the 1st Embodiment of this invention. It is a block diagram which shows the control structure of the control system X which concerns on the 1st Embodiment of this invention. It is a block diagram which shows the control structure of the ICS part 10 which concerns on the 1st Embodiment of this invention. It is a block diagram which shows the example of control structure of the ICS part 10 used for the development / learning process which concerns on the 1st Embodiment of this invention. It is a flowchart of the development / learning process which concerns on the 1st Embodiment of this invention. It is a flowchart of the model creation process which concerns on the 1st Embodiment of this invention. It is a block diagram which shows the example of control structure of the ICS part 10 used for the FF-NN reverse dynamics process which concerns on the 1st Embodiment of this invention. It is a flowchart of FF-NN reverse dynamics processing concerning a 1st embodiment of the present invention. It is a conceptual diagram which shows the structural example of the neural network of the FF-NN part 120 which concerns on the 1st Embodiment of this invention. It is a conceptual diagram which shows the block of the neural network of the FF-NN part 120 which concerns on the 1st Embodiment of this invention. It is a block diagram which shows the control structural example of the ICS part 10 used for the FB-FN learning process which concerns on the 1st Embodiment of this invention. It is a flowchart of the FB-FN learning process which concerns on the 1st Embodiment of this invention. It is a conceptual diagram which shows the structure of the fuzzy neural network which concerns on the 1st Embodiment of this invention. FIG. 2 is a conceptual diagram showing an overall configuration of a fuzzy PD unit 131 and a fuzzy PI unit 132 according to the first embodiment of the present invention. It is a block diagram which shows the learning process of the fuzzy PD part 131 and the fuzzy PI part 132 which concern on the 1st Embodiment of this invention. It is a conceptual diagram which shows the example of the adjusted membership function based on the 1st Embodiment of this invention. It is a conceptual diagram which shows the setting method of the fuzzy rule of the fuzzy PD part 131 and the fuzzy PI part 132 which concern on the 1st Embodiment of this invention. It is a conceptual diagram which shows the calculation method of the fuzzy neural network gain of the FB-FN part 130 which concerns on the 1st Embodiment of this invention. It is a conceptual diagram which shows the gain matrix of the FB-FN part 130 which concerns on the 1st Embodiment of this invention. It is a block diagram which shows the concept of the gain scheduling by the fuzzy neural network of the FB-FN part 130 which concerns on the 1st Embodiment of this invention. It is an example of a screen which shows the fuzzy PD gain matrix of the fuzzy PD part 131 which concerns on the 1st Embodiment of this invention. It is an example of a screen which shows the fuzzy PD control surface of the fuzzy PD part 131 which concerns on the 1st Embodiment of this invention. It is a conceptual diagram for demonstrating the robust stability of the FB-FN part 130 which concerns on the 1st Embodiment of this invention. It is a conceptual diagram of the PWM control of the software servo control process which concerns on the 1st Embodiment of this invention. It is a model figure of the conventional PID control which concerns on Example 1 of the 1st Embodiment of this invention. It is a model figure in the case of only NN-PID control which concerns on Example 1 of the 1st Embodiment of this invention. It is a model figure of fuzzy neural network control concerning Example 1 of the 1st mode of the present invention. It is a model figure of the PWM part 150 which concerns on Example 1 of the 1st Embodiment of this invention. It is a model diagram of a SEIHUKIRIKAE block of the PWM unit 150 according to Example 1 of the first embodiment of the present invention. It is a model diagram of the PWM18-20 block of the SEIHUKIRIKAE block of the PWM unit 150 according to Example 1 of the first embodiment of the present invention. It is a model figure of the output part 170 which concerns on Example 11 of the 1st Embodiment of this invention. It is a model figure of the input part 180 which concerns on Example 1 of the 1st Embodiment of this invention. It is a model figure of the counter part 160 which concerns on Example 1 of the 1st Embodiment of this invention. It is a model figure of the COUNTER block of the counter part 160 which concerns on Example 1 of the 1st Embodiment of this invention. It is a model diagram of the Triggered Subsystem of the COUNTER block of the counter unit 160 according to Example 1 of the first embodiment of the present invention. It is a graph which shows the experimental result of control using the model of the conventional PID control which concerns on the 1st Embodiment of this invention. It is a graph which shows the experimental result of the fuzzy neural network control which concerns on Example 1 of the 1st Embodiment of this invention. It is a block diagram which shows the control structure of the control system Y which concerns on the 2nd Embodiment of this invention. It is a block diagram which shows the control structure of the ICS part 50 which concerns on the 2nd Embodiment of this invention. It is a block diagram which shows the control structural example of the ICS part 50 used for the FF-NN reverse dynamics process which concerns on the 2nd Embodiment of this invention. It is a flowchart of the FF-NN reverse dynamics process which concerns on the 2nd Embodiment of this invention. It is a conceptual diagram of the assembly process which concerns on the 2nd Embodiment of this invention. It is a flowchart of the assembly process which concerns on the 2nd Embodiment of this invention. It is a photograph which shows the external appearance of the control object 7 which concerns on Example 2 of the 2nd Embodiment of this invention. It is a block diagram which shows the control structure of the control object 7 which concerns on Example 2 of the 2nd Embodiment of this invention. It is a graph which shows the experimental result in the sponge which concerns on Example 2 of the 2nd Embodiment of this invention. It is a graph which shows the experimental result in the building block which concerns on Example 2 of the 2nd Embodiment of this invention. It is a graph which shows the experimental result in the stuffed toy which concerns on Example 2 of the 2nd Embodiment of this invention. It is a figure which shows the experimental result in each object which concerns on Example 2 of the 2nd Embodiment of this invention.

<First Embodiment>
[Configuration of Control System X According to First Embodiment of the Present Invention]
First, the control configuration of the control system X according to the first embodiment of the present invention will be described with reference to FIGS. 1A to 2.
FIG. 1A is an external view of an example of a configuration used when developing the control system X. FIG. In the control system X of the present embodiment, the intelligent controller unit 1 is connected to a control object 2 that is a table moving device and a control device 3 (not shown).
The intelligent controller unit 1 is a control part including an ICS unit 10 (Intelligent Control System Module, embedded intelligent controller) used for intelligent and robust control. In the example of FIG. 1A, an example is shown in which a development embedded computer capable of connecting an embedded module is used as the intelligent controller unit 1.
Further, the control target 2 is a specific control target portion that is affected by disturbance and requires robust control. In the example of FIG. 1A, an example using a table moving device used for development and simulation is shown. The table moving device includes a motor drive driver 20 such as a motor driver or a counter. The motor driving driver 20 moves the table 210 by the driving motor 230 and inputs a control response.
FIG. 1B is an external view of the ICS unit 10. The ICS unit 10 is an intelligent and robust embedded intelligent controller using an embedded computer, and after learning, which will be described later, is attached to an actual device to be controlled to perform intelligent and robust control.

FIG. 2 is a control block diagram showing a specific configuration of the control system X.
As an example, the control system X includes an intelligent controller unit 1, a control target 2, and a control device 3.

The intelligent controller unit 1 incorporates the ICS unit 10 into various devices that need to be controlled such as robots, automobiles, bicycles, aircrafts, heat pumps, and the like when manufacturing products. Here, as a configuration example at the time of development, for example, an example of using a built-in computer that is a system development tool and is mainly used in a development stage before being installed in an actual product will be described. As an embedded computer for development, for example, AD7011-EVA manufactured by A & D Corporation can be used. Moreover, the intelligent controller unit 1 can perform real-time (real-time) processing using a high-speed control unit, for example, in a predetermined sampling time, for example, about 1 millisecond to several tens of milliseconds.
The controlled object 2 is an element that electromagnetically adjusts the drive, such as a motor, actuator, or heater such as a robot, automobile, bicycle, aircraft, heat pump, etc., and a sensor that acquires the output state of this element when manufacturing a product. including. In the present embodiment, an example of a configuration using a DC (direct current) motor-driven table moving device as a driving device will be described as a configuration for performing a simulation of a driving system in which a general disturbance occurs.
The control device 3 can use a general PC / AT compatible machine, a PC (Personal Computer) such as a MAC standard machine, or the like. The control device 3 can install and execute control software using an OS (Operating System) such as Windows (registered trademark) or Linux (registered trademark). In the following, an example will be described in which a fuzzy / neural network is constructed by installing MATLAB (registered trademark) / Simlink (registered trademark) manufactured by MathWorks as control software on the PC of the control device 3. Absent. In the following, an example of creating a model definition and a model execution screen using the VCDdesigner (trademark) and ModelDef (trademark) software of A & D Corporation in the control device 3 will be described. Instead, it can be realized using various software.

<Configuration of intelligent controller unit 1>
More specifically, the intelligent controller unit 1 includes, for example, an ICS unit 10, a control unit 11, an auxiliary storage unit 14, a main storage unit 15, a logic circuit unit 16, an I / F unit 17, and connectors 18 and 19. Consists of including.
The ICS unit 10 is an iCS (intelligent control module) incorporating an EEPROM (Electrically Erasable Programmable Read-Only Memory), a NOR / NAND flash memory, an SRAM (Static Random Access Memory), and the like. The ICS unit 10 is a binary file which is a program (execution code) of C language or the like of the control unit 11 created by the control device 3, definition files and data regarding various models and parameters, and a logic circuit of an FPGA (Field Programmable Gate Array). A storage medium (recording medium) that stores data, firmware of the intelligent controller unit 1, BIOS (Basic Input / Output System), and the like is also provided. In the development / learning process described later, the program and data are read and executed by the control unit 11. Further, the ICS unit 10 may itself be configured as a single board computer provided with control calculation means as shown in the external view of FIG. 1A.
The control unit 11 is a CPU (Central Processing Unit, Central Processing Unit, Central Processing Unit), MPU (Micro Processing Unit) capable of high-speed arithmetic processing such as a RISC (Reduced Instruction Set Computer) method or a CISC method. , A digital signal processor (DSP), an application specific processor (ASIC), and other control means having a calculation / control capability.
The auxiliary storage unit 14 is mainly a storage medium (recording medium) for the OS, such as an SSD (Solid State Drive), a compact flash (registered trademark) disk, a flash memory disk such as an SD memory card or a microSD memory card, an HDD (Hard). Disk Drive), auxiliary storage means including a magnetic tape medium, an optical disk medium, and the like. As the OS stored in the auxiliary storage unit 14, it is preferable to use Linux (registered trademark) using Xenomai extension, which is a real-time module used by being incorporated in the kernel, or an OS capable of real-time processing such as ITRON. .
The main storage unit 15 is mainly an SDRAM (Synchronous Dynamic Random Access Memory), DDR (Double-Data-Rate), DDR2, DDR3, LPDDR3, WideIO, XDR (extreme) as a storage medium (recording medium) for control / calculation. This is a main storage means including a high-speed storage means used for main storage such as data rate). The main storage unit 15 stores each program and data of the ICS unit 10, sensor data acquired from the control target 2, and the like.
The logic circuit unit 16 is a logic circuit such as CPLD (Complex Programmable Logic Device) or FPGA, various analog circuits, and the like. The logic circuit unit 16 includes an interface circuit with the control target 2, an A / D (analog-to-digital) converter, a D / A (digital-to-analog) converter, and the like.
The I / F unit 17 is an interface such as USB (Universal Serial Bus), wireless USB, serial, parallel, and LAN (Local Area Network). The I / F unit 17 is mainly used to transfer programs, binary files, data, and the like from the control device 3 and transfer data such as control results to the control device 3.
The connector 18 is a connector or the like mainly connected to the control device 3. The connector 18 can be connected to the control device 3 by a LAN cross cable, for example. In addition, a display unit / input unit such as a touch panel display, a speaker, a microphone, or the like may be separately connected to the connector 18.
The connector 19 is a connector mainly connected to the control target 2 and may be connected to a driving amplifier, a sensor, or the like separately.

<Configuration of control object 2>
Further, the control target 2 includes a motor drive driver 20 and a table unit 21.
The motor drive driver 20 can use a driver IC (Integrated Circuit) for driving a general DC motor. This driver IC drives the motor 230 of the table unit 21 by a control signal from the intelligent controller unit 1. Further, the motor drive driver 20 acquires the output from the encoder 220 and transmits it to the intelligent controller unit 1.
The table unit 21 is an example of a moving table driven by a DC motor. The table unit 21 includes a table 210 that is a moving table (stage), a motor 230 such as a DC motor that drives the table 210, and an encoder 220 that detects the position of the table 210.

(Configuration of ICS unit 10)
Next, a detailed control configuration of the ICS unit 10 will be described with reference to FIG.
The ICS unit 10 mainly includes a PID unit 110, an FF-NN unit 120 (feed forward unit), an FB-FN unit 130 (feedback unit), a difference unit 140, a PWM unit 150, a counter unit 160, an output unit 170, and an input unit. 180, a control signal generator 190, and a disturbance simulator 195. The ICS unit 10 performs various calculations in each process by combining these components.

The PID unit 110 is a part including a model program and data for performing classic PID control.
The PID unit 110 includes a PD compensator unit 111 for performing PD (proportional and differential) compensation in feedback control. The PD compensator unit 111 is used to grasp the characteristics of the control target 2.

The FF-NN unit 120 includes a program and data of a model of an FF-NN (Feedforward neural network) compensator. The FF-NN unit 120 learns and uses the servo system inverse dynamics characteristics by a neural network in order to improve the phase delay of the controlled object.
For this learning, the FF-NN unit 120 learns the inverse dynamics characteristics of the PD control system of the PD compensator unit 111 offline by the BP (Backpropagation) method.
As the FF-NN unit 120, for example, a three-layer, three-input, four-intermediate unit, one-output neural network as shown in FIG. 9A can be used. Here, displacement r, velocity s (dr / dt), and acceleration s ² (d ² r / dt ² ) are input to the input layer (Input Layer), respectively. Further, the output value of the neural network of the FF-NN unit 120 is the manipulated variable U _NN .
Details of the learning method of the FF-NN unit 120 will be described later.

The FB-FN unit 130 is a part including a program and data of a model of an FB-FN (Fuzzy neural network Based non-linear error, fuzzy neural network) compensator.
The FB-FN unit 130 is a deviation signal e (t) = r between the target signal r (t) and the output signal y (t) based on the fuzzy neural network base for compensating the nonlinear dynamic characteristic of the controlled object 2. Learning is performed offline so that the square of (t) -y (t) is minimized. After these learnings, the control system X can be incorporated into a device and used for control. The FB-FN unit 130 performs output compensation by gain scheduling.

The FB-FN unit 130 mainly includes a fuzzy PD unit 131 (fuzzy PD unit) and a fuzzy PI unit 132 (fuzzy PI unit).
The fuzzy PD unit 131 is a part that performs PD compensation using a fuzzy neural network. The fuzzy PI unit 132 is a part that performs PI (proportional, integral) compensation using a fuzzy neural network similar to the fuzzy PI unit 131.
Each of the fuzzy PD unit 131 and the fuzzy PI unit 132 can use, for example, a two-input four-layer one-output fuzzy neural network as shown in FIG. 12A. Fuzzy PD 131 and the fuzzy PI unit 132 both in the input layer, respectively, and inputs the displacement error signal e _p as a position error, a speed error signal e _v of the speed error displacement time derivative as an input signal. On this, fuzzy PD section 131, as the evaluation function the square of the operating amount U _PD of the PD compensation PID unit 110 corrects the coupling weights with BP method. The fuzzy PI unit 132 corrects the coupling weight by the BP method using the square of the PI compensation operation amount U _PI of the PID unit 110 as an evaluation function. The fuzzy PD unit 131 outputs the sum of the respective inputs as an operation amount U _FN-PD . Further, the fuzzy PI unit 132 outputs an operation amount U _FN-PI .
Details of the learning method of the fuzzy neural network of the fuzzy PD unit 131 and the fuzzy PI unit 132 will be described later.

The difference unit 140 calculates the displacement r, speed s (dr / dt), and acceleration s ² (d ² r / dt ² ) of the operation signal (command signal) of the PD compensator unit 111 which is a conventional classical PD compensator. This is the part that contains the program and data to be executed. The difference unit 140 includes a speed calculation unit 141 that calculates the speed s and an acceleration calculation unit 142 that calculates the acceleration s ² . Further, the difference unit 140 may calculate the displacement error signal e _p, the speed error signal e _v.
The PWM unit 150 includes a program and data of a software PWM (Pulse Width Modulation) module for driving a DC motor. The motor 230 connected to the motor drive driver 20 can be PWM controlled by the output signal of the PWM unit 150.
The counter unit 160 includes a module program and data for an encoder counter. Using the value calculated by the counter unit 160, difference calculation unit 140 is displaced r, speed s, the acceleration s ^2, and the displacement error signal e _p, it is possible to obtain the speed error signal e _v.
The output unit 170 is a part that performs switching to transmit the output from the PWM unit 150 to the motor drive driver 20.
The input unit 180 is a part that inputs a signal from the control target 2 such as a signal of the encoder 220 input via the motor drive driver 20 and transmits the signal to the counter unit 160.
The control signal generator 190 is a part that includes a program and data of a module that acquires an instruction signal from an external device (not shown) and outputs a control signal for controlling the control target 2. The control signal generator 190 can also output a signal such as a sine wave for control simulation.
The disturbance simulating unit 195 is a part such as a signal generator for simulating a signal caused by disturbance, friction, or the like as a disturbance. The disturbance simulating unit 195 can output a random signal having a predetermined output.

[Development / learning process of ICS unit 10 of control system X]
Next, a development / learning process for learning the ICS unit 10 and optimizing the control object 2 in the control system X will be described with reference to FIGS. 4 and 5.
FIG. 4 is a block diagram showing main components and signal flow in this control process.
The control system X according to the first embodiment of the present invention performs fuzzy neural network (FNN) based control in order to solve the problems of the conventional control system. Here, fuzzy is characterized in that a complicated system can be regarded as “warm” like a human being and can be described linguistically by an if-then fuzzy rule. On the other hand, a feature of neural networks is that the relationship between input and output can be identified by learning.
In the control processing according to the first embodiment of the present invention, the controller 3 mainly uses the models of the FF-NN unit 120, the FB-FN unit 130, and the difference unit 140 of the ICS unit 10 in FIG. Create and develop. Then, the control unit 100 of the intelligent controller unit 1 executes a program for each part and the like to learn for feedback control of the controlled object 2 and configure an actual control system. This control system can drive the controlled object 2 by, for example, PWM control of software servo.

More specifically, the robust gain scheduling control system using the fuzzy neural network according to the first embodiment of the present invention performs the feedforward compensation learning the servo system inverse dynamics characteristics by the neural network. The FF-NN unit 120 is configured by two compensation elements, and the FB-FN unit 130 performs gain scheduling by a fuzzy neural network. This control system is a two-degree-of-freedom control system in which the target value response characteristic and the closed loop characteristic can be independently designed, and can follow the target input and remove the influence of disturbance. The FF-NN unit 120 learns the inverse dynamics characteristics of the conventional fixed PD control system off-line in order to reduce the steady-state deviation due to modeling errors and intermittent disturbances and improve the mechatronic position. Next, in order to compensate for the nonlinear dynamic characteristic, the FB-FN unit 130 has a deviation signal e (t) = r (t) −y (t) between the target signal r (t) and the output signal y (t). Study offline or online to minimize the square of. The proposed control system is configured after learning.
Details of the development / learning process will be described below with reference to the flowchart of FIG.

(Step S101)
First, a model corresponding to each part of the ICS unit 10 is created by the control device 3 using software such as MATLAB / Simulink, Model Def, and VCDdesigner.
The control device 3 compiles the created model into a program binary file and transfers it to the ICS unit 10. This binary file is stored in the auxiliary storage means of the ICS unit 10 together with a definition file for actually executing the binary file and various data.
The program or the like stored in the ICS unit 10 can be executed independently by the control unit 11 of the intelligent controller unit 1 to perform the following learning process or the like.

(Step S102)
Next, the control unit 11 of the intelligent controller unit 1 uses the FF-NN unit 120 to perform FF-NN inverse dynamics processing.
In this process, inverse dynamics for the nonlinear characteristic of the controlled object 2 is acquired, and learning for linearizing the system characteristic is performed.
Here, in order to improve the phase delay of the controlled object, the FF-NN unit 120 learns the inverse dynamics characteristics of the PD control system of the PD compensator unit 111 (FIG. 3) offline by the BP method.
The control by the FF-NN unit 120 can be expected to improve the phase delay.

(Step S103)
Next, the control unit 11 of the intelligent controller unit 1 performs an FB-FN learning process using the FB-FN unit 130.
Here, only by the control of the FF-NN unit 120 described above, a nonlinear error occurs with respect to a target signal (target trajectory signal) such as the trajectory of the motor 230 of the control target 2. This is because NN is limited to compensation for only a part of the nonlinear characteristic to be controlled as input compensation.
For this reason, in this step, learning for compensation of nonlinear dynamic characteristics using the FB-FN unit 130 is performed.
As described above, the FB-FN unit 130 calculates the deviation signal e (t) = r (t) −y (t) between the target signal r (t) acquired from the difference unit 140 and the output signal y (t). Learn to minimize the square. This learning can be executed in the online state during the control or in the offline state using the data after the control.
After these learnings, the control system X is configured.

In this way, by using the FF-NN unit 120 and the FB-FN unit 130 for learning, it is possible to achieve improvement in tracking performance and error reduction in control.
Therefore, generalization ability corresponding to various control systems can be provided, and intelligent control can be realized.

(Step S104)
Further, the control unit 11 of the intelligent controller unit 1 can perform software servo control processing using the PWM unit 150.
Specifically, in the control system X configured at the time of each learning described above and after learning, the control unit 11 of the intelligent controller unit 1 can perform software servo processing and control the actual control target 2. It is.
Thus, the development / learning process ends.

<Details of model creation process>
Here, with reference to the flowchart of FIG. 6, the details of the model creation processing for creating a model and transferring it to the ICS unit 10 of the intelligent controller unit 1 will be described.

(Step S111)
First, the control device 3 detects a user instruction and performs model creation processing. Specifically, a control system block diagram is created by MATLAB / Simulink software.
More specifically, MATLAB / Simulink is control system design support software. Here, MATLAB is an abbreviation of MATrix LAB Laboratory, and is numerical analysis software based on matrix / vector operations. MATLAB is an interpreter (interactive) format high-level programming language, and uses an “M file” described in the MATLAB language similar to the BASIC language. MATLAB has abundant toolboxes including a control toolbox and a communication toolbox, and can be used by users for analysis. Also, Simulink that can construct a control system simulator by inputting a block diagram is prepared. A RealTime Workshop for mounting a controller designed by MATLAB / Simulink using an embedded CPU is also commercialized, and can be suitably used in this embodiment.
Simulink is software that allows analog computer functions to be executed in a digital environment, and is a convenient tool for analyzing differential equations.

In addition, as a software for intelligent control, a neural network toolbox and a fuzzy logic toolbox can be used in combination with MATLAB / SIMULLINK as appropriate.
The neural network toolbox is a group of M files necessary for designing and simulating a neural network system that can be used for pattern recognition, identification, control, and the like for a nonlinear system. Neural Network Toolbox corresponds to initialization, transfer function, supervised network / learning rules, perceptron, backpropagation (BP), Leavenberg-Marquardt backpropagation, Radial basis network, recurrent network (Elman, Hopfield), etc. Yes.
Fuzzy Logic Toolbox is a program that can create and edit a fuzzy inference system. Fuzzy Logic Toolbox makes it easy to create a fuzzy inference system using graphical tools or command functions. By using SIMULLINK together, it is possible to test a fuzzy system in a block diagram simulation environment. In addition, the Fuzzy Logic Toolbox can generate either real-time or non-real-time code from the SIMULLINK environment as long as the Real-Time Workshop can be used.

  The created model is, for example, as shown in FIGS. 19A to 20C shown in Example 1 described later.

(Step S112)
Next, the control device 3 detects a user instruction and performs C code creation / compilation processing.
In this processing, for example, C code, object code, and execution code written in C language are automatically generated from a model created by Real-Time Workshop of MATLAB.

(Step S113)
Next, the control device 3 detects a user instruction and performs model definition / build processing.
Here, model def is used to define model parameters and signals.
By performing this model definition, it is possible to create a definition file that makes it possible to use parameters and signals on the model created by MATLAB / Simulink at the time of execution.

(Step S114)
Next, the control device 3 detects a user instruction and performs a VC (Virtual Console) screen creation process.
In this process, screen creation is performed using the VCDdesigner software. Specifically, a “VC screen” serving as a user interface displayed on the PC display device of the control device 3 and the display unit of the intelligent controller unit 1 can be created.
In addition, VCDesigner can design a VC screen for executing the created model.
This makes it possible to design a screen that is easy for the user to handle when the created model is executed.

(Step S115)
Next, the control device 3 detects a user instruction and performs an execution code transfer process.
More specifically, the control device 3 can execute a program by the VCDdesigner. After performing an execution test using an emulator or the like of the control device 3, the user transfers an executable code binary file, definition file, various data, and the like (hereinafter referred to as a program or the like).
Note that the control device 3 can also automatically transfer the generated execution code or the like if the control device 3 and the intelligent controller unit 1 are in a connected state when the VCDdesigner starts working.

(Step S116)
Next, the intelligent controller unit 1 performs an execution code acquisition / execution process.
Here, the control part 11 of the intelligent controller part 1 memorize | stores the program etc. which were acquired from the above-mentioned control apparatus 3 in the auxiliary storage means of the ICS part 10. FIG.
After acquiring the execution code, the execution code or the like can be executed by the intelligent controller unit 1 alone.

The VCDdesigner can also be used when executing the created program or the like.
Specifically, when the control device 3 and the intelligent controller unit 1 are in a connected state and the screen creation is completed, the user presses the execution button of the VCDdesigner. When the execution button is pressed, the designed screen is drawn on the display unit of the control device 3 and the intelligent controller unit 1 and can be executed.
In this execution, step execution, acquisition of data from the control target 2, and the like can also be performed. The intelligent controller unit 1 can also sequentially transmit acquired data and the like to the control device 3. The acquired data can also be transmitted after a predetermined time.

(Step S117)
Next, the control device 3 performs control result acquisition analysis processing.
Specifically, the control device 3 can check the acquired data, the state of the model, and the like with MATLAB / Simulink, VCDdesigner, or analysis software. As a result, it is possible to confirm the operation of the model and perform the learning process described below.
Note that when executing using the VCDdesigner, the user cannot give an instruction using the button on the screen of the intelligent controller unit 1, but gives an instruction on the user interface screen of the control device 3. After the execution test is performed by the control device 3, the intelligent controller unit 1 can be executed alone.
This completes the model creation process.

<Details of FF-NN reverse dynamics processing>
Next, details of the FF-NN inverse dynamics processing will be described with reference to FIGS. In this processing, learning of the FF-NN unit 120 is performed by feedforward using a signal resulting from classical PD control.
FIG. 7 is a block diagram showing a relationship between each part of the ICS unit 10 used for the FF-NN inverse dynamics process and a control target. Thus, in order to improve the dynamic characteristics of the controlled object, the FF-NN unit 120 that is a neural network is used. The neural network of the FF-NN unit 120 learns the inverse dynamics of the controlled object 2 that is a servo system online or offline, and uses it as a feedforward compensator.
The details of this FF-NN inverse dynamics process will be described step by step based on the flowchart of FIG.

(Step S201)
First, the control unit 11 of the intelligent controller unit 1 performs the FF-NN preparation process using the PD compensator unit 111 of the ICS unit 10 or the like.
Here, as shown in the block diagram of FIG. 7, the PD compensator unit 111, which is a conventional classical PD compensator, performs a test for grasping the characteristics of the controlled object 2, and the command signal displacement (y), speed (Dy / dx) and acceleration (d ² y / dt ² ) signals are acquired. In this acquisition, the speed is calculated by the speed calculation unit 141 of the difference unit 140, and the acceleration is calculated by the acceleration calculation unit 142. In addition, it is important for this test that non-linear characteristics such as friction and play are included by this operation.
In addition, the PD gain in the PD control is set to a small value at which the control target 2 can move.
The speed and acceleration signals may be calculated by the numerical calculation process by the difference unit 140 offline after obtaining the displacement signal.

(Step S202)
Next, the control unit 11 performs an FF-NN learning process using the FF-NN unit 120.
Here, first, the dynamic characteristics of the plant to which the position and velocity feedback law is applied can be expressed by the following equation (1) when the nonlinear characteristic is approximated as a quadratic system and linearized.

Here, u (t) is an input, y (t) is a position, and k _p and k _v are position and velocity feedback gains, respectively. However for the k _v, it is assumed to include the counter electromotive force constant k _e of the servomotor. Assuming that r (t) is a target trajectory, the inverse dynamics of the servo system is expressed by the following equation (2) by solving the above equation (1) as y (t) = r (t).

That is, the neural network of the FF-NN unit 120 learns the relationship of the displacement r (k), the speed dr / dt (k), and the acceleration d ² r / dt ² (k) to u (t). Therefore, the input I (k) to the network is a column vector of [3 * 1] and can be expressed by the following formula:

I (k) = [r ( k), k p / k v · dr / dt (k), J / k p · d 2 r / dt 2 (k)] T

The linear function f (x) = x is used for the input / output relationship of each unit. The output of the network is the estimated value φ _k of the plant input.
If the output of the intermediate layer is ψ _i (k), the output of the neural network can be expressed as the following equation (3).

Here, n is the number of units in the intermediate layer, and m is the number of units in the input layer.
Learning of the neural network is performed by the back propagation (BP) method by defining the following expression (4) as an evaluation function.

  That is, the expression (4) is obtained by subtracting the input of the signal from the plant that is the control target 2 from the estimated value of the plant input calculated from the neural network of the FF-NN unit 120, and squared to divide by 1/2. Ask for it.

[Configuration Example of Neural Network of FF-NN Unit 120]
Next, a configuration example of a neural network of the FF-NN unit 120 which is an FF-NN compensator will be described with reference to FIG. 9A. In the FF-NN unit 120, a neural network having three layers, three inputs, four intermediate units and one output is used.
The input signal is a signal of displacement r, speed dr / dt, and acceleration d ² r / dt ² . Thereby, the dynamic characteristic of the plant is approximated as a secondary system and linearized.
Specifically, the displacement, speed, and acceleration data obtained by the NN learning process can be input as the FF-NN unit 120 input. The data is a predetermined amount, for example, about 100 per second. In this case, the sampling frequency is 10 msec.

  As software necessary for learning the neural network of the FF-NN unit 120, an example of a neural network generation command for MATLAB shown below can be input and used.

P = [Displacement;
speed;
acceleration];
T = [teacher signal];
net = newff ([-5 5; -5 5; -5 5], [3 1], {'tansig', 'purelin'}, 'trainlm');
y = sim (net, P);
plot (P, T, P, y, 'o');
net.trainParam.show = 10;
net.trainParam.epochs = 20000;
net.trainParam.goal = (0.0037 ^ 2);
net = train (net, P, T);
y = sim (net, P);
plot (P, T, P, y, 'o');
gensim (net, -1);

return

Here, P = displacement, velocity, acceleration, and T = teacher signal.
However, the teacher signal is the operation signal y, and the time required for grasping the characteristics (about three cycles).

Referring to FIG. 9B, when learning of the neural network is completed by MATLAB, a neural network (NN) block is generated.
This NN block is used as a control block in the FF-NN unit 120.
Thus, the FF-NN inverse dynamics process is completed.

<Details of FB-FN learning process>
Next, the details of the FB-FN learning process will be described with reference to FIGS.
FIG. 10 is a block diagram illustrating a relationship among the units of the ICS unit 10 corresponding to the FB-FN learning process.
Specifically, the FB-FN unit 130 controls the fuzzy PD and the fuzzy PI using a fuzzy neural network.
A fuzzy neural network is a combination of neurological learning ability and fuzzy reasoning ability that are complementarily combined to make use of their strengths to compensate for their shortcomings.
The FB-FN unit 130 according to the present embodiment has a feature of reducing a control error by a non-linear deviation compensation characteristic by a fuzzy neural network that adaptively adjusts a control gain when a change and a speed control error occur due to non-linear friction. .
Here, the actual model of the controlled object 2 includes non-linear elements that vary in value of the moment of inertia or the like for each product and cannot be expressed by a quadratic model. That is, in the control of the control target 2, a nonlinear friction error of the control target occurs with respect to the target trajectory signal.
Therefore, feedback compensation is performed by the FB-FN unit 130 which is a fuzzy PID controller so that an action of adaptively reducing a deviation signal caused by modeling error, friction, unknown disturbance, and the like works. For this reason, the FB-FN unit 130, which is an FB-FN compensator, is trained using the acquired data, and the nonlinear error signal is reduced.
By learning the FB-FN unit 130 in this way, deviations due to trajectory errors and model errors can be suppressed by feedback using fuzzy PID. This makes it possible to realize an intelligent robust control system (IRCS) having both intelligence and robustness.
Hereinafter, the details of each step of the FB-FN learning process will be described with reference to the flowchart of FIG.

(Step S301)
First, the control unit 11 of the intelligent controller unit 1 performs fuzzy neural network setting processing of the FB-FN unit 130 of the ICS unit 10.
Here, with reference to FIGS. 12A to 12D, the configuration of the fuzzy neural network (FNN) according to the first embodiment of the present invention, which is used in common for the fuzzy PD unit 131 and the fuzzy PI unit 132, will be described. To do.
The fuzzy neural network enables learning by the BP method, and devise the combination of the hierarchical neural network to have a correspondence with the fuzzy inference rules, and the fuzzy neural network acquires the knowledge about the pattern related to the input / output relationship acquired by the neural network. It can be grasped as knowledge of the symbol called rule. Further, the difficulty of identifying fuzzy rules and adjusting the membership function, which are problems in the fuzzy inference system, can be improved by the learning function of the neural network.

First, a method for creating a fuzzy neural network using the Fuzzy Logic Toolbox will be described.
First, the menu of the Fuzzy Logic Toolbox is called. On the menu:
1) Membership function editor 2) FIS editor 3) Rule editor 4) Rule viewer 5) Surface viewer
Here, in the fuzzy inference system using the fuzzy neural network in this embodiment, a model is first created by selecting the FIS editor.
At that time,
a) If you type tipper, a window is displayed.
b) Use the membership function editor to set the membership function name, position, and type.
・ Range [0 10]
A Gaussian curve is set.
Gauss2mf
Create a membership function by combining two Gaussian curves.
For example, the following formula (5) can be used.

  In this way, the fuzzy inference system according to this embodiment can be created.

[Structure and configuration example of fuzzy neural network]
FIG. 12A shows the structure of a fuzzy neural network according to the first embodiment of the present invention.
In the example of FIG. 12A, the input layer of the layer (A), the displacement error signal e _p and the displacement time derivative of the speed error signal e _v is the input signal, and distributes the input signal to the unit follows layer .
Next, the intermediate layer of the (B) layer has a membership function as an internal function and uses a Gaussian function. The output of each unit is the value of the membership function and is called the antecedent part.
(C) layer is a consequent part inferring with fuzzy rules, (B) the input space is divided into nine by layer, and the fitness of the divided space is calculated by the formula of the antecedent part of fuzzy rules, (C) A value μ _i that is standardized by the sum of the antecedent part conformity obtained in all units of the layer is output.
The (D) layer is a linear unit that outputs the sum of inputs.
As described above, a neural network having a fuzzy rule structure can be constructed, which is called a fuzzy neural network. While this fuzzy neural network enables learning by the BP method, it is possible to have a correspondence relationship with fuzzy inference rules by devising the combination of hierarchical neural networks.

FIG. 12B is a conceptual diagram showing the overall configuration of the fuzzy PD unit 131 and the fuzzy PI unit 132 according to the first embodiment of the present invention.
In this way, by combining the fuzzy PD unit 131 and the fuzzy PI unit 132 in parallel, a fuzzy PID can be realized as a whole.

(Step S302)
Next, the control unit 11 performs an FNN learning process of the FB-FN unit 130 of the ICS unit 10.
FIG. 13 is a block diagram illustrating a learning process of the fuzzy PD unit 131 and the fuzzy PI unit 132 of the FB-FN unit 130.
In the learning of the fuzzy neural network, displacement error and speed error are input based on the data of the PD control characteristic grasp test by the PD compensator unit 111 described above. Then, the operation amount of the PID is obtained, and this is used as a teacher signal to learn the fuzzy PD unit 131 and the fuzzy PI unit 132 by a method combining the BP method and the least square method.
In the neural network of the FB-FN unit 130, feedback gain learning is performed using the square of the position error:

E _p = 1/2 e _p ²

To minimize. Further, k _ij = W _ij is set so that the condition of k _ij ≧ 0 (j = I, p, v) is always satisfied.
To be more specific, the input data used for learning, the position error e _p linear PD control and linear PI control gain adjustment as a servo system to operate smoothly and its a variation e _v, teachers The data is an output signal U of linear PD control and linear PI control. For learning of a large error (step), a square wave step signal is used as input data, and teacher data is output signal U of linear PD control and linear PI control similar to those described above.

Here, the fuzzy PD learning of the fuzzy PD unit 131 will be described with reference to FIG.
The fuzzy PD learning of the fuzzy PD unit 131 is a basis for learning to compensate for modeling errors and nonlinear friction. That is, the actual model of the controlled object 2 has variations such as moment of inertia, and there are nonlinear elements that cannot be expressed by the quadratic model.
These modeling errors and nonlinear friction errors are expressed as follows:

e = _Td + _fd (y)
Here, T _d includes a modeling error as an unknown disturbance.
Let f _d (y) be sgn (y ′) {(Fs−Fc) exp (− | y ′ | / e) + Fc}.

In the fuzzy PD learning, learning is performed as inverse dynamics by a teacher signal obtained as a feedback PD output signal of the deviation output signal of the modeling error and the nonlinear friction error expressed by the above formula.
Learning is a square in which _{E (u PD) = l /} 2u PD 2 the evaluation function u _PD as a nonlinear error learning method to correct the connection weights in the BP method.

Here, the fuzzy PI learning of the fuzzy PI unit 132 will be described with further reference to FIG.
The fuzzy PI learning of the fuzzy PI unit 132 is a basis for learning steady deviation due to Coulomb friction and unknown disturbance. That is, in the controlled object 2, a steady deviation occurs due to Coulomb friction. This is a positional deviation when the acceleration is zero. In addition, a steady deviation due to unknown disturbances also occurs.
Expressing these errors, we have the following formula:

e = e _FC + e _Td

Here, e _FC is a steady deviation e _Td caused by Coulomb friction, and a steady deviation caused by unknown disturbance. In the fuzzy PI learning, the stationary deviation due to the Coulomb friction and the unknown disturbance expressed by the above formula is learned as inverse dynamics by the teacher signal acquired as the feedback PI output signal.
As a non-linear error learning method, the coupling weight is corrected by the BP method using E (u _PI ) = ½u _PI ² that is the square of u _PI as an expression function.

(Step S303)
Next, the control unit 11 performs membership function setting processing of the FB-FN unit 130 of the ICS unit 10.
The membership function setting process will be described with reference to FIG.
As described above, the membership function is used for the antecedent part of the fuzzy neural network in the fuzzy PD part 131 and the fuzzy PI part 132. In this embodiment, a Gaussian function is used here for this membership function.
As a result of learning by the above-described FNN learning process, the shape of the membership function is determined, so that the parameters are automatically adjusted.
The lower part of FIGS. 14 and 12A is an example of this adjusted membership function.

(Step S304)
Next, the control unit 11 performs a fuzzy rule setting process on the FB-FN unit 130 of the ICS unit 10.
FIG. 15 is a conceptual diagram illustrating a fuzzy rule setting method of the fuzzy PD unit 131 and the fuzzy PI unit 132.
In the fuzzy neural network according to the first embodiment of the present invention, in order to adaptively adjust the control gain, a fuzzy rule for adjusting the output corresponding to the error input amount is set.
That is, in order to arrange the gains of the FB-FN unit 130 in a matrix, the fuzzy rules shown below are set in the consequent units in the fuzzy PD unit 131 and the fuzzy PI unit 132.

More specifically, the output U _FN in FIG. 12B is a direct plant input. In the fuzzy part, the input space is divided into nine, and the fitness of the divided space is given by the following expression.

_{μ i = A i1 (e p} ) A i2 (e v)

Here, i = 1, 2,..., 9, i1, i2 = 1, 2, 3
A _ij is a membership function and represents a fuzzy variable.
A _i1 and A _i2 represent the degrees of conformity to Positive big, Small, and Negative big, respectively.
However, the fitness is normalized as shown in the following formula (6).

Then, in order to adaptively adjust the control gain, a fuzzy rule for adjusting the output corresponding to the error input amount is set.
Here, as shown in FIG. 12C, 9 is set as the fuzzy division, and the fuzzy rules are set as follows.

The i-th fuzzy rule is:

_{^{R i: IF e p is A}} i1 and e v is A i2 THEN y = fi (e p, e v)

Where i = 1, 2,. . . , 9, i ₁ , i ₂ = ₁ , _2, 3

It is.
The u _FN-PD output is obtained as the following equation (7).

Here, the u _FN-PID output which is the addition output of each u _FN is obtained by the following equation (8).

  Further, fuzzy inference is obtained by the following equation (9).

Here, f is represented by the following formula (10), assuming that u _FN-PID and u _FN are equal.

However, k _ij ≧ 0 (j = I, p, v) is a feedback gain described later.
Here, in the control method of Expression (10), an operation of switching between the fuzzy PD and the fuzzy PI by an error response is performed. At this time, in order to quickly bring the response close to the target value with respect to the influence of the control system due to disturbance or parameter fluctuation, as shown in the table in FIG. 15, the response to the influence of the control system due to disturbance or parameter fluctuation is promptly made. when is big, | to approach the target value, PID control when the both displacement error signal e _p and the speed error signal e _v is Small, an absolute value of e _p is not Small e _v in | e _v In order to reduce the damping and obtain a fast response, it is preferable to use the P control and the other are the PD control.
Each feedback constitutes a control law by learning in advance with a neural network in the subsequent stage. Each feedback gain is learned in advance by a neural network in the subsequent stage. A control law is comprised by such a structure. That is, a control law adapted to the error can be configured.
FIG. 16A shows a calculation method of the fuzzy neural network gain which is this feedback.

FIG. 16B shows a gain matrix of the FB-FN unit 130.
Since the feedback gain k _ij is learned for each fuzzy-divided error space, it plays a role as a dynamic compensator depending on the error amount. Thus, in order to adaptively adjust the control gain, the above-mentioned i-th fuzzy rule for adjusting the output in accordance with the error input amount is set.
That is, as an effect of fuzzy reasoning, when the error is small, the gain is low. However, the optimum gain adjustment is performed by learning a neural network. In other words, if position and speed errors occur due to modeling errors, friction, unknown disturbances, etc., a large gain is obtained, and a reduction in errors can be expected. Also, for large errors as part of the C ₇ of Fig. 16B to be described later, the effect of the feedback can be expected. Such a configuration can be expected to be robust against disturbances and parameter fluctuations.

FIG. 16C shows a block diagram of gain scheduling by a fuzzy neural network.
The gain scheduling determined by _{C i = μ i (k ip} e p + k iv + k iI ∫f ep dt) wherein each gain of C ₁ -C ₉ in FIG. 16B is calculated.

FIG. 16D shows a fuzzy PD gain matrix of the fuzzy PD unit 131. This is C _p and C _v gain which is the street calculation shown in FIG. 16A in divided input space into nine windows input for fuzzy PD. The control gain is adaptively adjusted to the input error.
FIG. 16E shows a fuzzy PD control surface of the fuzzy PD unit 131. This represents how the control output for the inputs of C _p and C _v changes for the fuzzy PD. According to the control surface of FIG. 16E, it can be seen that the control surface has robustness and robust stability against a large step-like disturbance.

[The basis for robust stability]
Here, the basis of the robust stability of the FB-FN unit 130 according to the first embodiment of the present invention will be described with reference to FIG.
The FB-FN unit 130 needs to design a control system that guarantees control performance in the fluctuation range, assuming that the parameter fluctuation is not minute but occurs in a certain finite range. Therefore, it is preferable to assume a large number of reference points that cover the parameter fluctuation range and to reduce the parameter sensitivity at this point. Further, the ratio for decreasing the sensitivity at each reference point can be set according to the system requirements by using a membership function.
In order to ensure the robustness of the error of the fuzzy neural network controller, it is necessary to minimize the parameter sensitivity to the fluctuation on the assumption that the parameter fluctuates around a certain reference point. FIG. 17 shows the membership function and the control surface of the fuzzy PI.
Here, as shown in the control surface of the classical PI in FIG. 17, when e or e. Perturbs in the direction of e or e. Increase.
On the other hand, in the control surface of the fuzzy PI of FIG. 17, when the change amount of e or e · which is a control error is small, the fuzzy PI is not so affected. In other words, the control surface of fuzzy PI is almost parallel to ee ·, and if perturbed, e and e · will only enter a transient region where two or more rules are activated, and the control output will be affected. do not do. That is, it turns out that the function which reduces a steady-state deviation is working.
This explains the basis for the robustness of the fuzzy PID controller.

(Step S305)
Next, the control unit 11 performs fuzzy inference processing on the FB-FN unit 130 of the ICS unit 10.
In this fuzzy inference process, a singleton output membership function described above can be used to interpolate a plurality of linear controllers applied to various operating conditions of a dynamic nonlinear system.
In this fuzzy inference process according to the present embodiment, a linear gain can be obtained by smoothly interpolating the input space to obtain an effective gain scheduler, and a nonlinear system can be modeled.
In other words, a fuzzy inference system is used to tune from input / output data using an algorithm such as the BP method. Here we use a hybrid learning algorithm to identify the parameters of the inference system. As described above, the least square method and the back-propagation gradient descent method (BP method) are applied to the nonlinear parameters.
Thus, the FB-FN learning process is terminated.

[Example of fuzzy PD unit 131]
The following is an example of an M file of FUZZY_PD, which is a fuzzy PD section 131 created and learned:

[Input1]
Name = 'input1'
Range = [-0.2868 0.2804]
NumMFs = 3
MF1 = 'in1mf1': 'gauss2mf', [0.0478592834462299 -0.371325 0.0319561439415608 -0.217974284608918]
MF2 = 'in1mf2': 'gauss2mf', [0.0608664733050278 -0.0978076222858487 0.0555142362386605 0.0852892043051262]
MF3 = 'in1mf3': 'gauss2mf', [0.0138714984932358 0.20884527353354 0.0478592834462299 0.361225]

[Input2]
Name = 'input2'
Range = [-1.0472 1.0472]
NumMFs = 3
MF1 = 'in2mf1': 'gauss2mf', [0.177881957852323 -1.36136 0.178084149908005 -0.73222056738112]
MF2 = 'in2mf2': 'gauss2mf', [0.150801124394399 -0.302584545101134 0.1487581237134 0.303271696091688]
MF3 = 'in2mf3': 'gauss2mf', [0.178465822406237 0.732532304938343 0.177881957852323 1.36136]

[Output1]
Name = 'output'
Range = [-9.121 8.9183]
NumMFs = 9
MF1 = 'out1mf1': 'linear', [29.9502149922501 0.502141541225866 -0.0114363187815991]
MF2 = 'out1mf2': 'linear', [0.699802412554496 2.88172200515163 -3.69267448235734]
MF3 = 'out1mf3': 'linear', [2.42419964633644e-009 1.94710601114873e-009 -4.46135969104574e-008]
MF4 = 'out1mf4': 'linear', [29.8646225385097 0.522490654609605 0.00109243397029133]
MF5 = 'out1mf5': 'linear', [29.9872760919535 0.502942439065712 -0.000137735051368041]
MF6 = 'out1mf6': 'linear', [29.9929707105531 0.500833500934634 0.000342857488096674]
MF7 = 'out1mf7': 'linear', [1.73151335707348e-014 7.40165806871063e-014 9.36256379708729e-014]
MF8 = 'out1mf8': 'linear', [0.60924632277904 2.59412276828522 3.30948675859052]
MF9 = 'out1mf9': 'linear', [29.9947186987529 0.494510357242337 0.00687566067744605]


[Rules]
1 1, 1 (1): 1
1 2, 2 (1): 1
1 3, 3 (1): 1
2 1, 4 (1): 1
2 2, 5 (1): 1
2 3, 6 (1): 1
3 1, 7 (1): 1
3 2, 8 (1): 1
3 3, 9 (1): 1

[File example of fuzzy PI section 132]
The following also shows an example of an M file of FUZZY_PI, which is a fuzzy PI section 132 created and learned:

[System]
Name = 'FNN_PI'
Type = 'sugeno'
Version = 2.0
NumInputs = 2
NumOutputs = 1
NumRules = 9
AndMethod = 'prod'
OrMethod = 'probor'
ImpMethod = 'prod'
AggMethod = 'sum'
DefuzzMethod = 'wtaver'

[Input1]
Name = 'input1'
Range = [-0.2275 0.2108]
NumMFs = 3
MF1 = 'in1mf1': 'gauss2mf', [0.0372257745066239 -0.293245 0.0620222208404797 -0.145616701160321]
MF2 = 'in1mf2': 'gauss2mf', [0.0509989814349655 -0.0938664146504105 0.061773470568765 0.0711977845073082]
MF3 = 'in1mf3': 'gauss2mf', [0.0451928100610181 0.122287868166051 0.0372257745066239 0.276545]

[Input2]
Name = 'input2'
Range = [-1.0472 1.0472]
NumMFs = 3
MF1 = 'in2mf1': 'gauss2mf', [0.177881957852323 -1.36136 0.194858188187073 -0.725976617084914]
MF2 = 'in2mf2': 'gauss2mf', [0.184472713520267 -0.318507248879403 0.182769544067602 0.316484420470913]
MF3 = 'in2mf3': 'gauss2mf', [0.189973693378099 0.727208320141822 0.177881957852323 1.36136]

[Output1]
Name = 'output'
Range = [-7.2538 6.6378]
NumMFs = 9
MF1 = 'out1mf1': 'linear', [43.4662552481124 15.4046594018716 18.8528571837822]
MF2 = 'out1mf2': 'linear', [-72.3728068591604 -8.21719705183032 -35.8241392764794]
MF3 = 'out1mf3': 'linear', [-0.746178588917306 -0.0969956028725818 -409.511443589854]
MF4 = 'out1mf4': 'linear', [43.6763652955075 4.47961924229057 -1.2979557058011]
MF5 = 'out1mf5': 'linear', [-188.845013085459 16.3752420684739 2.26580569880905]
MF6 = 'out1mf6': 'linear', [52.6504877364809 13.6637475028015 -2.81582668956971]
MF7 = 'out1mf7': 'linear', [1.36664656988905 -101.944112740175 561.037900146712]
MF8 = 'out1mf8': 'linear', [-37.7754372896572 -2.37931907009803 20.5924135196645]
MF9 = 'out1mf9': 'linear', [77.3693139104994 7.01860325733576 -18.4498686710174]

[Rules]
1 1, 1 (1): 1
1 2, 2 (1): 1
1 3, 3 (1): 1
2 1, 4 (1): 1
2 2, 5 (1): 1
2 3, 6 (1): 1
3 1, 7 (1): 1
3 2, 8 (1): 1
3 3, 9 (1): 1

<Details of software servo control processing>
Next, the software servo control process according to the first embodiment of the present invention will be described with reference to FIG.
In the software servo control processing according to the first embodiment of the present invention, a DC motor driving PWM (pulse width modulation) module and an up / down counter module of a position detecting encoder are created by software and realized. As a result, it is possible to save hardware components and to take measures against delays in operation time and noise caused by hardware.
The PWM module of the PWM unit 150 is softwareized by design. That is, in this embodiment, a logic circuit is created by MATLAB / Simulink and a PWM module is realized. That is, as shown in FIG. 18, the motor 230 can be driven by the control unit 11 changing and outputting the signal amplitude. Examples of this PWM module are shown in FIGS. 20A to 20C of Example 1 described later.
The COUNTER (counter) module is also made into software by design. That is, in this embodiment, a logic circuit is created by MATLAB / Simulink and a COUNTER module is realized. Examples of this COUNTER module are shown in FIGS. 20F to 20H of Example 1 described later.
In this way, intelligent and robust control can be realized by increasing the processing speed of the intelligent controller.

With the configuration described above, the following effects can be obtained.
First, there have been many problems with electronic control based on modern control theory that has been a problem in the industry. This is because the control gain is a fixed gain.
The control system X according to the first embodiment of the present invention can only solve this problem by using the ICS unit 10 which is an intelligent and robust embedded intelligent controller based on the next generation fuzzy neural network based control. it can.
The ICS unit 10 incorporates an FB-FN unit 130 that is a new intelligent robust gain compensator of fuzzy neural base control that adaptively variably adjusts the gain in response to a change in the target trajectory error. Thereby, sufficient target trajectory tracking control performance can be obtained with respect to system parameter fluctuations, sudden disturbances and target changes.

In addition, the control method of the control system X according to the first embodiment of the present invention can extend the conventional PID control method to a multivariable system.
That is, in the control method of the present embodiment, the nonlinear motion equation is directly used, linearization is performed by the neural network of the FF-NN unit 120, and disturbance suppression and non-interference are performed by the fuzzy neural network of the FB-FN unit 130. To do.
The control method of this embodiment devised a control law in which the gain adaptively changes with time so that the nonlinear system whose parameter changes with time always becomes critical damping, and the mutual interference of each axis in the multivariable system is devised. Make it non-interfering. As a result, multivariable control by intelligent control can be performed, and complete follow-up control of the manipulator can be realized.

In addition, the control system X according to the present embodiment can adjust the control gain variably and adapt it to unknown elements by combining fuzzy reasoning ability and neural network learning ability.
Actually, when the experiment comparing the fuzzy neural network control system constructed in Example 1 below with the conventional PID control and FF-NN control only system, the effectiveness of the fuzzy neural network control is high. It was.

Conventionally, there are many methods for designing manipulators, in which even if the controlled object is a multi-degree-of-freedom system, the control system is designed using an approximation formula and an independent control system is constructed for each joint axis. It was adopted in the product.
The control system X according to the present embodiment can solve this unsolved problem of multi-degree-of-freedom system design by intelligent control.
That is, in the control system X according to the present embodiment, a variable feedback gain corresponding to the position and force error is set in a fuzzy-divided space in order to construct a control system capable of robust and automatic gain adjustment. The control system can be used to design a multi-degree-of-freedom system.

Further, the control system X according to the present embodiment can automatically set a gain by learning a fuzzy neural network, and can realize robustness by intelligent control.
This realizes a practical intelligent and robust control system for manipulators and multivariable mechanical systems, and sets target values for robots, automobiles, electric bicycles, ships, aircraft, spacecrafts, rockets, temperature control, electric control, etc. Thus, the present invention can be applied to any device to be controlled for which conventional control for controlling is used.
In addition, the control system X according to the present embodiment can compensate for non-linear deviation by a fuzzy neural network that adaptively adjusts the control gain when displacement and speed control errors occur due to modeling errors, friction, unknown disturbances, etc. become.
In other words, an intelligent and robust control method using a fuzzy neural network is established, and a practical intelligent control system for a multivariable mechanical system can be realized.

Further, the control system X according to the present embodiment uses a model-based system development using an RTW (Real-Time Workshop) by a control device 3 such as a normal PC using an embedded computer as the intelligent controller unit 1. It is possible to develop a system in a short period of time.
Furthermore, by using software such as MATLAB / Simulink, the simulation can be easily performed, so that a great effect can be obtained in the cost reduction and time reduction of the system development.

In addition, the control system X according to the present embodiment incorporates a new intelligent robust gain compensator (fuzzy neural base control) that variably adjusts the gain adaptively to changes in the target trajectory error. Sufficient target trajectory tracking control performance can be obtained for fluctuations, disturbances, and target changes. Specifically, in order to construct a control system capable of robust and automatic gain adjustment, a control system in which variable feedback gains according to position and force errors are set in a fuzzy divided space is configured. .
In the control system X according to the present embodiment, the control system has a two-degree-of-freedom control configuration, the feedforward unit is configured by a neural network, and the plant characteristics are learned by parameters of displacement, speed, and acceleration, The inverse characteristic is acquired and the system is linearized.
The feedback unit is a compensator that reduces non-linear errors using a fuzzy neural network, and realizes perfect tracking performance.
These control systems are characterized in that they are executed in real time (real time) processing, for example, with a sampling time of 1 millisecond.

[Example 1]
Here, with reference to FIG. 19A to FIG. 21B, a description will be given of a first embodiment in which intelligent and robust control is actually performed using the configuration and each process of the ICS unit 10 described above.
In the first embodiment, the position of the controlled object 2 is controlled by the intelligent controller unit 1 including the ICS unit 10 that is an embedded computer. As the control object 2, a moving table driven by a DC motor was used.
The OS of the ICS unit 10 uses Linux (registered trademark) + Xenomai (trademark), and can be controlled in real time. The control device 3 which is a personal computer is directly connected to the intelligent controller unit 1 by a LAN cross cable and communicates by Ethernet (registered trademark).
The intelligent controller unit 1 sends a control signal from the D / A converter of the intelligent controller unit 1. This control signal is output to the motor 230 via the motor drive driver 20. When the motor 230 rotates, a pulse wave is output from the encoder 220. This pulse wave is acquired by the intelligent controller unit 1 by the A / D converter of the intelligent controller unit 1. The pulse wave is divided into 1/8, and the position information of the motor 230 can be fed back by counting the pulse wave.

In the present embodiment, TA7281P manufactured by Toshiba Semiconductor was used as the motor driver 20. As the motor 230, TG-47C-SM manufactured by Tsukasa Electric Co., Ltd. was used. TG-47C-SM has a rated voltage of 12 [V], a rated speed of 4400 [rpm], a rated current of 160 [mA], a no-load speed of 5250 [rpm], and a no-load current of 60 [mA].
As the encoder 220, a rotary encoder attached to the TG-47C-SM was used. The rotary encoder outputs 8 pulses per rotation.

(experimental method)
In this embodiment, a model for a control system X of a fuzzy neural network is constructed by MATLAB / Simulink. For comparison, a model of a conventional PID control and NN-PID-only control system was similarly constructed.
Models of these control systems are shown in FIGS. 19A to 19C. Moreover, the model figure of each site | part of this each control system is shown to FIG. 20A-FIG. 20H.

More specifically, FIG. 19A is a model diagram of conventional PID control. For the PID parameters of this model, the motor driver 20 was controlled with a proportional gain P = 12, an integral gain I = 0, and a differential gain D = 0.
FIG. 19B is a model diagram in the case of only NN-PID control. Learning of the neural network of the FF-NN unit 120 is performed based on PID control data, and learning is performed by the BP method using the displacement, speed, and acceleration of the command signal as inputs and the operation amount of the PID as a teacher signal. The neural network obtained in this way is fed forward to the manipulated variable of PID. PID parameters are the same as PID control.
FIG. 19C is a model diagram of fuzzy neural network control according to the present embodiment. FNN learning is performed based on the PID control data shown in FIG. 19A, with displacement error and speed error as inputs, and learning is performed using a combination of the BP method and the least squares method with the PID operation amount as a teacher signal. It was. At this time, the parameters of the NN-PID control in FIG. 19B were also used in the control system of the fuzzy neural network.

The control object 2 is a moving table that is moved by one DC motor 230, and the position is controlled by the constructed control system. The command signal was a sine wave with an amplitude of 5 and a period of 30 [s], and the motor was controlled with a sampling time of 1 [msec].
In addition, as a comparative operation test, performance was compared by executing three control systems with and without disturbance. The disturbance was reproduced on the model, and a value of +5 was added to the manipulated variable 10 seconds after execution.

(Experimental result)
The experimental results will be described with reference to FIGS. 21A and 21B.
FIG. 21A shows the result of control using a conventional PID control model (comparative example). In the conventional PID control, when there is no disturbance, there is a time delay even in the position control with one degree of freedom, and a slight error is observed. Further, in the conventional PID control, when a disturbance occurs, an error appears as a steady state difference, and the output is shifted upward according to the signal of the disturbance. Note that the control of only NN-PID in FIG. 19B (not shown) was able to eliminate the time delay somewhat by feedforward compensation. However, in the case of only NN-PID, when there is no disturbance, the control signal does not deviate so much, but when there is a disturbance, the result is the same as the conventional PID control, which is not robust and insufficient. .
FIG. 21B shows the result of using the fuzzy neural network control according to the present embodiment. In fuzzy neural network control, it is considered that it can cope with nonlinear elements to be controlled. That is, in the fuzzy neural network control, the control gain can be adaptively adjusted because the command signal can be tracked both with and without disturbance. Therefore, fuzzy neural network control is a robust control method and is effective for disturbances and nonlinear elements of control equipment.

<Second Embodiment>
[Configuration of Control System Y According to Second Embodiment of the Present Invention]
Next, a control system Y according to a second embodiment of the present invention will be described with reference to FIGS.
The control system Y of this embodiment is a “position / force hybrid” that continuously switches between position control and force control by “intelligent impedance control” using the intelligent robust control according to the first embodiment described above. By performing “/ compliance control”, a flexible motion characteristic can be realized in a robot or the like. As a result, the assembly work can be autonomously performed while adapting the robot or the like to the environmental characteristics.

First, the configuration of the control system Y according to the second embodiment will be described with reference to FIG. In FIG. 22, the same reference numerals as those in the control system X according to the first embodiment in FIG.
The intelligent controller unit 5 is an embedded computer for use in intelligent impedance control, like the ICS unit 10 according to the first embodiment. The detailed configuration of the intelligent controller unit 5 will be described later.
Similar to the control target 2 according to the first embodiment, the control target 6 is a specific control target part that is affected by a disturbance and requires position control and force control. In the present embodiment, an example in which an articulated arm robot is mainly used as the control target 6 will be described. The control target 6 is not limited to the robot, and any control target 6 can be used as long as it is a control target to which force is applied in various operations such as gripping or pressing an object.
The image / sound input unit 70 is a part for inputting an image and sound, which includes image pickup means such as a CCD or CMOS element, and sound input means such as a condenser microphone and an A / D converter. As the image / sound input unit 70, a general web camera or the like that can recognize the operation position of the robot 61 and does not need to have a high image resolution can be used.

<Configuration of control object 6>
The control target 6 includes a servo controller 60, a robot 61, and a force sensor 62.
The servo controller 60 is a circuit for controlling the operation of the robot 61 by various commands, like the motor drive driver 20 according to the first embodiment. The servo controller 60 receives a position command θ (k) related to inverse kinematics from the intelligent controller unit 5 and outputs an electrical signal related to a joint angle θ, which is an angle of each joint such as a hand and an arm, to the robot 61. . Further, the servo controller 60 calculates the position and angle of the corner joint from the signal related to the torque τ from the robot 61, and transmits it to the intelligent controller unit 5 as the displacement q.

The robot 61 is a multi-joint robot or the like that includes a one-axis robot hand and an arm that grips an object and moves a position in a three-dimensional space by four axes.
The robot 61 can perform various operations such as gripping an object, filling the object, creating a screw hole, screwing, and welding.
The robot 61 outputs torque τ, which is data related to the joint positions and loads of the motors and actuators of each joint, to the servo controller. As the robot 61, it is possible to use an arm robot having a number of axes of 5 and a movement accuracy of about 0.1 mm, for example, a micro robot “MOVEMASTER RV-M2” manufactured by Mitsubishi Electric Corporation.

  As the force sensor 62, for example, a six-axis force sensor can be used. For example, the force sensor 62 can detect a force and a moment for each of the X, Y, and Z axes. As the force sensor 62, for example, “6-axis force sensor WDF-6A200-4-UG2” manufactured by Wako Tech Co., Ltd. can be used.

(Configuration of ICS unit 50)
Next, a detailed control configuration of the ICS unit 50 will be described with reference to FIG.
In FIG. 23, the site | part of the same code | symbol as the ICS part 10 (FIG. 3) which concerns on 1st Embodiment has shown the same structure.
Specifically, the ICS unit 50 includes two systems of a position command system 51 (position control system fuzzy neural network) and a force command system 52 (force control system fuzzy neural network) as parts related to the processing of the FNN. In addition, a switching unit 135 that continuously switches these is provided.
In addition to this, the ICS unit 50 includes an image recognition unit 710 and a voice recognition unit 720.

The position command system 51 is a part that performs FNN processing for use in position control of the controlled object 6, similarly to the intelligent robust control in the ICS unit 10 (FIG. 3) according to the first embodiment.
As in the first embodiment, the position command system 51 uses the displacement q obtained from the position command θ (k) to the servo controller 60, and uses the displacement k (y) and speed related to inverse kinematics. Each signal of (dy / dx) and acceleration (d ² y / dt ² ) is calculated, and FNN processing is performed. The position command θ (k) is a value that expresses the coordinates for moving the object gripped by the robot 61 by, for example, the angle and position of each joint angle of the robot 61.

The force command system 52 is a part used for force control of the control target 6. The force command system 52 performs FNN processing using force data f obtained from the force sensor 62 of the control target 6 with the same configuration as the position command system 51.
Specifically, the force command system 52 receives signals (force (y), viscosity (dy / dx), stiffness (d ² y / dt ² )) relating to forward kinematics from the force data f from the force sensor 62. In this case, FNN processing is performed by calculating instead of displacement, velocity, and acceleration.
For this reason, the force command system 52 includes a PID unit 510 corresponding to the PID unit 110 of the position command system 51, a PD compensator unit 511 corresponding to the PD compensator unit 111, and an FF-NN unit corresponding to the FF-NN unit 120. 520 (feed forward means), FB-FN section 530 (feedback means) corresponding to the FB-FN section 130, fuzzy PD section 531 corresponding to the fuzzy PD section 131, fuzzy PI section 532 corresponding to the fuzzy PI section 132, difference A difference calculating unit 540 corresponding to the unit 140, a viscosity calculating unit 541 corresponding to the speed calculating unit 141, and a stiffness calculating unit 542 corresponding to the acceleration calculating unit 142.
The viscosity calculation unit 541 is a part that calculates the viscosity from the force f. This viscosity can be calculated from a 6 × 6 matrix when the force sensor 62 is a six-axis force sensor as described later.
The rigidity calculation unit 542 is a part that calculates rigidity from the force f. This stiffness can also be calculated from a 6 × 6 matrix.

The switching unit 135 is a part that continuously adjusts the output of the force control system 52 by position / force hybrid / compliance control in order to switch from position control to force control.
The switching unit 135 can change the output of the force control system 52 according to the value of the stiffness K _h (parameter α) obtained from the force f. This realizes a hybrid position / force hybrid / compliance control that continuously switches from position control to force control. Details of this position / force hybrid / compliance control will be described later.
In addition to the “normal mode” in which the output of the force control is changed depending on the value of the stiffness K _h (parameter α), the switching unit 135 receives the stiffness K _h (parameter Regardless of the value of α), it is possible to switch to the “disengagement mode” in which force control and position control are performed. With these modes, even if the attachment of components or the like fails, it is possible to flexibly cope with them.
Note that the switching unit 135 may adjust the output of the force command system 52 based on the acceleration value calculated from the displacement q. In addition, although the switching unit 135 can be changed to include not only the force control system 52 but also the output of the position control system 51, the change of the output of the force control system 52 is better when the position / force is hybrid / compliance. Control can be suitably realized.

The image recognition unit 710 is an operation part for performing image processing of the image data from the image / sound input unit 70 and setting a position command θ (k) for moving an object held by the robot 61. The image recognizing unit 710 recognizes a predetermined shape set by the control device 3 at the place where the object is placed, and sets the coordinates to be moved. On this basis, the control unit 11 (FIG. 22) can set θ (k) from the coordinates to be moved.
The coordinates to be moved can be set by the control device 3 or can be calculated and set by the difference unit 140 of the ICS unit 50.

  The voice recognition unit 720 is a part that recognizes voice by inputting a signal from the image / sound input unit 70. Thereby, the voice recognition unit 720 may transmit the recognition result to the control device 3 and set the position command θ (k).

[Development / learning process of ICS unit 50 of control system X]
Next, with reference to FIG. 24 and FIG. 25, in the control system Y according to the second embodiment of the present invention, the development / learning for learning the ICS unit 50 and optimizing the control of the controlled object 6 Processing will be described.
First, as the development / learning process of the ICS unit 50, the same process as the development / learning process (FIG. 5) according to the first embodiment is performed.
Then, the same processing as the FF-NN reverse dynamics processing in step S102 of the first embodiment is performed on the position command system 51, and then the dual command FF-NN (feed) is performed on the force command system 52. Forward neural network) Performs reverse dynamics processing.
The details of this dual-system FF-NN inverse dynamics process will be described below with reference to the flowchart of FIG.

<Details of dual-line FF-NN reverse dynamics processing>
(Step S211)
First, the control unit 11 of the intelligent controller unit 5 performs a position command system FF-NN preparation process. This process is the same as the FF-NN preparation process (FIG. 8) in step S201 according to the first embodiment, using the displacement q of the control target 6 instead of the displacement r of the control target 2, To the directive.

(Step S212)
Next, the control unit 11 of the intelligent controller unit 5 performs a position command system FF-NN learning process. Also for this process, the same process as the FF-NN learning process (FIG. 8) in step S202 according to the first embodiment is performed using the displacement q instead of the displacement r.

(Step S213)
Next, the control unit 11 of the intelligent controller unit 5 performs a force command system FF-NN preparation process using the PD compensator unit 511 of the ICS unit 50 and the like.
Here, as in step S211, the characteristics of the controlled object 6 are grasped, and learning for linearizing the characteristics of the control system Y is performed.
Specifically, using a PD compensator unit 511 similar to the PD compensator unit 111 which is a conventional classical PD compensator, a test for grasping the characteristics of the controlled object 6 is performed, and the force f corresponding to the force command is determined. Each signal of force (y), viscosity (dy / dx), and rigidity (d ² y / dt ² ) corresponding to each axis is acquired.
In this acquisition, the viscosity is calculated by the viscosity calculation unit 541 of the difference unit 540 and the rigidity is calculated by the rigidity calculation unit 542.
At this time, the PD gain is set to a small value at which the control target 6 can move.
Note that the viscosity and stiffness signals may be calculated by numerical calculation processing offline after obtaining the signal of the force f.

More specifically, the control unit 11 inputs force, viscosity, and rigidity data acquired in response to the above-described force command as an input of the neural network. As the data, as in the above-described position command system FF-NN learning process, it can be suitably learned if it is about 100 per second. That is, a sampling frequency of about 10 msec is suitable.
Further, the control unit 11 can prepare software necessary for learning of the neural network by inputting a neural network generation command for MATLAB, for example, as in the first embodiment.

At this time, the control unit 11 can prepare a neural network generation command with P = force, viscosity, stiffness, and T = teacher signal.
Further, the teacher signal can be the operation signal y, and the time required for grasping the characteristics can be set to about three cycles.

  When learning of the neural network is completed by MATLAB, an FF-NN unit 520 that is a block of the neural network is generated. This FF-NN unit 520 is used as a control block.

(Step S214)
Next, the control unit 11 of the intelligent controller unit 5 performs a force command system FF-NN learning process.
This process is performed for the force command using the force data f of the control target 6 as in step S212 described above.
Specifically, as this FB-FN learning process, as in the first embodiment, FB-FN learning related to a position command and a force command using a singleton output membership function is performed.
This is an inference process that interpolates a plurality of linear controllers applied to various operating conditions of a dynamic nonlinear system. By this inference, a linear gain can be obtained by smoothly interpolating the input space, and an effective gain scheduler can be obtained to model a nonlinear system.
Specifically, the FB-FN unit 130, which is a fuzzy PID controller, learns using the position command or data described above, and performs feedback compensation related to the position. Similarly, feedback compensation for force is performed using the FB-FN unit 530 which is a fuzzy PID controller. Thereby, the acquired nonlinear error signal is reduced.
More specifically, as in the first embodiment described above, tuning is performed from input / output data using back-propagation. At this time, a hybrid learning algorithm for identifying parameters of the inference system is used, and learning is performed on the nonlinear parameters using the least square method and the back-propagation gradient descent method.
Thus, the two-line FF-NN inverse dynamics process according to the second embodiment of the present invention is completed.

<Assembly processing according to the second embodiment of the present invention>
Next, referring to FIG. 26 and FIG. 27, the position / force hybrid / compliance control is performed using the intelligent impedance control according to the second embodiment of the present invention, and the parts are assembled to the robot 61. Details of the assembly process to be performed will be described.

Conventionally, the assembly work of an industrial robot involves contact between parts when the parts are assembled, and thus cannot be handled by simple position control, and force control is required. However, in the industry, there were few application examples of force control. This is because the force sensor is expensive and it is difficult to adjust parameters for force control.
Although there is an example in which compliance (softness) control is put to practical use as conventional force control, there is a problem that switching between position control and force control cannot be performed smoothly.
On the other hand, the control system Y of this embodiment uses the intelligent controller unit 5 to smoothly switch from position control to position / force hybrid / compliance control to force control by intelligent impedance control. It is possible to make it happen.

First, a specific example of position / force hybrid / compliance control according to the second embodiment of the present invention will be described with reference to FIG.
FIG. 26A shows the concept of attaching a pin 800, which is a component such as a work pin, to the attachment hole 810 by the robot 61 having the force sensor 62 using the position / force hybrid / compliance control of the present embodiment. Show. In the task operation of attachment, the intelligent controller unit 5 of the present embodiment first moves the pin 800 according to a position command to contact the attachment hole 810. Thereafter, the intelligent controller unit 5 adjusts the mounting of the pin 800 by position / force hybrid / compliance control that simultaneously performs position control and force control. After the adjustment of the position and force is completed, the intelligent controller unit 5 performs final attachment mainly by pressing only with a force command. That is, position control and force control can be continuously and smoothly switched by hybrid position / force / compliance control.
The intelligent controller unit 5 holds the pin 800 by the image / sound input unit 70, recognizes the position of the mounting hole 810, performs position control, and moves the robot 61 to the mounting hole 810. Here, when the pin 800 contacts the mounting hole 810, even if the position is slightly shifted, the intelligent controller unit 5 automatically switches to force control and can be mounted smoothly.
FIG. 26B is a conceptual diagram of a compliance (softness) model related to work pins when the force sensor 62 has six axes. Details of this compliance will be described later.
Hereinafter, with reference to FIG. 27, the details of the assembly process using the position control, the position / force hybrid / compliance control, and the force control according to the present embodiment will be described step by step.

(Step S501)
First, the control unit 11 of the intelligent controller unit 5 performs an approaching contact operation control process mainly using the position command system 51 of the ICS unit 50.
Specifically, the control unit 11 recognizes the position of the pin 800 by the image / sound input unit 70 in accordance with an instruction from the control device 3 and causes the robot 61 to hold it.
Thereafter, the control unit 11 controls the position of the robot 61 holding the pin 800 by the displacement q in the same manner as in the above-described software servo control process, and moves the robot 61 to the mounting hole 810 serving as the mounting position. At this time, the output from the force sensor 62 is almost only the output of the force f related to gripping.
When the pin 800 comes into contact with the mounting hole 810, the control unit 11 can detect the force f applied to the arm of the robot 61 or the robot hand.

(Step S502)
Next, the control unit 11 of the intelligent controller unit 5 performs a pressing operation control process for continuously switching from the position command system 51 of the ICS unit 50 to the force command system 52.
In this process, when the pin 800 comes into contact with the mounting hole 810, the control unit 11 obtains a compliance value 1 / K _d from the displacement q and the force f and inputs it to the switching unit 135.
Here, the control unit 11 can perform position / force hybrid / compliance control by continuously switching the position command system 51 to the force command system 52 by changing the compliance value 1 / _Kd . That is, the control unit 11 attaches the pin 800 by smoothly adjusting the position from the position control to the force control. As a result, it is possible to realize the same processing as that for adjusting the push-in by detecting a slight deviation from the position of the hole applied to the hand when the human inserts the pin into the hole.

  For the hybrid position / force hybrid / compliance control by the control unit 11, a model such as the following equation (11) is used.

In Expression (11), X: current angular joint position, X _d : position command θ (k), M _x : real inertia matrix, B _a : damper matrix, K _d : virtual spring matrix (virtual stiffness), F: force response, F _d : force command, S: switching matrix diag (S1,..., S6), K _f : force gain, E: unit matrix.
When the 6-axis force sensor 62 is used, S, M, B, and K are each a 6 × 6 matrix.

  Here, in the present embodiment, the following equation (12) is an intelligent impedance control equation corresponding to the intelligent robust control according to the first embodiment. The control unit 11 performs hybrid position / force / compliance control using this intelligent impedance control.

In the equation (12), M _x S ² is a predetermined value of the substantial rigidity set in advance according to the characteristics of the robot 61 and the pin 800 or the like.
B _a S is the virtual viscosity learned by the above-described force command system FF-NN learning process.
K _d is the virtual stiffness learned by the above-described force command system FF-NN learning process. A value obtained by dividing the virtual stiffness K _d by 1 and 1 / K _d is a compliance value indicating softness.

Here, the control unit 11 uses the switching unit 135 to calculate the stiffness K _h related to each axis from the force f measured by the force sensor 62. The rigidity K _h takes a value from 0 to K _hmax corresponding to the value of the virtual rigidity K _d described above.
Hereinafter, the value of the stiffness K _h is set as a parameter α for continuously switching between position control and force control. The control system Y of the present embodiment can continuously set three control modes of position control, force control, and position / force hybrid / compliance control only by adjusting this one parameter α.

The parameter α in the switching unit 135 is calculated as α = K _c / K _h + K _c .
K _h represents a virtual spring coefficient, and K _h = 0 to 4000 N / m.
Also, K _c is a constant, set to 1 / 10-1 / 20 of the K _h. That is, a value of about Kc = 100, 200, 500 N / m is used. The following example is set to 1/20 of the K _h, showing an example in which Kc = 200.
Further, when α = 1, force control is performed, and when α≈0, position control is performed. When 0 ≦ α <1, position and force compliance control is performed.
For example, when K _h = 4000 N / m, α = 200/4000 + 200≈0.05 is calculated, and position control is performed.
Further, in the case of K _h = 0 N / m, α = 200/0 + 200 = 1 is calculated, so that force control is performed.

More specifically, first, in the case of K _h = K _hmax (α = 0), the control unit 11 performs position control by setting the above equation (11) as the following equation (13). This corresponds to the stage of moving the pin 800 to the position of the mounting hole 810 by a position command.

Next, in the case of 0 <K _h <K _hmax (1>α> 0), the control unit 11 _changes the above formula (11) to the following formula (14) and performs hybrid position / force hybrid / compliance control. Do. This is the stage where the pin 800 starts to be pushed in while being adjusted by the force f even if the position is slightly shifted when the pin 800 is brought into contact with the mounting hole 810.

Finally, in the case of K _h = 0 (α = 1), the control unit 11 performs force control by setting the above equation (11) as the following equation (15). This corresponds to a stage where the pin 800 is pressed by a force command.

(Step S503)
Next, the control unit 11 performs an axis alignment mounting operation process mainly using the force command system 52 of the ICS unit 50.
Specifically, in the example of FIG. 26, if the above-described α is substantially 1 for a predetermined time, it is assumed that switching to force control has been completed, and the pin 800 is pushed into the mounting hole 810 as it is with force control. go.
At this time, the control unit 11 confirms whether the pin 800 is surely pushed into the mounting hole 810 from the displacement q.

(Step S504)
Here, the control unit 11 determines whether or not the pin 800 is attached in the attachment hole 810. Specifically, the control unit 11 determines that the force f is equal to or greater than a predetermined value, the change of the displacement q within a predetermined second is equal to or smaller than the predetermined value, and the displacement q corresponds to the coordinates of the mounting position after the mounting hole 810 is inserted. When the value is reached, it is determined as Yes as a successful attachment. In other cases, No is determined as an attachment failure.
In the case of Yes, the control part 11 advances a process to step S508.
In No, the control part 11 advances a process to step S505.
The control unit 11 may determine the attachment success from the image acquired by the audio / video input unit 70 or the like.

(Step S505)
When the attachment of the pin 800 fails, the control unit 11 performs a force release operation process.
Specifically, the control unit 11 cancels the force control in order to adjust the attachment position. Therefore, the switching unit 135 is instructed to enter the separation mode in which position control is performed even when the force f is equal to or greater than a predetermined value.

(Step S506)
Next, the control unit 11 performs a detachment operation process.
In this process, the control unit 11 issues a position command related to a new coordinate, for example, pulling up the pin 800.
At this time, it is preferable that the control unit 11 moves the pin 800 by using both position control and position / force hybrid / compliance control. Thereby, even when the mounting position is shifted and the pin 800 is caught in the hole at an unexpected angle, it can be removed smoothly.

(Step S507)
Next, the control unit 11 performs a rotation operation process.
Specifically, when the robot 61 reaches a new coordinate, the control unit 11 changes the switching unit 135 from “leave mode” to “normal mode”.
Then, the control unit 11 finely adjusts the coordinates related to the attachment position according to a predetermined pattern. For this reason, the position command θ (k) is rotated by a predetermined angle.
Then, the control part 11 returns a process to step S501, and retries attachment work.
If the attachment fails even after performing a predetermined number of retries, the control unit 11 can perform error processing.

(Step S508)
When the attachment of the pin 800 is successful, the control unit 11 performs an assembly operation process.
In this process, the control unit 11 performs final assembly of parts such as screwing, welding, rivet shooting, and further pressing to insert and locked the pin 800 inserted and locked. Perform this operation.
This completes the assembly process.

With the configuration described above, the following effects can be obtained.
In recent years, there has been a long-awaited demand for a control device that can flexibly respond to environmental changes in the industrial world.
Here, conventionally, in domestic and overseas robots and the like, position and force control are separately performed. However, since it is difficult to adjust optimal parameters for force control, there is a problem that switching between position and force control cannot be performed smoothly.
The control system Y according to the second embodiment of the present invention can switch between position control and force control smoothly and continuously by position / force hybrid / compliance control by intelligent impedance control. Become. In the intelligent impedance control, various optimum parameters can be automatically set as in the intelligent robust control according to the first embodiment. Thereby, a robot having flexible motion characteristics can be realized. The robot can autonomously perform assembling work by a flexible operation while adapting to environmental characteristics.
In addition, the control system Y according to the present embodiment can automatically set impedance control parameters by learning a fuzzy neural network, and can realize robust impedance control by intelligent control. Thereby, the intelligent impedance control system for the robot manipulator and the mechanical system can be put into practical use, and can be widely applied to industrial devices related to control. In particular, a robust autonomous intelligent system that adapts to the characteristics of the environment and acquires skills through supple movements allows an integrated robot to perform various assembly tasks without the need for adjustments.
In summary, the control system Y according to the present embodiment can smoothly and continuously switch the position / force hybrid / compliance control mode by intelligent impedance control by control using a fuzzy neural network. In the control system Y, setting of various optimum parameters is automated by offline learning. As a result, the robot can realize flexible motion characteristics, and can assemble autonomously by a flexible operation while adapting to the environmental characteristics.

[Example 2]
Here, as Example 2, an experimental example using the above-described intelligent impedance control will be described with reference to FIGS. 28A to 30.
In the second embodiment, in order to show the effects of position control, position / force hybrid / compliance control, and force control switching by intelligent impedance control, the control target is a simplified device instead of the robot 61 of the control target 6. 7 was used.
According to FIG. 28A and FIG. 28B, the control object 7 of Example 2 is controlled by the strain gauge 63 instead of the force sensor 62, using a moving table having the same configuration as the control object 2 of the first embodiment. This is an example.
In other words, in the second embodiment, the object 801 is sandwiched by position control between the moving table 210 and the metal plate 211 installed at the end, and the load is not applied to the control target 7 in the object 801 as well. And force hybrid / compliance control to force control.
This corresponds to one-axis control when the object 801 is gripped by the robot hand of the robot 61.

The rest of the configuration including the ICS unit 50 is the same as that of the control system Y according to the second embodiment described above.
That is, the OS of the ICS unit 50 uses Linux (registered trademark) + Xenomai (trademark), and can be controlled in real time. The control device 3 which is a personal computer is directly connected to the intelligent controller unit 5 by a LAN cross cable and communicates by Ethernet (registered trademark).

(experimental method)
First, the difference in characteristics of the object 801 to be grasped was learned by the two-line FF-NN inverse dynamics process (FIG. 25) according to the second embodiment described above.
Here, attention is paid to the difference in softness of the object 801 in this embodiment, and test objects having different softness are prepared as the object 801.
Specifically, commercially available polystyrene (reinforcing material for desktop personal computers, manufactured by Epson), building blocks, sponges, springs, chalk, and stuffed animals (Mistletoe ™ 2001, manufactured by Ty INC) were prepared. All were easily available and used general materials.
The result of the experiment was displayed on the control device 3 in real time using the above-described VCDdesigner. As a result, a voltage (V) that is a command signal for operating the motor 230 and a voltage (V) output from the strain gauge 63 are displayed.

29A to 29C are graphs showing the results of this experiment.
29A shows a representative example of the experimental result of the sponge, FIG. 29B shows the experimental result of the building block, and FIG. 29C shows the representative example of the experimental result of the stuffed animal. Thus, in this embodiment, it can be seen that each object is stably gripped after a predetermined time.

FIG. 30 is a summary showing the time until each object 801 switches from position control to force control and the time until the pressure stabilizes to 0.5V.
FIG. 30 shows the compliance (softness) characteristic of each object 801. That is, a hard object 801 such as a building block converges quickly as the output of the strain gauge 63 increases. In addition, a soft object 801 such as a sponge has a small repulsive force, and the output of the strain gauge becomes small. As a result, the convergence is delayed. According to such a result, it can be understood that the conventional PID control cannot theoretically cope with the softness of the object. On the other hand, by using the intelligent impedance control of the present embodiment, it is possible to suitably grip a fragile object 801 such as a choke or a soft object 801. That is, by automatically adjusting the gain by the learning function of the FNN, a necessary pressure can be applied within a predetermined time and an object having various physical properties can be gripped.
Further, it can be confirmed that the control time varies depending on the difference in the compliance characteristics of the object.

  It should be noted that the configuration, reaction, and operation of the above-described embodiment are examples, and it is needless to say that they can be appropriately modified and executed without departing from the spirit of the present invention.

  The present invention provides a control means for performing intelligent robust control and intelligent impedance control, and thus can be used for controlling various devices and is industrially applicable.

1, 5 Intelligent controller unit 2, 6, 7 Control target 3 Control device 10, 50 ICS unit 11 Control unit 14 Auxiliary storage unit 15 Main storage unit 16 Logic circuit unit 17 I / F unit 18, 19 Connector 20 Motor drive driver 21 Table unit 51 Position command system 52 Force command system 60 Servo controller 61 Robot 62 Force sensor 63 Strain gauge 70 Image audio input unit 110, 510 PID unit 111, 511 PD compensator unit 120, 520 FF-NN unit 130, 530 FB- FN unit 131, 531 Fuzzy PD unit 132, 532 Fuzzy PI unit 135 Switching unit 140, 540 Difference unit 141 Speed calculation unit 142 Acceleration calculation unit 150 PWM unit 160 Counter unit 170 Output unit 180 Input unit 190 Control signal generation unit 195 Disturbance simulation G part 210 Table 211 Metal plate 220 Encoder 230 motor 541 viscosity calculating section 542 rigid calculation unit 710 image recognition unit 720 a voice recognition unit 800 pin 801 object 810 mounting hole X, Y control system

Claims (11)

In a built-in intelligent controller that performs position control with a position command and performs force control with a force command,
A fuzzy neural network of a position control system for performing the position control;
A fuzzy neural network of a force control system for performing the force control;
A switching unit that changes the output of the force control system according to the value of the force sensor to be controlled, and continuously switches between the position control and the force control by hybrid hybrid / compliance control of the position and force. Embedded intelligence controller. In the position control system, comprising a feedforward means for learning the parameters of the displacement, speed, and acceleration of the controlled object by a neural network, and acquiring a reverse characteristic of the controlled object characteristic,
2. The force control system according to claim 1, further comprising feedforward means for learning parameters of the force, viscosity, and rigidity of the control object using a neural network and acquiring a reverse characteristic of the characteristic of the control object. Embedded intelligence controller. The position control system fuzzy neural network and the force control system fuzzy neural network each include a plurality of feedback means which are compensators for reducing nonlinear errors,
The feedback means includes
Fuzzy PD means for performing PD (proportional, differential) compensation of the non-linear dynamic characteristics of the controlled object;
The embedded intelligent controller according to claim 1 or 2, further comprising fuzzy PI means for performing PI (proportional, integral) compensation of the nonlinear dynamic characteristics of the controlled object in parallel. The feedback means of the fuzzy neural network of the position control system and the feedback means of the fuzzy neural network of the force control system are:
The built-in intelligent controller according to claim 3, wherein a variable feedback gain is set in accordance with a position and a force error in a fuzzy divided space. The fuzzy neural network of the position control system and the fuzzy neural network of the force control system are:
The embedded intelligence controller according to claim 3 or 4, wherein real-time processing is performed at a predetermined sampling time. The embedded intelligent controller according to any one of claims 3 to 5, wherein the fuzzy neural network of the position control system and the fuzzy neural network of the force control system perform software PWM control. A control system comprising the embedded intelligent controller according to any one of claims 1 to 6. Embedded intelligence controller according to any one of claims 1 to 6;
A control system comprising: a robot having a force sensor as the control target. In a control program for an embedded intelligent controller that performs position control with a position command and performs force control with a force command, the embedded intelligent controller
The position control is performed by a fuzzy neural network of a position control system,
Force control is performed by a fuzzy neural network of force control system,
A control program that changes the output of the force control system according to the value of the force sensor to be controlled, and continuously switches between the position control and the force control by position / force hybrid / compliance control. In a recording medium in which a control program of a built-in intelligent controller that performs position control by a position command and performs force control by a force command is recorded on a control target, the built-in intelligent controller
The position control is performed by a fuzzy neural network of a position control system,
Force control is performed by a fuzzy neural network of force control system,
A computer-readable recording program that changes the output of the force control system according to the value of the force sensor to be controlled and continuously switches between the position control and force control by hybrid hybrid / compliance control of position and force Recording medium. In the control method of the built-in intelligent controller that performs position control with a position command and performs force control with a force command with respect to the control target, in the built-in intelligent controller,
The position control is performed by a fuzzy neural network of a position control system,
Force control is performed by a fuzzy neural network of force control system,
A control method comprising: changing an output of the force control system according to a value of a force sensor to be controlled, and continuously switching between the position control and the force control by position / force hybrid / compliance control.

Images