JPH0436621A - Device and method for monitoring equipment - Google Patents

Abstract

PURPOSE:To detect an abnormal sound which is hidden behind a background noise and to recognize the extent of the abnormality by inputting signals from a filter device by frequency bands, judging whether or not a sound is normal, and outputting the result. CONSTITUTION:The acceleration of acoustic vibration propagated to a Bessel wall 1 is converted by a sensor 2 into an electric signal, which is converted by a preamplifier 3 into a voltage signal. The voltage signal is an analog signal which has an audio frequency distribution and separated by a BPF 4 into 10 frequency bands obtained by dividing a signal of 0 - 20 KHz by 10 and the mean intensity of a sound in each frequency range is found by a squaring device 5 and an averaging circuit 6. The intensity of the sound at every 10 frequency bands are inputted to a hierarchic neural network 7 consisting of an S layer (sensor layer: input), an A layer (Association layer: intermediate layer), and an intermediate layer (Response layer: output layer); and the network 7 judges whether or not the sound is normal, the result is outputted, and the extent of the abnormality is recognized.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to a device monitoring device and a monitoring method for monitoring the presence or absence of an abnormality within the device, and particularly to a method for monitoring equipment that monitors the presence or absence of abnormalities within the device, and in particular, it relates to a device monitoring device and a monitoring method for monitoring the presence or absence of an abnormality within the device. The present invention relates to a device monitoring device and a monitoring method that analyze and determine whether or not an abnormality has occurred.

[Conventional technology]

If equipment parts fall off in various circulation systems consisting of pipes through which steam and liquid flow in nuclear reactor equipment, such as the nuclear reactor and the steam generation section connected to it, these fallen parts (loose parts) can damage various equipment. Problems may occur, such as being exposed to water or the flow of internal fluid being obstructed. Nuclear reactor technology requires greater safety than other technical fields, and it is necessary to reduce the occurrence of loose parts as much as possible.
Furthermore, if loose parts occur, it is necessary to quickly detect this fact and to accurately detect the location where the loose parts occur and the state of movement of the loose parts. For this reason, various countermeasures have been considered.

An example of this is a loose parts monitoring system (LPMS), the outline of which will be explained using FIGS. 9 to 11.

Loose parts in the reactor primary cooling system
When this occurs and collides with the walls that make up the system, producing abnormal noise, the piezoelectric accelerometers (sensors) 1 to 12 placed on the flow path walls at various locations in the system, as shown in Figure 9, detect the vibrations are detected and converted into electrical signals.

This electrical signal is amplified by line drivers 13-24 and then sent to loose parts detectors 25-36. Here, the magnitude of this signal is measured, and if the signal exceeds a specified value, a signal (voltage pulse) indicating loose parts detection is sent to the digital loose parts locator 37.

The digital loose parts locator 37 determines whether loose parts are generated correctly or incorrectly based on the time difference between the times when the loose parts detection signals of each channel are generated.

As a result, if it is judged to be appropriate, mask alarm 3
8 generates an alarm sound. Furthermore, the data recorder 39 is automatically activated at the same time as the alarm is generated, and the signal data of the loose bart collision sound is recorded and used for subsequent data analysis.

Using the data obtained as described above, the location of loose parts is detected using the method shown in FIG.

In other words, the difference in distance between the loose part P1 and the sensors A and C can be determined from the detection time difference between the detection signals of sensor A and sensor C, so the curve ff1 obtained based on this can be determined.
l (hyperbola with constant time difference), sensor B and sensor C
The curve obtained by the time difference of the detection signal in the detection signal of! 2 and detect the impact position P1 of the loose bar.

FIG. 11 shows a method for measuring the delay time of the detection signal in each channel having sensors A, B, C, etc. For example, in the signal waveform IW at sensor A and the signal waveform ■W2 at sensor C, This is determined by measuring the time difference t at the time when the signal magnitude suddenly increases.

In the conventional nuclear coupler loose parts monitoring system described above, the value of the impact waveform of loose parts detected by a detector (for example, an accelerometer) installed in each device of the primary cooling system such as a nuclear reactor or steam generator However, if the noise exceeds a certain level compared to the normal noise generated in each device (for example, the operating sound of a pump or motor, or the sound of fluid flowing, these are called background noise), a high alarm will be issued. I am planning to issue it. In addition, the loose parts monitoring device has a built-in device (hereinafter referred to as a locator) that determines whether the detection signals from the detectors attached to each device are correct. It had the ability to make decisions. The criteria for determining whether it is correct or incorrect are: (a) If the number of high alarm warnings received within 50 milliseconds (mmsec) is 1 or less, it is considered an erroneous signal.

The reason is that the speed of sound in steel is 3 m/millisecond, and sound travels a distance of 150 m in 50 msec. The distance between the detectors attached to each device is about 20 m at most, and if a loose barge is occurring, a large number of signals should be emitted from nearby detectors within a short period of time. (b) If three or more alarm signals are received within 0.5 milliseconds, it will be considered a false signal. Due to the arrangement of the detectors, it is highly unlikely that more than two alarms will be received within 0.5 milliseconds, and this is due to the electrical noise between the cables connecting each detector to the control board. This is because there is a high possibility that it is a pulse signal generated by induction.

In the above cases (a) and (b), it was decided to reset the signal conditioner and detector, and at the same time reset the central alarm, locator, etc., and cancel the data.

The signal that has cleared the correct/incorrect judgment section is further judged whether the signal is correct or incorrect in the pattern analysis described below.

In other words, detectors (sensors) such as accelerometers are attached to each primary system equipment such as a nuclear reactor or a heat exchanger at required positions, and when loose parts occur, depending on the location of the loose parts, The combination of sensors that will detect this will be determined. Therefore, the closest sensor will first emit a detection signal, and then the second closest sensor will emit a detection signal. Once the mounting position of each sensor is determined, the time difference (maximum delay time) between the signals transmitted by both sensors is determined.

That is, the sensor channel (CH) that first transmits the signal;
The combination of sensor channels to be transmitted next and the maximum delay time between both channels are stored in advance as an event table in the device's internal memory, and the actually received signal patterns are compared to determine whether the signals are correct or incorrect. be.

An example of an event table is shown in Table 1.

Table 1 Using the above event table, pattern analysis is performed according to the procedure shown in FIG.

In other words, when loose parts occur in the flow path of a nuclear power plant and collide with walls in piping, etc., the combination of multiple detectors that detect the collision sound and the detection delay time pattern depend on the location. is determined by a simulated impact test, and compared with the loose bar alarm generation channel when the abnormal noise occurs, and if it does not match, it is canceled as a false alarm.

Furthermore, as shown in Japanese Patent Application No. 63-040379 (Fig. 12), the presence or absence of loose parts is determined using a combination of three or more positive signals. When the value is greater than or equal to a certain specified value, the loose parts detector device transmits a loose parts detection signal. In other words, only abnormal sounds that are stronger than a certain ratio compared to background noise are detected, and this alarm causes F
ER (Fafse Alarm EvafuaLi
on and Recording Unit)
21, a validity check is performed to determine whether loose parts are generated overall. Impact sounds that are two to three times stronger than the hack ground noise are transmitted as loose bart detection signals, but abnormal sounds that are lower than that are overlooked.

Therefore, in places with large background noise, the detectable impact energy becomes large and the LP
The detection capability of MS is significantly reduced.

In particular, boiling water reactors have greater background noise than pressurized water reactors, and countermeasures are required to improve the ability to detect abnormal sounds. If an attempt is made to increase the detection ability by setting the specified value to less than twice the strength of the background noise, false alarms will occur frequently, the reliability of the LPMS will decrease, and the LPMS will not be able to fully demonstrate its function.

[Problem to be solved by the invention]

The above-mentioned conventional technology does not take into consideration fluctuations in background noise or minute impact sounds, and has a problem in detection ability.

An object of the present invention is to provide a <name> that can detect abnormal sounds buried in background noise and self-recognize the degree of abnormality.

[Means to solve the problem]

The above object is achieved by dividing the acoustic signal from the sensor into frequency ranges and inputting them into a neural network to self-recognize normal sounds, abnormal sounds, etc. That is,
The problem with the conventional technology is that multiple sensor devices are attached to equipment and receive acoustic vibrations generated within the equipment and transmit signals, and that it is determined whether or not the sound inside the equipment is normal based on the signals from the sensor devices. In the equipment monitoring device, the equipment monitoring device has a filter device that divides the signal from each sensor device into a plurality of frequency bands, and a device that inputs the signal from the filter device for each of the plurality of frequency bands, A device monitoring device characterized by comprising a neural network device that determines whether or not the device is normal and outputs the result; a device that is attached to the wall of the device and receives acoustic vibrations generated within the device and transmits a signal. A plurality of sensors and 1. In an equipment monitoring device that has a judgment device that determines whether or not the sound inside the equipment is normal based on signals from the sensor, a filter device that divides the signal from each sensor device into a plurality of frequency ranges, and a filter device that divides the signal from each sensor device into a plurality of frequency ranges. A neural network device is provided, which inputs the above-mentioned signals in each frequency range and determines whether the sound is a normal sound or an abnormal sound. A device characterized in that the neurons between each layer are connected by synapses, and the middle layer neurons have a logistic function as input/output conversion, and the output layer neurons have a piecewise linear function as human output conversion. A process of transmitting signals from a monitoring device and multiple sensors attached to the equipment wall based on the sound generated within the equipment, and a judgment device receiving the signals from each of the above sensors to determine whether the sound is normal or not. A device monitoring method comprising: a step of dividing a signal from each sensor into a plurality of frequency ranges using a band-pass filter; and a step of inputting the signal into a neural network determination device for each divided frequency range. A device monitoring method comprising: a pre-learned judgment function in a neural network judgment device; and a step of judging whether the sound is a normal sound or an abnormal sound based on the input signal and outputting a judgment result. Solved by

[Effect]

The abnormality monitoring device using acoustic recognition according to the present invention can detect minute abnormal sounds by configuring a neural network optimally, learning normal sounds, and self-recognizing them as "normal." Moreover, false alarms are no longer issued.

For sound input, by dividing the frequency range into multiple components and using the sound intensity, it is possible to determine the degree of slight abnormality even if the overall sound intensity is the same as hack ground noise. .

[Example] As a specific example of the present invention, the overall configuration is shown in FIG.

The acoustic vibration that has propagated to the vessel wall 1 is converted into an electrification signal by the sensor 2, and then the preamplifier 3
is converted into a voltage signal by This voltage signal is an analog signal having an audible frequency distribution, and is divided by a bandpass filter 4 into 10 frequency ranges obtained by dividing a plurality of signals in the range of 0 to 20 KHz into 10.
The multiplier 5 and the averaging circuit 6 can determine the average intensity of sound for each frequency range.

The sound intensities for each of these 10 frequency ranges are used as input to the hierarchical neural network 70.

This neural network 7 has five layers (S e ns
or layer: input layer), A layer (association layer:
It is a hierarchical network consisting of an intermediate layer) and an R layer (response layer: output layer).

In this example, the 8th layer 8 has 0 neurons, the A layer 9 has 4 neurons, and the 8th layer 10 has 1 neuron, and these neurons 11, 14, 15 are hierarchically connected to synapses 12, 13.
The configuration was made by combining the two.

Here, neurons refer to human nerve cells,
Control circuits refer to circuits that perform similar functions. Synapses are the connections between neurons that allow them to exchange signals. The processing mechanism in the brain is executed by the model shown in layer A of FIG. The input to neuron j is the signal strength Xi from each neuron and the synaptic connection weight ωi
It is the sum of the products multiplied by .

When Σ ωi Xi becomes greater than the threshold of that neuron j, there is an output from neuron j.

Although signal processing by a single neuron is relatively simple, many networks have parallel and high-speed decision functions.

The third layer receives the input signal, the A layer transforms the input signal pattern, and the R layer generates power as a response for the network. 20 In order to put a neural network into practical use as an acoustic abnormality monitoring device, it is first essential to establish the convergence of learning in order to recognize normal acoustics as “normal.” , when a predetermined input pattern is input to each neuron in the three layers, the value of the output from the R layer becomes the correct output, and the connection weight of the synapse between the S shoulder and the A layer, the connection weight between the A layer and the R layer, Learning involves modifying the synaptic connection weights). The following are the main parameters to consider for this purpose.

(1) Input/output functions f and g of the A-layer and 8-layer neurons 14 and 15.

(2) The number m of neurons 11 in the A layer 9.

(3) Synapse 12 between layer 8 and layer A 90 and layer A 8
and the initial value of the connection weight of the synapse 13 between the 8th layer 10 and the modification parameter ε.

In the learning process, if one input pattern is input and the output value of the network is different from the correct output, based on the difference between that output value and the correct output (called a teacher signal),
The ω between the R layer and the A layer is corrected, and then the ω between the A layer and the S layer is corrected.

Below margin ω(t+1) = ω(1)+ε (T-0) ・f9ε
is called a modification parameter.

If the previous connection weight ω is directly corrected using the value (T-0) based on the error, the correction width will be too large and the output of the network will diverge with respect to the number of learning times.

To prevent this, multiply by 0.

About the learning method Figure 2 and the following formula (% formula %) ■ If the sum of the inputs to the n units of the R layer is i, then ω ; the weight of the connection from the A11m unit to the n unit of the 8th layer is 0 ; Output from m units of layer A = g(i... (2) Input/output function of each unit g (x) = 1 (x≧xo) 0 (x≦x+) ax()(.

<x<x.

The weight of the connection at time (t+1) is given by the input signal ~
The signal flow of the output signal is S layer = A layer - R layer. Modification of the synaptic connection weight ω during learning is performed in the R-A-S order, contrary to this signal flow. As an expression, it takes the form of error backpropagation. This is because the basis for modifying ω is the difference between the output in the R layer and the teacher signal T. As a result of determining the parameters (1) to (3) above for the convergence of normal sound learning for the purpose of abnormal sound recognition, it was possible to realize the application of the neural network of the present invention.

First, the input/output function f of the neuron 14 of the A layer 9 is
By using the logistic function shown in the figure, even if the distribution of sound intensity for each frequency range as an input is large, a numerical value between 0 and 1 is output, and in the case of a large input value (absolute value), Compared to the input value near 0, the slope is gentler, so the 3rd layer 8 in the learning process
Modification of the weight of the synaptic connection 12 between and A layer 9 ((5)
, (6)) are performed smoothly, the convergence of learning is stable and fast.

That is, if we take the logistic function f in Figure 3, (
5) The right side of Eq. This is a relationship as shown in FIG. 6A, so when the output 0 of neuron m is around 0.5, which is between 0 and 1, fo is large. In other words, in this case, even if the amount of synaptic correction is increased, it is difficult to diverge. on the other hand,
When the output O is close to O or l, the learning tends to diverge because the values are close to both ends of FIG. 6A. Since fo is small at this time, the amount of synaptic correction is made small.

Next, in the conventional neural network, the input/output function g of the neuron 15 of the 8th layer 10 is the same function f as that of the A layer 9.
However, in the acoustic recognition according to the embodiment of the present invention, the input/output function g of the R layer is Input/output function f of layer A: Not the same as the logistic function, but the fourth
By using the piecewise linear function shown in the figure, learning and abnormal noise detection became possible.

A case where the input/output function g of the R layer is a piecewise linear function shown in FIG. 4 will be explained. If the input/output function g of this neuron is as shown in Figure 6B, then the differential function g° of g is the 6th
It will be as shown in Figure C.

Therefore, equation (4) becomes as follows, and the amount of modification of the coupling weight ω~” between the A layer and the R layer is (3
) is the correction amount calculated by the formula. This is 1 of the logistic function f
The value to be corrected is clearer than correcting a value in the form of a subdifferential molecule (which becomes a network). The output value of the 9th layer neuron is expressed as an analog numerical value between O (normal) and 1 (completely abnormal). This determines the degree of abnormal noise, so if it is not linear, this degree will be unclear.

Regarding the number m of neurons in the A layer 9, which is the intermediate layer, it has become possible to put it into practical use by determining the optimal range of the number of neurons in order to learn acoustic recognition quickly and stably.

In other words, when the number of neurons m is small, the learning time per round is short, but the number of times until convergence increases, and when m is too large, the learning time per round becomes long, but the number of times tends to decrease. Therefore, the optimal number of neurons in the intermediate layer A to realize stable and fast convergence learning was 3 to 6 in this embodiment. From this, it was found that the optimal number of intermediate layer neurons is 30 to 60% of the number of input layer neurons, since the number is 3 to 6 for every 10 neurons in the input layer. This optimum range is shown in FIG. Regarding the number m of neurons in this middle layer A,
The optimal range changes depending on the number of neurons in the input layer A layer and the output layer R layer, as well as the input data to be trained and the teacher signal, but considering the hardware configuration of the neural network, use the middle layer neurons as much as possible. A few meters is more practical.

In the monitoring device incorporating this neural network,
In contrast to the normal sound shown by the solid line in Figure 6, even minute abnormal noises whose intensity over the entire frequency range as shown by the dotted line in Figure 6 is the same as the solid line can be reliably recognized as abnormal noise. Ta.

Another embodiment of the invention is shown in FIG. The acoustic voltage signals from the sensor 2 and preamplifier 3 have a frequency distribution over the entire audible range, but the background noise of the system to be diagnosed may have characteristics depending on the part of the system. , Abnormal sounds also have frequency characteristics depending on the cause of the abnormality.

This monitoring device is characterized by providing a pre-filter 16 in order to pay attention to these mutual characteristics and limit sounds in a certain frequency range in advance or automatically. It is possible to learn about sounds and further improve detection ability.

Furthermore, it is possible to realize a monitoring device that estimates the operating status of the entire system to be diagnosed based on the information indicating the degree of abnormality, which is the output of the neural network for each sensor according to the present invention.

The neural network of the present invention described above includes:
Although it is possible to construct it using electronic circuits or use nerve cells in biological systems, it is currently preferable to use electronic circuits as permanent equipment.

Further, according to the present invention, the sound is identified as normal or abnormal, and after the sound is determined to be abnormal, the position of an object such as a loose part is confirmed using a conventional locator.

〔Effect of the invention〕

According to the present invention, an acoustic signal from a sensor attached to a system to be diagnosed is divided into frequencies, input to the input layer, converted in the intermediate layer, and the sound input to the sensor from the output layer is normal. We have achieved a monitoring device incorporating a neural network that has the ability to self-determine whether something is abnormal, and if so, the extent of the abnormality, making it possible to detect minute differences in abnormal sounds buried in background noise without generating false alarms.

[Brief explanation of the drawing]

Figure 1 is an example diagram of the present invention, Figure 2 is a diagram explaining learning of the neural network used in the present invention, and Figure 3 is a diagram explaining the learning of the neural network used in the present invention.
The figure shows an example of the input/output functions of the middle layer neurons, FIG. 4 shows the example of the input/output functions of the output layer neurons, FIG. 5 shows the optimal number of middle layer neurons, and FIG. 6 shows the example of the input/output functions of the output layer neurons. FIG. 6A is a diagram showing the effect of the present invention. FIG. 6A is a diagram showing the relationship between the output of the intermediate layer neuron and the differential of the input/output function. FIGS. 6B and 6C are the relationship diagram of the input/output function of the output layer neuron and its differential function. FIG. 7 is a diagram showing another embodiment of the present invention, and FIGS. 8 to 12 are explanatory diagrams of the prior art. l... Bessel wall, 2... sensor, 3... preamplifier, 4... band pass filter, 5... squarer,
6... Averaging circuit, 7... Hierarchical neural network, 8... Input layer, 9... Middle layer, 10...
Output layer, 11.14.15... Neuron, 12.1
3...Synapse. Figure 5 Layer A layer R layer Figure 6A Figure 6B Figure 6C Figure Figure 10

Claims (4)

[Claims] (1) A plurality of sensor devices attached to equipment that receive acoustic vibrations generated within the equipment and transmit signals, and a device that determines whether or not the sound inside the equipment is normal based on the signals from the sensor devices. The equipment monitoring device has a filter device that divides each signal from each sensor device into a plurality of frequency bands, and a filter device that inputs the signals from the filter device for each of the plurality of frequency bands, and detects whether the sound is normal or not. What is claimed is: 1. A device monitoring device comprising: a neural network device that determines whether or not the device is running, and outputs the result. (2) The neural network device includes a neural network consisting of three layers, an input layer, a middle layer, and an output layer, each having a predetermined number of neurons, and the number of neurons in the middle layer is 30 to 60% of the number of neurons in the input layer. The equipment monitoring device according to claim (1), characterized in that: (3) A plurality of sensors installed on the equipment wall that receive acoustic vibrations generated within the equipment and transmit signals, and a judgment device that determines whether the sound inside the equipment is normal based on the signals from the sensors. A device monitoring device having a filter device that divides the signal from each sensor device into a plurality of frequency ranges, and inputs the divided signals for each frequency range to determine whether the sound is a normal sound or an abnormal sound. The neural network device consists of an input layer, an intermediate layer, and an output layer each having a required number of neurons, and the neurons between each layer are connected by synapses,
An equipment monitoring device characterized in that the intermediate layer neuron is provided with a logistic function as an input/output conversion, and the output layer neuron is provided with a piecewise linear function as an input/output conversion. (4) A process of transmitting signals from multiple sensors attached to the equipment wall based on the sound generated within the equipment, and a judgment device receiving the signals from each of the above sensors to determine whether the sound is normal or not. A device monitoring method comprising the steps of dividing the signal from each sensor into a plurality of frequency ranges using a bandpass filter, and inputting the signal into a neural network determination device for each divided frequency range, 1. A device monitoring method comprising: a pre-learned determination function in a neural network determination device; and a step of determining whether the sound is a normal sound or an abnormal sound based on the input signal and outputting a determination result.