CN108897309B - Aero-Engine Sensor Failure diagnosis and partition method based on fuzzy membership - Google Patents
Aero-Engine Sensor Failure diagnosis and partition method based on fuzzy membership Download PDFInfo
- Publication number
- CN108897309B CN108897309B CN201810767048.3A CN201810767048A CN108897309B CN 108897309 B CN108897309 B CN 108897309B CN 201810767048 A CN201810767048 A CN 201810767048A CN 108897309 B CN108897309 B CN 108897309B
- Authority
- CN
- China
- Prior art keywords
- sensor
- signal
- fault
- redundancy
- fusion
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
- 238000000034 method Methods 0.000 title claims abstract description 35
- 238000003745 diagnosis Methods 0.000 title claims abstract description 31
- 238000005192 partition Methods 0.000 title abstract 2
- 230000004927 fusion Effects 0.000 claims abstract description 39
- 238000002955 isolation Methods 0.000 claims abstract description 13
- 230000006870 function Effects 0.000 claims description 64
- 230000005284 excitation Effects 0.000 claims description 32
- 238000013528 artificial neural network Methods 0.000 claims description 31
- 238000004422 calculation algorithm Methods 0.000 claims description 19
- 238000004458 analytical method Methods 0.000 claims description 15
- 238000012549 training Methods 0.000 claims description 15
- 239000000446 fuel Substances 0.000 claims description 10
- 238000012886 linear function Methods 0.000 claims description 4
- 238000007500 overflow downdraw method Methods 0.000 claims description 3
- 230000007812 deficiency Effects 0.000 abstract 1
- 230000009977 dual effect Effects 0.000 abstract 1
- 230000005484 gravity Effects 0.000 abstract 1
- 230000008569 process Effects 0.000 description 13
- 238000004088 simulation Methods 0.000 description 8
- 238000004364 calculation method Methods 0.000 description 7
- 238000012937 correction Methods 0.000 description 6
- 238000005259 measurement Methods 0.000 description 6
- 238000012360 testing method Methods 0.000 description 6
- 239000011159 matrix material Substances 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000010219 correlation analysis Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 239000000295 fuel oil Substances 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000009434 installation Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B23/00—Testing or monitoring of control systems or parts thereof
- G05B23/02—Electric testing or monitoring
- G05B23/0205—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults
- G05B23/0218—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterised by the fault detection method dealing with either existing or incipient faults
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B23/00—Testing or monitoring of control systems or parts thereof
- G05B23/02—Electric testing or monitoring
- G05B23/0205—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults
- G05B23/0259—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterized by the response to fault detection
- G05B23/0275—Fault isolation and identification, e.g. classify fault; estimate cause or root of failure
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Automation & Control Theory (AREA)
- Combined Controls Of Internal Combustion Engines (AREA)
Abstract
The invention discloses a kind of Aero-Engine Sensor Failure diagnosis and partition method based on fuzzy membership.In view of the deficiencies of the prior art, resolving ideas of the invention is to be merged based on fuzzy logic and fault threshold to analytic redundancy and hardware remaining, and sensor fault diagnosis is carried out based on fusion signal, once signal degree of membership subaverage, then its signal specific gravity shared in fusion signal just declines, after fault threshold, excluded from fusion signal completely, to realize self-adapted tolerance and Fault Isolation.Compared with prior art, the present invention can be realized self-adapted tolerance and Fault Isolation under dual channel sensor single channel fault condition.
Description
Technical Field
The invention belongs to the field of system control and simulation in aerospace propulsion theory and engineering, and particularly relates to a fuzzy membership-based method for diagnosing and isolating faults of an aero-engine sensor.
Background
In an aircraft engine control system, sensor failures account for over 80% of the total failures. In order to improve the safety and reliability of engine control systems, sensor fault diagnosis technology has attracted extensive attention and has developed into an application-type subject across various subjects in recent 40 years.
For the key sensors of the engine control system, a multi-channel multi-sensor hardware redundancy technology is often adopted, for the sensors with three or more hardware redundancies, an election method is adopted to vote whether the sensors have faults, and for the dual-channel sensors, once the measurement deviation of two channels exceeds a threshold value, which sensor has the fault cannot be judged. The extra hardware redundancy will undoubtedly increase the load on the sensor installation, as well as increase the engine weight and cost. For this reason, with the development of intelligent algorithms since the nineties of the last century, researchers at home and abroad are concerned with the use of intelligent algorithms for the analytical redundancy modeling of sensors. The analytic redundancy sensor model participates in fault diagnosis of the dual-channel sensor, and the problem that the dual-channel sensor is not easy to distinguish when a fault occurs can be effectively solved.
The neural network becomes the primary choice of an analytic redundancy sensor model due to the global approximation capability, the nonlinear mapping characteristic and the high self-organizing and self-learning capabilities of the neural network, so that a novel aeroengine sensor fault diagnosis algorithm represented by the neural network is deeply researched. However, the engine has large flight envelope and variable work, the off-line trained neural network is difficult to adapt to uncertainty in the work of the engine, the on-line trained neural network has the problems of complex network structure, more input nodes and hidden layer nodes and poor real-time performance, the diagnosis of the fault is also focused on the estimation of the output signal of the sensor, and no special research work is carried out on the isolation of the fault.
Disclosure of Invention
The invention aims to solve the technical problem of overcoming the defects in the prior art, provides a fuzzy membership degree-based aeroengine sensor fault diagnosis and isolation method, and can realize self-adaptive fault tolerance and fault isolation under the condition of single-channel fault of a dual-channel sensor.
The invention specifically adopts the following technical scheme to solve the technical problems:
a fault diagnosis and isolation method for an aircraft engine sensor based on fuzzy membership, wherein the sensor is provided with two hardware redundancy signal channels; firstly, information fusion is carried out on hardware redundancy signals of two sensors and an analysis redundancy signal output by an analysis redundancy model of the sensor to obtain a fusion signal, and then fault judgment is carried out on two hardware redundancy signal channels according to the fusion signal; the information fusion method specifically comprises the following steps:
in the formula,is a fusion signal;a first hardware redundancy signal and a second hardware redundancy signal respectively;is the resolution redundancy signal;sequentially obtaining membership degrees of a first hardware redundancy signal, a second hardware redundancy signal and an analytic redundancy signal;
wherein μ (x) is a membership function; k is a radical of1、k2Is the coefficient of membership adjustment, α, β are the index of membership adjustment, mu0Is a preset adjustment value; dNT、dDT、dBTRespectively preset sensor deviation threshold, drift fault threshold and bias fault threshold, and dNT<dDT<dBT；
The fault judgment is carried out on the two hardware redundancy signal channels according to the fusion signalThe method comprises the following steps: calculating the deviation between the fused signal and the channels of the sensorIf there is a deviationConsidering that the m channel of the sensor i has no fault; if it isConsidering that the m channel of the sensor i has drift fault; if it isThe m channel of sensor i is considered to have a bias fault.
Preferably, the analysis redundancy model is a three-layer BP neural network which performs initial off-line training by using a least square method and performs online updating of the network output layer weights by using a recursive least square algorithm.
Further preferably, in the hidden layer of the BP neural network, the first half nodes and the second half nodes use power functions and polynomial functions, respectively, as excitation functions.
Further preferably, the output layer excitation function of the BP neural network is a linear function.
Further preferably, the input of the BP neural network only contains the fuel flow and the output signal of the other sensor with the highest correlation.
Still further preferably, the input of the BP neural network is formed by the fuel flow at the current time k, k-1 and k-2 and the output signals of the other sensor with the highest correlation at the time k-1 and k-2.
Preferably, dNT＝0.002，dDT＝0.01，dBT＝0.03，α＝0.6，β＝5，k1＝k2＝0.4，μ0＝0.6。
Compared with the prior art, the technical scheme of the invention and the further improvement or preferred technical scheme thereof have the following beneficial effects:
(1) the constructed diagnosis system has strong fault isolation capability: and fusing the resolution redundancy and the hardware redundancy based on the fuzzy logic and the fault threshold, wherein once the membership degree of the signal is lower than the average value, the proportion of the signal in the fused signal is reduced, and the signal is completely eliminated from the fused signal after exceeding the fault threshold.
(2) The applicability of the constructed sensor analysis redundancy model is strong: the invention updates the network weight on line based on the threshold value, ensures the adaptability of the network in various working states in the envelope, and is suitable for advanced single-chip microcomputers and 51 series single-chip microcomputers which do not support exponential operation by adopting a combined excitation mode of power functions and polynomial functions.
(3) The constructed sensor analysis redundancy model has good real-time performance: the invention simplifies the structure of the diagnosis system by utilizing correlation analysis, has less input of the analytical redundancy model and the number of nodes of the hidden layer, and adopts a recursive least square algorithm to update the weight of the network output layer, thereby avoiding the time consumption of iterative solution.
Drawings
FIG. 1 is an example of an aircraft engine sensor fault diagnostic system constructed in accordance with the present invention;
FIG. 2 is a schematic diagram of a hidden layer structure of a neural network using a combined excitation function;
FIG. 3 is a membership function image;
FIG. 4 is N1A sensor signal change curve of a single-channel drift fault diagnosis test of the sensor;
FIG. 5 is N1A sensor signal relative error curve of a single-channel drift fault diagnosis test of the sensor;
FIG. 6 is N1A sensor signal change curve of a single-channel offset fault diagnosis test of the sensor;
FIG. 7 is N1The relative error curve of the sensor signal of the single-channel offset fault diagnosis test of the sensor.
Detailed Description
Aiming at the defects of the prior art, the solution idea of the invention is to fuse the resolution redundancy and the hardware redundancy based on the fuzzy logic and the fault threshold, and to diagnose the sensor fault based on the fusion signal, once the signal membership is lower than the average value, the proportion of the signal in the fusion signal is reduced, and after the signal membership exceeds the fault threshold, the signal is completely removed from the fusion signal, thereby realizing the self-adaptive fault tolerance and fault isolation.
Specifically, the invention relates to a fuzzy membership based aircraft engine sensor fault diagnosis and isolation method, wherein the sensor is provided with two hardware redundancy signal channels; firstly, information fusion is carried out on hardware redundancy signals of two sensors and an analysis redundancy signal output by an analysis redundancy model of the sensor to obtain a fusion signal, and then fault judgment is carried out on two hardware redundancy signal channels according to the fusion signal; the information fusion method specifically comprises the following steps:
in the formula,is a fusion signal;a first hardware redundancy signal and a second hardware redundancy signal respectively;is the resolution redundancy signal;sequentially obtaining membership degrees of a first hardware redundancy signal, a second hardware redundancy signal and an analytic redundancy signal;
wherein μ (x) is a membership function; k is a radical of1、k2Is the coefficient of membership adjustment, α, β are the index of membership adjustment, mu0Is a preset adjustment value; dNT、dDT、dBTRespectively preset sensor deviation threshold, drift fault threshold and bias fault threshold, and dNT<dDT<dBT；
And performing fault judgment on the two hardware redundancy signal channels according to the fusion signal, specifically as follows: calculating the deviation between the fused signal and the channels of the sensorIf there is a deviationConsidering that the m channel of the sensor i has no fault; if it isConsidering that the m channel of the sensor i has drift fault;if it isThe m channel of sensor i is considered to have a bias fault.
In order to ensure the adaptability of the network in various working states within the envelope and improve the real-time performance, preferably, the analytic redundancy model is a three-layer BP neural network which performs initial off-line training by using a least square method and performs online updating of the weights of the network output layer by using a recursive least square algorithm. And further adopting a combined excitation function in the hidden layer, namely in the hidden layer of the BP neural network, the first half part of nodes and the second half part of nodes respectively use power functions and polynomial functions as excitation functions.
In order to reduce the complexity of the system and improve the real-time performance of online fault diagnosis, the invention further utilizes the correlation to greatly simplify the analytical redundancy model, namely the input of the BP neural network only comprises the fuel flow and the output signal of one other sensor with the highest correlation.
For the public understanding, the technical scheme of the invention is explained in detail by a specific embodiment and the accompanying drawings:
the present embodiment was conducted for a trouble diagnosis study of a turbofan engine control system that includes five sensors. The sensor signals to be diagnosed include: fan speed (N)1) Speed of compressor (N)2) Total pressure (P) at the outlet of the compressor3) Low pressure turbine outlet temperature (T)46) Low pressure turbine outlet pressure (P)46). The five sensors to be diagnosed are sequentially marked as sensors No. 1 to 5.
First, an analytical redundancy model needs to be designed for each sensor. In previous studies, the inputs to each model included the fuel flow W at the current times k, k-1, and k-2fAnd the throat area A of the nozzle8(6 variables), and 4 other sensor signals (8 variables) at the first k-1 and k-2 moments, there are 14 inputs, and the number of hidden layer nodes is usuallyAbout 2 times of the number of nodes of the input layer, if 25 is selected, each analytic redundancy model adopts three layers of BP neural networks, hidden layer bias is set, the output layer does not set bias, 14 multiplied by 25+25+25 is 400 neural network weight parameters, 5 diagnosis modules have 2000 parameters, and the real-time performance of the diagnosis system is not beneficial.
In the throttling working state of the engine, the nozzle area is kept unchanged at some time, and an unchanged input is redundant to the system input, so that the invention firstly eliminates A from the input8Signal, only retaining WfThe signals are known to have high correlation among 5 sensors to be diagnosed through correlation analysis, so that only 1 of 4 sensor signals is reserved as the input of the current sensor to be diagnosed, and the input signals are 3 fuel variable signals and 2 sensor signals, which are 5 inputs in total, as shown in fig. 1. The number of the hidden layer nodes is 15, each analysis redundancy model has 5 multiplied by 15+15+15 as 105 parameters, the scale of the model parameters is reduced by 3.8 times, the storage requirement and the calculation workload can be effectively reduced, and the real-time performance of the diagnosis system is improved.
In the embodiment, the sensors to be diagnosed are numbered according to the correlation, and each sensor analysis redundancy model takes fuel and the sensor with the subsequent serial number as input according to the numbering sequence. Sensor resolution redundancy model inputs such as number 1 include fuel flow WfAnd sensor 2 information, and sensor 5 analytic redundancy model input includes fuel flow WfAnd sensor No. 1 information. The fuel oil comprises multi-sensor measurement mean values at the current moment k, the moment k-1 and the moment k-2, each numbered sensor comprises the multi-sensor measurement mean values at the moment k-1 and the moment k-2, and each resolution redundancy model contains n-5 input quantities.
In this embodiment, a power function and polynomial function combined excitation mode as shown in fig. 2 is adopted in the hidden layer, 15 nodes are arranged in the hidden layer, the first 7 hidden layer nodes use the power excitation function, and the last 8 hidden layer nodes use the polynomial expansion excitation function, which takes into account the accuracy and the real-time performance of the network. A linear excitation function is taken at the output layer.
The currently widely used hidden layer excitation functions are single-pole and double-pole S functions, when a network falls into an error function flat area, the smaller derivative of the excitation function will cause the network weight and threshold updating speed to be slow, so that more time is consumed in the network training process based on gradient descent, and the S functions all contain exponential functions exThe operation is difficult to realize in some 51 series single-chip microcomputers for testing. For a non-linear function, it can be approximated by taylor expansion at x ═ 0:
it can be seen that a non-linear function can be represented by a combination of power functions. If the power function is taken as the hidden layer excitation function, for the mth hidden layer node, the excitation function is as follows:
gm(x)＝xm-1 (2)
as the number of hidden layer nodes increases, the order of the excitation function increases, and for smaller input signals, the output approaches zero due to higher power. Therefore, the expression of a nonlinear system is single by simply adopting a power excitation function, and the system precision is also influenced by insufficient excitation for smaller input and further singularity trend of an output matrix of a hidden layer.
If e in the unipolar and bipolar S functions is replaced by polynomial expansionxThe exponential operation increases the calculation workload of the hidden layer, thereby affecting the real-time performance of the system. The invention provides a hidden layer excitation function setting method of a mixed power function and a polynomial function, combines the real-time performance of the power function and the high precision of the polynomial function, adopts the power excitation function at the hidden layer node which is ordered in the front and adopts the polynomial excitation function at the hidden layer node which is ordered in the back, and can effectively avoid the singularity which can appear when the simple power function is excitedAnd the phenomenon is considered, and the accuracy and the real-time performance of the algorithm are considered.
In order to verify the feasibility of taking the combined function as the excitation function, the algorithm is tested on three sets of Benchmark datasets, and the output layer weight is identified by adopting a least square algorithm after the network weight is randomly initialized. The number of hidden layer nodes is set to be 15, the first 7 hidden layer nodes use power excitation functions, namely the power functions have the highest order of 6, the last 8 hidden layer nodes use polynomial expansion excitation functions, and the comparison result of the test with the simple power excitation function and the polynomial function is shown in table 1.
Table 1 comparison table of learning effect of Benchmark dataset under different excitation functions
As can be seen from table 1, the neural network algorithm using the polynomial function as the excitation function has high accuracy, but poor real-time performance; the training time is short when the power function is used as an excitation function, but large errors may occur at individual working points, which is shown in that the maximum errors of 3 groups of data sets are far greater than those of other algorithms. The combined excitation function integrates the advantages of the two functions, so that the training of the network has high precision and good real-time performance.
The invention adopts BP neural network to establish the analytic redundancy model of each sensor. Based on a turbofan engine component level model, the typical working point dynamic data under the ground working state is acquired in an off-line mode, and a least square method is adopted for off-line training of the neural network.
The data is collected from the slow vehicle, the throttle state and the maximum state, and 0.2 percent of the collected data is addedThe white noise of the analog sensor, the measurement noise of the analog sensor, and the generation of a dual-channel sensor analog signalFor each sensor's analytical redundancy model, the input sensor signal is takenAnd all input and output are normalized to be between-1 and 1, and an analytic redundancy model with 5 inputs and 15 hidden layer nodes is established.
The data normalization method is as follows:
wherein,for the data after the ith sensor two-channel mean normalization,as a sensor yiThe maximum and minimum possible measurements.
And performing offline training of the neural network according to the normalized data. Randomly generating the connection weight from the input layer to the hidden layer, calculating the output matrix of the hidden layer based on the input and the combined excitation function, recording as O, if T is the target output matrix, and W is the connection weight from the hidden layer to the output layer, some
OW＝T (4)
And updating the network weight by adopting a least square algorithm. The least square algorithm directly calculates the weight of the output layer through generalized inverse calculation, and the calculation formula is as follows
The invention carries out the fusion of hardware redundancy signals and resolution redundancy signals based on fuzzy membership, and the realization process is as follows:
(1) designing a membership function based on a threshold value:
according to the noise and the measurement error in the working process of the sensors, the allowable deviation threshold value d of the two sensors under the normal working condition is setNTSetting a sensor drift fault threshold dDTBy using intermediate membership functions
In this example, d is takenNT＝0.002，dDT＝0.01，dBT＝0.03，α＝0.6，β＝5，k1＝k2＝0.4，μ0The membership function curve is shown in fig. 3 as 0.6.
(2) And (3) calculating membership between sensor redundancies:
for sensor channel 1 signalDegree of membership thereof
For sensor channel 1 signalDegree of membership thereof
Resolving a redundancy signal for a sensorDegree of membership thereof
It can be seen that when the difference between the redundancy signals of the sensors is small, the difference is less than the threshold value dNTThe membership degree between the signals is 1, when the difference is larger, the membership degree is reduced, and when the difference is larger than the drift fault threshold value dDTThe degree of membership decreases to zero.
(3) Signal fusion between sensor redundancies:
fusing the sensor hardware redundancy signal and the analysis redundancy signal according to the membership degree to obtain a fusion signal
It can be seen that when the membership degrees are all 1, each sensor occupies equal proportion in the fusion signal, when the membership degree of a certain signal is reduced, the proportion occupied by the sensor in the fusion signal is reduced, and when the membership degree is reduced to zero, the influence of the sensor is eliminated from the fusion signal, so that the isolation of the fault sensor signal is effectively realized.
After the fusion signal is obtained, the sensor fault diagnosis can be carried out based on the fusion signal and the hardware redundancy signal, and the implementation process is as follows: calculating the deviation between the fused signal and the channels of the sensorIf there is a deviationConsidering that the m channel of the sensor i has no fault; if it isConsidering that the m channel of the sensor i has drift fault; if it isThe m channel of sensor i is considered to have a bias fault.
The invention further carries out online updating of the weight of the sensor analysis redundancy model based on the deviation between the fusion signal and the analysis redundancy signal, and the method specifically comprises the following steps:
(1) calculating the deviation between the fusion signal and the output signal of the sensor analysis redundancy model, and judging whether the network correction weight is needed:
because the working state of the aircraft engine is changeable and is influenced by degradation, environmental change and the like, the off-line training neural network is difficult to meet the working requirements of the engine under the full envelope and full state, and therefore on the premise that the sensor can provide effective signals (not all the multiple channels fail), the invention adopts an on-line updating mode for the off-line training neural network weight to adapt to the uncertainty of the engine. During on-line use, the deviation between the neural network output and the fusion signal is calculatedIf there is a deviation dNT<ei<dDTIf the network estimation output is large, the generalization capability of the network is reduced, and the weight of the network needs to be updated.
(2) Updating the weight of a network output layer based on a recursive least square algorithm:
although the least square algorithm adopted in the off-line training process has a high convergence rate, the calculated amount is large, and if the least square algorithm is adopted in the on-line training process, the real-time performance of the fault diagnosis system is directly influenced. Therefore, in the online calculation process of the output layer of the neural network, the method uses a recursive least square method to replace the traditional least square method so as to reduce the network training time. The main effect of the off-line training of the least square algorithm is to enable the neural network updated on line to obtain a better initial weight.
For each sensor resolution redundancy model, the hidden layer output under k sets of input data can be represented as:
whereinThe number of tier nodes is implied.
The output layer adopts a linear excitation function, and then the network output is as follows:
whereinThe connection right of the hidden layer to the output layer. While in practice it is desirable that the network output is consistent with the fused value, i.e. it is desirableUsing the connection weight as an unknown number, the least square solution of the equation isIf the k +1 th group of data is tested to have dNT<ei,k+1<dDTIf the network weight needs to be adjusted, thenAccording to a recursive least squares algorithm, there are
Since the network has been initialized by offline training, PkThe initial value is calculated off line, matrix inversion operation is avoided in the recursion process, the weight of the network is obtained through the recursion calculation, the iterative correction process is not needed, and the real-time performance of the diagnosis system is greatly improved.
In order to verify the effectiveness of the sensor fault diagnosis scheme, the sensor fault simulation diagnosis is carried out under the conditions that the height H is 0km and the Mach number Ma is 0.
(1) Drift fault simulation
Firstly, simulating and diagnosing the drift fault of the sensor by N1For a single-channel fault of a sensor as an example, the analytic redundancy model of the single-channel fault is input as W at k, k-1 moment and k-2 momentfThe sensor measures the mean value and the k-1 time and the k-2 time N2The sensor measures the mean value, N within 20ms of each calculation step1The sensors drift downwards at 2rpm, the maximum drift amount is 3%, and each sensor signal contains 0.2% of zero-mean white noise.
FIGS. 4 and 5 show N1Sensor S1Channel failure, S2The NN in the graph represents the output N of the resolution redundancy model1Sensor signal, Fuse stands for N after fusion1A sensor signal. In the simulation process, 38 corrections are carried out on the weight values through recursive least squares, and the corrections ensure the precision of the analytic redundancy model. Fig. 4 shows the sensor signal, the analytical redundancy model output signal and the fusion signal for the two channels, and fig. 5 shows the error of each signal with respect to the fusion signal. As can be seen from fig. 4 and 5, when the channel 1 starts to drift, the drift amount is small, and the drift fault threshold value is not exceeded until about 38s, so that before that, the fusion value of the system is diagnosed, and the sensor signals and the analytic redundancy model outputs of the two channels are comprehensively consideredOut of and S2The measured values of the channels deviate. Diagnosing the fault after 38S, isolating the fault channel signal not as the input of the analytic redundancy model, fusing the signal and the neural network signal with S2The deviation therebetween is reduced.
(2) Bias fault simulation
When simulating bias fault, N is used1For single channel failure of sensor, simulation S1The channel is biased 3% downwards, and the other sensors have no diagnosis effect when in fault. In the simulation process, 30 times of correction is carried out on the weight value through recursion least square, and the correction times are smaller than those in the offset process, because when the deviation in the initial stage of the drift process is small, the weight value is corrected without considering that a fault occurs.
The sensor signals, the analytical redundancy model output signal, and the fused signal for the two channels are shown in fig. 6, and the deviation of each signal from the fused signal is shown in fig. 7. Because the bias fault exceeds the fault threshold value at the beginning, the bias fault is quickly isolated by the fusion signal, and the fusion signal and the output of the resolution redundancy model are eliminated as the output of the resolution redundancy model, so that the fusion signal and the output of the resolution redundancy model are not influenced by the single-channel bias fault.
According to diagnosis and simulation of two fault modes, the method for diagnosing and isolating the faults of the aero-engine sensor based on the fuzzy membership can effectively realize diagnosis and isolation of the faults of the single-channel sensor.
Claims (7)
1. A fault diagnosis and isolation method for an aircraft engine sensor based on fuzzy membership, wherein the sensor is provided with two hardware redundancy signal channels; the method is characterized in that firstly, information fusion is carried out on hardware redundancy signals of two sensors and an analysis redundancy signal output by an analysis redundancy model of the sensor to obtain a fusion signal, and then fault judgment is carried out on two hardware redundancy signal channels according to the fusion signal; the information fusion method specifically comprises the following steps:
in the formula,is a fusion signal;a first hardware redundancy signal and a second hardware redundancy signal respectively;is the resolution redundancy signal;sequentially obtaining membership degrees of a first hardware redundancy signal, a second hardware redundancy signal and an analytic redundancy signal;
wherein μ (x) is a membership function; k is a radical of1、k2Is the coefficient of membership adjustment, α, β are the index of membership adjustment, mu0Is a preset adjustment value; dNT、dDT、dBTRespectively preset sensor deviation threshold, drift fault threshold and bias fault threshold, and dNT<dDT<dBT；
And performing fault judgment on the two hardware redundancy signal channels according to the fusion signal, specifically as follows: computing fusion letterDeviation between the signals and the channels of the sensorIf there is a deviationConsidering that the m channel of the sensor i has no fault; if it isConsidering that the m channel of the sensor i has drift fault; if it isThe m channel of sensor i is considered to have a bias fault.
2. The method of claim 1, wherein the analytical redundancy model is a three-layer BP neural network that uses a least square method for initial off-line training and a recursive least square algorithm for online updating of network output layer weights.
3. The method of claim 2, wherein in an implicit layer of the BP neural network, the first half nodes and the second half nodes use power functions and polynomial functions, respectively, as excitation functions.
4. The method of claim 2, wherein the output layer excitation function of the BP neural network is a linear function.
5. The method of claim 2, wherein the inputs to the BP neural network comprise only the fuel flow and the output signal of the one other sensor with the highest correlation.
6. The method of claim 5, wherein the input of the BP neural network is composed of the fuel flow at the current time k, k-1, k-2 and the output signal of the other sensor with the highest correlation at the time k-1, k-2.
7. The method of claim 1, wherein d isNT＝0.002，dDT＝0.01，dBT＝0.03，α＝0.6，β＝5，k1＝k2＝0.4，μ0＝0.6。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810767048.3A CN108897309B (en) | 2018-07-13 | 2018-07-13 | Aero-Engine Sensor Failure diagnosis and partition method based on fuzzy membership |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810767048.3A CN108897309B (en) | 2018-07-13 | 2018-07-13 | Aero-Engine Sensor Failure diagnosis and partition method based on fuzzy membership |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108897309A CN108897309A (en) | 2018-11-27 |
| CN108897309B true CN108897309B (en) | 2019-09-06 |
Family
ID=64348869
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810767048.3A Expired - Fee Related CN108897309B (en) | 2018-07-13 | 2018-07-13 | Aero-Engine Sensor Failure diagnosis and partition method based on fuzzy membership |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108897309B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11635734B2 (en) | 2019-01-10 | 2023-04-25 | Dalian University Of Technology | Interval error observer-based aircraft engine active fault tolerant control method |
| CN110030263B (en) * | 2019-04-16 | 2021-04-16 | 青岛大学 | Fault-tolerant control system and method for displacement sensor of active electromagnetic bearing |
| CN110610027B (en) * | 2019-08-13 | 2021-01-19 | 清华大学 | Aero-engine resolution redundancy calculation method based on short-time data |
| CN114235420B (en) * | 2021-11-29 | 2024-04-16 | 中国航发沈阳发动机研究所 | Aeroengine high-voltage rotating speed signal voting method based on 4 redundancy |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1683254A (en) * | 2005-03-17 | 2005-10-19 | 哈尔滨工业大学 | Intelligent monitoring and control method for coagulation process based on multisource information fusion technology |
| CN1920511A (en) * | 2006-08-01 | 2007-02-28 | 东北电力大学 | Fusion diagnosing method of centrifugal pump vibration accidents and vibration signals sampling device |
| CN101576604A (en) * | 2009-01-04 | 2009-11-11 | 湖南大学 | Method for diagnosing failures of analog circuit based on heterogeneous information fusion |
| CN101661075A (en) * | 2009-06-08 | 2010-03-03 | 浙江大学 | Power system failure diagnostic method based on neural network and fuzzy integral |
| EP2287785A1 (en) * | 2009-08-19 | 2011-02-23 | University Of Leicester | Fuzzy inference apparatus and methods, systems and apparatuses using such inference apparatus |
| CN104281899A (en) * | 2013-07-03 | 2015-01-14 | 云南电力调度控制中心 | Novel fault diagnosis method based on information fusion |
| CN105184063A (en) * | 2015-08-26 | 2015-12-23 | 云南电网有限责任公司电力科学研究院 | Multi-criteria fusion line selection method based on information fusion technique and fuzzy theory |
-
2018
- 2018-07-13 CN CN201810767048.3A patent/CN108897309B/en not_active Expired - Fee Related
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1683254A (en) * | 2005-03-17 | 2005-10-19 | 哈尔滨工业大学 | Intelligent monitoring and control method for coagulation process based on multisource information fusion technology |
| CN1920511A (en) * | 2006-08-01 | 2007-02-28 | 东北电力大学 | Fusion diagnosing method of centrifugal pump vibration accidents and vibration signals sampling device |
| CN101576604A (en) * | 2009-01-04 | 2009-11-11 | 湖南大学 | Method for diagnosing failures of analog circuit based on heterogeneous information fusion |
| CN101661075A (en) * | 2009-06-08 | 2010-03-03 | 浙江大学 | Power system failure diagnostic method based on neural network and fuzzy integral |
| EP2287785A1 (en) * | 2009-08-19 | 2011-02-23 | University Of Leicester | Fuzzy inference apparatus and methods, systems and apparatuses using such inference apparatus |
| CN104281899A (en) * | 2013-07-03 | 2015-01-14 | 云南电力调度控制中心 | Novel fault diagnosis method based on information fusion |
| CN105184063A (en) * | 2015-08-26 | 2015-12-23 | 云南电网有限责任公司电力科学研究院 | Multi-criteria fusion line selection method based on information fusion technique and fuzzy theory |
Non-Patent Citations (1)
| Title |
|---|
| 一种基于模糊神经网络的融合故障诊断方法;苏羽等;《计算机工程》;20040930;第30卷(第17期);第5-6、170页 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108897309A (en) | 2018-11-27 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108897309B (en) | Aero-Engine Sensor Failure diagnosis and partition method based on fuzzy membership | |
| CN110118128B (en) | Fault diagnosis and fault-tolerant control method for micro gas turbine sensor | |
| CN109635318B (en) | Intelligent analysis redundancy design method for aero-engine sensor | |
| CN106055770B (en) | A kind of Fault Diagnosis of Aircraft Engine Gas Path method based on sliding mode theory | |
| CN109709934B (en) | Fault diagnosis redundancy design method for flight control system | |
| CN111474919B (en) | Aeroengine control system sensor fault diagnosis method based on AANN network group | |
| CN109242147A (en) | Signal fused fan condition prediction technique based on Bp neural network | |
| CN114564000B (en) | Active fault tolerance method and system based on intelligent aircraft actuator fault diagnosis | |
| CN113283004A (en) | Aero-engine degradation state fault diagnosis method based on transfer learning | |
| CN111506049B (en) | Multiple fault diagnosis method for aero-engine control system based on AANN network system | |
| CN111860791A (en) | Aero-engine thrust estimation method and device based on similarity transformation | |
| CN112801267A (en) | Multiple fault diagnosis device for aircraft engine with dynamic threshold value | |
| CN102788955A (en) | Remaining lifetime prediction method of ESN (echo state network) turbine generator classification submodel based on Kalman filtering | |
| CN110985216A (en) | Intelligent multivariable control method for aero-engine with online correction | |
| CN114235420B (en) | Aeroengine high-voltage rotating speed signal voting method based on 4 redundancy | |
| CN112749789A (en) | Aero-engine multiple fault diagnosis device based on self-association neural network | |
| CN112906855A (en) | Dynamic threshold variable cycle engine multiple fault diagnosis device | |
| Li et al. | Multiple fault diagnosis of aeroengine control system based on autoassociative neural network | |
| Zou et al. | Fault diagnosis of shield machine based on SOM-BP neural network fusion | |
| Wang et al. | Gas flow path fault diagnosis and sensor fault diagnosis for aeroengine based on fusion filter algorithm | |
| CN113609747A (en) | Sensor resolution redundancy method of engine auxiliary power unit control system based on COSELM algorithm | |
| Li et al. | Fault diagnosis and reconstruction for sensor of aeroengine control system based on AANN network | |
| Vladov | Algorithms for diagnostic and parameter of failures of channels of measurement of TV3-117 aircraft engine automatic control system in flight modes based on neural network technologies | |
| CN111830827A (en) | Two-degree-of-freedom [ mu ] controller for fault-tolerant gain scheduling of aero-engine | |
| Hu et al. | An online deep reinforcement learning based parameter identification method for HVDC system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20190906 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |