CN114254555A

CN114254555A - Deep learning-based counteraction wheel fault detection and health assessment system and method

Info

Publication number: CN114254555A
Application number: CN202111372043.9A
Authority: CN
Inventors: 陈子涵; 张田青
Original assignee: China Academy of Space Technology CAST
Current assignee: China Academy of Space Technology CAST
Priority date: 2021-11-18
Filing date: 2021-11-18
Publication date: 2022-03-29

Abstract

The invention discloses a system and a method for fault detection and health evaluation of a reaction wheel based on deep learning, which can solve the problems that the difficulty of fault detection of the reaction wheel is high, and the health state of the reaction wheel cannot be evaluated by fully utilizing normal data when the fault data of the reaction wheel is less. The reaction wheel fault detection and health assessment system comprises an observer, a detector and an evaluator; the observer is used for tracking the dynamic state of the reaction wheel and generating a real-time residual error; the detector generates a self-adaptive threshold value of the current time in real time, compares the relation between a residual error and the threshold value and detects whether the reaction wheel is in fault; the evaluator is configured to periodically evaluate a health status of the reaction wheel; the evaluator acquires and records residual errors, and residual error data of a period of time form a residual error vector; extracting the characteristics of the residual error vector; and calculating the spatial topological structure of the feature vector, calculating the spatial deviation degree, and finally converting the spatial deviation degree into the health degree (CV), thereby realizing the health evaluation.

Description

Deep learning-based counteraction wheel fault detection and health assessment system and method

Technical Field

The invention relates to the technical field of fault diagnosis, in particular to a reaction wheel fault detection and health assessment system and method based on deep learning.

Background

Attitude and Orbit Control Systems (AOCS) are an important component of spacecraft. AOCS is used to control the stability and orientation of a spacecraft to a predetermined direction during operation when the spacecraft is subjected to external disturbing moments. The space environment where the spacecraft is located is severe, the spacecraft is often interfered by various moments, and many spacecrafts fail before tasks are not finished. Over 30% of spacecraft failures are due to AOCS, with nearly 50% of AOCS failures being due to actuators. 60% of actuator-induced faults are caused by reaction wheels and control moment gyros. Fault detection and health assessment of the reaction wheel is one of the important means to improve the reliability and safety of the AOCS of the spacecraft. The early fault detection of the reaction wheel can timely find the fault of the spacecraft and then quickly carry out corresponding processing, so that the safety of the spacecraft is effectively improved. Spacecraft faults can be effectively avoided by health assessment of the reaction wheels so that measures are taken before the reaction wheels fail.

The problem of fault detection of reaction wheels, one of the important components of AOCS, has been extensively studied in the past few decades and these methods can be divided into two categories: model-based methods and data-driven methods. The model-based method is to perform fault detection by establishing an analytical model of the reaction wheel, and needs to consider the model structure of the system and the measurement of the actual process, and the accuracy of fault detection depends on the accuracy of modeling. The method has the advantages that the fault detection result is accurate, and the defects that the modeling process is complex and the model is difficult to verify. Data-driven methods are often used in situations where an accurate analytical model cannot be obtained. Such methods utilize large amounts of historical data, using algorithms such as neural networks or fuzzy logic as function approximators to construct an analytical model of the system. They learn reaction wheel dynamics models from input-output data and failure knowledge from historical data. However, the main problem with these methods is the high requirements on historical data and computational complexity. The more successful prior art combines the two methods, constructs an observer by using methods such as a kalman filter or a neural network, and estimates the state of the system from available input and output measurement values, thereby realizing fault detection. However, the conventional neural network or kalman filter based method has the following disadvantages: (1) it is difficult to model the highly nonlinear dynamics of the momentum wheel and to obtain the time series relationship of its input and output. (2) When the reaction wheel experiences complex working conditions and various disturbances, the problems of more fault detection interference, more false alarms and great difficulty under the condition of multiple working conditions are caused. (3) In the field of fault diagnosis and health management, compared with fault detection, health assessment has more practical guiding significance. However, for the health assessment problem of the reaction wheel, at present, because the fault data are less and the fault evolution rule is not completely clear, the research is less, and the health state of the reaction wheel cannot be judged by fully utilizing the normal data.

Disclosure of Invention

In view of the above, the invention provides a system and a method for fault detection and health assessment of a reaction wheel based on deep learning, which can solve the problems of high interference, high false alarm and high difficulty in fault detection of the reaction wheel; and the problem that the health status of the reaction wheel is not fully evaluated by utilizing normal data when the fault data of the reaction wheel is less.

In order to solve the above-mentioned technical problems, the present invention has been accomplished as described above.

A deep learning based reaction wheel fault detection and health assessment system comprising an observer, a detector, an evaluator;

the observer is used for tracking the dynamic state of the reaction wheel and is composed of a long-time memory network (LSTM); the observer acquires a combination of input X and multi-order delay output of the reaction wheel; simulating the relationship of the input and output of the reaction wheel by an LSTM deep neural network of an observer, the estimated output of the reaction wheel being output by the LSTM deep neural network

Based on actual output of the reaction wheelY sum estimate output

Real-time calculation of the actual and estimated output residuals of the reaction wheel, R ═ R (1), R (2), …, R (N)]；

The detector is used for detecting whether the reaction wheel is in fault; the detector consists of a threshold generation module and a fault discrimination module, wherein the threshold generation module is realized based on an LSTM deep neural network, and the fault discrimination module compares the magnitude relation between residual errors and self-adaptive thresholds in real time; the detector receives the input X of the reaction wheel and the estimated output of the observer

Residual error R, outputting fault state of reaction wheel; at the time t, a threshold generating module generates an adaptive threshold epsilon (t) of the current time t in real time; and the fault discrimination module compares the residual error r (t) at the time t with an adaptive threshold epsilon (t) in real time: if the residual error is larger than the threshold value, judging that the reaction wheel has a fault, and if the residual error is smaller than or equal to the threshold value, judging that the reaction wheel is normal in state;

the evaluator is configured to periodically evaluate a health status of the reaction wheel; the evaluator comprises a feature extraction module, a self-organizing map (SOM) neural network module and a health degree calculation module; the characteristic extraction module acquires and records residual errors r (T), and residual error data in a period of time T form a residual error vector R (T) ═ r (1), r (2), …, r (T); extracting the characteristics of the residual error vector; inputting the characteristics into the SOM neural network module, and then inputting the output result of the SOM neural network module into the health degree calculation module to obtain a health degree (CV) result; the degree of health is a dimensionless scalar quantity in the range of 0-1 that characterizes the state of health of the reaction wheel; and finally evaluating the health state of the reaction wheel according to the health degree.

Preferably, the self-organizing map (SOM) neural network module trains the SOM neural network using the time domain features of the residual vectors in the normal state of the reaction wheel, and outputs the spatial topology of the time domain features of the residual vectors in the normal state of the reaction wheel after the training is completed, and records the spatial topology of the time domain features of the residual vectors in the normal state of the reaction wheelRecording the position of the Best Matching Unit (BMU) in the space topological structure, and recording the weight vector corresponding to the BMU as U₀。

A deep learning-based reaction wheel fault detection and health assessment method implemented based on a deep learning-based reaction wheel fault detection and health assessment system as previously described, the method comprising:

step S1: the observer acquires input and output of the reaction wheel in real time, wherein the input is an input torque command signal of the reaction wheel, and the output is an output torque signal of the reaction wheel; the observer generates an estimated output of the reaction wheel in real time based on the input torque command signal and the output torque signal of the reaction wheel, calculates a difference between an actual output of the reaction wheel and the estimated output of the observer, and defines the difference as a residual error;

step S2: generating, by a threshold generation module of the detector, an adaptive threshold corresponding to an operating state of the reaction wheel and related to time based on an input torque command signal of the reaction wheel and an estimated output generated by the observer; comparing, by a fault discrimination module of the detector, the residuals in real time with an adaptive threshold: if the residual error is greater than the adaptive threshold, determining that a reaction wheel is faulty; if the residual error is less than or equal to the self-adaptive threshold, judging that the reaction wheel state is normal;

step S3: inputting the residual error to an evaluator, which outputs a degree of health (CV) of the reaction wheel, concurrently with fault detection; the degree of health is a dimensionless scalar quantity in the range of 0-1 that characterizes the state of health of the reaction wheel; and analyzing the health state of the reaction wheel according to the health degree, and judging whether the performance of the reaction wheel is degraded.

Preferably, the step S1 includes:

simulating the dynamics of the reaction wheel by using an observer based on an LSTM deep neural network using an input torque command signal X and an output torque signal Y of the reaction wheel, and combining Y according to the input and multi-order delay output of the reaction wheel^(-2)Generating estimated output of reaction wheel in real time

Calculating a difference between the actual output of the reaction wheel and the estimated output of the observer, the difference being defined as a residual R; the residual r (t) at the time t is calculated by the formula

Preferably, the step S2 includes:

generating an adaptive threshold value of the moment of the reaction wheel by using an input signal of the moment t of the reaction wheel and an output signal of an observer and using a threshold value generation module based on an LSTM deep neural network in a detector;

the adaptive threshold at time t is calculated by:

ε(t)＝r₀(t)+β

wherein r is₀(t) represents the residual error at the time t under the condition that the reaction wheel works normally, and beta represents a correction coefficient; the correction coefficient is used for compensating residual fluctuation caused by factors such as time drift parameters and disturbance of the reaction wheel;

the fault discrimination module of the detector compares this threshold epsilon (t) with the residual r (t): if the residual error is greater than the threshold, the reaction wheel is considered to be faulty, otherwise it is considered to be in a normal state.

Preferably, the step S3 includes:

step S31: according to the residual errors, the evaluator records the residual errors within a period of time T to form a residual error vector R (T) ([ r (1), r (2), …, r (T)), and the characteristic extraction module calculates the time-domain characteristics of the residual error vector and takes the time-domain characteristics as the time-domain characteristics of the residual error vector to be evaluated for health;

step S32: inputting the time domain characteristics of the residual vector to be assessed for health into an SOM neural network module of an evaluator, outputting the space topological structure of the time domain characteristics of the residual vector at the moment by the SOM neural network module in the evaluator, recording the position of a BMU in the space topological structure, and recording the weight vector corresponding to the BMU as U₁；

Calculate U₀And U₁The spatial distance is recorded as a Minimum Quantization Error (MQE); MQE, representing the deviation relation between the residual vector to be evaluated and the residual vector when the reaction wheel works normally, namely the deviation degree of the characteristic space corresponding to the current running state and the normal state of the reaction wheel respectively;

M＝||U₁-U₀||

wherein M represents an MQE value;

the evaluator's health calculation module normalizes MQE to yield a CV value:

where b is a normalization coefficient.

The health CV is between 0.9 and 1 when the reaction wheel is operating normally.

Preferably, the feature values corresponding to the time domain features include a root mean square value, a peak value, and an average absolute value, and the calculation method is as follows:

root mean square value:

peak value:

B＝max(|r(t)|)

average absolute value:

wherein T represents the length of the residual sequence recorded by the evaluator; r (t) denotes a residual at time t; a represents the root mean square value of the residual vector; b denotes the peak of the residual vector; c denotes the average absolute value of the residual vector.

Has the advantages that:

(1) aiming at the conditions of high nonlinearity degree, complex disturbance and various working conditions of the reaction wheel, an LSTM deep neural network is utilized to model the complex dynamics characteristics of the reaction wheel, the time sequence relation of input and output of the reaction wheel is simulated, and a high-precision observer is established, so that estimation output is generated in real time, and corresponding residual errors are output.

(2) Conventional observer-based fault detection methods typically incorporate a fixed threshold to determine if the residual error exceeds a limit, and in order to detect a fault in the reaction wheel, the present invention uses an adaptive threshold method to detect the state of the reaction wheel. Since observer-based fault detection is achieved by comparing the difference between the residual and the threshold, the threshold is a key factor that affects the accuracy of fault detection. If the threshold is definitively too high, some faults may not be detected; false alarms can occur more often if the threshold is definitively too low. To avoid the situation, the invention designs a generation method of the adaptive threshold, and uses an LSTM deep neural network to design a threshold generation module of the detector, and the threshold generation module outputs the adaptive threshold.

The invention utilizes the LSTM deep neural network to dig the self-adaptive relationship between the output of the reaction wheel and the corresponding fault threshold value under different working conditions and disturbance, and establishes a detector based on the self-adaptive threshold value, thereby timely judging whether the reaction wheel has faults or not. By introducing the self-adaptive threshold, the accuracy of fault detection is effectively improved, and the detection false alarm rate is reduced.

(3) The SOM neural network learns the topological structure of the data while learning the data features, can express the multidimensional features in a one-dimensional or two-dimensional space, and simultaneously reserves the topological structure of an input feature space; the SOM neural network is an unsupervised machine learning algorithm and thus can be trained using only normal data. The SOM neural network used in the present invention consists of an input layer and a competition layer. The input layer is a one-dimensional vector and the competition layer is a two-dimensional planar array.

And (3) excavating a spatial topological relation between the performance degradation of the reaction wheel and the health characteristic parameters in a normal state by using the SOM neural network, and establishing a health degradation rule model of the reaction wheel, thereby realizing the health evaluation of the reaction wheel. When the health characteristic modeling is carried out by utilizing the evaluator based on the SOM neural network, the health evaluation of the reaction wheel can be realized only by data when the reaction wheel normally runs and complex and diversified fault data are not needed, and the engineering practicability of the health evaluation is effectively improved.

Drawings

FIG. 1 is an architecture diagram of a reaction wheel fault detection and health assessment system;

FIG. 2 is a schematic flow chart of a deep learning based fault detection and health assessment method for a reaction wheel according to the present invention;

FIG. 3 is a schematic diagram of a detector configuration;

FIG. 4 is a schematic diagram of an evaluator architecture;

FIG. 5 is a reaction wheel simulation model;

FIG. 6 is a graph of a swept sinusoidal signal of the present invention;

FIG. 7 is a graph of a Monte Carlo sinusoidal signal of the present invention;

FIG. 8 is a schematic diagram of the Monte Carlo sinusoidal signal input being the reaction wheel output and the observer residual of the present invention;

FIG. 9 is a schematic of experiment 1 residual and adaptive threshold of the present invention;

FIG. 10 is a graph of experiment 2 residual and adaptive threshold of the present invention;

FIG. 11 is a graph of experiment 3 residual and adaptive threshold of the present invention;

FIG. 12 is a graph of experiment 4 residual and adaptive threshold of the present invention;

FIG. 13 is a graph of experiment 5 residual and adaptive threshold of the present invention;

FIG. 14 is a graph showing the health of the tests 1 to 5 of the present invention.

Detailed Description

The invention is described in detail below with reference to the figures and examples.

The invention discloses a reaction wheel fault detection and health assessment system and method based on deep learning, and the reaction wheel fault detection and health assessment system comprises an observer, a detector and an evaluator, as shown in FIG. 1.

The observer and detector of the present invention constitute a reaction wheel failure detection section, and the observer and estimator constitute a reaction wheel health evaluation section.

The reaction wheel input is a torque command signal using X ═ X (1), X (2), …, X (n)]Represents; the output of the reaction wheel is a torque signal using Y ═ Y (1), Y (2), …, Y (n)]Wherein N is the length of the signal sequence; at time t, the input to the reaction wheel is x (t), the output is y (t), 3 ≦ t ≦ N, and the first-order delay output of the reaction wheel is denoted y^(-1)Y (t-1), representing the reaction wheel output torque at time t-1; the second order delay output of the reaction wheel is denoted as y^(-2)Y (t-2) representing the reaction wheel torque output torque at time t-2; the first order delayed output of the reaction wheel is therefore denoted as Y^(-1)＝[y(1),y(2),…,y(N-1)]With the second order delay output denoted as Y^(-2)＝[y(1),y(2),…,y(N-2)]。

The observer is used to track the dynamic state of the reaction wheel and is formed by a long-time memory network (LSTM). The observer acquires a combination of input X and multi-order delay output of the reaction wheel; generally, first and higher order delay outputs or combinations thereof may be used, with the present invention being a combination of a first order delay output and a second order delay output Y^(-2)＝[Y^(-1)；Y^(-2)]The technical process is illustrated by way of example; simulating the relationship of the input and output of the reaction wheel by an LSTM deep neural network of an observer, the estimated output of the reaction wheel being output by the LSTM deep neural network

Based on actual Y and estimated Y outputs of the reaction wheels

Real-time calculation of the actual and estimated output residuals of the reaction wheel, R ═ R (1), R (2), …, R (N)]。

The detector is used for detecting whether the reaction wheel is in fault; the detector consists of a threshold generation module and a fault discrimination module, wherein the threshold generation module is realized based on a long-time memory network (LSTM), and the fault discrimination module compares the magnitude relation between residual errors and self-adaptive thresholds in real time;the detector receives the input X of the reaction wheel and the estimated output of the observer

Residual error R, outputting fault state of reaction wheel; at the time t, a threshold generating module generates an adaptive threshold epsilon (t) of the current time t in real time; and the fault discrimination module compares the residual error r (t) at the time t with an adaptive threshold epsilon (t) in real time: and if the residual error is larger than the threshold value, judging that the reaction wheel has a fault, and if the residual error is smaller than or equal to the threshold value, judging that the reaction wheel is in a normal state.

The self-organizing map (SOM) neural network module trains an SOM neural network by using the time domain characteristics of the residual vector in the normal state of the reaction wheel, the spatial topological structure of the time domain characteristics of the residual vector in the normal state of the reaction wheel can be output after the training is finished, the position of a Best Matching Unit (BMU) in the spatial topological structure is recorded, and the weight vector corresponding to the BMU is recorded as U₀。

As shown in FIG. 2, a deep learning based reaction wheel fault detection and health assessment method of the present invention is based on a reaction wheel fault detection and health assessment system, the method comprising:

step S2: generating, by a threshold generation module of the detector, an adaptive threshold corresponding to an operating state of the reaction wheel and related to time based on an input torque command signal of the reaction wheel and an estimated output generated by the observer; comparing, by a fault discrimination module of the detector, the residuals in real time with an adaptive threshold: if the residual error is greater than the adaptive threshold, determining that a reaction wheel is faulty; if the residual error is less than or equal to the self-adaptive threshold, judging that the reaction wheel is normal;

step S3: and inputting the residual error into an evaluator while detecting the fault, outputting the health degree of the reaction wheel by the evaluator, analyzing the health state of the reaction wheel according to the health degree, and judging whether the performance of the reaction wheel is degraded or not.

The invention utilizes the input and output relation of the reaction wheel to establish an observer and a detector to realize the fault detection of the reaction wheel. The specific process of fault detection is shown in fig. 3. The method utilizes the residual error to establish an evaluator so as to realize the health evaluation of the reaction wheel. The specific process of health assessment is shown in fig. 3.

The step S1 includes: simulating the dynamics of the reaction wheel by using an observer based on an LSTM deep neural network using an input torque command signal X and an output torque signal Y of the reaction wheel, and combining Y according to the input and multi-order delay output of the reaction wheel^(-2)Generating estimated output of reaction wheel in real time

The difference between the actual output of the reaction wheel and the estimated output of the observer is calculated, which is defined as the residual R.

To more accurately describe the non-linear timing relationship between the reaction wheel input and output, the present invention constructs an observer using an LSTM deep neural network. As shown in fig. 2, at time t, x (t) represents the input to the reaction wheel, y (t) represents the output of the reaction wheel,

representing the estimated output of the observer.

The LSTM deep neural network is a deep cyclic neural network architecture, can identify the relationship between related important signals in a time sequence, can effectively capture long-term time dependence between the important signals, and is suitable for modeling and predicting time sequence data. For reaction wheels, LSTM is particularly well suited to identify small state changes (fault) information of reaction wheels during their real-time operation. The core idea of the LSTM deep neural network is the memory unit module (LSTM neuron). By increasing the number of memory unit modules, the depth of the LS TM deep neural network is gradually increased. The threshold generation modules of the observer and detector in the invention are two LSTM deep neural networks of different structures. Calculating estimated output value of reaction wheel by observer based on LSTM deep neural network

The residual is defined as the difference between the reaction wheel output and the observer estimated output value:

in the above equation, r (t) represents a residual error at time t.

When the reaction wheel is working properly, the residual error is close to 0, typically less than 0.1. Since only noise and modeling errors affect the residual at this time. However, when the reaction wheel is not working properly, the residual error increases accordingly.

The step S2 includes: an adaptive threshold for the moment of the reaction wheel is generated using a threshold generation module in the detector based on an LSTM deep neural network using the input signal of the reaction wheel and the output signal of the observer. The operating state of the reaction wheel may be estimated from the actual inputs to the reaction wheel and the estimated outputs generated by the observer to generate an adaptive threshold based on the operating state of the reaction wheel, the threshold varying with the operating state of the reaction wheel. The fault discrimination module of the detector compares the residual error with the adaptive threshold in real time: if the residual error is greater than the threshold, the reaction wheel is considered to be faulty, otherwise it is considered to be in a normal state.

Considering that the threshold is typically affected by the input, output, load, environmental disturbances, etc. of the reaction wheel, the present invention calculates the adaptive threshold by the input of the reaction wheel and the output of the observer. As shown in FIG. 3, the threshold generation module in the detector uses as inputs the input to the reaction wheel and the output of the observer, the output of which is the adaptive threshold. When the LSTM deep neural network of the threshold generation module is trained, the self-adaptive threshold at the time t is calculated by the following formula:

ε(t)＝r₀(t)+β (5)

in the above formula r₀(t) represents a residual error at time t when the reaction wheel is operating normally, and β represents a correction coefficient. The correction coefficient is used for compensating residual fluctuation caused by factors such as parameters and disturbance of the reaction wheel drifting along with time and the like so as to enhance the robustness of the adaptive threshold.

Since the adaptive threshold calculation uses the input command of the reaction wheel and the output of the observer as the detector inputs, the inputs are not affected by the failure or failure of the reaction wheel. When the reaction wheel is normal, the threshold value is larger than the residual error due to the setting of the correction coefficient; when the reaction wheel is malfunctioning, the threshold value is still based on the normal condition of the reaction wheel and does not vary significantly in magnitude. However, at this point, the actual residual error will become significantly larger and exceed the threshold due to the reaction wheel failure, and the failure may be detected. In addition, in terms of working conditions, a large amount of working condition information is contained in the input of the reaction wheel, the disturbance (friction increase caused by temperature and the like) of the reaction wheel, the output of the observer and the residual error of the observer, and based on the identification and state tracking capability of the LSTM deep neural network, the threshold value of the output of the LSTM deep neural network of the threshold value generation module of the detector can change along with the change of the working conditions, so that the purpose of self-adapting threshold value changing is achieved.

As shown in fig. 4, the step S3 includes: and transmitting the residual error to an evaluator while detecting the fault, periodically recording the residual error by the evaluator to obtain a residual error vector, further analyzing the health state of the reaction wheel according to the residual error vector, and judging whether the reaction wheel has certain performance degradation. And assuming that a residual vector is obtained at the time t, firstly, the characteristic extraction module calculates the characteristic value of the residual vector, then, the SOM neural network module calculates the spatial topological structure of the residual vector, the structure is the output of the SOM neural network, and the health degree calculation module calculates the health degree of the reaction wheel according to the output of the SOM neural network. The step S3 specifically includes:

step S31: from the residuals, the evaluator records the residuals over a period of time T, constituting a residual vector r (T) ═ r (1), r (2), …, r (T), and the feature extraction module calculates the temporal features of the residual vectors. The characteristic values comprise root mean square values, peak values and average absolute values, and the specific calculation mode is as follows:

root mean square value:

peak value:

B＝max(|r(t)|) (7)

average absolute value:

Step S32: inputting the time domain features of the residual vector into an SOM neural network module of an evaluator, wherein the SOM neural network module in the evaluator outputs the spatial topology of the residual vector; when the SOM neural network module is designed, the time domain characteristics of the residual vector when the reaction wheel is in a normal state are used for training the SOM neural network, and training is carried outOutputting the space topological structure of the time domain characteristics of the residual vector in the normal state of the reaction wheel after the training is finished, recording the position of a Best Matching Unit (BMU) in the space topological structure, and recording the weight vector corresponding to the BMU as U₀(ii) a During health assessment, inputting the time domain characteristics of the residual vector to be assessed into the SOM neural network module to obtain the spatial topological structure of the time domain characteristics of the residual vector at the moment, recording the position of a BMU (BMU) in the spatial topological structure, and recording the weight vector corresponding to the BMU as U₁(ii) a Then calculate U₀And U₁The spatial distance is recorded as a Minimum Quantization Error (MQE); MQE shows the deviation relationship between the residual vector to be evaluated and the residual vector when the reaction wheel is working normally, i.e. the deviation degree of the feature space corresponding to the current operating state and the normal state of the reaction wheel. The change in MQE can thus be used to indicate a state of health degradation process for the reaction wheel.

M＝||U₁-U₀|| (9)

Where M represents an MQE value.

MQE, which indicates that the reaction wheel performance degradation continues to increase, the present invention uses the degree of health to indicate the health of the reaction wheel in order to visually indicate the health of the reaction wheel. A greater degree of health indicates a greater likelihood that the reaction wheel is in good condition, and a lower degree of health indicates that the reaction wheel is in degraded performance or has failed. The evaluator's health calculation module normalizes MQE to yield a CV value:

where b is a normalization coefficient.

The health degree result output by the evaluator is represented by a CV value, CV is a dimensionless scalar quantity ranging from 0 to 1, 0 represents complete failure, and 1 represents that the performance is not degraded. The residuals are first calculated using an observer based on the LSTM deep neural network to estimate the output of the reaction wheel. Feature values are then extracted from the residuals, followed by calculation of the minimum quantization error using an SOM neural network based evaluator (MQE), and finally the health of the reaction wheel is calculated by means of normalization at MQE.

The deep learning based method for fault detection and health assessment of reaction wheels of the present invention is described below in conjunction with specific embodiments.

1. Simulation model design and algorithm model training

The embodiment of the invention adopts the simulation data of the reaction wheel for verification, the simulation model of the reaction wheel is established by MATLAB/SIMULINK, and comprises a motor torque control module, an electromotive force torque module, a motor disturbance module and a friction torque module, the simulation model is shown in figure 5, and parameters of each part are shown in table 1.

TABLE 1 simulation model parameters

In view of the uncertainty of the reaction wheel input when performing real-time fault detection, the training data set and the test data set should contain typical signals in the corresponding frequency, amplitude range when training the observer, detector and evaluator. Suppose that the input signal frequency range of the reaction wheel is f₀,f₁]The amplitude range is [ A ]₀,A₁]. The invention uses sweep frequency sine signal training, and the amplitude and the frequency of the sweep frequency sine signal linearly increase along with time in the corresponding amplitude and frequency range. The invention trains an observer model by using a large number of sweep signals, so that the observer model can simulate the input-output relationship of a reaction wheel under various working conditions, and the sweep signals of 120 seconds are shown in FIG. 6. To verify the training effect, the present invention uses a monte carlo sinusoidal signal, whose amplitude and frequency vary randomly from cycle to cycle within the corresponding amplitude and frequency ranges, as shown in fig. 7. The output signal of the reaction wheel is obtained by inputting a swept sine signal to the reaction wheel, the observer is trained using the input-output signal, and after training is completed, the observer is verified using a monte carlo sine signal, as shown in fig. 8. The result shows that the observer can well simulate the input-output relationship of the reaction wheel.

Typical failure modes for reaction wheels are: motor gain loss, voltage fluctuation, increased disturbance torque, and increased coulomb friction. The method verifies the fault detection algorithm and the health evaluation algorithm provided by the invention through different fault degrees for the gain loss fault of the reaction wheel injection motor. In the simulation experiment, the data sampling rate was 100 points/second. Based on the system architecture of the invention, after the training of the observer is completed, the fault detection and health assessment algorithm can be verified by using the determined input instruction:

x(t)＝sin(πt) (11)

2. verifying fault detection algorithm by using normal data and fault data of simulation model

In order to verify the effectiveness of the fault detection algorithm provided by the invention, 4 fault injection tests are carried out, and the fault degree is deepened in sequence. Each set of test simulation time was set to 20 seconds, fault injection was performed at 0 second, and the test details are shown in table 2.

Table 2 fault injection test detailed parameters

The LSTM deep neural network of the observer designed in the invention is trained by using an adam algorithm, and the network structure is as follows:

sequence input layer (3D characteristic)

LSTM neuron layer (200 neurons)

Dropout layer (probability 20%)

LSTM neuron layer (100 neurons)

Dropout layer (probability 20%)

Full connecting layer (50 dimensional characteristic)

Dropout layer (probability 20%)

Full connecting layer (1 dimension characteristic)

Regression output layer

The LSTM deep neural network of the threshold generation module of the detector designed by the invention is trained by using the adam algorithm, and the network structure is as follows:

sequence input layer (2D characteristic)

LSTM neuron layer (200 neurons)

Dropout layer (probability 20%)

Full connecting layer (50 dimensional characteristic)

Dropout layer (probability 50%)

Full connecting layer (1 dimension characteristic)

Regression output layer

The results of trials 1-5 are shown in fig. 9-13, where the solid black line represents the residual and the dotted black line represents the adaptive threshold.

Test 1 shows a state where the reaction wheel is normal. As shown in FIG. 9, the residual error in this experiment is well below the threshold, indicating that the reaction wheel is working properly. And meanwhile, the threshold value is automatically adjusted according to the change of the working state.

Test 2 shows the state of the reaction wheel motor gain loss fault. At this time, km is 90% of the normal case. As can be seen from fig. 10, the residual increases, and some of the time instants are closer to the threshold, but still do not exceed the threshold.

Test 3 shows the state of the reaction wheel motor gain loss fault. At this time k_mIs 80% of normal state. As can be seen from fig. 11, the residual increases, part of the time residual is below the threshold, and part of the time residual is above the threshold. This test shows that the adaptive threshold can accurately detect the fault at this time, while the fixed threshold may not detect the fault.

Test 4 represents a state of reaction wheel motor gain loss failure. At this time k_mThe content was 70% of the normal value. As can be seen from fig. 12, the residual error increases significantly at this time, and most of the time, the residual error significantly exceeds the threshold value, and a failure occurs significantly.

Test 5 represents a state of reaction wheel motor gain loss failure. At this time k_mThe content is 60% of the normal value. As can be seen from fig. 13, the residual error increases significantly at this time. Like experiment 4, the residual error of the experiment at most moments obviously exceeds the threshold value, and the residual error is far larger than the residual errors of experiments 1 to 4 in magnitude, which is a particularly serious condition of faults.

The test results of the group show that the observer and the detector for the reaction wheel, which are designed by the invention, can not only detect the fault in time when the fault occurs, but also display the fault degree of the reaction wheel to a certain extent.

3. Verifying health assessment algorithm by using normal data and fault data of simulation model

To verify the effectiveness of the health assessment algorithm proposed in the present invention, the data of the fault injection test shown in table 2 was used for verification, and the effect of the health assessment when faults of different severity occurred in the reaction wheel was shown by the fault data of different degrees.

The training parameters of the SOM neural network of the evaluator were set to:

the number of neurons in the competition layer is 8 × 8

The topological structure is arranged as a regular hexagon

The number of training iterations is 60

After the evaluator training is complete, the following assumptions are made when performing the health assessment:

the reaction wheel has a health degree of 0.98 in a normal state

The health degree of the reaction wheel is between 0.75 and 0.90 when the reaction wheel is in a performance degradation state

Health of reaction wheel under fault condition less than 0.75

For comparison, in this example, the test data of tests 1 to 5 shown in table 2 are pieced together. As shown in fig. 14, 0 to 20 seconds indicate the health degree of

test

1, 20 to 40 seconds indicate the health degree of

test

2, 40 to 60 seconds indicate the health degree of test 3, 60 to 80 seconds indicate the health degree of

test

4, and 80 to 100 seconds indicate the health degree of test 2. In FIG. 14, the solid black line represents reaction wheel health, the dashed horizontal black line represents fault thresholds, and the dashed vertical black line represents demarcations for different tests.

As shown in fig. 14, when the reaction wheel is in the completely normal state, the degree of health is in the vicinity of 0.98, indicating that the reaction wheel is normal.

When gain loss k of motor _m90% of normal times, the reaction wheel health is near 0.81, greater than the fault threshold of 0.75, indicating that the reaction is occurringThe action wheel performance degrades but remains in a normal state. As can be seen from a comparison of fig. 10, the residual approaches the threshold but does not exceed the threshold. The fault detection and health assessment results are consistent.

When gain loss k of motor_mAt 80% of normal times, the reaction wheel health is near 0.73, slightly less than the failure threshold of 0.75, indicating that the reaction wheel is in a failed state but not deeply failing. As can be seen from a comparison of fig. 11, the residual increases, and some of the time residuals are lower than the threshold, and some of the time residuals are higher than the threshold, which is in a critical state where the performance has just degraded to a fault. The fault detection and health assessment results are consistent.

When gain loss k of motor_mAt 70% of normal times, the reaction wheel health is near 0.69, which is less than the fault threshold of 0.75, indicating that the reaction wheel is in a fault state, and the fault degree is deepened and the fault is obvious. As can be seen from a comparison of fig. 12, the residual error is significantly increased at this time, and most of the time, the residual error significantly exceeds the threshold value, so that a fault is significantly occurred. The fault detection and health assessment results are consistent.

When gain loss k of motor_mAt 60% of normal times, the reaction wheel health is near 0.66, which is evident from the failure threshold of 0.75, indicating that the reaction wheel is in a failure state, and the failure degree is very deep and the failure is significant. As can be seen from comparison with fig. 13, the residual error is significantly increased at this time, and is much larger than the residual errors of experiments 1 to 4 in terms of magnitude, which is a particularly serious fault. The fault detection and health assessment results are consistent.

The test results in this group show that the reaction wheel evaluator designed by the invention can not only detect the occurrence of the fault, but also display the degradation degree of the reaction wheel performance, and show the degradation degree in the form of health degree.

Through the above detailed demonstration of the description and the embodiment of the fault detection and health assessment method designed by the invention, it can be understood that the invention provides a perfect fault detection and health assessment method for reaction wheels. The invention verifies each algorithm by using the test data, and the verification result shows that the fault detection and health evaluation effects are accurate. According to the characteristics of the reaction wheel, based on the state observer theory and combined with the characteristics of the LSTM deep neural network and the SOM neural network, the invention designs an effective fault detection and health assessment method for the reaction wheel, and has good engineering practicability.

The above embodiments only describe the design principle of the present invention, and the shapes and names of the components in the description may be different without limitation. Therefore, a person skilled in the art of the present invention can modify or substitute the technical solutions described in the foregoing embodiments; such modifications and substitutions do not depart from the spirit and scope of the present invention.

Claims

1. A deep learning based reaction wheel fault detection and health assessment system, comprising an observer, a detector, an evaluator;

Based on actual Y and estimated Y outputs of the reaction wheels

Calculating in real time the actual output and estimated output residual R ═ R (1), R (2),.., R (N) for the reaction wheel]；

The detector is used for detecting whether the reaction wheel is in fault; the detector consists of a threshold generation module and a fault discrimination module, wherein the threshold generation module generates a thresholdThe generation module is realized based on an LSTM deep neural network, and the fault judgment module compares the magnitude relation between residual errors and self-adaptive thresholds in real time; the detector receives the input X of the reaction wheel and the estimated output of the observer

Residual error R, outputting fault state of reaction wheel; at the time t, a threshold generating module generates an adaptive threshold epsilon (t) of the current time t in real time; the fault judging module compares a residual error r (t) at the time t with a self-adaptive threshold epsilon (t) in real time, judges that the reaction wheel has a fault if the residual error is greater than the threshold value, and judges that the state of the reaction wheel is normal if the residual error is less than or equal to the threshold value;

the evaluator is configured to periodically evaluate a health status of the reaction wheel; the evaluator comprises a feature extraction module, a self-organizing map (SOM) neural network module and a health degree calculation module; the characteristic extraction module acquires and records residual errors r (T), and residual error data in a period of time T form a residual error vector R (T) ([ r (1), r (2),.;, r (T)) ]; extracting the characteristics of the residual error vector; inputting the characteristics into the SOM neural network module, and then inputting the output result of the SOM neural network module into the health degree calculation module to obtain a health degree (CV) result; the degree of health is a dimensionless scalar quantity in the range of 0-1 that characterizes the state of health of the reaction wheel; and finally evaluating the health state of the reaction wheel according to the health degree.

2. The system of claim 1, wherein the self-organizing map (SOM) neural network module trains the SOM neural network using time domain features of the residual vectors in the normal state of the reaction wheel, outputs a spatial topology of the time domain features of the residual vectors in the normal state of the reaction wheel after the training is completed, and records a position of a Best Matching Unit (BMU) in the spatial topology, and a weight vector corresponding to the BMU is denoted as U₀。

3. A deep learning based reaction wheel fault detection and health assessment method implemented based on a deep learning based reaction wheel fault detection and health assessment system according to any of claims 1-2, characterized in that the method comprises:

step S2: generating, by a threshold generation module of the detector, an adaptive threshold corresponding to an operating state of the reaction wheel and related to time based on an input torque command signal of the reaction wheel and an estimated output generated by the observer; comparing, by a fault discrimination module of the detector, the residuals in real time with an adaptive threshold: if the residual error is greater than the adaptive threshold, determining that a reaction wheel is faulty; if the residual error is less than or equal to the adaptive threshold, determining that the reaction wheel is not faulty;

4. The method according to claim 3, wherein the step S1 includes:

Calculating a difference between the actual output of the reaction wheel and the estimated output of the observer, the difference being defined as a residual R; time tThe residual r (t) of (a) is calculated by the formula

5. The method according to claim 4, wherein the step S2 includes:

the adaptive threshold at time t is calculated by:

ε(t)＝r₀(t)+β

6. The method according to claim 5, wherein the step S3 includes:

step S31: according to the residual errors, the evaluator records the residual errors within a period of time T to form a residual error vector R (T) ([ r (1), r (2),.. multidot., r (T)) ], and the feature extraction module calculates the time domain features of the residual error vector and takes the time domain features as the time domain features of the residual error vector to be evaluated for health;

Calculate U₀And U₁Is spaced fromThe distance is recorded as a Minimum Quantization Error (MQE); MQE, representing the deviation relation between the residual vector to be evaluated and the residual vector when the reaction wheel works normally, namely the deviation degree of the characteristic space corresponding to the current running state and the normal state of the reaction wheel respectively;

M＝||U₁-U₀||

wherein M represents an MQE value;

the evaluator's health calculation module normalizes MQE to yield a CV value:

where b is a normalization coefficient.

7. The method of claim 6, wherein the feature values corresponding to the time domain features comprise root mean square values, peak values, and mean absolute values, and are calculated by:

root mean square value:

peak value:

B＝max(|r(t)|)

average absolute value: