Reaction wheel fault detection and health assessment system and method based on deep learning

Abstract

The present invention discloses a reaction wheel fault detection and health assessment system and method based on deep learning, which can solve the problem that the reaction wheel fault detection is difficult and the normal data cannot be fully utilized to assess its health status when the reaction wheel fault data is small. The reaction wheel fault detection and health assessment system includes an observer, a detector, and an evaluator; the observer is used to track the dynamic state of the reaction wheel and generate a real-time residual; the detector generates an adaptive threshold of the current time in real time, compares the relationship between the residual and the threshold, and detects whether the reaction wheel is faulty; the evaluator is used to periodically evaluate the health status of the reaction wheel; the evaluator obtains and records the residual, and composes the residual data of a period of time into a residual vector; extracts the features of the residual vector; calculates the spatial topological structure of the feature vector, and calculates the spatial deviation, and finally converts it into a health degree (CV), thereby realizing health assessment.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a reaction wheel fault detection and health assessment system and method based on deep learning. Background Art

[0002] The attitude and orbit control system (AOCS) is an important part of the spacecraft. When the spacecraft is affected by external disturbance torque, the AOCS is used to control the stability and orientation of the spacecraft to the predetermined direction during operation. The space environment where the spacecraft is located is harsh and is often disturbed by various torques. Many spacecraft fail before completing their missions. More than 30% of spacecraft failures are caused by AOCS, of which nearly 50% are caused by actuators. 60% of the actuator-induced failures are caused by reaction wheels and control torque gyroscopes. Fault detection and health assessment of reaction wheels is one of the important means to improve the reliability and safety of spacecraft AOCS. Early fault detection of reaction wheels can detect spacecraft failures in a timely manner, and then quickly make corresponding treatments, thereby effectively improving the safety of spacecraft. Spacecraft failures can be effectively avoided by performing health assessments on reaction wheels and taking measures before they fail.

[0003] As one of the important components of AOCS, the fault detection problem of reaction wheel has been widely studied in the past few decades. 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. This method needs to consider the model structure of the system and the measurement of the actual process. The accuracy of fault detection depends on the accuracy of modeling. The advantage of this method is that the fault detection result is accurate, and the disadvantage is that the modeling process is relatively complicated and the verification of the model is relatively difficult. Data-driven methods are often used in situations where an accurate analytical model cannot be obtained. This type of method uses a large amount of historical data and uses algorithms such as neural networks or fuzzy logic as function approximators to construct an analytical model of the system. They learn the reaction wheel dynamics model from the input and output data and learn fault knowledge from historical data. However, the main problem of these methods is that they have high requirements for historical data and computational complexity. The more successful existing technology is to combine the above two methods, use methods such as Kalman filters or neural networks to construct observers, estimate the state of the system from the available input and output measurements, and then realize fault detection. However, the methods based on traditional neural networks or Kalman filters have the following shortcomings: (1) It is difficult to simulate the highly nonlinear dynamic characteristics of the momentum wheel, and it is difficult to obtain the time series relationship between its input and output. (2) When the reaction wheel experiences complex working conditions and multiple disturbances, it will cause many interferences, false alarms, and difficulty in fault detection under multiple working conditions. (3) In the field of fault diagnosis and health management, health assessment has more practical guiding significance than fault detection. However, due to the lack of fault data and the unclear fault evolution law, there is little research on the health assessment of the reaction wheel, and normal data has not been fully utilized to judge the health status of the reaction wheel. Summary of the invention

[0004] In view of this, the present invention provides a reaction wheel fault detection and health assessment system and method based on deep learning, which can solve the problems of reaction wheel fault detection with many interferences, many false alarms and high difficulty; and the problem of failure to fully utilize normal data to assess its health status when there is less reaction wheel fault data.

[0005] In order to solve the above technical problems, the present invention is implemented as follows.

[0006] A deep learning-based reaction wheel fault detection and health assessment system, the reaction wheel fault detection and health assessment system comprising an observer, a detector, and an evaluator;

[0007] The observer is used to track the dynamic state of the reaction wheel, and the observer is composed of a long short-term memory network (LSTM); the observer obtains the combination of the input X of the reaction wheel and the multi-order delayed output; the relationship between the input and output of the reaction wheel is simulated by the LSTM deep neural network of the observer, and the estimated output of the reaction wheel is output by the LSTM deep neural network Based on the actual output Y of the reaction wheel and the estimated output Calculate the actual output of the reaction wheel and the estimated output residual R = [r(1), r(2), ..., r(N)] in real time;

[0008] The detector is used to detect whether the reaction wheel fails; the detector is composed of a threshold generation module and a fault discrimination module, the threshold generation module is implemented based on the LSTM deep neural network, and the fault discrimination module compares the magnitude relationship between the residual and the adaptive threshold in real time; the detector receives the input X of the reaction wheel and the estimated output of the observer The residual R outputs the fault status of the reaction wheel; at time t, the threshold generation module generates the adaptive threshold ε(t) of the current time t in real time; the fault discrimination module compares the residual r(t) at time t with the adaptive threshold ε(t) in real time: if the residual is greater than the threshold, it is determined that the reaction wheel has a fault; if the residual is less than or equal to the threshold, it is determined that the reaction wheel is in a normal state;

[0009] The evaluator is used to periodically evaluate the health status of the reaction wheel; the evaluator includes a feature extraction module, a self-organizing map (SOM) neural network module, and a health calculation module; the feature extraction module obtains and records the residual r(t), and composes the residual data within a period of time T into a residual vector R(T)=[r(1), r(2),…, r(T)]; extracts the features of the residual vector; inputs the features into the SOM neural network module, and then inputs the output result of the SOM neural network module into the health calculation module to obtain a health (CV) result; the health is a dimensionless scalar in the range of 0-1 that characterizes the health status of the reaction wheel; and finally, the health status of the reaction wheel is evaluated according to the health.

[0010] Preferably, the self-organizing map (SOM) neural network module uses the time domain characteristics of the residual vector of the reaction wheel in the normal state to train the SOM neural network. After the training is completed, the spatial topological structure of the time domain characteristics of the residual vector of the reaction wheel in the normal state can be output, and the position of the best matching unit (BMU) in the spatial topological structure is recorded. The weight vector corresponding to the BMU is recorded as U0.

[0011] A deep learning-based reaction wheel fault detection and health assessment method, the method is implemented based on the deep learning-based reaction wheel fault detection and health assessment system as described above, the method comprising:

[0012] Step S1: the observer acquires the input and output of the reaction wheel in real time, wherein the input is the input torque command signal of the reaction wheel, and the output is the output torque signal of the reaction wheel; based on the input torque command signal and the output torque signal of the reaction wheel, the observer generates the estimated output of the reaction wheel in real time, calculates the difference between the actual output of the reaction wheel and the estimated output of the observer, and defines the difference as a residual;

[0013] Step S2: Based on the input torque command signal of the reaction wheel and the estimated output generated by the observer, the threshold generation module of the detector generates an adaptive threshold corresponding to the working state of the reaction wheel and related to time; the fault discrimination module of the detector compares the residual and the adaptive threshold in real time: if the residual is greater than the adaptive threshold, it is determined that the reaction wheel is faulty; if the residual is less than or equal to the adaptive threshold, it is determined that the reaction wheel is in normal state;

[0014] Step S3: while fault detection is in progress, the residual is input into an evaluator, which outputs the healthiness (CV) of the reaction wheel; the healthiness is a dimensionless scalar in the range of 0-1 that characterizes the health status of the reaction wheel; the health status of the reaction wheel is analyzed based on the healthiness to determine whether the performance of the reaction wheel is degraded.

[0015] Preferably, the step S1 comprises:

[0016] The input torque command signal X and output torque signal Y of the reaction wheel are used to simulate the dynamic characteristics of the reaction wheel using an observer based on the LSTM deep neural network. (-2) , generates the estimated output of the reaction wheel in real time Calculate the difference between the actual output of the reaction wheel and the estimated output of the observer, which is defined as the residual R; the calculation formula of the residual r(t) at time t is

[0017] Preferably, the step S2 comprises:

[0018] Using the input signal of the reaction wheel at time t and the output signal of the observer, the threshold generation module based on the LSTM deep neural network in the detector is used to generate the adaptive threshold of the reaction wheel at that time;

[0019] The adaptive threshold at time t is calculated as follows:

[0020] ε(t)=r0(t)+β

[0021] Among them, r0(t) represents the residual at time t when the reaction wheel works normally, and β represents the correction coefficient; the correction coefficient is used to compensate for the residual fluctuation caused by factors such as the parameters and disturbances of the reaction wheel drifting over time;

[0022] The fault identification module of the detector compares the threshold ε(t) with the residual r(t): if the residual is greater than the threshold, it is considered that the reaction wheel is faulty, otherwise it is considered that it is in a normal state.

[0023] Preferably, the step S3 comprises:

[0024] Step S31: Based on the residual, the evaluator records the residual within a period of time T to form a residual vector R(T)=[r(1), r(2), ..., r(T)], and the feature extraction module calculates the time domain features of the residual vector and uses it as the time domain features of the residual vector to be evaluated;

[0025] Step S32: input the time domain features of the residual vector to be health-evaluated into the SOM neural network module of the evaluator, and the SOM neural network module in the evaluator outputs the spatial topological structure of the time domain features of the residual vector at this time, records the BMU position in the spatial topological structure, and the weight vector corresponding to the BMU is recorded as U1;

[0026] Calculate the spatial distance between U0 and U1, and record the spatial distance as the minimum quantization error (MQE); MQE represents the deviation relationship between the residual vector to be evaluated and the residual vector when the reaction wheel works normally, that is, the offset of the feature space corresponding to the current operating state and the normal state of the reaction wheel;

[0027] M=||U1-U0||

[0028] Wherein, M represents the MQE value;

[0029] The health calculation module of the evaluator normalizes MQE to obtain the CV value:

[0030]

[0031] Where b is the normalization coefficient.

[0032] When the reaction wheel is working properly, the health CV is between 0.9 and 1.

[0033] Preferably, the characteristic values ​​corresponding to the time domain characteristics include root mean square value, peak value, and average absolute value, which are calculated as follows:

[0034] RMS value:

[0035]

[0036] Peak:

[0037] B=max(|r(t)|)

[0038] Mean absolute value:

[0039]

[0040] Where T represents the length of the residual sequence recorded by the evaluator; r(t) represents the residual at time t; A represents the root mean square value of the residual vector; B represents the peak value of the residual vector; and C represents the average absolute value of the residual vector.

[0041] Beneficial effects:

[0042] (1) In view of the high nonlinearity, complex disturbances and diverse working conditions of the reaction wheel, the LSTM deep neural network is used to model the complex dynamic characteristics of the reaction wheel, simulate the time series relationship between its input and output, and establish a high-precision observer to generate estimated outputs in real time and output the corresponding residuals.

[0043] (2) Conventional observer-based fault detection methods usually introduce a fixed threshold to determine whether the residual exceeds the limit. In order to detect the fault of 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 affecting the accuracy of fault detection. If the threshold is defined too high, some faults may not be detected; if the threshold is defined too low, false alarms will often occur. In order to avoid this situation, the present invention designs a method for generating an adaptive threshold, uses an LSTM deep neural network to design a threshold generation module of the detector, and the threshold generation module outputs an adaptive threshold.

[0044] The present invention uses LSTM deep neural network to mine the adaptive relationship between the output of the reaction wheel and the corresponding fault threshold under different working conditions and disturbances, and establishes a detector based on adaptive threshold, so as to timely judge whether the reaction wheel has a fault. By introducing adaptive threshold, the accuracy of fault detection is effectively improved and the false alarm rate of detection is reduced.

[0045] (3) The SOM neural network focuses on learning the topological structure of the data while learning the data features, and can represent multi-dimensional features in one-dimensional or two-dimensional space while retaining the topological structure of the input feature space; the SOM neural network is an unsupervised machine learning algorithm, so it 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.

[0046] The SOM neural network is used to mine the spatial topological relationship between the health characteristic parameters of the reaction wheel when it is degraded and in normal state, and a model of the health degradation law of the reaction wheel is established, thereby realizing the health assessment of the reaction wheel. When using the evaluator based on the SOM neural network to model the health characteristics, only the data of the normal operation of the reaction wheel is required, and the health assessment of the reaction wheel can be realized without complex and diverse fault data, which effectively improves the engineering practicality of the health assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 The architecture diagram of the reaction wheel fault detection and health assessment system;

[0048] Figure 2 It is a flowchart of the reaction wheel fault detection and health assessment method based on deep learning of the present invention;

[0049] Figure 3 is a schematic diagram of the detector structure;

[0050] Figure 4 It is a schematic diagram of the evaluator structure;

[0051] Figure 5 It is a simulation model of reaction wheel;

[0052] Figure 6 This is a frequency swept sinusoidal signal diagram of the present invention;

[0053] Figure 7 This is a Monte Carlo sine signal diagram of the present invention;

[0054] Figure 8 The Monte Carlo sinusoidal signal input of the present invention is a schematic diagram of the reaction wheel output and the observer residual;

[0055] Fig. 9 This is a schematic diagram of the residual and adaptive threshold of Experiment 1 of the present invention;

[0056] Fig.10 This is a schematic diagram of the residual and adaptive threshold of Experiment 2 of the present invention;

[0057] Fig.11 This is a schematic diagram of the residual and adaptive threshold of Experiment 3 of the present invention;

[0058] Fig.12 This is a schematic diagram of the residual and adaptive threshold of Experiment 4 of the present invention;

[0059] Fig.13 This is a schematic diagram of the residual and adaptive threshold of Experiment 5 of the present invention;

[0060] Fig.14 This is a schematic diagram of the health of experiments 1-5 of the present invention. DETAILED DESCRIPTION

[0061] The present invention is described in detail below in conjunction with the accompanying drawings and embodiments.

[0062] The present invention is a reaction wheel fault detection and health assessment system and method based on deep learning, such as Figure 1 As shown, the reaction wheel fault detection and health assessment system includes an observer, a detector, and an evaluator.

[0063] The observer and the detector of the present invention constitute a reaction wheel fault detection part, and the observer and the evaluator constitute a reaction wheel health evaluation part.

[0064] The input of the reaction wheel is a torque command signal, represented by X=[x(1), x(2), ..., x(N)]; the output of the reaction wheel is a torque signal, represented by Y=[y(1), y(2), ..., y(N)], where N is the length of the signal sequence; at time t, the input of the reaction wheel is x(t), and the output is y(t), 3≤t≤N, and the first-order delayed output of the reaction wheel is y (-1) =y(t-1), represents the reaction wheel output torque at time t-1; the second-order delayed output of the reaction wheel is expressed as y (-2) =y(t-2), represents the reaction wheel torque output torque at time t-2; therefore, the first-order delayed output of the reaction wheel is expressed as Y (-1) =[y(1),y(2),…,y(N-1)], the second-order delay output is represented by Y (-2) =[y(1),y(2),…,y(N-2)].

[0065] The observer is used to track the dynamic state of the reaction wheel, and the observer is composed of a long short-term memory network (LSTM). The observer obtains the combination of the input X of the reaction wheel and the multi-order delayed output; generally speaking, the first-order and higher-order delayed outputs or their combinations can be used. The present invention uses the combination Y of the first-order delayed output and the second-order delayed output (-2) =[Y (-1) ; Y (-2) ] is used as an example to illustrate the technical process; the relationship between the input and output of the reaction wheel is simulated by the LSTM deep neural network of the observer, and the estimated output of the reaction wheel is output by the LSTM deep neural network Based on the actual output Y of the reaction wheel and the estimated output The actual output of the reaction wheel and the estimated output residual R = [r(1), r(2), ..., r(N)] are calculated in real time.

[0066] The detector is used to detect whether the reaction wheel fails; the detector is composed of a threshold generation module and a fault discrimination module, the threshold generation module is implemented based on a long short-term memory network (LSTM), and the fault discrimination module compares the magnitude relationship between the residual and the adaptive threshold in real time; the detector receives the input X of the reaction wheel and the estimated output of the observer The residual R outputs the fault status of the reaction wheel; at time t, the threshold generation module generates the adaptive threshold ε(t) of the current time t in real time; the fault discrimination module compares the residual r(t) at time t with the adaptive threshold ε(t) in real time: if the residual is greater than the threshold, it is determined that the reaction wheel has a fault; if the residual is less than or equal to the threshold, it is determined that the reaction wheel is in normal state.

[0067] The evaluator is used to periodically evaluate the health status of the reaction wheel; the evaluator includes a feature extraction module, a self-organizing map (SOM) neural network module, and a health calculation module; the feature extraction module obtains and records the residual r(t), and composes the residual data within a period of time T into a residual vector R(T)=[r(1), r(2),…, r(T)]; extracts the features of the residual vector; inputs the features into the SOM neural network module, and then inputs the output result of the SOM neural network module into the health calculation module to obtain a health (CV) result; the health is a dimensionless scalar in the range of 0-1 that characterizes the health status of the reaction wheel; and finally, the health status of the reaction wheel is evaluated according to the health.

[0068] The self-organizing map (SOM) neural network module uses the time domain features of the residual vector of the reaction wheel in the normal state to train the SOM neural network. After the training is completed, the spatial topological structure of the time domain features of the residual vector of the reaction wheel in the normal state can be output, and the position of the best matching unit (BMU) in the spatial topological structure is recorded. The weight vector corresponding to the BMU is recorded as U0.

[0069] like Figure 2 As shown, a reaction wheel fault detection and health assessment method based on deep learning of the present invention is based on a reaction wheel fault detection and health assessment system, and the method includes:

[0070] Step S1: the observer acquires the input and output of the reaction wheel in real time, wherein the input is the input torque command signal of the reaction wheel, and the output is the output torque signal of the reaction wheel; based on the input torque command signal and the output torque signal of the reaction wheel, the observer generates the estimated output of the reaction wheel in real time, calculates the difference between the actual output of the reaction wheel and the estimated output of the observer, and defines the difference as a residual;

[0071] Step S2: Based on the input torque command signal of the reaction wheel and the estimated output generated by the observer, the threshold generation module of the detector generates an adaptive threshold corresponding to the working state of the reaction wheel and related to time; the fault discrimination module of the detector compares the residual and the adaptive threshold in real time: if the residual is greater than the adaptive threshold, it is determined that the reaction wheel is faulty; if the residual is less than or equal to the adaptive threshold, it is determined that the reaction wheel is normal;

[0072] Step S3: while fault detection is in progress, the residual is input into an evaluator, which outputs the health of the reaction wheel, analyzes the health status of the reaction wheel based on the health, and determines whether the performance of the reaction wheel is degraded.

[0073] The present invention utilizes the input and output relationship of the reaction wheel to establish an observer and a detector to realize the reaction wheel fault detection. The specific process of fault detection is as follows: Figure 3 The present invention uses the residual to establish an evaluator to achieve the health assessment of the reaction wheel. The specific process of health assessment is as follows Figure 3 shown.

[0074] The step S1 comprises: using the input torque command signal X and the output torque signal Y of the reaction wheel, using an observer based on the LSTM deep neural network to simulate the dynamic characteristics of the reaction wheel, and outputting the combination Y according to the input and multi-order delay of the reaction wheel. (-2) , generates the estimated output of the reaction wheel in real time The difference between the actual output of the reaction wheel and the estimated output of the observer is calculated and defined as the residual R.

[0075] In order to more accurately describe the nonlinear time series relationship between the reaction wheel input and output, the present invention uses an LSTM deep neural network to construct an observer. Figure 2 As shown, at time t, x(t) represents the input of the reaction wheel, and y(t) represents the output of the reaction wheel. represents the estimated output of the observer.

[0076] The LSTM deep neural network is a deep recurrent neural network architecture that can identify the relationship between relevant important signals in a time series, can effectively capture the long-term time dependency between important signals, and is suitable for modeling and predicting time series data. For the reaction wheel, LSTM is particularly suitable for identifying its tiny state change (fault) information during the real-time operation of the reaction wheel. 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 gradually increases. The threshold generation modules of the observer and the detector in the present invention are two LSTM deep neural networks with different structures. The observer based on the LSTM deep neural network calculates the estimated output value of the reaction wheel The residual is defined as the difference between the reaction wheel output and the observer's estimated output:

[0077]

[0078] In the above formula, r(t) represents the residual at time t.

[0079] When the reaction wheel works properly, the residual is close to 0, usually less than 0.1, because only noise and modeling errors affect the residual. However, when the reaction wheel does not work properly, the residual will increase accordingly.

[0080] The step S2 includes: using the input signal of the reaction wheel and the output signal of the observer, and using the threshold generation module based on the LSTM deep neural network in the detector to generate the adaptive threshold of the reaction wheel at that moment. The working state of the reaction wheel can be estimated by the actual input of the reaction wheel and the estimated output generated by the observer, so as to generate an adaptive threshold according to the working state of the reaction wheel, and the threshold changes with the working state of the reaction wheel. The fault discrimination module of the detector compares the residual and the adaptive threshold in real time: if the residual is greater than the threshold, the reaction wheel is considered to be faulty, otherwise it is considered to be in a normal state.

[0081] Considering that the threshold is usually affected by the input, output, load, environmental disturbance, etc. of the reaction wheel, the present invention calculates the adaptive threshold through the input of the reaction wheel and the output of the observer. Figure 3 As shown, the threshold generation module in the detector uses the input of the reaction wheel and the output value of the observer as input, and its output is the adaptive threshold. When the LSTM deep neural network of the threshold generation module is trained, the adaptive threshold at time t is calculated by the following formula:

[0082] ε(t)=r0(t)+β (5)

[0083] In the above formula, r0(t) represents the residual at time t when the reaction wheel works normally, and β represents the correction coefficient. The correction coefficient is used to compensate for the residual fluctuation caused by factors such as the drift of the reaction wheel parameters and disturbances over time to enhance the robustness of the adaptive threshold.

[0084] Since the input command of the reaction wheel and the output of the observer are used as the detector input when calculating the adaptive threshold, its input is not affected by whether the reaction wheel is faulty or not. When the reaction wheel is normal, due to the setting of the correction coefficient, the threshold is greater than the residual; when the reaction wheel fails, the threshold is still based on the normal condition of the reaction wheel, and its order of magnitude will not change significantly. However, at this time, due to the reaction wheel failure, the actual residual will become significantly larger and exceed the threshold, and the fault can be detected. In addition, in terms of working conditions, since a large amount of working condition information is contained in the input of the reaction wheel, the disturbance of the reaction wheel (increased friction caused by temperature, etc.), the observer output and the observer residual, based on the recognition and state tracking capabilities of the LSTM deep neural network, the threshold of the LSTM deep neural network output of the threshold generation module of the detector will change with the change of working conditions, achieving the purpose of adaptive variable threshold.

[0085] like Figure 4 As shown, the step S3 includes: while detecting the fault, passing the residual to the evaluator, the evaluator periodically records the residual to obtain a residual vector, further analyzing the health status of the reaction wheel according to the residual vector, and judging whether the reaction wheel has a certain performance degradation. Assuming that the residual vector has been obtained at time t, first the feature extraction module calculates the eigenvalue of the residual vector, and then the SOM neural network module calculates the spatial topological structure of the residual vector, which is the output of the SOM neural network, and the health calculation module calculates the health of the reaction wheel according to the output of the SOM neural network. The step S3 specifically includes:

[0086] Step S31: Based on the residual, the evaluator records the residual within a period of time T to form a residual vector R(T)=[r(1), r(2),…, r(T)], and the feature extraction module calculates the time domain features of the residual vector. The feature values ​​include the root mean square value, peak value, and average absolute value, and the specific calculation method is:

[0087] RMS value:

[0088]

[0089] Peak:

[0090] B=max(|r(t)|) (7)

[0091] Mean absolute value:

[0092]

[0093] Where T represents the length of the residual sequence recorded by the evaluator; r(t) represents the residual at time t; A represents the root mean square value of the residual vector; B represents the peak value of the residual vector; and C represents the average absolute value of the residual vector.

[0094] Step S32: input the time domain features of the residual vector into the SOM neural network module of the evaluator, and the SOM neural network module in the evaluator outputs the spatial topological structure of the residual vector; when designing the SOM neural network module, the time domain features of the residual vector when the reaction wheel is in a normal state are used to train the SOM neural network, and after the training is completed, the spatial topological structure of the time domain features of the residual vector when the reaction wheel is in a normal state can be output, and the position of the best matching unit (BMU) in the spatial topological structure is recorded, and the weight vector corresponding to the BMU is recorded as U0; during health assessment, the time domain features of the residual vector to be health assessed are input into the SOM neural network module to obtain the spatial topological structure of the time domain features of the residual vector at this time, and the position of the BMU in the spatial topological structure is recorded, and the weight vector corresponding to the BMU is recorded as U1; then the spatial distance between U0 and U1 is calculated, and the spatial distance is recorded as the minimum quantization error (MQE); MQE represents the deviation relationship between the residual vector to be evaluated and the residual vector when the reaction wheel is working normally, that is, the offset degree of the feature space corresponding to the current operating state and the normal state of the reaction wheel. Therefore, the change of MQE can be used to represent the health degradation process of the reaction wheel.

[0095] M=||U1-U0|| (9)

[0096] Wherein, M represents the MQE value.

[0097] When MQE increases, it indicates that the performance degradation of the reaction wheel continues to worsen. In order to intuitively display the health status of the reaction wheel, the present invention uses health to indicate the health status of the reaction wheel. The greater the health, the greater the possibility that the reaction wheel is in a good state, and the lower the health, the reaction wheel is in performance degradation or has failed. The health calculation module of the evaluator normalizes MQE to obtain the CV value:

[0098]

[0099] Where b is the normalization coefficient.

[0100] The health results output by the evaluator are expressed as CV values, which are dimensionless scalars ranging from 0 to 1, where 0 indicates complete failure and 1 indicates no performance degradation. First, an observer based on an LSTM deep neural network is used to estimate the output of the reaction wheel to calculate the residual. Then, the eigenvalues ​​are extracted from the residual, and then the minimum quantization error (MQE) is calculated using an evaluator based on a SOM neural network. Finally, the health of the reaction wheel is calculated by normalizing the MQE.

[0101] The following describes the reaction wheel fault detection and health assessment method based on deep learning of the present invention in conjunction with specific embodiments.

[0102] 1. Simulation model design and algorithm model training

[0103] The embodiment of the present invention uses the simulation data of the reaction wheel for verification. The simulation model of the reaction wheel is established using MATLAB / SIMULINK, including a motor torque control module, an electromotive force torque module, a motor disturbance module and a friction torque module. The simulation model is as follows: Figure 5 The parameters of each part are shown in Table 1.

[0104] Table 1 Simulation model parameters

[0105]

[0106] Taking into account the uncertainty of the reaction wheel input when performing real-time fault detection, when training the observer, detector and evaluator, the training data set and the test data set should contain typical signals within the corresponding frequency and amplitude ranges. Assume that the frequency range of the input signal of the reaction wheel is [f0, f1], and the amplitude range is [A0, A1]. The present invention uses a swept frequency sine signal for training, and the amplitude and frequency of the swept frequency sine signal increase linearly with time within the corresponding amplitude and frequency range. The present invention uses a large number of swept frequency signals to train the observer model, so that it can simulate the input-output relationship of the reaction wheel under various working conditions. Figure 6 The signal shown is a 120-second sweep signal. To verify the training effect, the present invention uses a Monte Carlo sine signal for verification. The amplitude and frequency of the Monte Carlo sine signal change randomly within the corresponding amplitude and frequency range every cycle, such as Figure 7 As shown. By inputting a swept frequency sine signal to the reaction wheel, its output signal is obtained, and the observer is trained using this input-output signal. After the training is completed, the Monte Carlo sine signal is used for verification, as shown in Figure 8 The results show that the observer can simulate the input-output relationship of the reaction wheel very well.

[0107] Typical failure modes of the reaction wheel include: motor gain loss, voltage fluctuation, increased disturbance torque, and increased Coulomb friction. The present invention injects a motor gain loss fault into the reaction wheel, and verifies the fault detection algorithm and health assessment algorithm proposed in the present invention through different fault degrees. In the simulation test, the data sampling rate is 100 points / second. Based on the system architecture of the present invention, after the observer training is completed, a certain input instruction can be used to verify the fault detection and health assessment algorithm:

[0108] x(t)=sin(πt) (11)

[0109] 2. Use the normal data and fault data of the simulation model to verify the fault detection algorithm

[0110] In order to verify the effectiveness of the fault detection algorithm proposed in the present invention, four fault injection tests were conducted here, and the fault severity was successively deepened. The simulation time of each test group was set to 20 seconds, and the fault injection was performed at the 0th second. The test details are shown in Table 2.

[0111] Table 2 Detailed parameters of fault injection test

[0112]

[0113] The LSTM deep neural network of the observer designed in the present invention is trained using the Adam algorithm, and the network structure is as follows:

[0114] Sequence input layer (3D features)

[0115] LSTM neuron layer (200 neurons)

[0116] Dropout layer (probability 20%)

[0117] LSTM neuron layer (100 neurons)

[0118] Dropout layer (probability 20%)

[0119] Fully connected layer (50-dimensional features)

[0120] Dropout layer (probability 20%)

[0121] Fully connected layer (1-dimensional features)

[0122] Regression Output Layer

[0123] The LSTM deep neural network of the threshold generation module of the detector designed in the present invention is trained using the Adam algorithm, and the network structure is as follows:

[0124] Sequence input layer (2D features)

[0125] LSTM neuron layer (200 neurons)

[0126] Dropout layer (probability 20%)

[0127] Fully connected layer (50-dimensional features)

[0128] Dropout layer (probability 50%)

[0129] Fully connected layer (1-dimensional features)

[0130] Regression Output Layer

[0131] The results of experiments 1 to 5 are as follows Figure 9-13 Shown, the black solid line in the figure represents the residual, and the black dotted line represents the adaptive threshold.

[0132] Test 1 shows the state of the reaction wheel when it is normal. Fig. 9 As shown in the figure, the residual in this test is far below the threshold, indicating that the reaction wheel works normally. It also shows that the threshold is automatically adjusted according to the change of working status.

[0133] Test 2 shows the state of reaction wheel motor gain loss fault. At this time, km is 90% of normal. Fig.10 It can be seen that the residual increases at this time, and some moments are closer to the threshold, but still do not exceed the threshold.

[0134] Test 3 represents the state of reaction wheel motor gain loss fault. m 80% of normal. Fig.11 It can be seen that the residual increases at this time, and the residual is lower than the threshold at some times, and higher than the threshold at some times. This experiment shows that the adaptive threshold can accurately detect the fault at this time, while the fixed threshold may not be able to detect the fault.

[0135] Test 4 represents the state of reaction wheel motor gain loss fault. m 70% of normal. Fig.12 It can be seen that the residual increases significantly at this time, and most of the time the residual exceeds the threshold, indicating that a fault has occurred.

[0136] Test 5 represents the state of reaction wheel motor gain loss fault. m 60% of normal. Fig.13 It can be seen that the residual increases significantly at this time. Like test 4, the residual of this test significantly exceeds the threshold at most times. From the order of magnitude, the residual is much larger than the residuals of tests 1 to 4, which is a particularly serious fault.

[0137] This group of test results show that the reaction wheel observer and detector designed by the present invention can not only detect the fault in time when it occurs, but also can display the degree of the reaction wheel fault to a certain extent.

[0138] 3. Use the normal data and fault data of the simulation model to verify the health assessment algorithm

[0139] In order 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 is used for verification, and the effect of the health assessment when the reaction wheel has faults of different severity is displayed through fault data of different degrees.

[0140] The training parameters of the SOM neural network of the estimator are set as:

[0141] The number of neurons in the competition layer is 8×8

[0142] The topology is set to a regular hexagon

[0143] The number of training iterations is 60

[0144] After the evaluator is trained, the following assumptions are made when performing health assessment:

[0145] The health of the reaction wheel is 0.98 when it is in normal state.

[0146] When the reaction wheel is in a state of performance degradation, the health level is between 0.75 and 0.90.

[0147] The health of the reaction wheel is below 0.75 when it is in a faulty state

[0148] For the convenience of comparison, the test data of Tests 1 to 5 shown in Table 2 are spliced ​​in this embodiment. Fig.14 As shown, 0 to 20 seconds represents the health of test 1, 20 to 40 seconds represents the health of test 2, 40 to 60 seconds represents the health of test 3, 60 to 80 seconds represents the health of test 4, and 80 to 100 seconds represents the health of test 2. Fig.14 In the figure, the solid black line represents the health of the reaction wheel, the horizontal black dashed line represents the fault threshold, and the vertical black dashed line represents the boundary between different tests.

[0149] like Fig.14 As shown, when the reaction wheel is in a completely normal state, the health degree is around 0.98, indicating that the reaction wheel is normal.

[0150] When the motor gain loss k m When it is 90% of the normal state, the health of the reaction wheel is around 0.81, which is greater than the fault threshold of 0.75, indicating that the performance of the reaction wheel is degraded, but still in normal condition. Fig.10It can be seen that the residual is close to the threshold, but does not exceed the threshold. The fault detection and health assessment results are consistent.

[0151] When the motor gain loss k m When it is 80% of the normal state, the health of the reaction wheel is around 0.73, slightly less than the fault threshold of 0.75, indicating that the reaction wheel is in a fault state, but the degree of fault is not serious. Fig.11 It can be seen that the residual increases at this time, and the residual is lower than the threshold at some times, and higher than the threshold at some times, which is in the critical state where the performance just degrades to failure. The fault detection and health assessment results are consistent.

[0152] When the motor gain loss k m When it is 70% of the normal state, the health of the reaction wheel is around 0.69, which is less than the fault threshold of 0.75, indicating that the reaction wheel is in a fault state at this time, the fault degree has deepened, and a fault has obviously occurred. Fig.12 It can be seen that the residual increases significantly at this time, and most of the time the residual exceeds the threshold, indicating that a fault has occurred. The fault detection and health assessment results are consistent.

[0153] When the motor gain loss k m When it is 60% of the normal state, the health of the reaction wheel is around 0.66, which is significantly higher than the fault threshold of 0.75, indicating that the reaction wheel is in a fault state at this time, the fault is very serious, and the fault is significant. Fig.13 It can be seen that the residual increases significantly at this time. In terms of magnitude, the residual is much larger than the residuals of test 1 to test 4, which is a particularly serious fault. The fault detection and health assessment results are consistent.

[0154] This group of test results show that the reaction wheel evaluator designed by the present invention can not only detect the occurrence of faults, but also display the degree of performance degradation of the reaction wheel and display it in the form of health.

[0155] Through the above description of the fault detection and health assessment method designed by the present invention and the detailed demonstration of the embodiment, it can be seen that the present invention proposes a complete fault detection and health assessment method for the reaction wheel. The present invention verifies each algorithm using test data, and the verification results show that the fault detection and health assessment effects are accurate. According to the characteristics of the reaction wheel, based on the state observer theory, combined with the characteristics of the LSTM deep neural network and the SOM neural network, the invention designs an effective reaction wheel fault detection and health assessment method, which has good engineering practicality.

[0156] The above specific embodiments only describe the design principle of the present invention. The shapes and names of the components in the description may be different and are not limited. Therefore, those skilled in the art in the field of the present invention may modify or replace the technical solutions recorded in the above embodiments; and these modifications and replacements do not deviate from the creative purpose and technical solutions of the present invention and should all fall within the protection scope of the present invention.

Claims

1. A reaction wheel fault detection and health assessment method based on deep learning, the method is implemented based on a reaction wheel fault detection and health assessment system based on deep learning, characterized in that: The reaction wheel fault detection and health assessment system includes an observer, a detector, and an evaluator; The observer is used to track the dynamic state of the reaction wheel, and the observer is composed of a long short-term memory network (LSTM); the observer obtains the combination of the input X of the reaction wheel and the multi-order delayed output; the relationship between the input and output of the reaction wheel is simulated by the LSTM deep neural network of the observer, and the estimated output of the reaction wheel is output by the LSTM deep neural network Based on the actual output Y of the reaction wheel and the estimated output Calculate the actual output of the reaction wheel and the estimated output residual R = [r(1), r(2), ..., r(N)] in real time; The detector is used to detect whether the reaction wheel fails; The detector is composed of a threshold generation module and a fault discrimination module. The threshold generation module is implemented based on the LSTM deep neural network. The fault discrimination module compares the magnitude relationship between the residual and the adaptive threshold in real time. The detector receives the input X of the reaction wheel and the estimated output of the observer. Residual R, outputs the fault status of the reaction wheel; At time t, the threshold generation module generates an adaptive threshold ε(t) of the current time t in real time; the fault discrimination module compares the residual r(t) at time t with the adaptive threshold ε(t) in real time, and if the residual is greater than the threshold, it is determined that the reaction wheel has a fault; if the residual is less than or equal to the threshold, it is determined that the reaction wheel is in a normal state; The evaluator is used to periodically evaluate the health status of the reaction wheel; the evaluator includes a feature extraction module, a self-organizing map (SOM) neural network module, and a health calculation module; the feature extraction module obtains and records the residual r(t), and forms a residual vector R(T)=[r(1), r(2),…, r(T)] with the residual data within a period of time T; extracts the features of the residual vector; inputs the features into the SOM neural network module, and then inputs the output result of the SOM neural network module into the health calculation module to obtain a health (CV) result; Health is a dimensionless scalar ranging from 0 to 1 that characterizes the health status of the reaction wheel; Finally, the health status of the reaction wheel is evaluated according to the health degree; The self-organizing map (SOM) neural network module uses the time domain features of the residual vector of the reaction wheel in the normal state to train the SOM neural network. After the training is completed, the spatial topological structure of the time domain features of the residual vector of the reaction wheel in the normal state can be output, and the position of the best matching unit (BMU) in the spatial topological structure is recorded. The weight vector corresponding to the BMU is recorded as U0. The method includes: Step S1: the observer acquires the input and output of the reaction wheel in real time, wherein the input is the input torque command signal of the reaction wheel, and the output is the output torque signal of the reaction wheel; based on the input torque command signal and the output torque signal of the reaction wheel, the observer generates the estimated output of the reaction wheel in real time, calculates the difference between the actual output of the reaction wheel and the estimated output of the observer, and defines the difference as a residual; Step S2: Based on the input torque command signal of the reaction wheel and the estimated output generated by the observer, the threshold generation module of the detector generates an adaptive threshold corresponding to the working state of the reaction wheel and related to time; the fault discrimination module of the detector compares the residual and the adaptive threshold in real time: if the residual is greater than the adaptive threshold, it is determined that the reaction wheel is faulty; if the residual is less than or equal to the adaptive threshold, it is determined that the reaction wheel is not faulty; Step S3: while fault detection is in progress, the residual is input into an evaluator, which outputs the health of the reaction wheel, and the health state of the reaction wheel is analyzed according to the health to determine whether the performance of the reaction wheel is degraded; The step S1 comprises: The input torque command signal X and output torque signal Y of the reaction wheel are used to simulate the dynamic characteristics of the reaction wheel using an observer based on the LSTM deep neural network. (-2) , generates the estimated output of the reaction wheel in real time Calculate the difference between the actual output of the reaction wheel and the estimated output of the observer, which is defined as the residual R; the calculation formula of the residual r(t) at time t is The step S2 comprises: Using the input signal of the reaction wheel at time t and the output signal of the observer, the threshold generation module based on the LSTM deep neural network in the detector is used to generate the adaptive threshold of the reaction wheel at that time; The adaptive threshold at time t is calculated as follows: ε(t)=r0(t)+β Wherein, r0(t) represents the residual at time t when the reaction wheel works normally, and β represents the correction coefficient; the correction coefficient is used to compensate for the residual fluctuation caused by factors such as the parameters and disturbances of the reaction wheel drifting over time; the fault discrimination module of the detector compares the threshold ε(t) with the residual r(t): if the residual is greater than the threshold, the reaction wheel is considered to be faulty, otherwise it is considered to be in a normal state; The step S3 comprises: Step S31: Based on the residual, the evaluator records the residual within a period of time T to form a residual vector R(T)=[r(1), r(2), ..., r(T)], and the feature extraction module calculates the time domain features of the residual vector and uses it as the time domain features of the residual vector to be evaluated; Step S32: input the time domain features of the residual vector to be health-evaluated into the SOM neural network module of the evaluator, and the SOM neural network module in the evaluator outputs the spatial topological structure of the time domain features of the residual vector at this time, records the BMU position in the spatial topological structure, and the weight vector corresponding to the BMU is recorded as U1; Calculate the spatial distance between U0 and U1, and record the spatial distance as the minimum quantization error (MQE); MQE represents the deviation relationship between the residual vector to be evaluated and the residual vector when the reaction wheel works normally, that is, the offset of the feature space corresponding to the current operating state and the normal state of the reaction wheel; M=||U1-U0|| Wherein, M represents the MQE value; The health calculation module of the evaluator normalizes MQE to obtain the CV value: Where b is the normalization coefficient.

2. The method according to claim 1, characterized in that The characteristic values ​​corresponding to the time domain characteristics include root mean square value, peak value, and average absolute value, which are calculated as follows: RMS value: Peak: B = max(r(t)) Mean absolute value: Where T represents the length of the residual sequence recorded by the evaluator; r(t) represents the residual at time t; A represents the root mean square value of the residual vector; B represents the peak value of the residual vector; and C represents the average absolute value of the residual vector.

Images