A method for acquiring fault characteristics of aerospace control systems based on automatic encoder
By processing and training the data of the spacecraft control system based on an automatic encoder, the problem of difficulty in extracting the fault characteristics of the spacecraft control system is solved, and efficient and accurate fault diagnosis is achieved.
Patent Information
- Application Number
- CN202210501494.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-09
- Publication Date
- 2025-05-13
- Estimated Expiration
- 2042-05-09
AI Technical Summary
After the spacecraft control system fails, the output of the faulty components is not obvious due to closed-loop control, and the fault characteristics are often implicitly hidden between the outputs of multiple components and are difficult to extract.
The automatic encoder-based method is used to preprocess and network training the data samples of the spacecraft control system. The unsupervised learning and dimensionality reduction capabilities of the automatic encoder are used to extract the implicit fault characteristics between the outputs of multiple components.
It realizes efficient and accurate diagnosis of spacecraft control system failures, can fully explore the correlation between high-dimensional data, and extract implicit fault characteristics.
Smart Images
Figure CN115017607B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a method for acquiring fault characteristics of an aerospace control system based on an automatic encoder, and belongs to the field of aerospace. Background Art
[0002] The spacecraft control system is responsible for attitude and orbit control tasks, and is a typical closed-loop system, fault-tolerant system, and information system. On the one hand, the spacecraft control system has a complex functional structure and a high proportion of on-orbit failures; on the other hand, the spacecraft control system has a large number of single machines and rich information, and the information is strongly coupled and redundant, which facilitates fault diagnosis and on-orbit reconstruction. Therefore, fault diagnosis technology has become an effective means to overcome the inherent reliability deficiencies of products at the system level and to improve the safe, reliable and stable operation capabilities of spacecraft control systems. It has been widely used in on-orbit management at home and abroad.
[0003] Due to the closed-loop characteristics and information redundancy of the spacecraft control system, the outputs of its multiple components are coupled. This results in the output of the faulty component not being obvious due to the closed-loop control when a spacecraft control system fails, and the fault characteristics are often hidden between the outputs of multiple components and difficult to extract. Summary of the invention
[0004] The technical problem solved by the present invention is: to overcome the shortcomings of the prior art, utilize the unsupervised learning and dimensionality reduction capabilities of the autoencoder, and propose a method for acquiring fault features of a spacecraft control system based on the autoencoder, so as to fully explore the correlation between high-dimensional data and extract the implicit fault features between the outputs of multiple components, thereby realizing efficient and accurate diagnosis of spacecraft control system faults.
[0005] The technical solution of the present invention is:
[0006] A method for acquiring fault characteristics of aerospace control systems based on an automatic encoder, comprising:
[0007] Preprocessing of data samples for spacecraft control systems;
[0008] Build an autoencoder and use the preprocessed data samples to train the network;
[0009] The number of hidden layer nodes of the trained autoencoder is optimized to obtain the fault characteristics of the aerospace control system based on the autoencoder.
[0010] Furthermore, the preprocessing of the data samples of the spacecraft control system specifically includes:
[0011] For spacecraft control systems, obtain the outputs of sensors, controllers, and actuators at the current moment;
[0012] Based on the output of the above components, construct data sample X;
[0013] Normalize the data sample X;
[0014] Based on the normalization process, all elements in X are processed into values in the range of (0,1) to obtain the normalized data sample X * .
[0015] Furthermore, the sensors include star sensors, infrared earth sensors, and inertial attitude sensors; and the actuators include thrusters and control torque gyroscopes.
[0016] Furthermore, the output of the i-th star sensor is the measured three-axis attitude of the spacecraft The output of the i-th infrared earth sensor is the measured spacecraft roll and pitch axis attitude The output of the i-th inertial attitude sensor is the measured angular velocity ω relative to the inertial space IMUi ;
[0017] The output of the controller includes the control torque T of the three axes C =(T Cx ,T Cy ,T Cz );
[0018] The thruster data includes the jet time t of the three axes J =(t J ,t Jy ,t Jz );
[0019] The output Θ of the i-th control moment gyro CMGi The frame angle position θ obtained by measurement CMGi and frame angular velocity Composition, that is
[0020] Furthermore, the data samples are constructed as follows:
[0021] X={Φ SS1 ,…,Φ SSN ,Φ IRES1 ,…,Φ IRESN ,ω IMU1 ,…,ω IMUN ,T C ,t J ,Θ CMG1 ,…,Θ CMGN}
[0022] The subscripts SSN, IRESN, IMUN, and CMGN are the number of star sensors, infrared earth sensors, inertial attitude sensors, and control moment gyros, respectively.
[0023] Normalize the data sample X to obtain the corresponding normalized data as follows:
[0024]
[0025]
[0026] where ω IMUMax is the measured saturation value of the inertial attitude sensor;
[0027] Where T CMax is the limit value of the controller output;
[0028] where t JMax The maximum start-up time of a single control cycle of the thruster;
[0029] in is the maximum angular velocity limit of the control moment gyro frame;
[0030] Normalized data sample X * for:
[0031] X * ={Φ * SS1 ,…,Φ * SSN ,Φ * IRES1 ,…,Φ * IRESN ,ω * IMU1 ,…,ω * IMUN ,T * C ,t * J ,Θ * CMG1 ,…,Θ * CMGN}.
[0032] Furthermore, the autoencoder is a 3-layer neural network, wherein the output layer has the same size as the input layer, and the input layer x i is the normalized data sample X * The elements in the output layer For x iThe reconstructed value of the hidden layer h is the fault feature that needs to be extracted.
[0033] Furthermore, the transformation process from the input layer to the hidden layer is called encoding, and the transformation process from the hidden layer to the output layer is called decoding;
[0034] Assume f and g represent the encoding and decoding functions respectively, then the encoding and decoding processes are expressed as follows:
[0035] h=f(X * )=S f (W E X * +p)
[0036] Y * =g(h)=S g (W D h+q)
[0037] In the formula, S f and S g Take it as the sigmoid function, the function form is shown as follows:
[0038]
[0039] W E Represents the weight matrix between the input layer and the hidden layer; W D represents the weight matrix between the hidden layer and the output layer, p represents the bias vector of the hidden layer; q represents the bias vector of the output layer, Y * is the output value of the reconstructed network.
[0040] Furthermore, the cost function of the autoencoder network is defined as:
[0041]
[0042] Where N represents the number of training samples, ||·||2 is the bi-norm; based on the cost function J AE , the parameters W of the autoencoder network E , W D , p, q for training and updating:
[0043] pass Update the value of W D ;
[0044] pass Update the value of W E ;
[0045] pass Update p with the value of
[0046] pass Update the value of q;
[0047]
[0048]
[0049]
[0050]
[0051] in is the Hadamard product, δ is the learning rate; E is a vector whose elements are all 1;
[0052] The initial parameter W is as follows E , W D , p, q:
[0053] W E =0.1*randn(n,m)
[0054] W D =0.1*randn(m,n)
[0055] p=0.1*randn(1,n)
[0056] q=0.1*randn(1,m)
[0057] In the formula, randn() represents a function that can generate standard normal distribution random numbers, and randn(n,m) represents the generation of a standard normal distribution matrix with n rows and m columns.
[0058] Furthermore, the optimization of the number of hidden layer nodes of the trained autoencoder includes:
[0059] The number of nodes in the hidden layer of the autoencoder ranges from 1 to n, and is selected in increments of INT(n / 10), where INT(*) is the value function;
[0060] For each selection result of the number of nodes, 0.8 times the sample size is used for training, and the remaining 0.2 times the sample size is used for verification to calculate the distortion of the observer;
[0061] After calculating the distortion results under different numbers of nodes, set the lowest distortion as p1, and select the network with the least number of hidden nodes in the distortion range of p1 to 1.1*p1 as the final fault residual generation network.
[0062] Furthermore, the distortion S is defined by the following formula:
[0063]
[0064] The beneficial effects of the present invention compared with the prior art are:
[0065] (1) Aiming at the difficulty of extracting fault features of spacecraft control systems in the prior art, the present invention makes full use of the unsupervised learning and dimensionality reduction capabilities of autoencoders and proposes a method for acquiring fault features of spacecraft control systems based on autoencoders. The method can fully mine the correlation between high-dimensional data and extract the implicit fault features between the outputs of multiple components, thereby realizing efficient and accurate diagnosis of spacecraft control system faults.
[0066] (2) The present invention proposes a spacecraft control system fault feature acquisition framework based on an autoencoder, which fully utilizes the ability of the autoencoder to fully explore the correlation between high-dimensional data and accurately extract the implicit fault features between the outputs of multiple components.
[0067] (3) The present invention proposes a hidden layer node selection method based on an autoencoder, defines the distortion degree of the observer, and provides a method for selecting the number of hidden layer nodes of the autoencoder based on the distortion degree, which can effectively improve the network training efficiency and provide the practicality of deep learning in spacecraft fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS
[0068] Figure 1 This is a flow chart of a method for acquiring fault characteristics of a spacecraft control system based on an automatic encoder according to the present invention;
[0069] Figure 2 Obtaining autoencoder network structure diagram for spacecraft control system fault signature;
[0070] Figure 3 Flowchart for training autoencoder networks to obtain fault signatures for spacecraft control systems;
[0071] Figure 4 Experimental flow chart for autoencoder-based fault signature acquisition;
[0072] Figure 5 Schematic diagram of the classification results after fault features are extracted based on the automatic encoder. DETAILED DESCRIPTION
[0073] The specific embodiments of the present invention are further described in detail below with reference to the accompanying drawings.
[0074] In view of the difficulty in extracting fault features of spacecraft control systems in the prior art, the present invention makes full use of the unsupervised learning and dimensionality reduction capabilities of autoencoders, and proposes a method for acquiring fault features of spacecraft control systems based on autoencoders. The method can fully mine the correlation between high-dimensional data and extract the implicit fault features between the outputs of multiple components, thereby realizing efficient and accurate diagnosis of spacecraft control system faults.
[0075] The steps of the present invention are:
[0076] (1) Preprocess the data samples of the spacecraft control system.
[0077] (2) Build an autoencoder and use data samples to train the network.
[0078] (3) The number of hidden layer nodes of the trained autoencoder is optimized to obtain the fault characteristics of the aerospace control system based on the autoencoder.
[0079] like Figure 1 As shown, the specific implementation is as follows:
[0080] (1) Implementation method of step 1:
[0081] For the spacecraft control system, obtain the outputs of sensors, controllers, and actuators at the current moment. Sensors include star sensors, infrared earth sensors, and inertial attitude sensors, and actuators include thrusters and control moment gyroscopes. The output of the i-th star sensor is the measured three-axis attitude of the spacecraft The unit is radian (rad); the output of the i-th infrared earth sensor is the measured roll and pitch axis attitude of the spacecraft The unit is radian (rad); the output of the i-th inertial attitude sensor is the measured angular velocity ω relative to the inertial space IMUi , in radians per second (rad / s); the controller output includes the control torque T of the three axes C =(T Cx ,T Cy ,T Cz ), in Newton meters (Nm); the thruster data includes the jet time t of the three axes given by the controller J =(t J ,t Jy ,t Jz ), in seconds (s); the output of the i-th control moment gyro is the measured frame angular position θ CMGi and frame angular velocity Composition, that is The frame angle position θ CMGi The unit is radian (rad), the frame angular velocity The unit is radians per second (rad / s).
[0082] Based on the output of the above components, construct the following data samples:
[0083] X={Φ SS1 ,…,Φ SSN ,Φ IRES1 ,…,Φ IRESN ,ωIMU1 ,…,ω IMUN ,T C ,t J ,Θ CMG1 ,…,Θ CMGN}
[0084] The subscripts SSN, IRESN, IMUN, and CMGN represent the number of star sensors, infrared earth sensors, inertial attitude sensors, and control moment gyroscopes, respectively.
[0085] In order to train the samples, the samples are first normalized as follows:
[0086]
[0087]
[0088] where ω IMUMax is the measured saturation value of the inertial attitude sensor;
[0089] Where T CMax is the limit value of the controller output;
[0090] where t JMax The maximum start-up time of a single control cycle of the thruster;
[0091] in is the maximum angular velocity limit of the control moment gyro frame;
[0092] Based on the above normalization processing, all elements in X are processed into values in the range of (0,1), and the data samples after normalization are:
[0093] X * ={Φ * SS1 ,…,Φ * SSN ,Φ * IRES1 ,…,Φ * IRESN ,ω * IMU1 ,…,ω * IMUN ,T * C ,t * J ,Θ * CMG1 ,…,Θ * CMGN}
[0094] (2) Implementation method of step 2:
[0095] The automatic encoder used in this patent is a 3-layer neural network, and its network structure is as follows Figure 2 As shown, the output layer has the same size as the input layer, and the input layer x i For X * The elements in the output layer For x i The reconstructed value of the hidden layer h is the fault feature that needs to be extracted.
[0096] The transformation process from the input layer to the hidden layer is called encoding, and the transformation process from the hidden layer to the output layer is called decoding. Let f and g represent the encoding and decoding functions respectively, then the two processes can be expressed as follows:
[0097] h=f(X * )=S f (W E X * +p)
[0098] Y * =g(h)=S g (W D h+q)
[0099] In the formula, S f and S g Take it as the sigmoid function, the function form is shown as follows:
[0100]
[0101] W E Represents the weight matrix between the input layer and the hidden layer; W D represents the weight matrix between the hidden layer and the output layer, p represents the bias vector of the hidden layer; q represents the bias vector of the output layer, Y * is the output value of the reconstructed network.
[0102] The cost function of the autoencoder network is defined as:
[0103]
[0104] Where N represents the number of training samples and ||·||2 is the bi-norm. Based on the cost function, we use Figure 2 The training method shown has an effect on the parameters W of the autoencoder network. E , W D , p, q for training and updating:
[0105]
[0106]
[0107]
[0108]
[0109]
[0110]
[0111]
[0112]
[0113] in is the Hadamard product, δ is the learning rate, and E is a vector whose elements are all 1.
[0114] The initial parameter W is as follows E , W D , p, q:
[0115] W E =0.1*randn(n,m)
[0116] W D =0.1*randn(m,n)
[0117] p=0.1*randn(1,n)
[0118] q=0.1*randn(1,m)
[0119] In the formula, randn() represents a function that can generate standard normal distribution random numbers, and randn(n,m) represents the generation of a standard normal distribution matrix with n rows and m columns.
[0120] (3) Implementation method of step three:
[0121] In order to improve the accuracy of fault feature extraction of spacecraft control system, it is necessary to optimize the number of nodes in the hidden layer of the autoencoder. The optimization rules are as follows:
[0122] (1) The number of nodes in the hidden layer of the autoencoder is selected from 1 to n in increments of INT(n / 10), where INT(*) is the value function;
[0123] (2) For each selection result of the number of nodes, 0.8 times the sample size is used for training, and the remaining 0.2 times the sample size is used for verification to calculate the distortion of the observer. The distortion is defined by the following formula:
[0124]
[0125] (3) After calculating the distortion results for different numbers of nodes, in order to reduce network complexity and save training time, the lowest distortion is set to p, and the network with the least number of hidden nodes in the distortion range of p to 1.1p is selected as the final fault residual generation network.
[0126] In order to illustrate the effectiveness of the method of the present invention, the method of the present invention was experimentally verified using mathematical simulation data of the spacecraft control system. The experimental data includes five types of data samples, including normal operation and four failure modes, with a total of 5,000 samples collected, 1,000 samples for each of the normal operation and four failure modes, and each sample contains 2,340 data.
[0127] The method of the present invention reconstructs data by encoding and decoding data, and extracts the fault features contained in the data. Since the effect of the present invention cannot be directly observed only from the fault features, the Softmax classifier is further used to perform multi-fault classification on the fault features obtained by the present invention to achieve fault identification, and the accuracy of fault identification is used to verify the effectiveness of the present invention. The verification method is as follows: Figure 4 As shown. The output state of the Softmax classifier is divided into 5 groups, corresponding to the normal situation and four fault modes. The data dimension is reduced to 40 dimensions by using an automatic encoder, and then the softmax classifier is used for fault classification. Classification results Figure 5 As shown, the accuracy can reach 99.5%.
[0128] Parts of the present invention that are not described in detail belong to common knowledge among those skilled in the art.
Claims
1. A method for acquiring fault characteristics of aerospace control systems based on an automatic encoder, characterized in that include: Preprocessing of data samples for spacecraft control systems; Build an autoencoder and use the preprocessed data samples to train the network; The autoencoder is a three-layer neural network, where the output layer has the same size as the input layer, and the input layer x i is the normalized data sample X * The elements in the output layer y i =x i For x i The reconstruction value of the hidden layer h is the fault feature to be extracted; the transformation process from the input layer to the hidden layer is called encoding, and the transformation process from the hidden layer to the output layer is called decoding; Assume f and g represent the encoding and decoding functions respectively, then the encoding and decoding processes are expressed as follows: h=f(X * )=S f (W E X * +p) Y * =g(h)=S g (W D h+q) In the formula, S f and S g Take it as the sigmoid function, the function form is shown as follows: W E Represents the weight matrix between the input layer and the hidden layer; W D represents the weight matrix between the hidden layer and the output layer, p represents the bias vector of the hidden layer; q represents the bias vector of the output layer, Y * is the output value of the reconstructed network; The number of hidden layer nodes of the trained autoencoder is optimized to obtain the fault characteristics of the aerospace control system based on the autoencoder, specifically: The number of nodes in the hidden layer of the autoencoder ranges from 1 to n, and is selected in increments of INT(n / 10), where INT(*) is the value function; For each selection result of the number of nodes, 0.8 times the sample size is used for training, and the remaining 0.2 times the sample size is used for verification to calculate the distortion of the observer; After calculating the distortion results under different numbers of nodes, set the lowest distortion as p1, and select the network with the least number of hidden nodes in the distortion range of p1 to 1.1*p1 as the final fault residual generation network.
2. The method for acquiring fault characteristics of aerospace control systems based on an automatic encoder according to claim 1, characterized in that: The preprocessing of the data samples of the spacecraft control system specifically includes: For spacecraft control systems, obtain the outputs of sensors, controllers, and actuators at the current moment; Based on the output of the above components, construct data sample X; Normalize the data sample X; Based on the normalization process, all elements in X are processed into values in the range of (0,1) to obtain the normalized data sample X * .
3. The method for acquiring fault characteristics of aerospace control systems based on an automatic encoder according to claim 2, characterized in that: The sensors include star sensors, infrared earth sensors, and inertial attitude sensors; the actuators include thrusters and control torque gyroscopes.
4. The method for acquiring fault characteristics of aerospace control systems based on an automatic encoder according to claim 3 is characterized in that: The output of the i-th star sensor is the measured three-axis attitude of the spacecraft The output of the i-th infrared earth sensor is the measured spacecraft roll and pitch axis attitude The output of the i-th inertial attitude sensor is the measured angular velocity ω relative to the inertial space IMUi ; The output of the controller includes the control torque T of the three axes C =(T Cx ,T Cy ,T Cz ); The thruster data includes the jet time t of the three axes J =(t Jx ,t Jy ,t Jz ); The output θ of the i-th control moment gyro CMGi The frame angle position θ obtained by measurement CMGi and frame angular velocity Composition, that is 5. The method for acquiring fault characteristics of aerospace control systems based on an automatic encoder according to claim 4 is characterized in that: The specific construction of data samples is as follows: X={Φ SS1 ,…,F SSN ,F IRES1 ,…,F IRESN ,oh IMU1 ,…,oh IMUN ,T C ,t J ,I CMG1 ,…,Θ CMGN } The subscripts SSN, IRESN, IMUN, and CMGN are the number of star sensors, infrared earth sensors, inertial attitude sensors, and control moment gyros, respectively. Normalize the data sample X to obtain the corresponding normalized data as follows: where ω IMUMax is the measured saturation value of the inertial attitude sensor; Where T CMax is the limit value of the controller output; where t JMax The maximum start-up time of a single control cycle of the thruster; in is the maximum angular velocity limit of the control moment gyro frame; Normalized data sample X * for: X * ={Φ * SS1 ,…,F * SSN ,F * IRES1 ,…,F * IRESN ,oh * IMU1 ,…,oh * IMUN ,T * C ,t * J ,I * CMG1 ,…,Θ * CMGN }。 6. The method for acquiring fault characteristics of aerospace control systems based on an automatic encoder according to claim 1, characterized in that: The cost function of the autoencoder network is defined as: Where N represents the number of training samples, ||·||2 is the bi-norm; based on the cost function J AE , the parameters W of the autoencoder network E , W D , p, q for training and updating: pass Update the value of W D ; pass Update the value of W E ; pass Update p with the value of pass Update the value of q; in is the Hadamard product, δ is the learning rate; E is a vector whose elements are all 1; The initial parameter W is as follows E , W D , p, q: W E =0.1*randn(n,m) W D =0.1*randn(m,n) p=0.1*randn(1,n) q=0.1*randn(1,m) In the formula, randn() represents a function that can generate standard normal distribution random numbers, and randn(n,m) represents the generation of a standard normal distribution matrix with n rows and m columns.
7. The method for acquiring fault characteristics of aerospace control systems based on an automatic encoder according to claim 1, characterized in that: The distortion S is defined by the following formula: