Satellite attitude control system fault prediction method based on data driving

Abstract

The invention provides a satellite attitude control system fault prediction method based on data driving. The satellite attitude control system fault prediction method comprises the steps of collecting historical fault time sequence data of key parameters of a satellite attitude control system; dividing the data into to-be-predicted parameter historical data and other influence parameter historical data; compressing the historical data of the to-be-predicted parameter by using a mean value down-sampling method and performing preprocessing; historical data of other influence parameters are preprocessed, and feature data, closely related to the to-be-predicted parameters, of the historical data are extracted through an ECA-COV1D module; and inputting the mean value down-sampling processing and the parameter sequence extracted by the ECA-COV1D module into the LSTM network to predict the future state of the satellite attitude control system. According to the method, the amount of data input into the LSTM network can be effectively reduced, meanwhile, correlation and differences among other sensor parameters can be better reflected, the prediction precision is improved by adding the auxiliary loss function, and therefore the purpose of short-term fault prediction is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of satellite fault prediction, and in particular to a data-driven satellite attitude control system fault prediction method. Background Art

[0002] With the rapid development of the satellite and application industries, satellites have become the most widely used spacecraft. The attitude control system is the core subsystem of the satellite. Since its use cycle accompanies the entire life cycle of the satellite, its failure rate is significantly higher than that of other subsystems. Once a failure disaster occurs due to an abnormality, the loss is extremely serious.

[0003] In order to ensure the normal operation of satellites in orbit, it is necessary to improve the ability of satellite control systems to cope with faults. Fault prediction technology can use various reasoning strategies and intelligent algorithms to predict the future state of the system, the remaining effective life, and the failure time. Through "pre-prediction", it can make up for the shortcomings of "post-detection" and provide certain guarantees for the safe and stable operation of satellites in orbit.

[0004] At present, the main method for predicting the remaining life of a wheel is to use the momentum wheel lubricant as the main influencing factor for fault prediction and to establish a momentum wheel degradation model. However, since it requires in-depth analysis of the physical characteristics of the system and the prediction accuracy depends entirely on the accuracy of the established model, insufficient model accuracy will cause the prediction function to fail. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a data-driven satellite attitude control system fault prediction method, which can effectively reduce the amount of data input into the LSTM network, while better reflecting the correlation and difference between other sensor parameters, and improving the prediction accuracy by adding an auxiliary loss function, thereby achieving the purpose of short-term fault prediction.

[0006] The technical solution of the present invention is: a data-driven satellite attitude control system fault prediction method, comprising the following steps:

[0007] S1) Collect historical fault sequence data of key parameters of satellite attitude control system; and divide it into historical data of parameters to be predicted and historical data of other influencing parameters;

[0008] S2), using the mean downsampling method to compress the historical data of the parameters to be predicted and preprocess them to obtain a new sequence of parameters to be predicted;

[0009] S3), preprocessing the historical data of other influencing parameters and using the ECA-COV1D module to extract the characteristic data closely related to the parameters to be predicted as a new sequence of other sensor parameters;

[0010] S4), input the parameter sequence processed in step S2) and step S3) into the LSTM network to predict the future state of the satellite attitude control system.

[0011] Preferably, in step S1), the historical fault timing data includes telemetry data of six parameters, namely, roll angle, pitch angle, yaw angle, and angular acceleration of three attitude angles, and any one of the six parameters is selected as the historical data of the parameter to be predicted, and the remaining five parameters are selected as the historical data of other influencing parameters.

[0012] Preferably, in step S2), the historical data of the parameter to be predicted is compressed using a mean downsampling method, specifically:

[0013]

[0014] In the formula, μ represents the downsampled data; n is the number of historical data of the parameter to be predicted; x i is the historical data of the i-th parameter to be predicted.

[0015] Preferably, in steps S2) and S3), the preprocessing includes normalization and data conversion processing.

[0016] Preferably, in step S2), the normalization process adopts the Min-Max normalization method, and the Min-Max normalization method is expressed as:

[0017]

[0018] Among them, x n is the normalized time series data, x is the original time series data, max(x) and min(x) are the maximum and minimum values ​​of the original time series data respectively; the time series data after normalization is finally expressed as: T = [x1, x2, ..., x n ,].

[0019] Preferably, in step S2), the normalized time series data T is converted to obtain:

[0020]

[0021] In the formula, the number of rows of X, i.e., nw, is the number of samples of the training data; the number of columns of X, i.e., w, is the size of the prediction window; and Y represents the predicted value under the corresponding input sample.

[0022] Preferably, in step S3), the ECA-COV1D module includes a convolutional layer, a pooling layer, and an ECA module.

[0023] Preferably, in step S3), the ECA-COV1D module extracts the features of the local area through a convolutional layer. For the input of the convolutional layer of the l+1th layer of 1DCNN:

[0024]

[0025] Its convolution operation is expressed as:

[0026]

[0027] in, The jth output of the l+1th convolutional layer; N j is the jth convolution region of the l+1th layer; N j The corresponding element in N j The weight matrix in ; is the bias of the l+1th layer; f is the ReLU activation function, which is used to enhance the expressive power of the network; * represents the convolution operation.

[0028] Preferably, in step S3), the ECA-COV1D module performs a downsampling operation on the features after the convolution operation through a pooling layer, that is:

[0029]

[0030] Among them, y j For region R j The maximum value within, x i For region R j All values ​​within b j For bias.

[0031] Preferably, in step S3), the ECA module compresses the feature map obtained after the pooling operation into a feature vector of size 1×1×C by GAP without dimensionality reduction, and then introduces a parameter to adjust the kernel size k. The k calculation process is expressed as:

[0032]

[0033] Where γ and b are hyperparameters; Ψ(C) is a nonlinear mapping; log2(i) represents the logarithm of the number of input feature channels;

[0034] Preferably, in step S3), after determining the kernel size k, the feature vector is convolved in one dimension in the channel direction; finally, the weight value of each channel is learned by the Sigmoid activation function, and the weight value is multiplied by the corresponding element of the initial input feature map to obtain the output feature map. Before the output feature map with weights is used as the input of the LSTM neural network together with the downsampled parameter sequence to be predicted, a channel merging operation is also required. By comparing the element values ​​at the same element position of different channels and selecting the maximum value at the element position as the merged channel element, the main features in each channel can be highlighted and the generalization ability of the input data can be improved.

[0035] Preferably, in step S4), the LSTM network selectively adds new information and forgets previously accumulated information to achieve the function of long-term memory by introducing a new internal state and gating mechanism. Each memory unit of the LSTM includes three gates, namely, a forget gate f t , input gate i t and output gate o t , the calculation method of the three gates is expressed as:

[0036]

[0037] Among them, σ(·) is the sigmoid function, Its output interval is (0,1); x t is the input of the current unit; h t-1 is the hidden state of the previous unit; W * and b * (*∈{f,o,i}) is the weight matrix and bias corresponding to each of the three gates.

[0038] As a preferred embodiment, in step S4), the input gate i t By and Candidate Status The element product operation is performed to control how much information of the current unit needs to be saved, and the forget gate f t Then, by comparing the internal state c of the previous unit t-1 An element-wise multiplication is performed to control how much of the information of the previous unit needs to be forgotten; therefore, the calculation method for updating the internal state of the current unit is expressed as:

[0039]

[0040] Among them, the candidate state It is composed of the hidden state of the previous unit and the input information of the current unit; that is:

[0041]

[0042] Among them, Wc and b c are the corresponding weight matrix and bias, activation function Its output interval is (-1,1);

[0043] The output gate o t Control c t How much information needs to be output to the hidden state h t , the final output of the current unit of the LSTM network is:

[0044] h t =o t ⊙tanh(c t );

[0045] And, by adding a fully connected layer after the hidden state, the hidden state is mapped to the final output result, that is:

[0046]

[0047] Among them, W y is the weight matrix, h t is the hidden state of the last unit of the LSTM network, b y is the bias vector.

[0048] Preferably, in step S4), the LSTM network introduces an auxiliary loss function during the training process to correct the error by calculating the true value y at the next moment. t+1 As a positive sample, introduce the true value y at the previous moment t As a negative sample; the auxiliary loss function L a The expression is:

[0049]

[0050] In the formula, T represents, σ ​​represents the sigmoid function, and h t is the hidden state of the last unit of the LSTM network; y t+1 The true value at the next moment; y t The true value at the previous moment;

[0051] By adding an auxiliary loss function, the true value y at the next moment is used t+1 To supervise the current hidden state h t Learning can introduce additional supervision signals at each time step, helping the gradient to propagate better and improving the accuracy of the overall prediction;

[0052] Therefore, the global error loss function is expressed as:

[0053] L=L T+αL a ;

[0054] Among them, α is a hyperparameter, which indicates the importance of the feature information extracted at each time step to the final prediction result; L T is the MSE loss function; L a is the auxiliary loss function.

[0055] The beneficial effects of the present invention are:

[0056] 1. The present invention can effectively reduce the amount of data input into the LSTM network, while better reflecting the association and difference between other sensor parameters, and improve the prediction accuracy by adding an auxiliary loss function, thereby achieving the purpose of short-term fault prediction;

[0057] 2. The present invention introduces additional supervisory signals at each time step to help the gradient propagate better and improve the accuracy of the overall prediction;

[0058] 3. The present invention reduces network parameters through the ECA-COV1D module; and through channel merging, the main features in each channel can be highlighted, thereby improving the generalization ability of input data; without dimensionality reduction, the influence of feature maps in different channels on the prediction effect can be adaptively learned; thereby improving the accuracy of the prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a flow chart of the method of the present invention;

[0060] Figure 2 It is a structural block diagram of the ECA-COV1D module of the method of the present invention;

[0061] Figure 3 It is a structural block diagram of the ECA mechanism of the method of the present invention;

[0062] Figure 4 This is a structural diagram of the LSTM network memory unit of the method of the present invention. DETAILED DESCRIPTION

[0063] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings:

[0064] like Figure 1 As shown, this embodiment provides a data-driven satellite attitude control system fault prediction method, comprising the following steps:

[0065] S1) Collect historical fault sequence data of key parameters of satellite attitude control system; and divide it into historical data of parameters to be predicted and historical data of other influencing parameters;

[0066] In this embodiment, the satellite attitude control system is usually equipped with an inertial measurement unit for measuring various attitude parameters of the satellite, such as sensors such as accelerometers and gyroscopes, which are used to measure linear acceleration and angular velocity, respectively. The historical fault time series data collected in this embodiment includes telemetry data of six parameters, namely, roll angle, pitch angle, yaw angle, and angular acceleration of three attitude angles, and any one of the six parameters is selected as the historical data of the parameter to be predicted, and the remaining five parameters are used as the historical data of other influencing parameters.

[0067] S2), using the mean downsampling method to compress the historical data of the parameters to be predicted and preprocess them to obtain a new sequence of parameters to be predicted;

[0068] Telemetry data has high dimensions, changes slowly, and contains a large amount of repeated historical information. If the data at the original sampling frequency is used for prediction, the change trend of key parameters can only be predicted within a very limited step length, which cannot meet the needs of fault prediction. Therefore, in order to reduce redundancy, shorten training time, and improve prediction ability, this embodiment uses the mean downsampling method to compress the historical data of the parameters to be predicted, specifically:

[0069]

[0070] In the formula, μ represents the downsampled data; n is the number of historical data of the parameter to be predicted; x i is the historical data of the i-th parameter to be predicted.

[0071] At the same time, the parameters of telemetry data have different dimensions, and the ranges of different feature values ​​vary greatly, which often makes it difficult for the model to converge and reduces the training efficiency of the model. Through normalization, the data can be limited to a smaller range while ensuring that the features have similar scales. This embodiment uses the Min-Max normalization method to normalize the collected historical data of the parameters to be predicted; specifically:

[0072]

[0073] Among them, x n is the normalized time series data, x is the original time series data, max(x) and min(x) are the maximum and minimum values ​​of the original time series data respectively; the time series data after normalization is finally expressed as: T = [x1, x2, ..., x n ,].

[0074] In addition, after the normalization process, data conversion is required; data conversion is to change the form of the original data without affecting its basic characteristics. In this embodiment, the data conversion is specifically as follows:

[0075]

[0076] In the formula, the number of rows of X, i.e., nw, is the number of samples of the training data; the number of columns of X, i.e., w, is the size of the prediction window; and Y represents the predicted value under the corresponding input sample.

[0077] S3), preprocessing the historical data of other influencing parameters and using the ECA-COV1D module to extract the characteristic data closely related to the parameters to be predicted as a new sequence of other sensor parameters;

[0078] In this embodiment, since the collected historical data of other influencing parameters are two-dimensional data, this embodiment needs to be normalized and data converted, and the operation method is the same as in step S2). In this embodiment, the observation time window is data within m hours, the data sampling frequency is n seconds, the number of sensor parameters is 5, and the total length of the data is L, then the shape is K×L. The two-dimensional data is split vertically and changed into multiple one-dimensional data. At this time, the data shape is K×1×L.

[0079] like Figure 2 As shown, the ECA-COV1D module described in this embodiment includes a convolution layer, a pooling layer, and an ECA module.

[0080] The ECA-COV1D module extracts the features of the local area through the convolution layer. Different convolution kernels represent different feature extractors. For the input of the convolution layer of the l+1th layer of 1DCNN:

[0081]

[0082] Its convolution operation is expressed as:

[0083]

[0084] in, The jth output of the l+1th convolutional layer; N j is the jth convolution region of the l+1th layer; N j The corresponding element in N t The weight matrix in ; is the bias of the l+1th layer; f is the ReLU activation function, which is used to enhance the expressive power of the network; * represents the convolution operation.

[0085] The ECA-COV1D module performs downsampling operation on the features after convolution operation through the pooling layer, namely:

[0086]

[0087] Among them, y t For region Rt The maximum value within, x i For region R t All values ​​within b t For bias.

[0088] like Figure 3 As shown, the ECA module compresses the feature map obtained after the pooling operation into a feature vector of size 1×1×C through GAP without dimensionality reduction, and then introduces a parameter to adjust the kernel size k. The calculation process of k is expressed as:

[0089]

[0090] Where γ and b are hyperparameters, which are 2 and 1 respectively; Ψ(C) is a nonlinear mapping; log2(C) represents the logarithm of the number of input feature channels;

[0091] After determining the kernel size k, the feature vector is convolved in one dimension in the channel direction to achieve communication between each channel and its adjacent channels; finally, the weight value of each channel is learned through the Sigmoid activation function, and the weight value is multiplied by the corresponding element of the initial input feature map to obtain the output feature map. In the output feature map, the relevant features of some channels are further strengthened, while the irrelevant features of other channels are also weakened. In order to reduce network parameters, before the weighted output feature map is used together with the downsampled parameter sequence to be predicted as the input of the LSTM neural network, a channel merging operation is required. By comparing the element values ​​at the same element position of different channels and selecting the maximum value at the element position as the merged channel element, the main features in each channel can be highlighted and the generalization ability of the input data can be improved.

[0092] S4), inputting the parameter sequence processed in step S2) and step S3) into the LSTM network to predict the future state of the satellite attitude control system;

[0093] like Figure 4 As shown, the LSTM network selectively adds new information and forgets previously accumulated information to achieve the function of long-term memory by introducing a new internal state and gating mechanism. Each memory unit of the LSTM includes three gates, namely, the forget gate f t , input gate i t and output gate o t , the calculation method of the three gates is expressed as:

[0094]

[0095] Among them, σ(·) is the sigmoid function, Its output interval is (0,1); x tis the input of the current unit; h t-1 is the hidden state of the previous unit; W * and b * (*∈{f,o,i}) is the weight matrix and bias corresponding to each of the three gates.

[0096] As a preferred embodiment, in step S4), the input gate i t By and Candidate Status The element product operation is performed to control how much information of the current unit needs to be saved, and the forget gate f t Then, by comparing the internal state c of the previous unit t-1 An element-wise multiplication is performed to control how much of the information of the previous unit needs to be forgotten; therefore, the calculation method for updating the internal state of the current unit is expressed as:

[0097]

[0098] Among them, ⊙ represents the XOR operation; candidate state It is composed of the hidden state of the previous unit and the input information of the current unit; that is:

[0099]

[0100] Among them, W c and b c are the corresponding weight matrix and bias, activation function Its output interval is (-1,1);

[0101] The output gate o t Control c t How much information needs to be output to the hidden state h t , the final output of the current unit of the LSTM network is:

[0102] h t =o t ⊙tanh(c t );

[0103] And, by adding a fully connected layer after the hidden state, the hidden state is mapped to the final output result, that is:

[0104]

[0105] Among them, W y is the weight matrix, h t is the hidden state of the last unit of the LSTM network, b y is the bias vector.

[0106] Preferably, in step S4), the LSTM network introduces an auxiliary loss function during the training process to correct the error by calculating the true value y at the next moment. t+1 As a positive sample, introduce the true value y at the previous moment t As a negative sample; the auxiliary loss function L a The expression is:

[0107]

[0108] In the formula, T represents, σ ​​represents the sigmoid function, and h t is the hidden state of the last unit of the LSTM network; y t+1 The true value at the next moment; y t The true value at the previous moment;

[0109] By adding an auxiliary loss function, the true value y at the next moment is used t+1 To supervise the current hidden state h t Learning can introduce additional supervision signals at each time step, helping the gradient to propagate better and improving the accuracy of the overall prediction;

[0110] Therefore, the global error loss function is expressed as:

[0111] L=L T +αL a ;

[0112] Among them, α is a hyperparameter, which indicates the importance of the feature information extracted at each time step to the final prediction result; L T is the MSE loss function; L a is the auxiliary loss function.

[0113] S5), evaluating the prediction results through evaluation indicators;

[0114] In this example, MAE, MSE and R 2 The prediction results are evaluated by three evaluation indicators; among them, the smaller the values ​​of MAE and MSE, the higher the accuracy of the prediction method, and R 2 The value range of is (0,1), and the closer it is to 1, the better the fitting effect. The calculation methods of the three evaluation indicators are as follows:

[0115]

[0116] Where N is the number of samples; f(x k ) is the predicted output value of the kth sample; y k is the true value of the kth sample; is the mean of N samples.

[0117] The above embodiments and descriptions are only for illustrating the principles and best embodiments of the present invention. Without departing from the spirit and scope of the present invention, the present invention may be subject to various changes and improvements, all of which fall within the scope of the present invention to be protected.

Claims

1. A data-driven satellite attitude control system fault prediction method, characterized in that: The following steps are involved: S1) Collect historical fault sequence data of key parameters of satellite attitude control system; and divide it into historical data of parameters to be predicted and historical data of other influencing parameters; S2), using the mean downsampling method to compress the historical data of the parameters to be predicted and preprocess them to obtain a new sequence of parameters to be predicted; S3), preprocessing the historical data of other influencing parameters and using the ECA-COV1D module to extract the characteristic data closely related to the parameters to be predicted as a new sequence of other sensor parameters; S4), input the parameter sequence processed in step S2) and step S3) into the LSTM network to predict the future state of the satellite attitude control system.

2. The method for predicting satellite attitude control system faults based on data drive according to claim 1, characterized in that: In step S1), the historical fault timing data includes telemetry data of six parameters, namely, roll angle, pitch angle, yaw angle, and angular acceleration of three attitude angles, and any one of the six parameters is selected as the historical data of the parameter to be predicted, and the remaining five parameters are used as the historical data of other influencing parameters.

3. The method for predicting satellite attitude control system faults based on data drive according to claim 2, characterized in that: In step S2), the historical data of the parameter to be predicted is compressed using the mean downsampling method, specifically: In the formula, μ represents the downsampled data; n is the number of historical data of the parameter to be predicted; x i is the historical data of the i-th parameter to be predicted.

4. The method for predicting satellite attitude control system faults based on data drive according to claim 1, characterized in that: In steps S2) and S3), the preprocessing includes normalization and data conversion processing; The normalization process described herein adopts the Min-Max normalization method.

5. The method for predicting satellite attitude control system faults based on data drive according to claim 4, characterized in that: The normalized time series data T is transformed to obtain: In the formula, the number of rows of X, i.e., nw, is the number of samples of the training data; the number of columns of X, i.e., w, is the size of the prediction window; and Y represents the predicted value under the corresponding input sample.

6. The method for predicting satellite attitude control system faults based on data drive according to claim 1, characterized in that: In step S3), the ECA-COV1D module includes a convolutional layer, a pooling layer, and an ECA module; The ECA-COV1D module extracts features of local areas through convolutional layers; The ECA-COV1D module downsamples the features after the convolution operation through the maximum pooling layer.

7. The method for predicting satellite attitude control system faults based on data drive according to claim 6, characterized in that: In step S3), the ECA module compresses the feature map obtained after the pooling operation into a feature vector of size 1×1×C by GAP without dimensionality reduction, and then introduces a parameter to adjust the kernel size k. The calculation process of k is expressed as: Where γ and b are hyperparameters; Ψ(C) is a nonlinear mapping; log2(C) represents the logarithm of the number of input feature channels; After determining the kernel size k, perform one-dimensional convolution on the feature vector in the channel direction; finally, learn the weight value of each channel through the Sigmoid activation function, multiply the weight value with the corresponding element of the initial input feature map to obtain the output feature map; Before using the weighted output feature map and the downsampled parameter sequence to be predicted as the input of the LSTM neural network, a channel merging operation is required; By comparing the element values ​​at the same element position of different channels, the maximum value at the element position is selected as the merged channel element.

8. The method for predicting satellite attitude control system faults based on data drive according to claim 1, characterized in that: In step S4), the LSTM network selectively adds new information and forgets previously accumulated information to achieve the function of long-term memory by introducing a new internal state and gating mechanism. Each memory unit of the LSTM includes three gates, namely, the forget gate f t , input gate i t and output gate o t , the calculation method of the three gates is expressed as: Among them, σ(·) is the sigmoid function, Its output interval is (0, 1); x t is the input of the current unit; h t-1 is the hidden state of the previous unit; W * and b * (*∈{f, o, i}) is the weight matrix and bias corresponding to each of the three gates.

9. The method for predicting satellite attitude control system faults based on data drive according to claim 8, characterized in that: In step S4), the input gate i t By and Candidate Status The element product operation is performed to control how much information of the current unit needs to be saved, and the forget gate f t Then, by comparing the internal state c of the previous unit t-1 An element-wise multiplication is performed to control how much of the information of the previous unit needs to be forgotten; therefore, the calculation method for updating the internal state of the current unit is expressed as: Among them, ⊙ represents the XOR operation; candidate state It is composed of the hidden state of the previous unit and the input information of the current unit; that is: Among them, W c and b c are the corresponding weight matrices and biases respectively, tanh(x) is the activation function; its output interval is (-1, 1); The output gate o t Control c t How much information needs to be output to the hidden state h t , the final output of the current unit of the LSTM network is: h t =o t ⊙tanh(c t ); And, by adding a fully connected layer after the hidden state, the hidden state is mapped to the final output result, that is: Among them, W y is the weight matrix, h t is the hidden state of the last unit of the LSTM network, b y is the bias vector.

10. The method for predicting satellite attitude control system faults based on data drive according to claim 9, characterized in that: In step S4), the LSTM network introduces an auxiliary loss function during the training process to correct the error by t+1 As a positive sample, introduce the true value y at the previous moment t As a negative sample; the auxiliary loss function L a The expression is: In the formula, T represents, σ ​​represents the sigmoid function, and h t is the hidden state of the last unit of the LSTM network; y t+1 The true value at the next moment; y t The true value at the previous moment; By adding an auxiliary loss function, the true value y at the next moment is used t+1 To supervise the current hidden state h t Learning can introduce additional supervision signals at each time step, helping the gradient to propagate better and improving the accuracy of the overall prediction; Therefore, the global error loss function is expressed as: L=L T +αL a ; Among them, α is a hyperparameter, which indicates the importance of the feature information extracted at each time step to the final prediction result; L T is the MSE loss function; L a is the auxiliary loss function.

Images