Satellite fault prediction model training and fault prediction method based on confidence estimation

Abstract

The application relates to a satellite fault prediction model training and fault prediction method based on confidence estimation, and belongs to the field of satellite fault prediction, which comprises the following steps: acquiring original satellite telemetry data and performing data preprocessing; inputting the original satellite telemetry data into a satellite fault prediction model trained by an unknown satellite fault prediction model training method based on confidence estimation, calculating an output feature vector and a bias vector; calculating the harmonic mean of each dimension of the bias vector to obtain a confidence estimation value of the input data; judging whether the confidence estimation value is greater than a confidence threshold value; if the confidence estimation value is less than the confidence threshold value, the satellite state is determined as unknown fault; if the confidence estimation value is greater than or equal to the confidence threshold value, the satellite state is determined as no fault or known fault. The application can simultaneously predict the known faults and unknown faults of a satellite platform and a load according to satellite telemetry data, reduces the influence of data noise on a satellite fault prediction model, and improves the reliability of the fault prediction model.

Description

TECHNICAL FIELD

[0001] The present application relates to a satellite fault prediction method based on confidence estimation in the field of satellite fault prediction, in particular to the prediction of unknown satellite faults. BACKGROUND

[0002] With the rapid development of satellite technology, the functions and structures of satellite payloads and platforms become more complex, the equipment becomes more precise, and the telemetry information becomes more comprehensive, which puts higher requirements on the safety and reliability of satellite payloads and platforms. Some traditional fault prediction methods have been unable to meet the growing satellite support needs.

[0003] Currently, in the field of satellite fault prediction, existing satellite fault prediction methods are all for known faults. The satellite fault prediction method based on expert system is limited by the coverage of the knowledge base and cannot accurately predict unknown faults. The satellite fault prediction method based on data-driven is limited by the model training data and cannot give a reliable prediction for unknown faults. The prediction output of the existing method is mostly point estimation. Although they have strong discriminant ability within the boundary, for any data input, the model will always give an output within a range of values. When the satellite has an unknown fault, the existing method cannot tell whether the model is determined or not for the input data, and cannot predict unknown faults. SUMMARY

[0004] To solve the above problems, the present application proposes a satellite fault prediction model training and fault prediction method based on confidence estimation. In the high sensitivity application field of satellite fault prediction, the present application improves the reliability of the satellite fault prediction method and predicts unknown satellite faults.

[0005] The present application adopts the following technical solutions:

[0006] A satellite fault prediction model training method based on confidence estimation, comprising the following steps:

[0007] Step S1: Obtain the original satellite telemetry data and initialize the confidence neural network parameters;

[0008] Step S2: Perform data preprocessing on the original satellite telemetry data to obtain noise-free data, and add Gaussian noise and Laplace noise to obtain noisy data;

[0009] Step S3: Input the noisy data and noise-free data into the confidence neural network to obtain the corresponding feature vector and bias vector;

[0010] Step S4: Calculate the harmonic mean of each dimension of the bias vector corresponding to the noisy data to obtain the confidence estimate S of the noisy data.

[0011] Step S5: Calculate the first loss function for the training process based on the feature vectors corresponding to the noisy data;

[0012] Step S6: Calculate the second loss function for the training process based on the confidence estimate and the feature vectors and bias vectors corresponding to the noisy and noiseless data;

[0013] Step S7: Summing the first loss function and the second loss function yields the total loss function of the model training process, and the parameters of the confidence neural network are optimized using the backpropagation algorithm;

[0014] Step S8: Repeat steps S2 to S7 until the total loss function converges, and obtain the trained unknown satellite fault prediction model.

[0015] Specifically, step S3 is as follows:

[0016] Noisy and noise-free data are input into a confidence neural network, and the network outputs a feature vector of the noisy data. Sum of deviation vectors Obtain the feature vector of noise-free data Sum of deviation vectors in It is an eigenvector The i-th dimension, It is the deviation vector The i-th dimension, f i σ is the i-th dimension of the feature vector f. i It is the i-th dimension of the deviation vector σ, i = 1, ..., D, where D is the length of the eigenvector and the deviation vector.

[0017] The first loss function L1 mentioned in step S5 is:

[0018]

[0019] In the formula, The loss function is Softmax. The loss function is centered, and both α1 and α2 are non-negative multipliers;

[0020] Softmax loss function for:

[0021]

[0022] Central loss function for:

[0023]

[0024] wherein, is the corresponding class center in the batch of satellite telemetry data samples, the class center has the same dimension as the feature vector , 2' is the two-norm of the matrix, and N is the number of the batch of samples.

[0025] wherein, the second loss function L2 in step S6 is:

[0026] L2=S*L1+βL LMS

[0027] wherein, L1 is the first loss function, L LMS is the MLS distance loss function, and β is a non-negative multiplier;

[0028] The MLS distance loss function is:

[0029]

[0030] An unknown satellite fault prediction method based on confidence estimation, comprising the following steps:

[0031] Step S1: acquiring satellite telemetry data to be subjected to fault prediction, the satellite telemetry data containing information acquisition frames of a satellite platform and a load;

[0032] Step S2: performing data preprocessing on the acquired satellite telemetry data to obtain input data;

[0033] Step S3: performing forward propagation on the input data by an unknown satellite fault prediction model trained by an unknown satellite fault prediction model training method based on confidence estimation, the unknown satellite fault prediction model performing the following steps on the input data:

[0034] Step S31: performing forward propagation on the input data by the unknown satellite fault prediction model to obtain a feature vector and a bias vector wherein f i is the i-th dimension of the feature vector f, σ i is the i-th dimension of the bias vector σ, i=1,…,D, and D is the length of the feature vector and the bias vector;

[0035] Step S32: calculating the harmonic mean of each dimension of the bias vector σ to obtain a confidence estimation value

[0036] Step S4: judging whether the confidence estimation value S of the input satellite telemetry data is greater than a confidence threshold Th s ; if the confidence estimation value S is less than the confidence threshold Th sIf the confidence estimation value S is greater than or equal to the confidence threshold Th, it is determined that the current state of the satellite is unknown, there is an unknown fault, and step S5 is not executed. s If the confidence estimation value S is less than the confidence threshold Th, step S5 is continued.

[0037] Step S5: The feature vector output by the classification branch is input into a Softmax function to obtain a classification probability vector of each category, and the position of the maximum value in the classification probability vector is taken as the classification result. According to the classification result, the state of the satellite is determined to be a category of no fault or known fault.

[0038] Compared with the prior art, the present application has the following beneficial effects:

[0039] The unknown satellite fault prediction model training method based on confidence estimation and the unknown satellite fault prediction method based on confidence estimation are respectively used in the training process and the test process of the satellite fault prediction model. The satellite fault prediction model has more accurate fault prediction capability, reduces the influence of data noise on the satellite fault prediction model, and improves the reliability and robustness of the fault prediction model. In addition, by setting a threshold to determine the size of the confidence estimation value, it is determined whether the feature distribution of the input sample is shifted, so that the satellite platform and the load known fault can be predicted, and the unknown fault can also be predicted. In addition, the optimization of the first and second loss functions avoids the phenomenon of model training collapse. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 FIG. 1 is a flowchart of the unknown satellite fault prediction model training method based on confidence estimation according to an embodiment of the present application.

[0041] Figure 2 FIG. 3 is a principle block diagram of the unknown satellite fault prediction model training method based on confidence estimation according to an embodiment of the present application.

[0042] Figure 3 FIG. 4 is a flowchart of the unknown satellite fault prediction method based on confidence estimation according to an embodiment of the present application. DETAILED DESCRIPTION

[0043] The technical solutions of the present application will be described in detail below with reference to the accompanying drawings and preferred embodiments.

[0044] In one embodiment, as shown in FIGS. 1 and 2, the present application provides an unknown satellite fault prediction model training method based on confidence estimation, which comprises the following steps 1 to 8. Figure 1 Figure 2 The specific training process of the unknown satellite fault prediction model training method based on confidence estimation is as follows:

[0045] ​Step S1: Obtain original satellite telemetry data and initialize confidence neural network parameters;

[0046] Obtain original satellite telemetry data and perform one-hot encoding, and initialize confidence neural network model parameters.

[0047] Step S2: Perform data preprocessing on the original satellite telemetry data to obtain noise-free data, and add Gaussian noise and Laplace noise to obtain noisy data;

[0048] All original satellite telemetry data is normalized and subjected to principal component analysis data preprocessing operation, and Gaussian noise and Laplace noise are added to the training data.

[0049] Step S3: Input the noisy data and noise-free data into the confidence neural network to obtain the corresponding feature vector and bias vector;

[0050] Specifically, the noisy data and noise-free data are input into the confidence neural network, and the confidence neural network outputs the feature vector and bias vector of the noisy data wherein is the i-th dimension of the feature vector is the i-th dimension of the bias vector , i=1,…,D, D is the length of the feature vector and the bias vector; the feature vector and bias vector of the noise-free data are obtained wherein f i is the i-th dimension of the feature vector f, σ i is the i-th dimension of the bias vector σ, i=1,…,D, D is the length of the feature vector and the bias vector.

[0051] Step S4: Calculate the harmonic mean of each dimension of the bias vector corresponding to the noisy data to obtain the confidence estimate of the data added with noise;

[0052] Calculate the harmonic mean of each dimension of the bias vector of the noisy data to obtain the confidence estimate value

[0053] Step S5: Calculate the first loss function of the training process according to the feature vector of the noisy data;

[0054] The first loss function is:

[0055]

[0056] wherein, is a Softmax loss function, is a center loss function, and both a1 and a2 are non-negative multipliers. The Softmax loss function is:

[0057]

[0058] The center loss function is:

[0059]

[0060] wherein, is a corresponding class center in a batch of satellite telemetry data samples, the class center has the same dimension as the feature vector , and 2' is a two-norm of a matrix, and N is a number of the batch of samples

[0061] Step S6: According to the confidence estimation value and the feature vectors and bias vectors corresponding to the noisy data and the noise-free data, a second loss function of the training process is calculated.

[0062] The second loss function L2 is:

[0063] L2 = S*L1 + βL LMS

[0064] wherein, L1 is the first loss function, and L LMS is an MLS distance loss function, and β is a non-negative multiplier. The MLS distance loss function L LMS is:

[0065]

[0066] Step S7: The first loss function and the second loss function are summed to obtain a total loss function of the model training process, and the model parameters are optimized through a back propagation algorithm.

[0067] The total loss function L of the model training process is calculated and the model parameters are optimized through a back propagation (BP) algorithm, wherein the total loss function L is calculated by the following formula:

[0068] L = L1 + L2

[0069] The model parameters are optimized through the back propagation algorithm.

[0070] Step S8: The above steps S2 to S7 are repeated until the total loss function converges, the model parameters of the training are obtained, and then an unknown satellite fault prediction model of the training is obtained.

[0071] The unknown satellite fault prediction model training method based on confidence estimation provided in the embodiment uses confidence estimation in the training process of the satellite fault prediction model, and the confidence estimation value can accurately reflect the distribution of the satellite telemetry data, so that the trained satellite fault prediction model can predict the unknown fault of the satellite, and the reliability and robustness of the satellite fault prediction model are improved. Meanwhile, the loss function of the satellite fault prediction model is optimized to avoid the training collapse phenomenon in the training process, and the accuracy of the known satellite fault prediction is improved.

[0072] In another embodiment, as shown in Figure 3 The present application provides an unknown satellite fault prediction method based on confidence estimation, which uses the unknown satellite fault prediction model trained by the unknown satellite fault prediction model training method based on confidence estimation to predict the known fault and unknown fault of the satellite, specifically including the following steps:

[0073] Step S1: acquiring satellite telemetry data to be used for fault prediction, wherein the satellite telemetry data contains information acquisition frames of the satellite platform and the load;

[0074] Step S2: performing data preprocessing on the acquired satellite telemetry data to obtain input data;

[0075] Step S3: performing forward propagation on the input data by the unknown satellite fault prediction model trained by the unknown satellite fault prediction model training method based on confidence estimation, wherein the unknown satellite fault prediction model performs the following steps on the input data:

[0076] Step S31: performing forward propagation on the input data by the unknown satellite fault prediction model to obtain a feature vector and a bias vector wherein f i is the i-th dimension of the feature vector f, σ i is the i-th dimension of the bias vector σ, i=1,…,D, and D is the length of the feature vector and the bias vector;

[0077] Step S32: calculating the harmonic mean of each dimension of the bias vector σ to obtain a confidence estimation value

[0078] Step S4: judging whether the confidence estimation value S of the input satellite telemetry data is greater than a confidence threshold Th s ; if the confidence estimation value S is less than the confidence threshold Th s , it is judged that the current state of the satellite is unknown, and there is an unknown fault, and step S5 is not executed; if the confidence estimation value S is greater than or equal to the confidence threshold Th sIf yes, then continue to step S5;

[0079] Step S5: The feature vector output by the classification branch is input into a Softmax function to obtain a classification probability vector of each class, and the position of the maximum value in the classification probability vector is taken as the classification result. According to the classification result, the satellite state is determined as the class of no fault or known fault.

[0080] The unknown satellite fault prediction method based on confidence estimation provided in the embodiment uses a satellite fault prediction model to predict satellite faults. The satellite fault prediction model is a model trained by the unknown satellite fault prediction model training method based on confidence estimation. The satellite fault prediction model outputs a confidence estimation value for input satellite telemetry data. The confidence estimation value can accurately reflect the distribution of satellite telemetry data, so that the trained satellite fault prediction model can predict unknown satellite fault conditions, and the reliability and robustness of the satellite fault prediction model are improved.

[0081] The technical features of the above-described embodiments can be combined in any manner. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described, but as long as the combinations of the technical features do not contradict, they should be considered as falling within the scope of the present disclosure.

[0082] The above-described embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it should not be understood as limiting the scope of the patent. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of the patent protection of the present application should be subject to the appended claims.

Claims

1. A method for training an unknown satellite fault prediction model based on a confidence estimation, characterized in that, The method comprises the following steps: Step S1: obtaining original satellite telemetry data and initializing confidence neural network parameters; Step S2: performing data preprocessing on the original satellite telemetry data to obtain noise-free data, and adding Gaussian noise and Laplacian noise to obtain noisy data; Step S3: inputting the noisy data and the noise-free data into the confidence neural network respectively to obtain corresponding feature vectors and bias vectors; Step S4: calculating the harmonic mean of each dimension of the bias vector corresponding to the noisy data to obtain a confidence estimate S of the noisy data; Confidence estimate is the noisy data bias vector, D is the length of the feature vector and bias vector. Step S5: calculating a first loss function of the training process according to the feature vector corresponding to the noisy data; Step S6: calculating a second loss function of the training process according to the confidence estimate value and the feature vectors and bias vectors corresponding to the noisy data and the noise-free data; Step S7: summing the first loss function and the second loss function to obtain a total loss function of the model training process, and optimizing the confidence neural network parameters through a back propagation algorithm; Step S8: repeating steps S2 to S7 until the total loss function converges, to obtain a trained unknown satellite fault prediction model. 2.The confidence estimation based unknown satellite fault prediction model training method according to claim 1, characterized in that, Step S3 is specifically: the noisy data and the noise-free data into a confidence neural network, the confidence neural network outputs a feature vector of the noisy data and a bias vector a feature vector of the noise-free data and a bias vector wherein is an i-th dimension of the feature vector is an i-th dimension of the bias vector i is an i-th dimension of the feature vector f, σ i is an i-th dimension of the bias vector σ, i = 1, …, D, D is a length of the feature vector and the bias vector.​​ 3.The confidence estimation based unknown satellite fault prediction model training method of claim 2, wherein, The first loss function L1 in step S5 is: In the formula, is a Softmax loss function, is a center loss function, and both a1 and a2 are non-negative multipliers; Softmax loss function is: Center loss function is: wherein, is the corresponding class center in the batch of satellite telemetry data samples, the class center has the same dimension as the feature vector , and ′ is the two-norm of the matrix, and N is the number of the batch of samples.

4. The method of claim 2, wherein the method further comprises: The second loss function L2 in step S6 is: L2 = S*L1 + βL LMS In the formula, L1 is the first loss function, L LMS is the MLS distance loss function, and β is a non-negative multiplier. The MLS distance loss function is:

5. A method for unknown satellite fault prediction based on confidence estimation, characterized in that, The method comprises the following steps: Step S1: obtaining satellite telemetry data to be subjected to fault prediction, the satellite telemetry data containing information acquisition frames of a satellite platform and a load; Step S2: performing data preprocessing on the obtained satellite telemetry data to obtain input data; Step S3: performing forward propagation on the input data through an unknown satellite fault prediction model trained by the confidence estimate-based unknown satellite fault prediction model training method according to any one of claims 1-4, the unknown satellite fault prediction model performing the following steps on the input data: Step S31: forward propagation of the input data through the unknown satellite fault prediction model to obtain a feature vector and a bias vector where f i is the i-th dimension of the feature vector f, σ i is the i-th dimension of the bias vector σ, i = 1, …, D, D is the length of the feature vector and the bias vector; Step S32: Calculate the harmonic mean of each dimension of the bias vector σ to obtain the confidence estimate Step S4: judging whether the confidence estimation value S of the input satellite telemetry data is greater than the confidence threshold Th s ; if the confidence estimation value S is less than the confidence threshold Th s , judging that the current state of the satellite is unknown, there is an unknown fault, and step S5 is not executed; if the confidence estimation value S is greater than or equal to the confidence threshold Th s , step S5 is continued to be executed; Step S5: using the feature vector output by the classification branch to obtain a classification probability vector of each category through a Softmax function, taking the position of the maximum value in the classification probability vector as a classification result, and judging the satellite state to be a category of no fault or known fault according to the classification result.

Images