Spacecraft maneuver detection method based on causal deep learning

Abstract

The application relates to a spacecraft maneuver detection method based on causal deep learning. The method comprises the following steps: calculating a detection residual; constructing a structural causal equation of the detection residual, learning a causal parameter through window sliding, obtaining a causal residual sequence; adopting a secondary window to slide and extract a minimum average processing effect sequence from the causal residual sequence, extracting a maneuver causal feature parameter based on the minimum average processing effect sequence; constructing a Hybrid Transformer network, inputting the maneuver causal feature parameter into the Hybrid Transformer network, and obtaining a maneuver detection preliminary result; and performing self-adaptive optimization on the hyperparameters of the Hybrid Transformer network according to a Bayesian optimization algorithm, and outputting a final spacecraft maneuver detection result according to the optimized network. The method can realize noise suppression, accurate feature extraction and efficient detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of space situation awareness, in particular to a spacecraft maneuver detection method based on causal deep learning. BACKGROUND

[0002] The rapid detection of space target orbit maneuver is a prerequisite for emergency scheduling triggering. In actual detection, radar and optical data are generally affected by factors such as multipath effect, atmospheric disturbance and short-arc observation, and there are problems such as large noise and weak effective features. In addition, the emergency scene also requires the detection algorithm to complete the discrimination within seconds, which puts higher requirements on the timeliness of the method.

[0003] The existing maneuver detection algorithm has the following limitations: first, it is severely dependent on prior information, such as Kalman filter algorithm which needs to preset the covariance matrix; second, the spatiotemporal feature capturing ability is insufficient, the maneuver event in the detection data is a complex pattern combining local mutation and sequence evolution, and the traditional method is difficult to effectively consider local feature extraction and long-range dependence relationship modeling, and the weak signal recognition ability is limited; third, the calculation efficiency is difficult to meet the timeliness requirement, the traditional method uses sequence recursion and dynamic fusion calculation, and the emergency response ability is insufficient. SUMMARY

[0004] Therefore, it is necessary to provide a spacecraft maneuver detection method based on causal deep learning which can realize noise suppression, accurate feature extraction and efficient detection in view of the above technical problems.

[0005] A spacecraft maneuver detection method based on causal deep learning, the method comprises:

[0006] Obtaining actual detection data of radar and optical; based on counterfactual reasoning, assuming that the target has not occurred maneuver, using TLE data and SGP4 / SDP4 model to calculate theoretical detection data, comparing the actual detection data with the theoretical detection data to obtain detection residual;

[0007] After constructing the structural causal equation of the detection residual, the causal parameters are obtained by window sliding learning, and a causal residual sequence is obtained;

[0008] The minimum average processing effect sequence is extracted from the causal residual sequence by using a secondary window, and the maneuver causal feature parameters are extracted based on the minimum average processing effect sequence;

[0009] A Hybrid Transformer network is constructed, the maneuver causal feature parameters are input into the Hybrid Transformer network, and are processed through a causal convolution layer, a Transformer encoder, a projection LSTM layer and a full connection layer in sequence to obtain a preliminary maneuver detection result;

[0010] According to the Bayesian optimization algorithm, the hyperparameters of the Hybrid Transformer network are adaptively optimized, and the final spacecraft maneuver detection result is output according to the optimized Hybrid Transformer network.

[0011] The above-mentioned spacecraft maneuver detection method based on causal deep learning realizes effective separation of maneuver signals and noise through counterfactual reasoning and structural causal equation, and suppresses detection noise interference; through secondary window sliding to extract MATE sequence and multi-dimensional features, local mutation and sequence evolution characteristics of spacecraft maneuver are accurately captured; through the Hybrid Transformer network, the advantages of causal convolution, Transformer and projection LSTM are fused, local feature extraction and long-range dependence modeling are considered, and the calculation efficiency is improved; through Bayesian optimization, the hyperparameters are adaptively optimized, and the model convergence speed and generalization ability are improved. The scheme of the present application does not need to rely on prior information, while ensuring the detection accuracy, it meets the requirement of second-level timeliness, effectively solves the technical difficulties of the existing method. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 FIG. 1 is a flowchart of a spacecraft maneuver detection method based on causal deep learning in one embodiment;

[0013] Figure 2 FIG. 2 is a flowchart of orbit maneuver causal feature parameter selection in one embodiment;

[0014] Figure 3 FIG. 3 is a schematic diagram of obtaining detection residuals based on counterfactual reasoning in one embodiment;

[0015] Figure 4 FIG. 4 is a schematic diagram of MATE calculation in another embodiment;

[0016] Figure 5 FIG. 5 is a flowchart of feature selection sliding window in one embodiment;

[0017] Figure 6 FIG. 6 is a Hybrid Transformer network structure diagram in one embodiment;

[0018] Figure 7 FIG. 7 is a causal convolution network structure diagram in one embodiment. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0020] In one embodiment, asFigure 1 As shown, a spacecraft maneuver detection method based on causal deep learning is provided, comprising the following steps:

[0021] Step 102, actual detection data of radar and optical is obtained; based on counterfactual reasoning, it is assumed that the target does not occur maneuver, theoretical detection data is calculated by using TLE data and SGP4 / SDP4 model, actual detection data is compared with theoretical detection data, and detection residual error is obtained.

[0022] Selecting the optical detection data and the radar detection data as maneuver detection indicators, the reasons are as follows: The parameters such as can directly reflect the position of the target. The parameters such as pitch angle rate and azimuth angle rate only describe the dynamic change of the target state, and are not direct state performance. In actual radar and optical detection system, there will be errors in the data. For the absolute position change, the noise usually fluctuates less. For the rate, the noise is large, which affects the accuracy of maneuver detection. Especially in high-frequency data acquisition, the estimation of the rate is affected by the data noise, resulting in unstable results.

[0023] The process of selecting the orbit maneuver causal characteristics is as shown in Figure 2 , which consists of three parts. The input is the detection data or and the target orbit database, and the output is the maneuver causal characteristic parameter.

[0024] The detection residual error is obtained based on counterfactual reasoning, and the basic idea is as shown in Figure 3 . Three elements are involved: 1) actual event: target maneuver, orbit parameter change, detection data dynamic change; 2) assumption condition: target does not occur maneuver; 3) preset value result: normal evolution of orbit, theoretical detection data is obtained by orbit prediction.

[0025] Suppose the theoretical detection data is , the actual detection data is , and the detection residual error can be expressed as:

[0026] .

[0027] By calculating the difference between the actual detection data and the theoretical detection data, the detection residual error containing the maneuver signal characteristics can be obtained, and the preliminary separation of the maneuver signal and the normal orbit signal is realized.

[0028] Step 104, the causal parameters are obtained by constructing the structural causal equation of the detection residual error and learning through window sliding, and the causal residual error sequence is obtained.

[0029] The sequence data formed by arranging the detection residuals in time sequence constructs a time series of detection residuals, which contains residual information at different times and the correlation in the time dimension. The structural causal equation is used to mine the internal causal relationship between the detection residuals, and then identify the residual changes caused by spacecraft maneuvers. The window sliding learning refers to setting a fixed length time window, sliding on the time series of detection residuals, and analyzing the residual data in each window to learn the causal correlation parameters (i.e., causal parameters) between the residuals.

[0030] The core logic of constructing the structural causal equation (SCE) is that the detection residuals can be decomposed into two independent multivariate Gaussian random variables, corresponding to the maneuver-related residuals and the noise-related residuals. When the spacecraft maneuvers, the distribution of the maneuver-related random variables will change, thereby causing the distribution of the overall detection residuals to change; when the spacecraft does not maneuver, the distribution of the maneuver-related random variables remains stable, and the distribution of the detection residuals is mainly determined by the noise. The SCE can accurately depict this causal relationship, thereby effectively distinguishing the residual changes caused by maneuvers from the residual fluctuations caused by noise, and providing accurate causal residual data for subsequent feature extraction.

[0031] In step 106, a secondary window is used to slide and extract a minimum average treatment effect sequence from the causal residual sequence, and a maneuver causal feature parameter is extracted based on the minimum average treatment effect sequence.

[0032] The theoretical residual refers to the residual component obtained based on the normal orbit evolution law, and the causal residual refers to the residual component directly related to the maneuver obtained by SCE separation. The MATE sequence is the full name of the minimum average treatment effect sequence, and the average treatment effect (ATE) is an index for measuring the degree of influence of the maneuver on the residual. Since the ATE value of a single point is prone to errors due to noise interference, a secondary window sliding strategy is used to extract the MATE sequence in this step: the first window sliding is to set a sliding window in the causal residual sequence, and calculate the ATE value in each window; the second optimization is to select the minimum ATE value in each window as the MATE value of the window, and through this way, abnormal values caused by noise are filtered, and a more stable MATE sequence is obtained.

[0033] The feature parameters extracted based on the MATE sequence refer to mining statistical information and derived information capable of representing the maneuvering features from the MATE sequence. Ten types of basic features are extracted, including four types of direct statistical features and six types of sliding window average features. Due to the differences in dimensions, resolutions and the like between the optical detection data and the radar detection data, the MATE sequence of the optical detection data will be expanded to 20-dimensional features on the basis of the ten types of basic features, and the MATE sequence of the radar detection data will be expanded to 40-dimensional features, so as to fully adapt to the characteristics of different detection data and improve the comprehensiveness of feature representation. By extracting multi-dimensional features, the local mutation features of the maneuvering event and the evolution features on the time sequence can be comprehensively captured, thereby providing rich and accurate feature inputs for subsequent model detection.

[0034] In step 108, a Hybrid Transformer network is constructed, the maneuvering causal feature parameters are input into the Hybrid Transformer network, sequentially pass through a causal convolution layer, a Transformer encoder, a projection LSTM layer and a full connection layer for processing, and a maneuvering detection preliminary result is obtained.

[0035] The Hybrid Transformer network is a hybrid architecture neural network designed by the present application. The Hybrid Transformer combines the global modeling of the Transformer, the local feature extraction of the causal convolution and the hidden state dimension reduction capability of the projection LSTM, thereby improving the accuracy of the orbit maneuvering detection. The input is the features of the radar or optical detection data, and the output is the maneuvering detection result. The network structure is as shown in Figure 6 The input data is first processed by the causal convolution layer to obtain features ensuring the causal relationship. Then, the features are transmitted to the Transformer encoder to capture the long-range dependence relationship of the data through the self-attention mechanism. Next, the parameter dimension is reduced through the projection LSTM to improve the calculation efficiency. Finally, the features are mapped to the classification results through the full connection layer, and the model hyperparameter optimization is performed through the Bayesian optimization algorithm. The causal convolution layer is a convolution operation designed for time sequence data. Its core feature is to ensure that the output only depends on the input data at the current time and the past time, and does not depend on the data at the future time, thereby preserving the causality of the time sequence and avoiding damage to the causal structure. The network structure is as shown in Figure 7The main role of this layer is to extract the local mutation features in the MATE sequence features, providing a basis for subsequent global feature modeling. The Transformer encoder is a network module based on self-attention mechanism, and the core components include position encoding, multi-head attention mechanism, feedforward neural network, residual connection and layer normalization, etc. Position encoding is used to add position information to sequence features, ensuring that the model captures the sequential relationship in the time dimension. Multi-head attention mechanism can calculate the correlation between positions in the sequence in parallel, effectively capturing long-range dependency features. Residual connection can alleviate the gradient vanishing problem caused by increasing network depth, and layer normalization can stabilize the output distribution of each layer to speed up model convergence. The main role of this layer is to capture the long-range dependencies in the MATE sequence features, and adapt to the sequence evolution characteristics of the maneuvering event.

[0036] The projection LSTM layer is an improvement on the traditional LSTM (Long Short-Term Memory Network), which adds a projection layer between the hidden state and the output to map the high-dimensional hidden state to a low-dimensional space. Its core role is to reduce feature dimension, reduce model parameter quantity, improve computational efficiency, and adapt to the needs of subsequent classification tasks. The fully connected layer is a classic module in neural networks, which maps the low-dimensional features output by the projection LSTM layer to a binary classification space of maneuvering or non-manipulating, thereby obtaining the preliminary results of maneuver detection. Through the synergistic effect of the above four layers of network, the Hybrid Transformer network can simultaneously capture local features and global dependencies, improve detection accuracy while ensuring computational efficiency, and meet the timeliness requirements of emergency scenarios.

[0037] Step 110, according to the Bayesian optimization algorithm, the hyperparameters of the Hybrid Transformer network are adaptively optimized, and the final spacecraft maneuver detection results are output according to the optimized Hybrid Transformer network.

[0038] The hyperparameters refer to parameters that need to be preset before training the Hybrid Transformer network, such as learning rate, regularization coefficient, number of attention heads in the Transformer encoder, hidden state dimension of the projection LSTM layer, etc. The values of these parameters directly affect the training effect and detection performance of the model. The Bayesian optimization algorithm models the historical parameter evaluation results, predicts the potential performance of unexplored parameters, and then selects the parameters with the most optimization potential for the next step of evaluation. Compared with traditional particle swarm optimization (PSO) and other algorithms, the Bayesian optimization algorithm has the advantages of high optimization efficiency and fast convergence speed. The Bayesian optimization algorithm continuously iterates and optimizes the values of the hyperparameters to minimize the prediction error of the Hybrid Transformer network, thereby improving the generalization ability and detection accuracy of the model. Finally, the classification results output by the model after hyperparameter optimization are the final spacecraft maneuver detection results, which can accurately determine whether the spacecraft has performed an orbital maneuver.

[0039] The above-mentioned spacecraft maneuver detection method based on causal deep learning realizes effective separation of maneuver signals and noise through counterfactual reasoning and structural causal equations, and suppresses detection noise interference; through secondary window sliding to extract MATE sequences and multi-dimensional features, local mutations and sequence evolution characteristics of spacecraft maneuvers are accurately captured; through the Hybrid Transformer network, the advantages of causal convolution, Transformer and projection LSTM are integrated, local feature extraction and long-range dependence modeling are considered, and the calculation efficiency is improved; through Bayesian optimization, hyperparameters are adaptively optimized, and the convergence speed and generalization ability of the model are improved. The scheme of the present application does not need to rely on prior information, and meets the requirement of second-level timeliness while ensuring detection accuracy, effectively solving the technical difficulties of existing methods.

[0040] In one of the embodiments, a structural causal equation for constructing a detection residual is constructed, including:

[0041] Let the multivariate Gaussian random variable corresponding to the detection residual be X, which is decomposed into two independent multivariate Gaussian random variables and , and The mean values of and are , , and The covariance matrix between and is When is fixed, the conditional expectation formula of

[0042] ;

[0043] The above equation shows that if the random variables and are subject to a multivariate Gaussian distribution, then the expectation of , can be expressed as a linear function of .

[0044] Assuming the Markov order is , for the probability density function , the following formula is obtained:

[0045] ;

[0046] Let , ;

[0047] The following is obtained: ;

[0048] Since , the structural causal equation of the detection residual is , where is the coefficient matrix, is the noise, denotes the th parameter value at the th step, denotes the th parameter value at the th step.

[0049] If the random variables are subject to a multivariate Gaussian distribution, can be linearly represented by and . Where is the coefficient matrix, is the noise. If the distribution of does not change, the distribution of will not change in theory. Target maneuver, the distribution will change, the distribution of will also change. Therefore, the change in can be used to detect target maneuvers.

[0050] In one embodiment, a quadratic window is used to slide from the causal residual sequence to extract the minimum average processing effect sequence, including:

[0051] A sliding window is selected in the causal residual sequence, and MATE is calculated in the sliding window in combination with the theoretical residual and the structural causal equation. Then, the window is slid backward, and the process of calculating MATE is repeated to obtain a MATE sequence, i.e., a minimum average treatment effect sequence.

[0052] In a specific embodiment, since the observation noise has randomness, calculating the ATE value of a single point will cause errors. This embodiment is based on , and a secondary sliding window is used to select the minimum ATE in the window to obtain a MATE sequence, and the mean, skewness, kurtosis and other parameters of the MATE sequence are extracted as maneuvering detection feature parameters. Figure 4 The calculation process of MATE is given.

[0053] In one of the embodiments, the maneuvering causal feature parameters include the mean of the MATE sequence, the variance of the MATE sequence, the skewness value of the MATE sequence, the kurtosis value of the MATE sequence, the maximum mean of the MATE sequence, the maximum variance of the MATE sequence, the mean of the mean of the MATE sequence, the mean of the variance of the MATE sequence, the difference between the maximum mean and the minimum mean of the MATE sequence, and the difference between the maximum variance and the minimum variance of the MATE sequence; the maneuvering causal feature parameters are extracted based on the minimum average treatment effect sequence, including:

[0054] First, a sliding window is selected in the causal residual sequence, and MATE is calculated in the sliding window in combination with the theoretical residual and the preceding formula. Then, the window is slid backward, and the process is repeated to obtain , i.e., a MATE sequence. Then, 10 features are extracted based on the MATE sequence, denoted as . For the MATE sequence of optical detection data, 20 features are obtained. For the MATE sequence of radar detection data, 40 features are obtained. The first four types of features are based on the mean, variance, skewness and other statistical values.

[0055] The calculation method of the mean of the MATE sequence is as follows:

[0056] ;

[0057] wherein, is the sequence length, is the value of the MATE at the i th position; and

[0058] The calculation method of the variance of the MATE sequence is as follows:

[0059] ;

[0060] The calculation method of the skewness value of the MATE sequence is as follows:

[0061] ​;

[0062] The peak value of the MATE sequence is calculated as follows:

[0063] .

[0064] In one embodiment, the next six feature parameters are obtained by sliding window averaging, the basic principle is as follows: a specified number of MATEs are averaged, and each time the window is slid, the window moves one step to the right, until all MATEs appear in at least one window, as shown in the flowchart Figure 5 The purpose of sliding window averaging is to concentrate more computing resources on the average value, which helps to detect complex maneuvers that are difficult to detect at the end of the time span. At the same time, the sliding window can also smooth the data and reduce noise interference, and the definitions of the features are as follows:

[0065] During the sliding process, the maximum mean value of the MATE sequence is calculated as follows:

[0066] ;

[0067] wherein, is the size of the sliding window, represents the maximum value of the mean value of all sliding windows;

[0068] During the sliding process, the maximum variance of the MATE sequence is calculated as follows:

[0069] ;

[0070] wherein, is the mean value of the MATE in the th window.

[0071] In one embodiment, during the sliding process, the average value of the mean value of the MATE sequence is calculated as follows:

[0072] ;

[0073] During the sliding process, the average value of the variance of the MATE sequence is calculated as follows:

[0074] .

[0075] In one embodiment, during the sliding process, the difference between the maximum mean value and the minimum mean value of the MATE sequence is calculated as follows:

[0076] ;

[0077] During the sliding process, the difference between the maximum variance and the minimum variance of the MATE sequence is calculated as follows:

[0078] .

[0079] In one embodiment, the causal convolutional layer ensures that in time... Output in real time Only depends on the input sequence arrive Partially independent of future inputs , The formula for calculating the output at time t is:

[0080] ;

[0081] in, for Output at any moment For time steps in the sequence The value at time, It is the first convolution kernel Each weight.

[0082] In a specific embodiment, causal convolution consists of an input layer, hidden layers, and an output layer. Its working principle is as follows: assuming an input sequence... convolution kernel ,in Let be the length of the convolution kernel. Traditional convolution operations are computed based on a sliding time window of the input, while causal convolution ensures computation within a time window. Output in real time Only depends on the input sequence arrive Partially independent of future inputs The calculation formula is:

[0083] ;

[0084] In the formula, for Output at any moment For time steps in the sequence The value at time, It is the first convolution kernel Each weight. Features processed by causal convolution can retain their potential temporal features, ensuring that their causal structure is not damaged, and providing meaningful input for subsequent Transformer modules.

[0085] In one embodiment, the formula for calculating the projection layer of the projected LSTM layer is:

[0086] ;

[0087] in, It is the projection weight matrix. is the target output dimension, is the dimension of the LSTM hidden state, is the bias term, is the hidden state.

[0088] In a specific embodiment, the projected LSTM is an extension of the traditional LSTM by adding a projection layer between the hidden state and the output, adjusting the output dimension to match the requirements of the classification task, and mapping to a low-dimensional space. The calculation steps are as follows: assuming the input sequence is , the hidden state is , the cell state is , the projection weight matrix is , and the target output is .

[0089] Calculate the forgetting gate , the input gate , the candidate cell state , the cell state , the output gate and the hidden state related parameters;

[0090] Calculate the projected output, which maps the hidden state to the output space:

[0091] ;

[0092] where is the projection weight matrix, is the target output dimension, is the dimension of the LSTM hidden state, is the bias term.

[0093] According to the projected output , then combined with the full connection layer to perform the maneuver detection task.

[0094] In the Hybrid Transformer architecture, the features processed by the Transformer are input into the projected LSTM layer. The projected LSTM maps the features to a low-dimensional space through hidden state projection, and then maps them to the category space through the fully connected layer. Finally, the detection result is output through the activation function.

[0095] In one embodiment, the hyperparameters of the Hybrid Transformer network are adaptively optimized according to the Bayesian optimization algorithm, including:

[0096] Step 1: Randomly generate groups of parameters as network input, calculate the sum of squared errors between network predicted value and actual value under each group of parameters:

[0097] ;

[0098] In the formula, For the first The sum of squared errors of the group parameters, , They are the first The true and predicted values ​​of the group parameters Indicates the first step that needs optimization. One learning rate and one regularization coefficient;

[0099] Step 2: Obtain the estimated function based on the sum of squared errors. minimum value and input parameters; where:

[0100] .

[0101] Step 3: Calculation Corresponding function value Evaluate the probe point pair estimation function The impact, and select the point with the greatest impact for the next step of exploration; among which:

[0102] .

[0103] Step 4: If the obtained minimum value does not meet the preset value or the maximum number of iterations is not reached, return to step 2; otherwise, output the parameter settings that minimize the network prediction error.

[0104] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0105] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0106] The above-described embodiments are merely illustrative of several embodiments of the present application, which are described in more detail and in a specific manner, but should not be construed as limiting the scope of the present application. It should be noted that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A spacecraft maneuver detection method based on causal deep learning, characterized in that, The method includes: Acquire actual radar and optical detection data; based on counterfactual reasoning assuming the target did not maneuver, calculate theoretical detection data using TLE data and SGP4 / SDP4 model, and compare actual detection data with theoretical detection data to obtain detection residuals; After constructing the structural causal equation for the detection residuals, causal parameters are obtained through window sliding learning, resulting in a causal residual sequence. A quadratic window is used to slide and extract the minimum average treatment effect sequence from the causal residual sequence, and the dynamic causal feature parameters are extracted based on the minimum average treatment effect sequence; A Hybrid Transformer network is constructed, and the causal feature parameters of the maneuver are input into the Hybrid Transformer network. The network is then processed sequentially through a causal convolutional layer, a Transformer, a projected LSTM layer, and a fully connected layer to obtain preliminary results of maneuver detection. The hyperparameters of the Hybrid Transformer network are adaptively optimized using the Bayesian optimization algorithm, and the final spacecraft maneuver detection results are output based on the optimized Hybrid Transformer network. The motorized causal feature parameters include the MATE sequence mean, MATE sequence variance, MATE sequence skewness, MATE sequence kurtosis, MATE sequence maximum mean, MATE sequence maximum variance, average of the MATE sequence mean, average of the MATE sequence variance, difference between the MATE sequence maximum and minimum mean, and difference between the MATE sequence maximum and minimum variance. The motorized causal feature parameters are extracted based on the minimum average treatment effect sequence, including: The average value of the MATE sequence is calculated as follows: in, For sequence length, For the first The value of each MATE; The variance of the MATE sequence is calculated as follows: ； The calculation method for the MATE sequence skewness value is as follows: ； The calculation method for the kurtosis value of the MATE sequence is as follows: ； During the sliding process, the maximum mean of the MATE sequence is calculated as follows: in, To adjust the sliding window size, This indicates taking the maximum value among all sliding window values; During the sliding process, the maximum variance of the MATE sequence is calculated as follows: in, For the first The average MATE value within each window; During the sliding process, the average value of the MATE sequence mean is calculated as follows: ； During the sliding process, the average variance of the MATE sequence is calculated as follows: ； During the sliding process, the difference between the maximum and minimum mean of the MATE sequence is calculated as follows: ； During the sliding process, the difference between the maximum and minimum variances of the MATE sequence is calculated as follows: .

2. The method according to claim 1, characterized in that, Constructing the structural causal equation for the detection residuals includes: Let X be the multivariate Gaussian random variable corresponding to the detection residual, and decompose it into two independent multivariate Gaussian random variables. and , and The means are respectively and The variance-covariance matrices are respectively , , and The covariance matrix between them is and Then fixed hour The conditional expectation formula is: Assume the Markov order is For probability density function The following formula is obtained: make , , get ; because The structural causal equation for the detected residual is then: ,in, It is a coefficient matrix. For noise, Indicates the first The parameter of the first Step value, Indicates the first The parameter of the first Step value.

3. The method according to claim 1, characterized in that, The minimum average treatment effect sequence is extracted from the causal residual sequence using a quadratic window sliding method, including: A sliding window is selected in the causal residual sequence. Within the sliding window, the MATE is calculated by combining the theoretical residual and the structural causal equation. Then the window is slid backward and the process of calculating MATE is repeated to obtain the MATE sequence, which is the minimum average treatment effect sequence.

4. The method according to claim 1, characterized in that, The causal convolutional layer ensures that in time... Output in real time Only depends on the input sequence arrive Partially independent of future inputs , The formula for calculating the output at time t is: in, for Output at any moment For time steps in the sequence The value at time, It is the first convolution kernel Each weight.

5. The method according to claim 1, characterized in that, The formula for calculating the projection layer of the projection LSTM layer is as follows: in, It is the projection weight matrix. It is the target output dimension. It is the dimension of the hidden state of the LSTM. It is a bias term. It is in a hidden state.

6. The method according to claim 1, characterized in that, The hyperparameters of the HybridTransformer network are adaptively optimized using the Bayesian optimization algorithm, including: Step 1: Randomly generate Group parameters As input to the network, calculate the sum of squared errors between the network predictions and actual values ​​for each set of parameters: In the formula, For the first The sum of squared errors of the group parameters, , They are the first The true and predicted values ​​of the group parameters Indicates the first step that needs optimization. One learning rate and one regularization coefficient; Step 2: Obtain the estimated function based on the sum of squared errors. minimum value and input parameters; where, ； Step 3: Calculation Corresponding function value Evaluate the probe point pair estimation function The impact is assessed, and the point with the greatest impact is selected for the next step of exploration; among them... ； Step 4: If the obtained minimum value does not meet the preset value or the maximum number of iterations is not reached, return to step 2; otherwise, output the parameter settings that minimize the network prediction error.

Images