Satellite telemetering anomaly detection method and system based on multi-scale time domain convolutional network

Abstract

The invention discloses a satellite telemetering anomaly detection method and system based on a multi-scale time domain convolutional network, belongs to the technical field of time sequence anomaly detection, and solves the problems that an existing space-time fusion method is insufficient in single-scale time feature capture capability and simple in multi-scale feature fusion mode. The method comprises the steps of collecting satellite telemetry data, constructing a data sequence and a sample matrix, and performing preprocessing and feature extraction to obtain a spatial feature vector; constructing a multi-layer MSTCN network, and performing time evolution feature extraction and feature fusion on the spatial feature vector to obtain a fused feature vector; activating the fusion feature vector to obtain feature output of the telemetry data; calculating residual statistical magnitude of the actual data and the predicted data, calculating a mean value and a standard deviation of the residual statistical magnitude, and setting an abnormal threshold value and a range thereof; and calculating an anomaly score of the satellite telemetry data, comparing the anomaly score with a preset threshold value, and judging anomaly. The method is suitable for anomaly detection scenes with different dimensions and different types of telemetry data.

Description

Technical Field

[0001] This invention belongs to the field of time series anomaly detection technology, specifically relating to a satellite telemetry anomaly detection technology based on multi-scale temporal convolutional networks. Background Technology

[0002] Satellite telemetry systems are a core component of satellite status monitoring and fault diagnosis. By collecting multi-dimensional parameters such as attitude, orbit, and power supply, they enable real-time perception of the satellite's health status. However, the on-orbit operating environment of satellites is complex, with environmental interference factors such as space radiation and temperature fluctuations. Telemetry parameters may become abnormal due to sensor drift, circuit aging, actuator jamming, and other reasons. Failure to detect these abnormalities in a timely manner could lead to catastrophic failures. Therefore, efficient methods for detecting satellite telemetry anomalies are crucial to ensuring the safe operation of satellites.

[0003] Existing satellite telemetry anomaly detection methods are mainly divided into time-correlated methods and spatiotemporal fusion methods. Time-correlated methods rely on the temporal evolution characteristics of a single parameter, but ignore the spatial correlation between multi-dimensional parameters (such as the coupling relationship between attitude angle and orbital parameters). This may lead to incomplete feature representation of the model, thereby affecting the accuracy and other performance aspects of anomaly detection.

[0004] Spatiotemporal fusion methods improve detection performance by jointly extracting spatial correlation and temporal evolution features, but limitations remain. First, single-scale temporal feature capture is insufficient, meaning it can only capture evolutionary features at a single time scale. However, anomalies in satellite telemetry parameters, such as short-term sudden failures and long-term slow drifts, may be distributed across different time scales. For example, high-frequency vibrations of attitude angles correspond to short time scales, while slow decreases in battery voltage correspond to long time scales. Therefore, a single scale cannot provide comprehensive coverage, affecting the accuracy of anomaly detection. Second, multi-scale feature fusion methods are simplistic. Some methods achieve multi-scale feature extraction by stacking networks with different parameters, but fine-grained information is easily lost during inter-layer feature transfer, and the weights of features at different scales cannot be dynamically adjusted, thus affecting the accuracy of anomaly detection. Summary of the Invention

[0005] This invention provides a satellite telemetry anomaly detection method and system based on a multi-scale temporal convolutional network (TCN). Its purpose is to address the shortcomings of existing spatiotemporal fusion methods, such as insufficient single-scale temporal feature capture capability and simplistic multi-scale feature fusion methods. By fusing convolutional branches with different dilation rates within the same temporal convolutional network (TCN) layer, multi-scale temporal evolution features are dynamically captured. Combined with multi-dimensional spatial correlation features from satellite telemetry, accurate anomaly detection of satellite telemetry parameters is achieved. First, a spatial feature extraction module captures the spatial correlation between multi-dimensional telemetry parameters. Then, a multi-scale temporal evolution feature is extracted using a multi-scale temporal convolutional network (MSTCN). Finally, anomaly determination is achieved based on residual analysis.

[0006] Firstly, the purpose of this invention is to provide a satellite telemetry anomaly detection method based on multi-scale temporal convolutional networks, comprising the following steps:

[0007] S1: Collect satellite telemetry data, construct data sequences and sample matrices, perform preprocessing and feature extraction, and obtain spatial feature vectors;

[0008] S2: Construct a multi-layer MSTCN network, and use the multi-layer MSTCN network to perform temporal evolution feature extraction and feature fusion on the spatial feature vector to obtain a fused feature vector;

[0009] S3: The fused feature vector is activated using a linear rectified function to obtain the feature output of the telemetry data;

[0010] S4: Calculate the residual statistics between the actual data and the predicted data, calculate the mean and standard deviation of the residual statistics, and set the outlier threshold and its range;

[0011] S5: Calculate the anomaly score of the satellite telemetry data based on the characteristics of the telemetry data, compare the anomaly score with a preset threshold, and determine the anomaly.

[0012] Furthermore, a preferred embodiment is provided: S1 includes:

[0013] S11: Standardize the satellite telemetry data sequences in the sample matrix;

[0014] S12: Divide the standardized data sequence to obtain the sample input window;

[0015] S13: Perform a nonlinear transformation on the sample input window to obtain the feature interaction matrix;

[0016] S14: Concatenate the row and column vectors of the feature interaction matrix to obtain the spatial feature vector.

[0017] Furthermore, a preferred embodiment is provided: the MSTCN network includes convolutional layers and scale attention layers, wherein the convolutional layers include multiple parallel dilated convolutional branches, each branch employing a different dilation rate to capture temporal evolution features at short, medium, and long time scales, respectively.

[0018] Furthermore, a preferred solution is provided: for a spatial feature vector, the output of each dilated convolution is represented as:

[0019]

[0020] Where F = [f1, f2, ..., f T ] represents the spatial eigenvector, C k It is the kernel size of the k-th branch, d k is the dilation factor of the k-th branch, represents the number of intervals for each element of the convolution kernel, and t represents time step.

[0021] Furthermore, a preferred embodiment is provided: the multiple parallel dilated convolution branches are connected by a residual block structure, and 1×1 convolutions are introduced to ensure that the number of channels in the added feature maps is consistent.

[0022] Furthermore, a preferred solution is provided: the feature fusion step specifically involves multiplying the output of each branch by its corresponding attention weight, and then summing the results to obtain a fused feature vector.

[0023] Furthermore, a preferred embodiment is provided: the fused feature vector is represented as follows:

[0024]

[0025] in, This is the fused temporal feature vector, containing feature information at short, medium, and long time scales; ⊙ represents element-wise multiplication, and ψ... s For the output of multiple parallel dilated convolution branches, θ s For attention scores, f Csk It is the k-th convolution kernel of the s-th branch.

[0026] Secondly, the purpose of this invention is to propose a satellite telemetry anomaly detection system based on a multi-scale temporal convolutional network. This system is implemented based on a satellite telemetry anomaly detection method based on a multi-scale temporal convolutional network as described in any one or more of the above-mentioned schemes. The system includes:

[0027] Data processing module: used to acquire satellite telemetry data, construct data sequences and sample matrices, perform preprocessing and feature extraction, and obtain spatial feature vectors;

[0028] Feature fusion module: used to construct a multi-layer MSTCN network, and to perform temporal evolution feature extraction and feature fusion on the spatial feature vector using the multi-layer MSTCN network to obtain a fused feature vector;

[0029] Vector activation module: used to activate the fused feature vector using a linear rectified function to obtain the feature output of the telemetry data;

[0030] Threshold setting module: used to calculate the residual statistics between actual data and predicted data, calculate the mean and standard deviation of the residual statistics, and set the outlier threshold and its range;

[0031] Anomaly detection module: used to calculate the anomaly score of satellite telemetry data based on the characteristics of the telemetry data, compare the anomaly score with a preset threshold, and determine anomalies.

[0032] Thirdly, the present invention aims to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, and when the processor runs the computer program stored in the memory, the processor executes a satellite telemetry anomaly detection method based on a multi-scale temporal convolutional network according to any one or more of the above-described schemes.

[0033] Fourthly, the present invention aims to provide a computer-readable storage medium for storing a computer program that executes a satellite telemetry anomaly detection method based on a multi-scale temporal convolutional network as described in any one or more of the above-described schemes.

[0034] Compared with the prior art, the advantages of the present invention are:

[0035] 1) The method described in this invention designs an MSTCN network to achieve simultaneous capture of multi-scale temporal evolution features. Multiple parallel dilated convolutional branches are set up in the same TCN layer to capture temporal evolution features at short, medium, and long time scales, respectively. MSTCN synchronously captures multi-scale features within the same layer, avoiding feature loss between layers and solving the problem that a single scale cannot cover all anomalous patterns.

[0036] 2) The method described in this invention employs a SAM mechanism. Attention scores for each dilated convolutional branch are calculated using a fully connected layer, and the outputs of each branch are weighted and fused. This mechanism dynamically allocates weights based on the temporal scale of the abnormal pattern, solving the problems of existing methods' simple fusion methods and inability to focus on key scales.

[0037] Finally, using actual satellite attitude control system telemetry data as one of the application objects, the proposed method was verified and evaluated by injecting data on possible anomalies such as transverse bias and constant gain in the telemetry sequence. The results show that this invention achieves high accuracy (98.6%) and low false alarm rate (3.5%) in detecting anomalies of different dimensions and types in telemetry data.

[0038] This invention is applicable to anomaly detection scenarios of different dimensions and types in telemetry data. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0040] Figure 1 This is a block diagram of satellite telemetry anomaly detection based on multi-scale temporal convolutional networks as described in a specific embodiment of the present invention;

[0041] Figure 2 The yaw rate telemetry data and momentum wheel speed telemetry data described in the specific embodiments of the present invention;

[0042] Figure 3 This is the abnormal score distribution of yaw rate abnormal data as described in a specific embodiment of the present invention;

[0043] Figure 4 This is the distribution of abnormal scores for momentum wheel speed abnormality data as described in a specific embodiment of the present invention. Detailed Implementation

[0044] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0045] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0046] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0047] Implementation Method 1

[0048] This implementation proposes a satellite telemetry anomaly detection method based on a multi-scale temporal convolutional network (MTCN), employing a spatial-first, temporal-second processing architecture. First, a spatial feature extraction module captures the spatial correlations between multi-dimensional telemetry parameters. A multi-dimensional feature interaction layer (MDFIL) is designed to capture the spatial correlations between different telemetry parameters through nonlinear transformations, generating spatial feature vectors containing cross-parameter associations. Second, a multi-scale temporal evolution feature is extracted using a multi-scale temporal convolutional network (MSTCN). Multiple parallel dilated convolutional branches are set within the same TCN layer to capture evolution features at short, medium, and long time scales, respectively. The outputs of each branch are dynamically weighted and fused using a scale attention mechanism (SAM) to preserve the integrity of the multi-scale features. Finally, the residual between the actual satellite telemetry data and the predicted data is calculated and compared with a threshold to complete anomaly detection.

[0049] Specifically, the following steps are included:

[0050] Step 1, Data Acquisition and Preprocessing:

[0051] Satellite telemetry data is collected, and data sequences and sample matrices are constructed. For a given satellite telemetry channel, its data sequence is an ordered set of elements consisting of recorded values ​​and recording times. The data sequence of telemetry channel i is abbreviated as S. i =[s i1 ,s i2 ,...,s it ], where s itLet be the telemetry value of telemetry channel i at time t. The sample input matrix from all dimensions of telemetry channels is abbreviated as:

[0052] A = [a1, a2, ..., a t ,...,a T ] T (1)

[0053] Among them, column vectors Let be the set of data from all telemetry channels at time t, represent the multidimensional spatial vector of the sample at time step t, and n-1 be the number of related dimensions of the telemetry dimension to be measured. T is the length of the sample.

[0054] Step 1.2: Since the data from each dimension of the telemetry channel have different measurement units, it is necessary to standardize the data. For the satellite telemetry data sequence a at time t, which includes the telemetry data of the dimension to be measured and other related dimensions... t Perform the standardization process as shown in equation (2):

[0055]

[0056] Where μ and σ are respectively a t The mean and standard deviation of the sequence are calculated. Next, the standardized data is divided into time windows of length t and entered into the detection window to facilitate data processing by the feature extraction module.

[0057] Step 1.3: For the standardized sample input window, to capture the spatial correlation between multi-dimensional parameters, an MDFIL is designed to achieve feature interaction between different parameters through nonlinear transformation. For the standardized vector a at each time step... t A nonlinear transformation is performed through a fully connected layer (FC) to generate the feature interaction matrix M. t As shown in equation (3):

[0058] M t =σ(a t ·W m +b m (3)

[0059] in, These are the linear transformation matrix and the linear transformation bias, respectively.

[0060] Step 1.4: For the feature interaction matrix M t By concatenating the column vectors, we obtain a spatial feature vector F containing cross-parameter correlations, as shown in equation (4):

[0061] F = Concat(M) t [:,1],...,Mt [:,n]), (4)

[0062] Step 2, Temporal Evolution Feature Extraction:

[0063] Step 2.1: Based on the spatial feature vectors obtained from MDFIL, further design MSTCN to extract the temporal evolution features of multidimensional telemetry sequences. K parallel dilated convolutional branches are set in the same TCN layer, each branch using a different dilation rate d1, d2, ..., d... K This captures the temporal evolution features at short, medium, and long time scales, respectively. For the spatial feature vector sequence F = [f1, f2, ..., f...],... T The output of each dilated convolution is shown in equation (5):

[0064]

[0065] Where C k It is the kernel size of the k-th branch, d k is the dilation factor of the k-th branch, representing the number of intervals for each element of the convolution kernel. Taking d1=1 (corresponding to the short-term branch), d2=2 (corresponding to the medium-term branch), and d3=4 (corresponding to the long-term branch) as examples, the receptive fields of each branch are 2d1+1=3, 2d2+1=5, and 2d3+1=9, respectively, which capture evolutionary features at different time scales (such as sudden anomalies at the short-term scale and slow drifts at the long-term scale).

[0066] Step 2.2: To dynamically adjust the weights of features at different scales, a SAM is designed. By learning the attention weights of each branch, weighted fusion of multi-scale features is achieved. ψ is output for each branch. s (k), the attention score θ is calculated via FC. s As shown in equation (6).

[0067] θ s =Softmax(FC(ψ) s (K))), (6)

[0068] Where, θ s Let be the attention weight vector for the s-th branch, used to measure the importance of the features in that branch.

[0069] Step 2.3: Multiply the output of each branch by the corresponding attention weight, and then sum them to obtain the fused feature vector p, as shown in equation (7):

[0070]

[0071] in, The fused temporal feature vector contains feature information at short, medium, and long time scales; ⊙ represents element-wise multiplication. By combining multi-void rate convolutional branches with SAM, MSTCN can simultaneously capture the multi-scale temporal evolution features of satellite telemetry data and dynamically adjust the weights of features at each scale.

[0072] For multi-layered MSTCN networks, a residual block structure is used to skip connections between adjacent layers. To ensure that the number of feature map channels on both sides of the skip connection layer matches, a 1×1 convolution is introduced to make the number of channels in the added feature maps consistent.

[0073] Step 3, Feature Vector Activation:

[0074] The feature output of the residual block is activated using the Rectified Linear Unit (ReLU) function shown in Equation (8), and finally the feature output of the telemetry data of the dimension to be measured is shown in Equation (9).

[0075] ReLU(p) = max(0,p), (8)

[0076]

[0077] Where Y = [y1, ..., y T ] represents the output of the detected dimension features obtained from the prediction of time evolution features, C 1×1 (F) represents a 1×1 convolution operation on F, and lunm represents the number of layers in MSTCN.

[0078] Step 4, Threshold setting:

[0079] The threshold is selected based on the mean μ and standard deviation σ of the residuals between the actual and predicted data during model training. Residuals exceeding a certain multiple of the standard deviation are defined as outlier thresholds, which are set to 3 times in this implementation. The upper bound of the threshold is λ. up With the lower bound of the threshold λ low As shown in equations (10) and (11):

[0080] λ up =μ+3σ (10)

[0081] λ low =μ-3σ (11)

[0082] Step 5, Anomaly Detection:

[0083] Step 5.1: Calculate the anomaly score of the telemetry data for the dimension to be measured at time step t.

[0084]

[0085] Step 5.2: Define an anomaly label set ξ = [ξ1, ξ2, ..., ξ] t If the abnormal score at time step t If the telemetry data deviates from the set anomaly threshold range, the telemetry data at that time step is judged as anomalous data, and the corresponding anomaly label ξ for time step t is changed accordingly. t Mark it as 1, otherwise mark it as 0.

[0086] Implementation Method 2

[0087] This embodiment is a further illustrative example of the satellite telemetry anomaly detection method based on a multi-scale temporal convolutional network described in Embodiment 1.

[0088] Step 1: This implementation method verifies the approach using a real-world telemetry dataset of the satellite attitude control system. The satellite telemetry dataset originates from the condition monitoring dataset during the satellite's operation in orbit. Fault injection is performed on two dimensions of data: the yaw rate (YAR) telemetry data and the momentum wheel speed (MWS) telemetry data of the attitude control system. A total of 3000 sets of normal and fault-containing abnormal data are analyzed. Figure 2 As shown in the figure. 2(a) to 2(d) represent normal YAR data and abnormal data with constant bias fault, normal YAR data and abnormal data with constant gain fault, normal MWS data and abnormal data with constant bias fault, and normal MWS data and abnormal data with constant gain fault, respectively.

[0089] Step 2: For the telemetry sequence of the dimension to be measured, i.e., the yaw rate sequence S yar and the momentum wheel speed sequence S mws The telemetry sequences of the attitude control system and other dimensions are all standardized as shown in equation (2). The anomaly detection process of the yaw rate sequence is used as a case study for derivation, and the anomaly detection process of the momentum wheel speed sequence is similar.

[0090] Step 3: Implement feature interaction between different parameters using MDFIL. For each time step of all telemetry sequences, the normalized vector a... t Generate the feature interaction matrix Myar t As shown in equation (13):

[0091] Myar t =σ(a t ·W m +b m (13)

[0092] Step 4: Concatenate the column vectors of the feature interaction matrix to obtain the spatial feature vector Fyar containing cross-parameter correlations, as shown in Equation (14):

[0093] Fyar = Concat(Myar) t [:,1],...,Myar t [:,25]), (14)

[0094] Step 5: Based on the spatial feature vector obtained from MDFIL, further extract the temporal evolution features of the multidimensional telemetry sequence through MSTCN, and finally obtain the feature output of the TCN network for the yaw rate sequence Syar as shown in Equation (16). As shown in Equation (15):

[0095]

[0096] Step 6: Determine the upper bound of the anomaly detection threshold λ according to equations (9) and (10). up With the lower bound of the threshold λ low Based on calculations, the upper and lower thresholds for anomaly detection of the roll rate sequence are 0.0076 and -0.0047, respectively. The upper and lower thresholds for anomaly detection of the momentum wheel speed sequence are 0.025 and -0.028, respectively.

[0097] Step 7: Calculate the anomaly score of the telemetry data of the dimension to be measured at time step t according to formula (10). Calculations show that the outlier score distributions of the 3000 sets of yaw rate data containing faults and the outlier score distributions of the momentum wheel speed data (i.e., the residuals between predicted and actual data) are as follows: Figure 3 and Figure 4 As shown.

[0098] Step 8: Define a prediction anomaly label set ξ = [ξ1, ξ2, ..., ξ] for each of the two sets of telemetry sequences. t If the abnormal score at time step t If the telemetry data deviates from the set anomaly threshold λ, the telemetry data at that time step is judged as anomalous data, and the corresponding anomaly label ξ for time step t is changed. t Mark 1 if the anomaly label is not marked, otherwise mark 0. Compare the anomaly label set obtained by predicting the two datasets with the actual data anomaly label set, and calculate the accuracy (ACC), false alarm rate (FAR), and F1 score (F1). The calculation results are shown in Table 1.

[0099] Table 1. Anomaly Detection Indicators for Yaw Rate and Momentum Wheel Speed

[0100]

[0101] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A satellite telemetry anomaly detection method based on multi-scale temporal convolutional networks, characterized in that, Includes the following steps: S1: Collect satellite telemetry data, construct data sequences and sample matrices, perform preprocessing and feature extraction, and obtain spatial feature vectors; S2: Construct a multi-layer MSTCN network, and use the multi-layer MSTCN network to perform temporal evolution feature extraction and feature fusion on the spatial feature vector to obtain a fused feature vector; S3: The fused feature vector is activated using a linear rectified function to obtain the feature output of the telemetry data; S4: Calculate the residual statistics between the actual data and the predicted data, calculate the mean and standard deviation of the residual statistics, and set the outlier threshold and its range; S5: Calculate the anomaly score of the satellite telemetry data based on the characteristics of the telemetry data, compare the anomaly score with a preset threshold, and determine the anomaly.

2. The satellite telemetry anomaly detection method based on a multi-scale temporal convolutional network according to claim 1, characterized in that, S1 includes: S11: Standardize the satellite telemetry data sequences in the sample matrix; S12: Divide the standardized data sequence to obtain the sample input window; S13: Perform a nonlinear transformation on the sample input window to obtain the feature interaction matrix; S14: Concatenate the row and column vectors of the feature interaction matrix to obtain the spatial feature vector.

3. The satellite telemetry anomaly detection method based on a multi-scale temporal convolutional network according to claim 1, characterized in that, The MSTCN network includes convolutional layers and scale attention layers. The convolutional layers include multiple parallel dilated convolutional branches, each with a different dilation rate, to capture temporal evolution features at short, medium, and long time scales, respectively.

4. The satellite telemetry anomaly detection method based on a multi-scale temporal convolutional network according to claim 3, characterized in that, For a spatial feature vector, the output of each dilated convolution is represented as: Where F = [f1, f2, ..., f T ] represents the spatial eigenvector, C k It is the kernel size of the k-th branch, d k is the dilation factor of the k-th branch, represents the number of intervals for each element of the convolution kernel, and t represents time step.

5. The satellite telemetry anomaly detection method based on a multi-scale temporal convolutional network according to claim 3, characterized in that, The multiple parallel dilated convolution branches are connected by a residual block structure, and 1×1 convolutions are introduced to ensure that the number of channels in the added feature maps is consistent.

6. The satellite telemetry anomaly detection method based on a multi-scale temporal convolutional network according to claim 4, characterized in that, The specific steps of feature fusion are as follows: multiply the output of each branch by the corresponding attention weight, and then sum them to obtain the fused feature vector.

7. The satellite telemetry anomaly detection method based on a multi-scale temporal convolutional network according to claim 4, characterized in that, The fused feature vector is represented as follows: in, This is the fused temporal feature vector, containing feature information at short, medium, and long time scales; ⊙ represents element-wise multiplication, and ψ... s For the output of multiple parallel dilated convolution branches, θ s To score attention, It is the k-th convolution kernel of the s-th branch.

8. A satellite telemetry anomaly detection system based on a multi-scale temporal convolutional network, characterized in that, The system is implemented based on a satellite telemetry anomaly detection method based on a multi-scale temporal convolutional network as described in any one of claims 1-7, and the system includes: Data processing module: used to acquire satellite telemetry data, construct data sequences and sample matrices, perform preprocessing and feature extraction, and obtain spatial feature vectors; Feature fusion module: used to construct a multi-layer MSTCN network, and to perform temporal evolution feature extraction and feature fusion on the spatial feature vector using the multi-layer MSTCN network to obtain a fused feature vector; Vector activation module: used to activate the fused feature vector using a linear rectified function to obtain the feature output of the telemetry data; Threshold setting module: used to calculate the residual statistics between actual data and predicted data, calculate the mean and standard deviation of the residual statistics, and set the outlier threshold and its range; Anomaly detection module: used to calculate the anomaly score of satellite telemetry data based on the characteristics of the telemetry data, compare the anomaly score with a preset threshold, and determine anomalies.

9. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes a satellite telemetry anomaly detection method based on a multi-scale temporal convolutional network according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program that executes a satellite telemetry anomaly detection method based on a multi-scale temporal convolutional network according to any one of claims 1-7.

Images