Satellite telemetry data anomaly detection method capable of automatically distinguishing data states

Abstract

The invention discloses a satellite telemetry data anomaly detection method capable of automatically distinguishing data states, which comprises the following steps of: 1, preprocessing original telemetry data to obtain time series data, and dividing the time series data into windows with fixed lengths; step 2, dynamically dividing the feature vectors of all windows by using a satellite telemetry data state recognition module based on adaptive state recognition, and outputting a final state recognition result; 3, constructing a satellite telemetry data anomaly detection module based on the time convolutional network, and outputting a detection result of a time sequence mode by taking the final state recognition result as input; and step 4, visualizing the detection result of the time sequence mode through the detection result and a visual interpretation module. The method has good generalization ability, can accurately identify the known normal state, the abnormal state and the unknown abnormal state without manual annotation, and significantly improves the analysis efficiency and the anomaly detection precision of the satellite telemetry data.

Description

Technical Field

[0001] The present invention relates to satellite anomaly detection, in particular to a satellite telemetry data anomaly detection method that automatically distinguishes data states. Background Art

[0002] Telemetry data plays a crucial role in satellite monitoring. It records a wide range of observational information collected by satellites during their long-term operations, covering fields such as meteorology, geology, and ecology. This data is often high-dimensional, nonlinear, dynamically changing, and diverse, providing a rich source of information for anomaly detection, revealing potential faults or abnormal conditions. However, the complexity and uncertainty of this data complicate detection, and traditional methods often fail to effectively capture underlying patterns, making anomaly detection a significant challenge. Therefore, accurately identifying anomalies from this vast and complex volume of telemetry data is a challenging task, requiring the adoption of more advanced and efficient detection methods.

[0003] Existing satellite anomaly detection methods have evolved from traditional rule-based methods to deep learning-driven intelligent detection systems. Traditional methods, such as the satellite fault diagnosis method based on a fault dictionary (CN101590918B) and the spacecraft on-orbit anomaly alarm and fault diagnosis system, rely on expert knowledge to build a fault pattern library and perform anomaly detection through pattern matching. Although these methods are relatively effective for known fault modes, they have weak generalization capabilities when faced with complex and unknown fault types and are difficult to adapt to new anomalies. With the development of data analysis technology, star sensor data anomaly detection (CN202010361728.2) and the satellite anomaly state detection system based on telemetry data (CN114218770A) have introduced statistical methods such as Mahalanobis distance and dynamic time warping (DTW), which have improved anomaly detection accuracy. However, these methods generally require strong assumptions about the data distribution and have difficulty processing high-dimensional and complex telemetry data. Recently, a study proposed a clustering-based satellite telemetry data fault diagnosis method (such as CN111875057A). The Gaussian mixture model structure is selected through the leaf node information criterion, and it is trained using normal multivariate telemetry data. The univariate Gaussian distribution is combined to judge the similarity of new data. Without clear prior knowledge, potential anomalies can be identified, providing new ideas for unsupervised detection of on-orbit faults.

[0004] As satellite anomaly detection has entered the intelligent detection stage, the network-controlled satellite communication network anomaly detection method (CN102487293A) combines data mining and artificial immune technology to primarily detect anomalies in satellite communication networks, enhancing ground station communication behavior analysis capabilities. However, its ability to detect physical anomalies in telemetry data is limited. Subsequently, the intelligent detection method for in-orbit spacecraft surface anomalies (CN108469234B) leverages three-dimensional reconstruction and computer vision technologies to automatically detect surface defects on in-orbit satellites, improving detection accuracy. However, this is still limited by computational resource consumption and false detection rates in complex environments.

[0005] In recent years, the application of deep learning in satellite anomaly detection has gradually developed, such as the anomaly diagnosis method for satellite telemetry multi-dimensional time series data (CN117688496B). This method uses graph attention network (GAT) and transfer entropy to enhance the relationship modeling ability between data, making it have higher detection accuracy in dynamic environments. However, this method still has problems such as reliance on manual labeling, insufficient generalization ability, and separate training for different tasks. In addition, traditional methods based on fault dictionaries are still difficult to adapt to unknown anomaly patterns. Autonomous fault diagnosis and intelligent maintenance systems, although focusing on long-term health management of spacecraft, have limited detection capabilities for telemetry anomalies.

[0006] In general, existing methods still have the following limitations in satellite telemetry anomaly detection: (1) Data sets rely on manual labeling, which is costly. Most deep learning methods rely on a large amount of labeled data for training, while satellite telemetry data often lack high-quality labels. The labeling process requires expert participation, which is time-consuming, labor-intensive and costly. In addition, the working conditions of different satellite missions may vary greatly, and labeled data is difficult to reuse directly, further exacerbating the difficulty of data labeling; (2) Lack of generalization ability. They can only identify known normal and abnormal patterns in the data set and cannot effectively detect unknown faults. Traditional methods are mostly based on rule matching or pattern recognition. They perform well for known abnormal patterns, but when faced with complex, unknown or atypical faults, their generalization ability is limited and they are prone to missed detection. In addition, due to the complex on-orbit environment of satellites, abnormal states that have never been encountered may occur. Existing methods find it difficult to effectively capture these unknown anomalies, thereby reducing the effectiveness of the system. The reliability of the system; (3) Some methods have poor interpretability, making it difficult to understand the decision-making basis of anomaly detection. Although traditional deep learning methods can improve the accuracy of anomaly detection, the results of black-box detection methods often lack interpretability and are difficult to provide a clear basis for anomaly detection. Operation and maintenance personnel find it difficult to understand the decision-making process of the model, resulting in anomaly detection results that are difficult to be directly used for fault diagnosis and subsequent processing. In addition, the lack of interpretability may also lead to the inability to effectively verify and correct misdetected anomalies, affecting the credibility of the system; (4) The lack of intuitive visual analysis limits the rapid response of operation and maintenance personnel to abnormal situations. Satellite telemetry data is usually high-dimensional time series data. Traditional methods are difficult to provide clear anomaly visualization methods, making it difficult for operation and maintenance personnel to intuitively understand the occurrence process and trend of anomalies. Most existing anomaly detection methods only output an anomaly score or detection result, lacking specific anomaly pattern analysis, which limits the ability to trace the fault and support decisions. Summary of the Invention

[0007] The purpose of the present invention is to provide a satellite telemetry data anomaly detection method that automatically distinguishes data states. The method has good generalization ability and can accurately identify known normal states, abnormal states, and unknown abnormal states without manual labeling, thereby significantly improving the analysis efficiency and anomaly detection accuracy of satellite telemetry data.

[0008] The present invention is achieved through the following technical solutions:

[0009] A satellite telemetry data anomaly detection method for automatically distinguishing data states comprises the following steps:

[0010] Step 1: Preprocess the original telemetry data to obtain time series data, and divide the time series data into windows of fixed length;

[0011] Step 2: Dynamically divide the feature vectors of all windows using the satellite telemetry data state recognition module based on adaptive state recognition, and output the final state recognition result;

[0012] The process of dividing the feature vectors of all input windows by the satellite telemetry data state recognition module based on adaptive state recognition is divided into an initialization phase and a state recognition learning phase, wherein: in the initialization phase, the number of states is dynamically adjusted by adaptively setting the initial similarity threshold; in the state recognition learning phase, feature characters are generated by extracting window features and state recognition, thereby converting time series data into a discrete representation form, and outputting a state recognition result, which is a time series;

[0013] Step 3: Construct a satellite telemetry data anomaly detection module based on a temporal convolutional network, which takes the final state recognition result as input and outputs the detection result of the time series pattern;

[0014] The satellite telemetry data anomaly detection module based on the temporal convolutional network is a network model that combines temporal feature learning and pattern recognition capabilities by integrating the spatial embedding method on the basis of the temporal convolutional neural network. It includes a feature extraction module and a judgment and recognition module. The feature extraction module extracts temporal features through multiple convolutional layers and pooling layers; the judgment and recognition module completes state judgment with the help of the fully connected layer and the output layer.

[0015] The processing process of the satellite telemetry data anomaly detection module based on the time convolutional network on the final state recognition result of the input is as follows: the feature extraction module performs deep feature extraction on the input time series through the time convolutional network, the extracted deep features first pass through two fully connected layers, the vector output by the last fully connected layer is divided into two paths, and respectively transmitted to the x-axis fully connected layer and the y-axis fully connected layer to obtain the probability coordinate x vector and the probability coordinate y vector, and the probability coordinate x vector and the probability coordinate y vector are combined into a feature coordinate vector and input into the judgment and recognition module, and the input time series pattern is judged according to the Euclidean distance between the feature coordinates of the input time series pattern and the standard coordinates of the pattern, whether the input time series pattern belongs to a known pattern or the input time series pattern is uncertain;

[0016] Step 4: Visualize the detection results of time series patterns through the detection results and visualization interpretation module.

[0017] Furthermore, the preprocessing of step 1 includes outlier removal, equal-interval interpolation and normalization operations.

[0018] Furthermore, the specific process of step 2 is:

[0019] Step 2.1, set the upper and lower bounds of the number of states in state recognition and the initial similarity threshold, initialize the initial position of the state center of each telemetry data, and define the maximum number of iterations;

[0020] Step 2.2: Assign each piece of data to the closest telemetry data state center according to the currently set similarity threshold. Adjust the position of the telemetry data state center in a timely manner after each data assignment until all data are processed and the initial state recognition result is generated.

[0021] Step 2.3: Determine whether the number of states in the first state recognition result is within the preset upper and lower bounds of the state number. If so, proceed to step 2.4; otherwise, adjust the similarity threshold and return to step 2.2 to perform the first state recognition again.

[0022] Step 2.4: Under the premise of fixing the current telemetry data state center, redistribute all data points to the telemetry data state center that is closest to it. After the redistribution is completed, calculate the mean of all data points in each state, update the telemetry data state center, and record the current iteration number.

[0023] Step 2.5: Compare the telemetry data state center of the current iteration result with that of the previous iteration result to see if there is any change. If the telemetry data state center no longer changes or the number of iterations reaches the preset maximum value, then the state recognition ends; otherwise, return to step 2.4 and continue to adjust the state recognition result;

[0024] Step 2.6: Return the final state recognition result that meets the upper and lower bounds of the set state number and the state recognition result is stable.

[0025] Furthermore, the temporal convolutional network of step 3 includes four one-dimensional causal convolutional layers, each of which uses a ReLU activation function, wherein: the first layer c1 is set with 64 channels, a convolution kernel size of 3, and a dilation rate of d=1; the second layer c2 is set with 128 channels, a convolution kernel size of 3, and a dilation rate of d=2; the third layer c3 is set with D1=8 channels, a convolution kernel size of 3, and a dilation rate of 4; the fourth layer c4 is set with D2=8 channels, a convolution kernel size of 3, and a dilation rate of d=8;

[0026] The second layer c2, the third layer c3 and the fourth layer c4 all have residual connections, and the outputs of the third layer c3 and the fourth layer c4 are tensors of size D1×L and D2×L respectively, where L is the number of time steps.

[0027] Furthermore, in step 3, the satellite telemetry data anomaly detection module based on the temporal convolutional network processes the input final state recognition result as follows:

[0028] Step 3.1: Set the initial core threshold T corresponding to each standard mode n and the rejection threshold T c , T n and T c Adaptively update as the training progresses;

[0029] Step 3.2: The feature extraction module uses a temporal convolutional network to perform deep feature extraction on the input time series.

[0030] Step 3.3: The extracted deep features are passed through two fully connected layers, fc1 and fc2, and the output vector of the fc2 layer is split into two paths, which are sent to the x-axis fully connected layer fc3 and the y-axis fully connected layer fc4 respectively.

[0031] The fully connected layer fc3 outputs the probability coordinate x value of the time series pattern on each known category, and the probability coordinate x values ​​of all categories constitute a probability coordinate x vector, which is expressed as:

[0032]

[0033] Where k∈K, k=1, 2, 3..., K represents the number of known categories, (i) represents the i-th window;

[0034] The fully connected layer fc4 outputs the probability coordinate y value of the time series pattern on each known category. The probability coordinate y values ​​of all categories constitute a probability coordinate y vector, which is expressed as:

[0035]

[0036] The probability coordinates of the time series pattern under K known categories are represented by the probability coordinate x vector and the probability coordinate y vector, thereby converting the input time series pattern into feature coordinates in the feature space;

[0037] Step 3.4: Combine the probability coordinate x vector and the probability coordinate y vector into a feature coordinate vector and input it into the judgment recognition module. Based on the Euclidean distance between the feature coordinates of the input time series pattern and the standard coordinates of the known pattern, it is determined whether the input time series pattern belongs to a known pattern or is uncertain. Specifically:

[0038] 1) When the Euclidean distance is less than the core threshold T n When , the input time series pattern is determined to belong to the known pattern;

[0039] 2) When the Euclidean distance is greater than the negative threshold T c When , the input time series pattern is determined not to belong to the known pattern;

[0040] 3) When the Euclidean distance is at the core threshold T nand the rejection threshold T c When , the input time series pattern is uncertain.

[0041] Furthermore, the calculation formula of the Euclidean distance in step 3.4 is expressed as:

[0042]

[0043] Where, d A represents the Euclidean distance between the standard coordinates of the known pattern A and the characteristic coordinates of the input time series pattern, x A The x-axis coordinate of the time series pattern of the input obtained by the discriminant recognition module, A The y-axis coordinate represents the time series pattern of the input obtained by the discriminative recognition module.

[0044] Furthermore, the specific process of step 4 is: through the detection results and the visual interpretation module, the characteristic coordinates of each state center, the input time series and the circular boundary with the negation threshold and the core threshold as the radius are drawn, so as to intuitively display the spatial relationship between the feature points and the state centers of each mode.

[0045] The present invention has the following beneficial technical effects:

[0046] 1) The present invention combines adaptive state recognition with time series characteristic modeling to organically integrate the state recognition process with the anomaly detection process, thereby achieving efficient state recognition and accurate anomaly detection of satellite telemetry data. Specifically, the satellite telemetry data state recognition module based on adaptive state recognition dynamically adjusts the state number range and similarity threshold to autonomously optimize the state recognition results, accurately grouping high-dimensional, nonlinear time series data and extracting characteristic patterns of known states and some abnormal states; the satellite telemetry data anomaly detection module based on time convolutional network is the core of anomaly detection, which uses time convolutional network to extract deep features and map the input sequence to The coordinate points in the feature space are combined with the Euclidean distance and dynamic threshold mechanism to determine the status of satellite data; through the synergistic effect of the above two modules, not only can the normal and abnormal states of known operations be accurately identified, but also unknown abnormal states can be effectively detected, which significantly improves the analysis efficiency of satellite telemetry data and the accuracy of anomaly detection; finally, through the detection results and visualization interpretation module, the output detection results are visualized, and the state trend of the current satellite telemetry data and the distance between the coordinates and the state center are intuitively presented, which helps to quickly diagnose and make decisions, improves the comprehensibility of the detection process and the traceability of the reasoning process, and greatly improves the accuracy and interpretability of data analysis. In short, when the present invention is used for the analysis of complex telemetry data of single variables and multiple variables, it relies on strong real-time performance, good model interpretability, excellent robustness and excellent scalability to significantly improve data utilization and the efficiency and accuracy of satellite telemetry data analysis, and provides strong technical support for satellite on-orbit operation status monitoring and anomaly detection.

[0047] 2) The present invention automatically learns the potential distribution of data through unsupervised learning and can perform classification without human intervention, forming a fully automatic detection process that does not require manual labeling. This not only reduces labor costs but also adapts to complex and changeable actual environments. It is particularly suitable for satellite telemetry tasks with huge data volumes and difficult-to-enumerate abnormal patterns, and has higher practicality and reliability in actual applications.

[0048] 3) The present invention breaks through the limitation of traditional methods that can only identify known abnormal patterns. It can not only accurately distinguish between known normal and abnormal states, but also realize the autonomous identification and detection of unknown abnormal patterns, thereby improving the intelligence and adaptability of the system. It can effectively deal with sudden and unrecorded abnormal situations that may occur during satellite operation, thereby meeting the needs of autonomous monitoring in complex space environments.

[0049] 4) Different from the traditional black-box detection method, the present invention uses a visualization module to draw the Euclidean distance between the abnormal point and the center of the normal state to intuitively quantify the degree of abnormality, making the detection results more transparent and credible, and giving the abnormal detection results powerful interpretability analysis.

[0050] 5) Compared with existing methods that only rely on statistical rules or fixed thresholds, the present invention can adaptively learn the distribution characteristics of satellite telemetry data and dynamically adjust the anomaly discrimination criteria, thereby more accurately identifying abnormal states. It is particularly suitable for high-dimensional and complex satellite telemetry data analysis.

[0051] 6) Compared with traditional classification methods, the adaptive satellite telemetry data state identification method proposed in the present invention has dynamic adjustment capabilities. It can adaptively optimize the state identification results by dynamically adjusting the state number range and similarity threshold. There is no need to manually set a fixed state number, which is more in line with the actual satellite telemetry data distribution.

[0052] 7) Unlike traditional methods that rely on fixed thresholds or single-layer discrimination, the satellite telemetry data anomaly detection method based on the temporal convolutional network of the present invention adopts a two-layer judgment mechanism with dynamic thresholds. It can not only accurately detect known abnormal states, but also adaptively identify unknown abnormal patterns, improving the adaptability and detection efficiency of complex satellite measurement data. It is more suitable for dynamic environmental changes in satellite systems, can effectively capture subtle anomalies, and achieve more efficient and robust anomaly detection, providing strong technical support for satellite on-orbit operation status monitoring and anomaly detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a flow chart of the present invention;

[0054] Figure 2 Schematic diagram of the structure of the satellite telemetry data state recognition module based on adaptive state recognition of the present invention;

[0055] Figure 3 This is a schematic diagram of the structure of a satellite telemetry data anomaly detection module based on a temporal convolutional network according to the present invention;

[0056] Figure 4 The result of the satellite telemetry data anomaly detection module based on the temporal convolutional network of the present invention is a visualization result of a known abnormal state;

[0057] Figure 5 The result of the satellite telemetry data anomaly detection module based on the temporal convolutional network of the present invention is a visualization result of the first known normal state;

[0058] Figure 6 The result of the satellite telemetry data anomaly detection module based on the temporal convolutional network of the present invention is a visualization result of the second known normal state;

[0059] Figure 7 The visualization result of the satellite telemetry data anomaly detection module based on temporal convolutional network is the unknown anomaly state. DETAILED DESCRIPTION

[0060] The present invention will be further described in detail below with reference to specific embodiments, which are intended to explain the present invention rather than to limit it.

[0061] refer to Figure 1 As shown, a satellite telemetry data anomaly detection method for automatically distinguishing data states includes the following steps:

[0062] Step 1: Preprocess the original telemetry data to obtain time series data, and divide the time series data into windows of fixed length;

[0063] The preprocessing includes outlier removal, equal-interval interpolation, and normalization. The outlier removal is used to remove outliers introduced by data acquisition equipment failure or environmental interference; equal-interval interpolation is used to fill gaps in the data to make the time series continuous and complete; and normalization is used to adjust the data's numerical range to a standardized interval, eliminating the impact of dimensional differences between different variables and making subsequent processing results more stable and reliable.

[0064] The time series data can better reflect the overall trend and local changes of the time series. The time series data is divided into windows of fixed length, i.e., subsequences. The length of the window is set according to application requirements and data characteristics. Each window must retain sufficient time information while maintaining the computational efficiency of data processing. When extracting features from the data in each window, the extracted feature types include the mean, standard deviation, peak value, and frequency component of the time series. These features can fully describe the statistical characteristics and dynamic behavior of the data in the window and constitute the feature vector corresponding to each window.

[0065] Step 2: Build Figure 2 The satellite telemetry data state recognition module based on adaptive state recognition shown in the figure dynamically divides the feature vectors of all windows and outputs the final state recognition result;

[0066] The processing process of the input data of the satellite telemetry data state recognition module based on adaptive state recognition is divided into an initialization stage and a state recognition learning stage, wherein: in the initialization stage, it dynamically adjusts the number of states by adaptively setting the initial similarity threshold, so that the state recognition result can accurately reflect the intrinsic distribution of the data and prevent the problem of too many or too few states; in the state recognition learning stage, feature characters are generated by extracting window features and state recognition. Feature characters are a summary and abstraction of the original data features, which convert complex time series data into a discrete representation form. This representation form can better handle the fuzzy matching problem existing in telemetry data and improve the interpretability of state recognition results. In order to further explore the abnormal state morphological characteristics in telemetry data, a variety of feature extraction methods are combined to form a comprehensive feature vector for in-depth analysis. For example, a key frame feature representation learning model based on mean absolute error (MAE) is used to extract key frame features in the time series, and then these features are subjected to state recognition to obtain discrete representations, thereby describing the characteristics and evolution laws of the key frames from multiple angles and levels.

[0067] The satellite telemetry data state recognition module based on adaptive state recognition dynamically divides the feature vectors of all windows and outputs the final state recognition result. The specific process is as follows:

[0068] Step 2.1. Initialize the parameters of the satellite telemetry data state recognition module based on adaptive state recognition: set the upper and lower bounds of the number of states in state recognition and the initial similarity threshold, initialize the initial position of each telemetry data state center, and define the maximum number of iterations to control the state recognition process;

[0069] Step 2.2, first state recognition: According to the currently set similarity threshold, each data is assigned to the closest telemetry data state center one by one. After each data assignment, the position of the telemetry data state center is adjusted in time until all data is processed and the initial state recognition result is generated;

[0070] Step 2.3, check the state number range: determine whether the state number in the first state recognition result is within the preset upper and lower limits of the state number. If so, proceed to step 2.4; otherwise, adjust the similarity threshold and return to step 2.2 to re-perform the first state recognition;

[0071] Typically, the similarity threshold is adjusted by decreasing it to increase the number of states;

[0072] Step 2.4. Adjust the state recognition results: Under the premise of fixing the current telemetry data state center, redistribute all data points to the telemetry data state center that is closest to it. After the redistribution is completed, calculate the mean of all data points in each state, update the telemetry data state center, and record the current iteration number.

[0073] Step 2.5: Determine the stability of the state recognition result: Compare the telemetry data state center of the current iteration result with the previous iteration result to see if there is any change. If the telemetry data state center no longer changes or the number of iterations reaches the preset maximum value, then end the state recognition; otherwise, return to step 2.4 and continue to adjust the state recognition result;

[0074] When the state center of the telemetry data no longer changes, it means that the state recognition result tends to be stable;

[0075] Step 2.6: Output the final state recognition result: Return the final state recognition result that meets the upper and lower bounds of the set state number and is stable, that is, the satellite telemetry data with the recognized state, including the final recognition state center and data distribution;

[0076] The satellite telemetry data with identified states is a time window with classification labels, which can be a multi-dimensional time series or a single-dimensional time series. The final state identification result is used to accurately describe the distribution structure of the telemetry data and dynamically model it.

[0077] Step 3: Build Figure 3 The satellite telemetry data anomaly detection module based on the temporal convolutional network shown in the figure takes the final state recognition result as input and outputs the detection result of the time series pattern;

[0078] The construction process of the satellite telemetry data anomaly detection module based on the temporal convolutional network is as follows: the spatial embedding method is integrated with the temporal convolutional neural network to optimize the structure of the traditional temporal convolutional neural network and construct a network model with both temporal feature learning and pattern recognition capabilities. That is, the satellite telemetry data anomaly detection module based on the temporal convolutional network includes a feature extraction module and a judgment and recognition module.

[0079] The feature extraction module extracts time series features through multiple convolutional layers and pooling layers. In the feature extraction stage, the convolutional layer adopts a combination of causal convolution and dilated convolution to maintain the causal relationship of the time series while expanding the receptive field and effectively capturing long-term dependency information. The introduction of residual connections further alleviates the gradient vanishing problem and improves the stability of feature extraction. The pooling layer is used as a dimensionality reduction method, and maximum pooling is used to extract key patterns. Global average pooling enhances the robustness of feature expression while reducing computational complexity.

[0080] The judgment and recognition module uses the fully connected layer and the output layer to complete the state judgment. In the judgment and recognition stage, the time series pattern is mapped to the feature space, converted into coordinate points using the spatial embedding method, and the Euclidean distance between it and the standard pattern is calculated. Anomaly detection and state recognition are performed based on the calculated distance;

[0081] The determination and recognition module adopts n classification to determine n known patterns. The deep convolutional neural network gives n probability coordinates composed of probability values, corresponding to the feature coordinates of n types of known patterns, and then calculates the Euclidean distance between the feature coordinates and the standard coordinates of each category (i.e., corresponding to the known pattern). The detection result of the input time series pattern is determined according to the size of the distance value, wherein the input of the network model is a time series pattern, i.e., a time series of fixed length, and the output results are divided into two cases, namely: the input time series pattern is determined to belong to a certain known pattern and the input time series pattern is uncertain;

[0082] The satellite telemetry data anomaly detection module based on the temporal convolutional network takes the satellite telemetry data with identified states as input and outputs the detection results of the time series pattern as follows:

[0083] Step 3.1. Initialize the parameters of the satellite telemetry data anomaly detection module based on the temporal convolutional network: set the initial core threshold T corresponding to each standard mode n and the rejection threshold T c , T n and T c Adaptively update as the training progresses;

[0084] Step 3.2, feature extraction: Use the temporal convolutional network to perform deep feature extraction on the input time series. The temporal convolutional network consists of four layers of one-dimensional causal convolutional layers. Each one-dimensional causal convolutional layer uses the ReLU activation function, where: the first layer c1 sets 64 channels, the convolution kernel size is 3, and the expansion rate is d=1; the second layer c2 sets 128 channels, the convolution kernel size is 3, and the expansion rate is d=2; the third layer c3 sets D1=8 channels, the convolution kernel size is 3, and the expansion rate is 4; the fourth layer c4 sets D2=8 channels, the convolution kernel size is 3, and the expansion rate is d=8; the c2, c3 and c4 layers all have residual connections, and the outputs of c3 and c4 are tensors of size D1×L and D2×L respectively, where L is the number of time steps. Each channel of each one-dimensional causal convolutional layer represents the time-varying characteristics of the object in a certain potential state;

[0085] Step 3.3, get coordinates: The extracted deep features pass through two fully connected layers fc1 and fc2, and then the output vector of the fc2 layer is split into two paths, which are sent to the x-axis fully connected layer fc3 and the y-axis fully connected layer fc4 respectively;

[0086] The fc3 layer outputs the probability coordinate x value of the time series pattern on each known category, and the probability coordinate x values ​​of all categories constitute a probability coordinate x vector, which is expressed as:

[0087]

[0088] Where k∈K, k=1, 2, 3..., K represents the number of known categories, (i) represents the i-th window;

[0089] The fc4 layer outputs the probability coordinate y value of the time series pattern on each known category. The probability coordinate y values ​​of all categories constitute a probability coordinate y vector, which is expressed as:

[0090]

[0091] The probability coordinate x vector and the probability coordinate y vector are used to represent the "probability coordinates" of the time series pattern under K known categories, thereby flattening the traditional output probability value into probability coordinates, that is, converting an input time series pattern into a coordinate point in the feature space, namely, the feature coordinates;

[0092] Step 3.4, distance calculation and judgment: Combine the probability coordinate x vector and the probability coordinate y vector into a feature coordinate vector and input it into the judgment recognition module. Based on the Euclidean distance between the feature coordinates of the input time series pattern and the standard coordinates of the pattern, determine whether the input time series pattern belongs to a known pattern;

[0093] For a known pattern A, the Euclidean distance d between its characteristic coordinates and the input time series pattern is A Expressed as:

[0094]

[0095] Where x A The x-axis coordinate of the time series pattern of the input obtained by the discriminant recognition module, A The y-axis coordinate of the time series pattern of the input obtained by the discriminant recognition module;

[0096] When the feature coordinates fall on the core of the known pattern A, that is, the distance d A Less than the core threshold T n When , the input time series pattern is determined to belong to pattern A; if the feature coordinates are far away from the known pattern A, that is, the distance d A Greater than the negative threshold T c When , the input time series pattern is determined not to belong to pattern A; if the distance between the feature coordinates and the known pattern A is between the core threshold and the negative threshold, that is, T n <d A <Tc When , the input time series pattern is uncertain;

[0097] Step 4: Visualize the detection results of the time series pattern through the detection results and visualization interpretation module. The specific process is as follows:

[0098] like Figures 4 to 7 As shown in the figure, when visualizing the detection results and visualization interpretation module, the center of each state, the characteristic coordinates of the input time series, and the circular boundary with the negative threshold and the core threshold as the radius are drawn to intuitively show the spatial relationship between the feature points and the center of each mode state;

[0099] If the characteristic coordinates of a time series fall within the core threshold circle of a known pattern center, the time series is judged to belong to the known pattern; if the characteristic coordinates of a time series are outside the negative threshold circle of all known patterns centers, the time series is judged to be an unknown anomaly; if the characteristic coordinates are between the core threshold and the negative threshold, it is classified or marked as pending according to the nearest neighbor strategy or set rules;

[0100] Through the visualization results, users can intuitively understand the classification basis and degree of anomaly of each time series sample, which enhances the interpretability and verifiability of the anomaly detection results. At the same time, it is also of great significance for evaluating the discrimination between different patterns, adjusting threshold setting strategies, and discovering potential new anomaly categories, thereby improving the accuracy, robustness, and practicality of the anomaly detection system.

[0101] Compared with traditional methods, this embodiment has significant advantages in satellite telemetry data anomaly detection and state identification. Through the adaptive satellite telemetry data state identification method, it can effectively process high-dimensional, multivariate telemetry data, automatically adapt the number of states, and accurately identify small-scale anomaly patterns without manual labeling. At the same time, the satellite telemetry data anomaly detection module based on the temporal convolutional network successfully combines the spatial and temporal characteristics of telemetry data, and can automatically and efficiently detect unknown anomaly states based on known patterns. In addition, this embodiment combines visualization analysis based on Euclidean distance to visualize the output anomaly detection results, which can intuitively present the degree of deviation between the anomaly points and the normal state distribution, thereby clearly revealing the basis for anomaly judgment, improving the comprehensibility of the detection process and the traceability of the reasoning process, and significantly improving the accuracy and interpretability of data analysis. In short, this embodiment improves the robustness and adaptability of the detection results, provides reliable technical support for real-time monitoring of the satellite's on-orbit operation status and potential fault identification, and significantly reduces the need for manual intervention.

[0102] The satellite telemetry data anomaly detection method for automatically distinguishing data states proposed in this embodiment focuses on the dynamic adjustment of the state center and similarity threshold in the early stage. During the first state identification process, the state center is updated in real time for each new data point. In subsequent adjustments, the state center remains unchanged until all data points are redistributed, at which time the center is updated. This improves the dynamic adaptability and global optimization capabilities of the state identification results, thereby better adapting to the distribution characteristics of actual data. The core of the later stage is to achieve efficient classification and anomaly detection of the satellite's on-orbit operation state through deep extraction of time series features and precise calculation of high-dimensional feature coordinates. It not only uses the early state identification results to accurately detect known operation states, but also successfully identifies unknown abnormal states through a two-layer judgment strategy of adaptive thresholds. The combination of core threshold and negation threshold can not only improve the high-precision judgment of known states, but also effectively expand to the identification of unknown abnormal states, providing a more intelligent and comprehensive solution for satellite telemetry data anomaly detection.

Claims

1. A satellite telemetry data anomaly detection method that automatically distinguishes data states, characterized in that: The steps include: Step 1: Preprocess the original telemetry data to obtain time series data, and divide the time series data into windows of fixed length; Step 2: Dynamically divide the feature vectors of all windows using the satellite telemetry data state recognition module based on adaptive state recognition, and output the final state recognition result; The process of dividing the feature vectors of all input windows by the satellite telemetry data state recognition module based on adaptive state recognition is divided into an initialization phase and a state recognition learning phase, wherein: in the initialization phase, the number of states is dynamically adjusted by adaptively setting the initial similarity threshold; in the state recognition learning phase, feature characters are generated by extracting window features and state recognition, thereby converting time series data into a discrete representation form, and outputting a state recognition result, which is a time series; Step 3: Construct a satellite telemetry data anomaly detection module based on a temporal convolutional network, which takes the final state recognition result as input and outputs the detection result of the time series pattern; The satellite telemetry data anomaly detection module based on the temporal convolutional network is a network model that combines temporal feature learning and pattern recognition capabilities by integrating the spatial embedding method on the basis of the temporal convolutional neural network. It includes a feature extraction module and a judgment and recognition module. The feature extraction module extracts temporal features through multiple convolutional layers and pooling layers; the judgment and recognition module completes state judgment with the help of the fully connected layer and the output layer. The processing process of the satellite telemetry data anomaly detection module based on the time convolutional network on the final state recognition result of the input is as follows: the feature extraction module performs deep feature extraction on the input time series through the time convolutional network, the extracted deep features first pass through two fully connected layers, the vector output by the last fully connected layer is divided into two paths, and respectively transmitted to the x-axis fully connected layer and the y-axis fully connected layer to obtain the probability coordinate x vector and the probability coordinate y vector, and the probability coordinate x vector and the probability coordinate y vector are combined into a feature coordinate vector and input into the judgment and recognition module, and the input time series pattern is judged according to the Euclidean distance between the feature coordinates of the input time series pattern and the standard coordinates of the pattern, whether the input time series pattern belongs to a known pattern or the input time series pattern is uncertain; Step 4: Visualize the detection results of time series patterns through the detection results and visualization interpretation module.

2. The satellite telemetry data anomaly detection method for automatically distinguishing data states according to claim 1 is characterized in that: The preprocessing of step 1 includes outlier removal, equal-interval interpolation and normalization.

3. The satellite telemetry data anomaly detection method for automatically distinguishing data states according to claim 1, characterized in that: The specific process of step 2 is: Step 2.1, set the upper and lower bounds of the number of states in state recognition and the initial similarity threshold, initialize the initial position of the state center of each telemetry data, and define the maximum number of iterations; Step 2.2: Assign each piece of data to the closest telemetry data state center according to the currently set similarity threshold. Adjust the position of the telemetry data state center in a timely manner after each data assignment until all data are processed and the initial state recognition result is generated. Step 2.3: Determine whether the number of states in the first state recognition result is within the upper and lower bounds of the preset number of states. If yes, proceed to step 2.

4. Otherwise, adjust the similarity threshold and return to step 2.2 to perform the first state recognition again; Step 2.4: Under the premise of fixing the current telemetry data state center, redistribute all data points to the telemetry data state center that is closest to it. After the redistribution is completed, calculate the mean of all data points in each state, update the telemetry data state center, and record the current iteration number. Step 2.5: Compare the telemetry data state center of the current iteration result with that of the previous iteration result to see if there is any change. If the telemetry data state center no longer changes or the number of iterations reaches the preset maximum value, then the state recognition ends; otherwise, return to step 2.4 and continue to adjust the state recognition result; Step 2.6: Return the final state recognition result that meets the upper and lower bounds of the set state number and the state recognition result is stable.

4. The satellite telemetry data anomaly detection method for automatically distinguishing data states according to claim 1, characterized in that: The temporal convolutional network of step 3 includes four one-dimensional causal convolutional layers, each of which uses the ReLU activation function, wherein: the first layer c1 is set with 64 channels, a convolution kernel size of 3, and a dilation rate of d=1; the second layer c2 is set with 128 channels, a convolution kernel size of 3, and a dilation rate of d=2; the third layer c3 is set with D1=8 channels, a convolution kernel size of 3, and a dilation rate of 4; the fourth layer c4 is set with D2=8 channels, a convolution kernel size of 3, and a dilation rate of d=8; The second layer c2, the third layer c3 and the fourth layer c4 all have residual connections, and the outputs of the third layer c3 and the fourth layer c4 are tensors of size D1×L and D2×L respectively, where L is the number of time steps.

5. The satellite telemetry data anomaly detection method for automatically distinguishing data states according to claim 1, characterized in that: The processing process of the satellite telemetry data anomaly detection module based on the temporal convolutional network on the input final state recognition result in step 3 is as follows: Step 3.1: Set the initial core threshold T corresponding to each standard mode n and the rejection threshold T c , T n and T c Adaptively update as the training progresses; Step 3.2: The feature extraction module uses a temporal convolutional network to perform deep feature extraction on the input time series. Step 3.3: The extracted deep features are passed through two fully connected layers, fc1 and fc2, and the output vector of the fc2 layer is split into two paths, which are sent to the x-axis fully connected layer fc3 and the y-axis fully connected layer fc4 respectively. The fully connected layer fc3 outputs the probability coordinate x value of the time series pattern on each known category, and the probability coordinate x values ​​of all categories constitute a probability coordinate x vector, which is expressed as: Where k∈K, k=1, 2, 3..., K represents the number of known categories, (i) represents the i-th window; The fully connected layer fc4 outputs the probability coordinate y value of the time series pattern on each known category. The probability coordinate y values ​​of all categories constitute a probability coordinate y vector, which is expressed as: The probability coordinates of the time series pattern under K known categories are represented by the probability coordinate x vector and the probability coordinate y vector, thereby converting the input time series pattern into feature coordinates in the feature space; Step 3.4: Combine the probability coordinate x vector and the probability coordinate y vector into a feature coordinate vector and input it into the judgment recognition module. Based on the Euclidean distance between the feature coordinates of the input time series pattern and the standard coordinates of the known pattern, it is determined whether the input time series pattern belongs to a known pattern or is uncertain. Specifically: 1) When the Euclidean distance is less than the core threshold T n When , the input time series pattern is determined to belong to the known pattern; 2) When the Euclidean distance is greater than the negative threshold T c When , the input time series pattern is determined not to belong to the known pattern; 3) When the Euclidean distance is at the core threshold T n and the rejection threshold T c When , the input time series pattern is uncertain.

6. The satellite telemetry data anomaly detection method for automatically distinguishing data states according to claim 5, characterized in that: The calculation formula of the Euclidean distance in step 3.4 is expressed as: Where, d A represents the Euclidean distance between the standard coordinates of the known pattern A and the characteristic coordinates of the input time series pattern, x A The x-axis coordinate of the time series pattern of the input obtained by the discriminant recognition module, A The y-axis coordinate represents the time series pattern of the input obtained by the discriminative recognition module.

7. The satellite telemetry data anomaly detection method for automatically distinguishing data states according to claim 1, characterized in that: The specific process of step 4 is: through the detection results and the visual interpretation module, the characteristic coordinates of each state center, the input time series, and the circular boundary with the negative threshold and the core threshold as the radius are drawn, so as to intuitively display the spatial relationship between the feature points and the state center of each mode.

Images