Abnormality detection method for power equipment system based on spatiotemporal feature segmentation reconstruction
By employing a spatiotemporal feature segmentation and reconstruction method, this paper utilizes temporal convolution and graph convolution to process multi-source data. Combining adaptive graph structure and prior knowledge of equipment, an unsupervised multi-source temporal graph convolutional network is constructed, which solves the problem of insufficient multi-source data fusion in power equipment systems and achieves efficient anomaly detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XI AN JIAOTONG UNIV
- Filing Date
- 2022-08-17
- Publication Date
- 2026-06-05
AI Technical Summary
Existing technologies lack the ability to extract and fuse features from multi-source monitoring data in power equipment systems, resulting in inaccurate anomaly detection results, especially in the absence of abnormal samples, making it difficult to effectively identify abnormal system states.
A spatiotemporal feature-based segmentation and reconstruction method is adopted. Multi-source data is processed through temporal convolution and graph convolution. An unsupervised multi-source temporal graph convolutional network is constructed by combining adaptive graph structure and device prior knowledge to extract spatiotemporal features and perform segmentation and reconstruction to improve the accuracy of anomaly detection.
It enables accurate identification of abnormal states in power equipment systems under conditions without abnormal samples, improves the multi-source data fusion capability and the efficiency of anomaly detection, and enhances the accuracy and sensitivity of detection results.
Smart Images
Figure CN115293280B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power equipment anomaly detection, and in particular to a power equipment system anomaly detection method based on spatiotemporal feature segmentation and reconstruction. Background Technology
[0002] Power equipment is the power source for the operation of critical equipment. Abnormalities in power equipment can affect the normal operation of the entire system, causing economic losses and even casualties. Condition monitoring of critical power equipment systems typically involves acquiring data from multiple sensors of various types. Analyzing the acquired monitoring data allows for anomaly detection and monitoring of the equipment's operational health. However, the complex characteristics, information redundancy, and nonlinear correlations of multi-source monitoring data pose challenges to multi-source data fusion. Therefore, research on methods for detecting abnormal system states using multi-source monitoring data is urgently needed.
[0003] Traditional anomaly detection algorithms, primarily based on machine learning methods such as Support Vector Machines, Isolation Forests, Singular Value Decomposition, and Principal Component Analysis, suffer from poor feature extraction and fusion capabilities when processing multi-source monitoring data. They fail to comprehensively utilize multi-source data information, making it difficult to fully characterize the health status of power equipment systems and resulting in lower accuracy in anomaly detection. Furthermore, in practical engineering scenarios, data on abnormal equipment states is scarce, and due to the uncertainty of the causes of anomalies, it is difficult to simulate abnormal system states and collect abnormal data. With the development and widespread application of deep learning, data-driven multi-source data fusion algorithms and unsupervised learning paradigms have made significant progress in the field of intelligent identification of equipment health status. Therefore, researching an intelligent anomaly detection method that can effectively fuse multi-source monitoring data even in the absence of abnormal samples is of great value in solving practical problems. Summary of the Invention
[0004] The purpose of this invention is to provide an anomaly detection method for power equipment systems based on spatiotemporal feature segmentation and reconstruction, in order to solve the problems existing in the prior art. This invention processes multi-source data and extracts its spatiotemporal features based on temporal convolution and graph convolution. It combines adaptive graph structure and prior knowledge of equipment objects to model the topological relationships between multi-source data, promoting the fusion of multi-source data. By utilizing the segmented spatiotemporal features, segmented reconstruction is performed on a convolutional autoencoder to improve the accuracy of anomaly detection results, and finally realize the identification of abnormal system states.
[0005] To achieve the above objectives, the present invention adopts the following technical solution:
[0006] The anomaly detection method for power equipment systems based on spatiotemporal feature segmentation and reconstruction includes the following steps:
[0007] Step 1: Obtain multi-source monitoring data from the equipment status monitoring platform, cut the multi-source monitoring data to a specific length to obtain multi-source time series samples, preprocess each multi-source time series sample to obtain a multi-source dataset, and divide it into training set and test set;
[0008] Step 2: Based on the prior knowledge of the power equipment's own structure, obtain the prior graph structure from the multi-source time series samples obtained in Step 1, and construct an adaptive graph structure learning module to output the adaptive graph structure.
[0009] Step 3: Based on the prior graph structure and the adaptive graph structure, construct an unsupervised multi-source temporal graph convolutional network model for system anomaly detection;
[0010] Step 4: Optimize the unsupervised multi-source temporal graph convolutional network model constructed in Step 3 using the training set from Step 1;
[0011] Step 5: Use the network model optimized in Step 4 to evaluate the system state of the training set, process the obtained state scores, and calculate the threshold.
[0012] Step 6: Use the network model trained in Step 4 to evaluate the system state of the test set, and output the abnormal state detection results based on the threshold calculated in Step 5.
[0013] Furthermore, the specific method of truncating to a specific length in step 1 is as follows: the multi-source monitoring data are divided into multi-source time series samples with a specific length L without overlap.
[0014] In step 1, the preprocessing uses a normalization operation, specifically: normalizing each channel of the multi-source time series samples individually, and dividing the multi-source dataset according to the source of the multi-source monitoring data; selecting a portion of the samples with normal system status as the training set, and using the remaining normal samples and samples with abnormal system status as the test set; the normalization formula is as follows:
[0015]
[0016] In the formula, For the normalized data, x ij Let x be the j-th data point in the i-th channel, and μ be x. i The mean of each data point, σ is x i Standard deviation of each data point, x i This represents the data for the i-th channel.
[0017] Furthermore, in step 2, a priori graph structure is established based on prior knowledge of the power equipment's own structure. This priori graph structure includes the adjacency matrix of the graph structure and the feature matrix of the nodes. For any input multi-source time series sample... n is the number of sample channels, and l is the sample length. Let the set of real numbers be represented as a prior graph structure sample. Node feature matrix Adjacency Matrix The specific method for constructing the adjacency matrix of the prior graph structure is as follows:
[0018] First, the locations of the sensors from each channel of the multi-source monitoring data on the power equipment are clearly identified. Second, the values of the adjacency matrix are determined based on the connection relationships between the structures containing the sensors, using a... ij =0 indicates that there is no connection between the location of the i-th channel sensor and the location of the j-th channel sensor, a ij =1 indicates the existence of a connection; obtain the value of the adjacency matrix:
[0019] A p =[a ij ].
[0020] Furthermore, in the adaptive graph structure learning module constructed in step 2, the probability of edges existing between nodes is parameterized. An adaptive graph structure is then learned through network learning to model the intrinsic relationships between multi-source data. For any input multi-source time series sample... n is the number of sample channels, l is the sample length, and its adaptive graph structure is represented as follows: Node feature matrix Adjacency Matrix The process of constructing the adjacency matrix of an adaptive graph structure includes: parameterizing the probability that edges exist between nodes. The edges between nodes are sampled using the Gumbel-softmax technique, and the state variables are used. Representing the states with and without edges respectively, we obtain the adjacency matrix A of the adaptive graph structure. a The specific construction process is as follows:
[0021] First, the probability that there are edges between parameterized nodes:
[0022]
[0023] Where N is the number of nodes, i.e. the number of multi-source data channels; Let be the probability that an edge exists. This represents the probability that no edge exists.
[0024] Secondly, sampling noise from a uniform distribution Calculate the corresponding Gumbel noise
[0025]
[0026]
[0027] In the formula, For uniformly distributed sampling noise, For the corresponding Gumbel noise;
[0028] Finally, according to Calculate the state variables of the edges
[0029]
[0030] In the formula, For uniformly distributed sampling noise, For the corresponding The Gumbel noise, where τ is a temperature parameter used to control the smoothness of the sampling;
[0031] Then, the adjacency matrix of the adaptive graph structure is obtained:
[0032]
[0033] A a =[a ij ].
[0034] Furthermore, the unsupervised multi-source temporal graph convolutional network model constructed in step 3 consists of two modules: the first is a spatiotemporal feature extraction network module, and the second is a segmentation and reconstruction anomaly detection module. The multi-source time series samples, along with the prior graph structure and adaptive graph structure in step 2, are input into the spatiotemporal feature extraction network module to obtain the spatiotemporal features of the samples. The spatiotemporal features are then input into the segmentation and reconstruction anomaly detection module to obtain the anomaly state score of the samples, thereby achieving anomaly detection.
[0035] Furthermore, the spatiotemporal feature extraction network module consists of multiple spatiotemporal feature extraction layers connected by skip connections; the spatiotemporal feature extraction layers include three parts: temporal convolution, graph attention convolution based on prior graph structure, and graph attention convolution based on adaptive graph structure. Both the graph attention convolution based on prior graph structure and the graph attention convolution based on adaptive graph structure include two steps: attention calculation and state update.
[0036] The attention calculation process is as follows:
[0037]
[0038]
[0039] In the formula, To input the features of node i in the convolutional layer of this graph, Let N be the input node j features, ω be the trainable weight matrix, and N be the number of features. i Let e be the set of neighboring nodes of node i. ij Let α be the similarity of features between nodes i and j. ij The attention weights are calculated;
[0040] The status update process is as follows:
[0041]
[0042] In the formula, σ is the activation function. Let i be the feature of the input node i. This refers to the updated output features of node i.
[0043] Furthermore, the segmentation and reconstruction anomaly detection module includes two encoders and one decoder; the two encoders have identical structures, each consisting of a convolutional layer, a batch normalization layer, and an activation function, respectively, while the decoder consists of a transposed convolutional layer, a batch normalization layer, and an activation function; the spatiotemporal features are output by the spatiotemporal feature extraction network. After segmentation, we get h u ,h d Input two encoders to obtain the hidden vector hid u and hid d Then, the same decoder is used to decode and reconstruct the two hidden vectors. The process is formulated as follows:
[0044] hid u =E u (h u )
[0045] hid d =E d (h d )
[0046] out u =D(hid) u )
[0047] out d =D(hid) d )
[0048] In the formula, E u E d D represents two encoding networks and one decoding network, respectively, and h u The first half of the spatiotemporal characteristics, h d For the latter half of the spatiotemporal features, hid u and hid d The encoded hidden vector, out u and out dThis is the output after decoding and reconstruction.
[0049] Furthermore, in step 4, when optimizing the network model using the training set, the optimization targets are the parameters in the network model constructed in step 3 and the parameters of the adaptive graph structure learning module constructed in step 2. The optimization objective is to minimize the training loss, and the training loss calculation process is as follows:
[0050] First, the input multi-source time series samples Segmentation to obtain Calculate reconstruction loss
[0051]
[0052] In the formula, x u ,x d These are the two segments obtained after splitting the input multi-source time series sample, out. u ,out d These are the two parts of the output obtained from the reconstruction;
[0053] Next, the adaptive graph sparsity loss is calculated.
[0054]
[0055] In the formula, Let N be the probability that nodes i and j have edges in the adaptive graph structure, and N be the number of sensors.
[0056] Next, calculate the spatiotemporal feature distribution loss. That is, the two parts of the hidden vector after encoding (hid) u and hid d Maximum average difference between:
[0057]
[0058] In the formula, hid u ,hid d These are the two encoded hidden vectors, φ(·) is a mapping from the original space to the regenerated Hilbert space, and s is the batch size used during training.
[0059] Finally, the training loss is obtained.
[0060]
[0061] In the formula, λ1>0, λ2<0, and λ1 and λ2 are preset parameters. For reconstruction error, For adaptive graph sparsity loss, The loss is the spatiotemporal feature distribution.
[0062] Furthermore, in step 5, state evaluation is performed using the training set data to obtain an abnormal state score, and a threshold is calculated based on the training set state score. The abnormal state score and threshold calculation are as follows:
[0063] Abnormal status score:
[0064]
[0065] In the formula, λ2<0, and λ2 is a preset parameter. For reconstruction error, Loss is based on the spatiotemporal characteristic distribution.
[0066] The threshold values are as follows:
[0067]
[0068] In the formula, t is the threshold for the abnormal state score, and N S The number of samples used in the threshold calculation, score i The score is given for the abnormal state of the i-th sample.
[0069] Furthermore, the abnormal state detection results in step 6 are represented as follows:
[0070]
[0071] In the formula, X i For the detection sample, t is the threshold for the abnormal state score, and state(X) i ) represents the state of the sample, and score represents the score for the abnormal state of the sample.
[0072] Compared with the prior art, the present invention has the following beneficial technical effects:
[0073] 1) This invention uses graph neural networks to fuse multi-source data information, represents each channel of multi-source data as graph nodes, and constructs a graph structure by combining the internal correlation of data and prior information. It uses graph attention convolution to aggregate the information of each node, and comprehensively and effectively fuse the features of multi-source data.
[0074] 2) This invention combines temporal convolution and graph convolution to extract spatiotemporal features from multi-source data, and uses segmentation features combined with a convolutional autoencoder for reconstruction, forcibly expanding the distance between the hidden features after encoding, so as to enhance the model's ability to identify normal data and improve its sensitivity to abnormal data.
[0075] 3) The present invention proposes an intelligent detection method for power equipment systems based on spatiotemporal feature segmentation and reconstruction, which can accurately identify abnormal states of equipment systems under the condition of no abnormal samples, and effectively solves the problem of low detection efficiency of abnormal detection methods due to the difficulty in fusing multi-source data. Attached Figure Description
[0076] The accompanying drawings are provided to further understand the invention and constitute a part of this invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.
[0077] Figure 1 This is a flowchart of the method of the present invention;
[0078] Figure 2 This is a structural diagram of the model constructed by the method of the present invention. Detailed Implementation
[0079] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0080] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0081] A system anomaly detection method based on spatiotemporal feature segmentation and reconstruction, such as Figure 1 As shown, it includes the following steps:
[0082] Step 1: Obtain multi-source monitoring data from the equipment status monitoring platform, truncate the data to a specific length to obtain multi-source time series samples, preprocess each sample to obtain a multi-source dataset, and divide it into training and test sets. The preprocessing mainly involves standard score normalization, with the following formula:
[0083]
[0084] In the formula, For the normalized data, x i· For the data of the i-th channel, x ij Let x be the j-th data point in the i-th channel, and μ be x. i· The mean of each data point, σ is x i· Standard deviation of each data point.
[0085] Normalization can accelerate the convergence speed of subsequent model training, reduce the impact of excessive numerical differences on model training, and at the same time, normalization can preserve the information of each part of the data, so that values with excessively small values will not lose the information they contain.
[0086] Step 2: Based on the multi-source time series samples obtained in Step 1, and combined with prior knowledge of the device's own structure and expert experience, a prior graph structure is constructed to form connections between nodes, and an adaptive graph structure learning module is constructed to obtain the adaptive graph structure.
[0087] Prior graph structures can preserve the influence of the mechanical equipment's own structure on data fusion during multi-source data fusion, and constrain the learning process of subsequent adaptive graph structures.
[0088] The adaptive graph structure learning module is updated during the model optimization process and outputs an adaptive graph structure for any input multi-source time series samples. n is the number of sample channels, l is the sample length, and its adaptive graph structure can be represented as follows: Node feature matrix Adjacency Matrix The process of constructing the adjacency matrix of an adaptive graph structure includes: parameterizing the probability p that there is an edge between nodes. ij The gumbel-softmax technique is used to sample the edges between nodes. Representing the states with and without edges respectively, we obtain the adjacency matrix A of the adaptive graph structure. a The specific construction process is as follows:
[0089] First, the probability that there are edges between parameterized nodes:
[0090]
[0091] Where N is the number of nodes, i.e. the number of multi-source data channels; Let be the probability that an edge exists. This represents the probability that no edge exists.
[0092] Secondly, sampling noise from a uniform distribution Calculate the corresponding Gumbel noise
[0093]
[0094]
[0095] In the formula, For uniformly distributed sampling noise, For the corresponding The Gumbel noise. The above two equations represent... and The corresponding calculation relationship.
[0096] Finally, according to Calculate the state variables of the edges
[0097]
[0098] In the formula, For uniformly distributed sampling noise, For the corresponding The Gumbel noise, where τ is a temperature parameter used to control the smoothness of the sampling;
[0099] Then, the adjacency matrix of the adaptive graph structure is obtained:
[0100]
[0101] A a =[a ij ]
[0102] Step 3: Construct an unsupervised multi-source temporal graph convolutional network model for system anomaly detection.
[0103] Model structure as follows Figure 2 As shown, the overall model is divided into two modules: the first is the spatiotemporal feature extraction network module, and the second is the segmentation and reconstruction anomaly detection module. Multi-source data samples and the two types of graph structures in step 2 are input into the spatiotemporal feature extraction network module to obtain the spatiotemporal features of the samples. The spatiotemporal features are then input into the segmentation and reconstruction anomaly detection module to obtain the anomaly status score of the samples, thereby realizing anomaly detection.
[0104] Based on the adaptive graph structure and prior graph structure obtained in step 2, spatiotemporal features of multi-source time-series data are extracted. The spatiotemporal feature extraction network module consists of multiple spatiotemporal feature extraction layers connected by skip connections; each spatiotemporal feature extraction layer includes three parts: temporal convolution, graph attention convolution based on the prior graph structure, and graph attention convolution based on the adaptive graph structure. Both the graph attention convolution based on the prior graph structure and the graph attention convolution based on the adaptive graph structure include two steps: attention calculation and state update.
[0105] The attention calculation process is as follows:
[0106]
[0107]
[0108] In the formula, Let i be the feature of the input node i. Let ω be the feature of the input node j, ω be a trainable weight matrix, and N be the number of nodes. i Let e be the set of neighboring nodes of node i. ij Let α be the similarity of features between nodes i and j. ij The attention weights are calculated;
[0109] The status update process is as follows:
[0110]
[0111] In the formula, σ is the activation function. To input the features of node i in the convolutional layer of this graph, This represents the node features output by the convolutional layer of the graph.
[0112] The segmentation and reconstruction anomaly detection network is based on a convolutional autoencoder and consists of two encoders and one decoder. The two encoders have identical structures, each composed of a convolutional layer, a batch normalization layer, and an activation function. The decoder consists of a transposed convolutional layer, a batch normalization layer, and an activation function. The spatiotemporal feature extraction network outputs spatiotemporal features. After segmentation, we get h u ,h d Input two encoders to obtain the hidden vector hid u ,hid d Then, the same decoder is used to decode and reconstruct the two hidden vectors. The process is formulated as follows:
[0113] hid u =E u (h u )
[0114] hid d=E d (h d )
[0115] out u =D(hid) u )
[0116] out d =D(hid) d )
[0117] In the formula, E u E d D represents the encoding / decoding network, h represents the encoding / decoding network, h represents the decoding ... u The first half of the spatiotemporal characteristics, h d For the latter half of the spatiotemporal features, hid u and hid d The encoded hidden vector, out u and out d This is the output after decoding and reconstruction.
[0118] Step 4: Optimize the deep network model constructed in Step 3 using the training set partitioned in Step 1.
[0119] The optimization targets are the parameters in the network model constructed in step 3 and the parameters of the adaptive graph structure learning module constructed in step 2.
[0120] The model optimization objective is to minimize the training loss, which consists of three parts: segmentation and reconstruction loss. Adaptive graph structure sparsity loss Spatiotemporal feature distribution loss The calculation process is as follows:
[0121] First, the input multi-source time series samples Segmentation to obtain Calculate reconstruction loss
[0122]
[0123] In the formula, x u ,x d These are the two segments obtained after splitting the input multi-source time series sample, out. u ,out d These are the two parts of the output obtained from the reconstruction.
[0124] Next, the adaptive graph sparsity loss is calculated.
[0125]
[0126] In the formula, Let N be the probability that nodes i and j have edges in the adaptive graph structure, and N be the number of sensors.
[0127] Next, calculate the spatiotemporal feature distribution loss. That is, the two parts of the hidden vector after encoding (hid) u and hid d Maximum average difference between:
[0128]
[0129] In the formula, hid u ,hid d These are the two encoded hidden vectors, φ(·) is a mapping from the original space to the regenerated Hilbert space, and s is the batch size used during training.
[0130] Finally, the training loss is obtained.
[0131]
[0132] In the formula, λ1>0, λ2<0, and λ1 and λ2 are preset parameters. To segment the reconstruction loss, To maintain the sparsity of the adaptive graph structure, The loss term is used to expand the feature space of normal samples.
[0133] Maintaining the sparsity of the adaptive graph structure aims to improve the efficiency of multi-source data fusion and reduce the impact of noise. Expanding the differences in spatiotemporal feature distribution aims to improve the decoder's understanding of the feature space of normal samples, thereby reducing the model's misclassification rate of normal samples.
[0134] Step 5: Use the training set data from Step 1 to perform state evaluation, obtain abnormal state scores, and calculate thresholds based on the training set state scores. The abnormal state scores and threshold calculations are as follows:
[0135] Abnormal status score:
[0136] In the formula, λ2<0, and λ2 is a preset parameter. To reconstruct the loss, The loss is the spatiotemporal feature distribution.
[0137] The threshold values are as follows:
[0138]
[0139] In the formula, t is the threshold for the abnormal state score, and N s The number of samples used in the threshold calculation, scorei The score is given for the abnormal state of the i-th sample.
[0140] Step 6: Based on the threshold obtained in Step 5, perform anomaly detection on the test set in Step 1. The anomaly detection results are shown below:
[0141]
[0142] In the formula, t is the threshold for the abnormal state score, and state(X) i ) represents the state of the sample, and score is the score for abnormal states.
[0143] The invention will be further described in detail below with reference to specific implementation examples:
[0144] In this experiment, multi-source monitoring data collected by 8 (dataset 1) and 19 (dataset 2) sensors at different or the same locations were used as raw data to detect abnormal system states of the equipment. Both datasets 1 and 2 used 1000 training samples and 2000 test samples. Following the method proposed in this invention, the samples were preprocessed to construct a prior graph structure, and training and testing were performed using the proposed network model. The final detection accuracy was 99.9% with a false negative rate of 0% and a false negative rate of 0.1% on dataset 1, and 100% accuracy was achieved on dataset 2.
[0145] This example also compares this method with five other anomaly detection methods on dataset 1. The other methods are described below:
[0146] 1) OCSVM: Single-class Support Vector Machine;
[0147] 2) CAE-1SVM: An anomaly detection method utilizing a convolutional encoder and a single-class support vector machine;
[0148] 3) CAE: Anomaly detection method based on multi-channel convolutional autoencoder;
[0149] 4) memAE: An anomaly detection method based on an autoencoder and equipped with a memory module;
[0150] 5) MSCRED: An anomaly detection method for multi-scale convolutional encoding and decoding based on two-dimensional matrix input;
[0151] The test results are shown in Table 1, which fully demonstrate the effectiveness of the proposed system anomaly state detection method based on spatiotemporal feature segmentation and reconstruction.
[0152] Table 1 Comparison of detection results using different methods
[0153]
[0154] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the scope of protection of the pending claims of the invention.
Claims
1. A method for anomaly detection in power equipment systems based on spatiotemporal feature segmentation and reconstruction, characterized in that, Includes the following steps: Step 1: Obtain multi-source monitoring data from the equipment status monitoring platform, cut the multi-source monitoring data to a specific length to obtain multi-source time series samples, preprocess each multi-source time series sample to obtain a multi-source dataset, and divide it into training set and test set; Step 2: Based on the prior knowledge of the power equipment's own structure, obtain the prior graph structure from the multi-source time series samples obtained in Step 1, and construct an adaptive graph structure learning module to output the adaptive graph structure. Step 3: Based on the prior graph structure and the adaptive graph structure, construct an unsupervised multi-source temporal graph convolutional network model for system anomaly detection. The unsupervised multi-source temporal graph convolutional network model consists of two modules: the first is a spatiotemporal feature extraction network module, and the second is a segmentation and reconstruction anomaly detection module. Multi-source time series samples and the prior graph structure and adaptive graph structure from Step 2 are input into the spatiotemporal feature extraction network module to obtain the spatiotemporal features of the samples. The spatiotemporal features are then input into the segmentation and reconstruction anomaly detection module to obtain the anomaly state score of the samples, thereby achieving anomaly detection. The spatiotemporal feature extraction network module consists of multiple spatiotemporal feature extraction layers connected by skip connections. Each spatiotemporal feature extraction layer comprises three parts: temporal convolution, graph attention convolution based on a prior graph structure, and graph attention convolution based on an adaptive graph structure. Both the graph attention convolution based on a prior graph structure and the graph attention convolution based on an adaptive graph structure include two steps: attention calculation and state update. The attention calculation process is as follows: In the formula, For the nodes of the input graph convolutional layer feature, For input nodes feature, For a weight matrix that can be trained, For nodes The set of neighboring nodes, For nodes Feature similarity, The attention weights are calculated; The status update process is as follows: In the formula, For activation function, For input nodes feature, For the updated output nodes feature; Step 4: Optimize the unsupervised multi-source temporal graph convolutional network model constructed in Step 3 using the training set from Step 1; Step 5: Use the network model optimized in Step 4 to evaluate the system state of the training set, process the obtained state scores, and calculate the threshold. Step 6: Use the network model trained in Step 4 to evaluate the system state of the test set, and output the abnormal state detection results based on the threshold calculated in Step 5.
2. The method for anomaly detection of power equipment systems based on spatiotemporal feature segmentation and reconstruction according to claim 1, characterized in that, Step 1, specifically the truncating to a specific length, involves truncating the multi-source monitoring data to a specific length. Multi-source time series samples are obtained by non-overlapping partitioning; In step 1, the preprocessing adopts a normalization operation, specifically: normalize each channel of the multi-source time series sample separately, and divide the multi-source dataset according to the source of the multi-source monitoring data. Select a portion of the samples with normal system status as the training set, and use the remaining normal samples and the samples with abnormal system status as the test set. The normalization formula is as follows: In the formula, For normalized data, For the first The first channel Data points, for Mean of each data point for Standard deviation of each data point For the first Data from each channel.
3. The method for anomaly detection of power equipment systems based on spatiotemporal feature segmentation and reconstruction according to claim 1, characterized in that, In step 2, a priori graph structure is established based on prior knowledge of the power equipment's own structure. This priori graph structure includes the adjacency matrix of the graph structure and the feature matrix of the nodes. For any input multi-source time series sample... , This represents the number of sample channels. For sample length, Let the set of real numbers be represented as a prior graph structure sample. Node feature matrix Adjacency matrix The specific method for constructing the adjacency matrix of the prior graph structure is as follows: First, the locations of the sensors from each channel of the multi-source monitoring data on the power equipment are clearly identified. Second, the values of the adjacency matrix are determined based on the connection relationships between the structures containing the sensors. Indicates the first The location of the first channel sensor is related to the first... There is no connection between the locations of the individual channel sensors. This indicates that a connection exists; obtain the value of the adjacency matrix: 。 4. The method for anomaly detection of power equipment systems based on spatiotemporal feature segmentation and reconstruction according to claim 3, characterized in that, In the adaptive graph structure learning module constructed in step 2, the probability of edges existing between nodes is parameterized. The network learns the adaptive graph structure to model the intrinsic relationships between multi-source data. For any input multi-source time series sample... , This represents the number of sample channels. Given a sample length, its adaptive graph structure is represented as follows: Node feature matrix Adjacency matrix The process of constructing the adjacency matrix of an adaptive graph structure includes: parameterizing the probability that edges exist between nodes. The gumbel-softmax technique is used to sample the edges between nodes, and state variables are used to perform the sampling. Representing the states with and without edges respectively, we obtain the adjacency matrix of the adaptive graph structure. The specific construction process is as follows: First, the probability that there are edges between parameterized nodes: in, This refers to the number of nodes, i.e., the number of multi-source data channels. Let be the probability that an edge exists. This represents the probability that no edge exists. Secondly, sampling noise from a uniform distribution Calculate the corresponding Gumbel noise : In the formula, For uniformly distributed sampling noise, For the corresponding Gumbel noise; Finally, according to The state variables of the edges are calculated. : In the formula, , This is a temperature parameter used to control the smoothness of the sampling. Then, the adjacency matrix of the adaptive graph structure is obtained: 。 5. The method for anomaly detection of power equipment systems based on spatiotemporal feature segmentation and reconstruction according to claim 1, characterized in that, The segmentation and reconstruction anomaly detection module includes two encoders and one decoder; the two encoders have the same structure, consisting of a convolutional layer, a batch normalization layer and an activation function, respectively, and the decoder consists of a transposed convolutional layer, a batch normalization layer and an activation function; Spatiotemporal features output by the spatiotemporal feature extraction network After segmentation, we get ,Will Input two encoders to obtain the hidden vectors. and Then, the same decoder is used to decode and reconstruct the two hidden vectors. The process is formulated as follows: In the formula, These represent two encoding networks and one decoding network, respectively. This is the first half of the spatiotemporal characteristics. This refers to the latter half of the spatiotemporal characteristics. and The encoded hidden vector, and This is the output after decoding and reconstruction.
6. The method for anomaly detection of power equipment systems based on spatiotemporal feature segmentation and reconstruction according to claim 5, characterized in that, When optimizing the network model using the training set in step 4, the optimization targets are the parameters of the network model constructed in step 3 and the parameters of the adaptive graph structure learning module constructed in step 2. The optimization objective is to minimize the training loss. The training loss calculation process is as follows: First, the input multi-source time series samples Segmentation to obtain Calculate the reconstruction loss : In the formula, These are the two segments obtained after segmenting the input multi-source time series sample. These are the two parts of the output obtained from the reconstruction; Next, the adaptive graph sparsity loss is calculated. : In the formula, For nodes in an adaptive graph structure The probability that an edge exists. Number of sensors; Next, calculate the spatiotemporal feature distribution loss. That is, the two hidden vectors after encoding and Maximum average difference between: In the formula, These are the two parts of the encoded hidden vector. This is a mapping from the original space to the regenerated Hilbert space. This refers to the batch size used during training. Finally, the training loss is obtained. : In the formula, , , These are preset parameters. For reconstruction error, For adaptive graph sparsity loss, The loss is the spatiotemporal feature distribution.
7. The method for anomaly detection of power equipment systems based on spatiotemporal feature segmentation and reconstruction according to claim 6, characterized in that, In step 5, the state is evaluated using the training set data to obtain an abnormal state score, and a threshold is calculated based on the training set state score. The abnormal state score and threshold calculation are as follows: Abnormal status score: In the formula, , These are preset parameters. For reconstruction error, Loss is based on the spatiotemporal characteristic distribution. The threshold values are as follows: In the formula, The threshold for scoring abnormal states. The number of samples used in the threshold calculation. For the first Abnormal state score for each sample.
8. The method for anomaly detection of power equipment systems based on spatiotemporal feature segmentation and reconstruction according to claim 7, characterized in that, The abnormal state detection results in step 6 are shown below: In the formula, For testing samples, The threshold for scoring abnormal states. For the state of the sample, Scoring of abnormal states in samples.