Box transformer substation monitoring method based on unsupervised learning

Abstract

The invention discloses a box-type transformer substation monitoring method based on unsupervised learning, and the method comprises the following steps: S1, collecting multi-source sensing data in the operation process of a box-type transformer, and generating a standardized input data set; s2, selecting a data sample in a normal operation state, constructing an unsupervised learning model, and obtaining a feature representation set; s3, inputting data, extracting current operation state characteristics, and recognizing a voiceprint abnormal state by combining outlier detection; s4, performing window sliding and statistical analysis on the time sequence data, and outputting a trend abnormal interval and an abnormal index type; s5, constructing a variable association graph structure, and identifying potential abnormal variables and propagation paths; s6, integrating the various types of abnormal information and the feature representation set, and constructing a voiceprint map library; and S7, integrating and processing the abnormal result and the voiceprint map database, and executing visual display and intelligent early warning. According to the invention, box transformer substation abnormity monitoring and early warning based on unsupervised learning are realized, and the fault identification accuracy and response efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment operating status monitoring and anomaly detection, and in particular to a box-type transformer monitoring method based on unsupervised learning. Background Art

[0002] In the existing power system operation and management, box-type transformers are key equipment in the power transmission and distribution process, and the stability of their operating status is directly related to the reliability and safety of the power supply system. At present, the monitoring of box-type transformers mainly relies on single-type sensor data such as temperature, current, and voltage, and adopts classification methods based on rule thresholds or supervised learning for fault diagnosis and early warning. However, these methods have many limitations. For example, traditional methods based on fixed thresholds cannot adapt to the dynamic changes in the diverse operating states of box-type transformers in complex environments and are prone to false positives or missed reports. Although supervised learning methods have certain learning capabilities, they rely on a large number of labeled fault samples and are limited by the difficulty of obtaining abnormal data. In practical applications, they have problems such as high labeling costs, incomplete sample categories, and poor generalization ability.

[0003] Furthermore, traditional box-type transformer monitoring systems often overlook the multimodal signal data generated by the box-type transformer during operation, particularly soundprint and vibration information, which often contain rich equipment status characteristics and potential abnormal signs. Existing technologies rarely comprehensively utilize multi-source sensory data for deep feature learning and multi-dimensional status assessment. They also lack the ability to model and integrate the correlations between multimodal information, resulting in weak recognition of complex abnormal patterns. Furthermore, most methods employ static model structures, which are unable to dynamically adjust and adaptively optimize according to changes in equipment status, and lack the ability to deploy online for long-term deployment and high-precision monitoring.

[0004] Therefore, how to provide a box-type transformer monitoring method based on unsupervised learning is an urgent problem that those skilled in the art need to solve. Summary of the Invention

[0005] One purpose of the present invention is to propose a box-type transformer monitoring method based on unsupervised learning. The present invention fully integrates technical means such as multi-source perception data modeling, unsupervised feature learning, anomaly identification, and multi-dimensional graph construction. It describes in detail the use of multi-channel encoders, sparse decoder networks, and time context prediction structures to build an unsupervised learning model, and combines voiceprint features, trend changes, and variable association information to achieve the entire process of accurate monitoring of the operating status of box-type transformers and abnormality warning. This method has the advantages of not relying on fault labels, high anomaly identification accuracy, strong adaptability to complex working conditions, and strong visual analysis capabilities, significantly improving the intelligence and reliability of box-type transformer operating status monitoring.

[0006] A box-type transformer monitoring method based on unsupervised learning according to an embodiment of the present invention includes the following steps:

[0007] S1. Collect multi-source sensory data during the operation of the box-type transformer, pre-process the multi-source sensory data, and generate a standardized input data set;

[0008] S2. Select data samples in normal operating state from the input data set, build an unsupervised learning model, train the data samples, and obtain a feature representation set;

[0009] S3. Input the input data set into the trained unsupervised learning model, extract the current operating status features, and compare the output of the unsupervised learning model with the feature representation set. Combined with the outlier detection algorithm, identify the abnormal state of the voiceprint, and output the voiceprint abnormality labeling result and feature deviation value;

[0010] S4. Perform continuous window sliding and statistical feature analysis on the time series data in the input data set, and output the trend anomaly interval and anomaly indicator type;

[0011] S5. Construct a variable association graph structure based on multi-source perception data, use graph structure analysis methods to identify correlation changes between various monitoring quantities, locate potential abnormal variables and propagation paths, and output variable association information;

[0012] S6. Construct a voiceprint atlas library by taking the voiceprint anomaly marking results, feature deviation values, trend anomaly intervals, anomaly indicator types, and potential anomaly variables into consideration with reference to the feature representation set.

[0013] S7. Integrate the voiceprint anomaly marking results, trend anomaly intervals, variable association information and voiceprint atlas library, perform multi-dimensional visualization through the monitoring platform, trigger abnormal alarm information to be pushed, and realize real-time monitoring and intelligent early warning of the operating status of the box-type transformer.

[0014] Optionally, the multi-source perception data specifically includes sound, vibration and temperature signals collected during the operation of the box-type transformer.

[0015] Optionally, the preprocessing of the multi-source perception data specifically includes performing denoising, normalization and time synchronization operations on the multi-source perception data.

[0016] Optionally, the S2 specifically includes:

[0017] S21, extracting samples collected during the no-alarm operation period of the box-type transformer from the input data set, and constructing a data subset under normal operating conditions, the data subset including the sound modal sample sequence X s ={x s1 ,x s2 ,…,xsn}、Vibration mode sample sequence X v ={x v1 ,x v2 ,…,x vn} and the temperature modal sample sequence X t ={x t1 ,x t2 ,…,x tn}, the samples of each time step i consist of sound mode, vibration mode and temperature mode, where n is the total number of samples;

[0018] S22, construct a multi-channel encoder network, set up a sound mode encoder, a vibration mode encoder, and a temperature mode encoder respectively, each encoder receives the corresponding sound mode, vibration mode, and temperature mode input, and outputs the sound mode, vibration mode, and temperature mode potential feature vector z si 、z vi and z ti ;

[0019] S23, perform weighted fusion of the latent feature vectors of the sound mode, vibration mode and temperature mode in the latent space to generate a unified fusion vector z i ;

[0020] S24. Build a sparse decoder network g φ , the input is the unified fusion vector z i , output the reconstructed modal value Corresponding to sound, vibration and temperature signals respectively, a sparsity constraint mechanism is introduced during the unsupervised learning model training process to restrict most dimensions in the latent feature vectors of sound mode, vibration mode and temperature mode to remain close to zero, activate a small number of key feature dimensions, and obtain encoder output with sparse representation capability;

[0021] S25. Define weighted reconstruction error loss function L recon , measures the reconstruction difference between the reconstructed modal value and the multi-source perception data, and assigns different weights according to the importance of the modality:

[0022]

[0023] Among them, α is the reconstruction loss weight coefficient of the sound mode, β is the reconstruction loss weight coefficient of the vibration mode, γ is the reconstruction loss weight coefficient of the temperature mode, and x si is the sound modal input feature vector of the i-th sample, x vi is the vibration mode input eigenvector of the i-th sample, x ti is the original input eigenvector of the temperature mode of the i-th sample;

[0024] S26, build a time context predictor, receive the current time step unified fusion vector z i Then output the predicted value of the next time step Forming a time trend modeling structure;

[0025] S27. Define the joint loss function as L total :

[0026]

[0027] Among them, μ is the prediction error weight coefficient, λ is the sparse constraint coefficient, and x i+1 is the true multimodal input data at the i+1th time step, R(z i ) The i-th unified fusion vector z generated by the encoder i Regularization term on ;

[0028] S28, using a subset of data selected from the input data set in normal operating state as the training set, iteratively optimize the multi-channel encoder network, sparse decoder network and temporal context predictor during the training process, and jointly minimize the joint loss function L total , update the unsupervised learning model parameters;

[0029] S29, during the training process, the unified fusion vector z of each sample in the latent space i Save as feature representation samples to form feature representation set Z;

[0030] S210: Using the trained unsupervised learning model parameters and feature representation set as the output of the unsupervised learning model.

[0031] Optionally, the S3 specifically includes:

[0032] S31, inputting each group of samples in the input data set into the multi-channel encoder network in the trained unsupervised learning model to extract the potential feature vectors of the sound mode, vibration mode and temperature mode;

[0033] S32, inputting the sound mode, vibration mode and temperature mode potential feature vectors into the sparse decoder network and the time context predictor respectively, and generating the reconstructed modal value and the next time step prediction value respectively;

[0034] S33, respectively calculate the sound modal potential feature vector z si , vibration mode potential eigenvector z vi , temperature mode potential eigenvector z ti The minimum Euclidean distance between the corresponding feature representation set is used to obtain the anomaly scores of the three modes, forming a ternary anomaly indicator set a i ;

[0035] S34. Based on the feature representation set obtained in the training phase, a density scoring function based on Euclidean distance is used to perform an outlier factor scoring on the unified fusion vector of the current sample to generate a preliminary anomaly score value;

[0036] S35, introduce the modal attention weighting mechanism, according to the importance distribution of sound mode, vibration mode and temperature mode under different working conditions, the three-element abnormal index set a i The deviation values ​​of each modal feature are adaptively weighted, dynamically calibrated in combination with the preliminary anomaly score, and the aggregated anomaly vector is output;

[0037] S36. Based on the aggregated anomaly vector, set the sliding window range [ik,i], build a time smoothing scoring curve, and set the double threshold rule θ high ,θ low ,dynamically judge short-term sudden anomalies and continuous trend anomalies, and generate anomaly score distribution graph, where,k,represents the length of the sliding window;

[0038] S37. Use the local position attention mechanism to add position information weights to the areas with significant scores in the abnormal score distribution map to form focus area labels;

[0039] S38: Output the voiceprint abnormality label result for the sample judged to be abnormal i =1, and the modal aggregation offset vector is recorded as the feature deviation value.

[0040] Optionally, the S4 specifically includes:

[0041] S41. Process the time series data in the input data set, and construct the multi-source perception data into multimodal time series samples in chronological order. Let the sample at each moment be in Represent the observation values ​​of acoustic signal, vibration signal and temperature signal respectively;

[0042] S42, set the sliding window length and sliding step size, perform continuous window sliding operation on the multimodal time series, and form a sliding window sample sequence W j , forming a sliding window set

[0043] S43. Introducing a multi-scale window analysis mechanism. Based on different sliding window lengths, multi-scale statistical features are extracted from the sliding window sample sequence in parallel to form a multi-scale feature vector set, thereby improving the ability to simultaneously perceive both slowly changing and sudden anomalies.

[0044] S44, calculating the difference variation between adjacent windows for the statistical feature sequence at each scale to obtain a multi-scale difference sequence, and performing normalization processing to eliminate the difference in feature dimensions between different scales;

[0045] S45. Introduce an adaptive anomaly discrimination strategy based on median absolute deviation, calculate the anomaly factor value for each scale difference sequence, mark the sliding window with anomaly factor value greater than the dynamic threshold, and output the multi-scale trend anomaly candidate interval;

[0046] S46. Using a fusion voting mechanism based on modal weights, the multi-scale trend anomaly candidate intervals are jointly judged to form a globally consistent trend anomaly interval set;

[0047] S47. Assign cluster labels to the signal modes involved in each trend anomaly interval, identify the dominant anomaly mode and mark it as an anomaly indicator type, and establish a corresponding mapping relationship between the trend anomaly interval and the anomaly indicator type.

[0048] Optionally, the S5 specifically includes:

[0049] S51. Based on the input data set, extract the time series feature vector of each monitoring variable node and construct a variable feature set;

[0050] S52. Introduce a structure-aware dynamic neighborhood construction mechanism. For each variable node, a structure neighborhood set is constructed based on the sliding correlation weight at any time step t to form a sparse and controllable time-sensitive graph structure.

[0051] S53, based on the constructed sparse controllable time-sensitive graph structure G (t) =(V,E (t) ,W (t) ), where V is the set of all monitoring variable nodes, E (t) Represents the edge set between all monitoring variable nodes at time step t, W (t) is the edge weight set at time step t, where the edge weight Indicates v i With v j The correlation at time point t is calculated by calculating the dynamic correlation index between each pair of monitoring variable nodes in the sliding window, retaining the edges with strong correlation to form a strongly coupled subgraph;

[0052] S54. Introduce a node state embedding mechanism and construct an embedding function based on the node self-attention mechanism. This function jointly encodes the change trend of the monitored variable node itself and the dynamic characteristics of the structural neighborhood. The embedding function combines the feature vector of the variable at time step t and the neighborhood information of the node, and generates an embedded representation of the node in the graph structure through the self-attention mechanism.

[0053] S55. Based on the strongly coupled subgraph, execute the abnormal sensitive path backtracking strategy. Starting from the abnormal candidate node set, use the time back propagation path to mine high-frequency interaction objects, identify the abnormal propagation chain between the abnormal candidate nodes, and mark the high-risk propagation channel.

[0054] S56. Construct variable interaction tensor Comprehensively consider node state embedding, edge weight volatility, and interaction strength, and combine the anomaly propagation chain to locate the anomaly source and propagation path:

[0055]

[0056] in, For node v i The embedded feature vector at time step t, Node v j The embedded feature vector at time step t, is the square of the Euclidean distance, Indicates that at time step t, node v i With v j The strength of the correlation between Indicates that at time step t-1 node v i With v j The strength of the correlation between

[0057] S57, introduce a multi-perspective causal impact scoring mechanism and construct a variable correlation scoring function Φ(v i ):

[0058]

[0059] Among them, Centrality (v i ) represents the centrality index of the variable in the graph, BackPathScore(v i ) represents the abnormal path backtracking score, λ1, λ2, λ3 are weight coefficients;

[0060] S58, based on the variable association scoring function Φ(v i ) output results, screen the key variables with significant fluctuations in variable dependencies, and combine the dominant propagation path with the structural embedding vector to output the final variable association information set.

[0061] Optionally, the S6 specifically includes:

[0062] S61. Construct a voiceprint atlas library by taking the voiceprint anomaly marking results, feature deviation values, trend anomaly intervals, anomaly indicator types, and potential anomaly variables as a reference with the feature representation set;

[0063] S62. Based on the voiceprint anomaly marking results and feature deviation values, each abnormal monitoring sample and the associated feature representation vector are stored as node information in the voiceprint atlas. Referring to the feature representation set, the feature vectors in the feature representation set are compared with the feature vectors of the abnormal monitoring sample to determine the abnormality level of each sample. The samples are then stored as reference nodes in the voiceprint atlas.

[0064] S63. Based on the abnormal trend interval and the abnormal indicator type, determine whether each sample is abnormal, mark each node in the voiceprint atlas library as normal or abnormal, and clarify the status category of each node in the voiceprint atlas library;

[0065] S64. Combining the potential abnormal variables and the feature representation set, define the relative position and attributes of each node in the voiceprint atlas, and construct a relationship mapping between nodes based on the relative positions and attributes;

[0066] S65. Dynamically update the voiceprint atlas library, update the abnormal marking results and trend abnormality intervals according to the real-time monitoring data, and adjust the relationship and weight between the nodes in the voiceprint atlas library;

[0067] S66. Use the updated voiceprint atlas library for anomaly detection and real-time monitoring. Through the multi-dimensional visualization of the voiceprint atlas library, display abnormal samples, abnormal propagation paths and characteristic information, and provide them to operation and maintenance personnel to assist in decision-making.

[0068] The beneficial effects of the present invention are:

[0069] By introducing an unsupervised learning model, the present invention solves the problems of traditional box-type transformer monitoring methods that rely on a large number of labeled samples, cannot adapt to complex operating conditions, and have difficulty in accurately identifying early anomalies. By collecting multi-source perception data during the operation of the box-type transformer and constructing an unsupervised learning model composed of a multi-channel encoder, a sparse decoder, and a time context prediction structure, efficient self-learning and feature representation of the box-type transformer operating status characteristics are achieved. The proposed multi-dimensional anomaly recognition method based on voiceprint anomalies, trend anomalies, and variable correlation changes improves the sensitivity and recognition accuracy of early abnormal states of box-type transformers; the constructed voiceprint atlas integrates information such as abnormal feature markers, trend deviations, and variable propagation paths, realizing systematic modeling and visual presentation of complex abnormal relationships. Finally, by integrating multiple types of abnormal information into the monitoring platform for visual display and alarm push, the box-type transformer operating status monitoring has stronger real-time, accuracy, and interpretability, effectively supporting the intelligent operation and maintenance of equipment and fault prevention, and improving the safety and stability of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0071] Figure 1 This is a flow chart of a box-type transformer monitoring method based on unsupervised learning proposed by the present invention;

[0072] Figure 2 This is a schematic diagram of the variable association graph structure construction and abnormal propagation chain identification process of a box-type transformer monitoring method based on unsupervised learning proposed in the present invention. DETAILED DESCRIPTION

[0073] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0074] refer to Figure 1 and Figure 2 A box-type transformer monitoring method based on unsupervised learning includes the following steps:

[0075] S1. Collect multi-source sensory data during the operation of the box-type transformer, pre-process the multi-source sensory data, and generate a standardized input data set;

[0076] S2. Select data samples in normal operating state from the input data set, build an unsupervised learning model, train the data samples, and obtain a feature representation set;

[0077] S3. Input the input data set into the trained unsupervised learning model, extract the current operating status features, and compare the output of the unsupervised learning model with the feature representation set. Combined with the outlier detection algorithm, identify the abnormal state of the voiceprint, and output the voiceprint abnormality labeling result and feature deviation value;

[0078] S4. Perform continuous window sliding and statistical feature analysis on the time series data in the input data set, and output the trend anomaly interval and anomaly indicator type;

[0079] S5. Construct a variable association graph structure based on multi-source perception data, use graph structure analysis methods to identify correlation changes between various monitoring quantities, locate potential abnormal variables and propagation paths, and output variable association information;

[0080] S6. Construct a voiceprint atlas library by taking the voiceprint anomaly marking results, feature deviation values, trend anomaly intervals, anomaly indicator types, and potential anomaly variables into consideration with reference to the feature representation set.

[0081] S7. Integrate the voiceprint anomaly marking results, trend anomaly intervals, variable association information and voiceprint atlas library, perform multi-dimensional visualization through the monitoring platform, trigger abnormal alarm information to be pushed, and realize real-time monitoring and intelligent early warning of the operating status of the box-type transformer.

[0082] The unsupervised learning-based transformer monitoring method described in this paper effectively improves the accuracy and real-time performance of transformer operating status identification by integrating multi-source sensor data analysis and a multi-stage anomaly detection process. This method not only eliminates the reliance on a large number of fault-labeled samples, addressing the labeling difficulties and poor model generalization capabilities of traditional monitoring technologies, but also, through training a multi-channel unsupervised learning model, obtains a high-quality feature representation set that comprehensively characterizes the normal operating status of the transformer. Combining outlier detection, trend analysis, and variable association graph construction, it achieves multi-dimensional joint identification of voiceprint anomalies, trend anomalies, and variable propagation anomalies. Furthermore, by constructing a voiceprint atlas library and performing atlas fusion and visualization, the interpretability and diagnostic accuracy of anomalies are significantly enhanced. Ultimately, this method transforms transformer monitoring from "single-point indicator monitoring" to "multimodal, multi-dimensional anomaly detection," offering advantages such as flexible deployment, strong scalability, and timely response, enhancing the intelligent operation and maintenance and fault prevention capabilities of power systems.

[0083] In this embodiment, the multi-source sensing data specifically includes sound, vibration and temperature signals collected during the operation of the box-type transformer.

[0084] In this embodiment, the preprocessing of the multi-source perception data specifically includes performing denoising, normalization and time synchronization operations on the multi-source perception data.

[0085] In this embodiment, S2 specifically includes:

[0086] S21, extracting samples collected during the no-alarm operation period of the box-type transformer from the input data set, and constructing a data subset under normal operating conditions, the data subset including the sound modal sample sequence X s ={x s1 ,x s2 ,…,x sn}、Vibration mode sample sequence X v ={x v1 ,x v2 ,…,x vn} and the temperature modal sample sequence X t ={x t1 ,x t2 ,…,x tn}, the samples of each time step i consist of sound mode, vibration mode and temperature mode, where n is the total number of samples;

[0087] S22, construct a multi-channel encoder network, set up a sound mode encoder, a vibration mode encoder, and a temperature mode encoder respectively, each encoder receives the corresponding sound mode, vibration mode, and temperature mode input, and outputs the sound mode, vibration mode, and temperature mode potential feature vector z si 、z vi and z ti ;

[0088] S23, perform weighted fusion of the latent feature vectors of the sound mode, vibration mode and temperature mode in the latent space to generate a unified fusion vector z i ;

[0089] S24. Build a sparse decoder network g φ , the input is the unified fusion vector z i , output the reconstructed modal value Corresponding to sound, vibration and temperature signals respectively, a sparsity constraint mechanism is introduced during the unsupervised learning model training process to restrict most dimensions in the latent feature vectors of sound mode, vibration mode and temperature mode to remain close to zero, activate a small number of key feature dimensions, and obtain encoder output with sparse representation capability;

[0090] S25. Define weighted reconstruction error loss function L recon , measures the reconstruction difference between the reconstructed modal value and the multi-source perception data, and assigns different weights according to the importance of the modality:

[0091]

[0092] Among them, α is the reconstruction loss weight coefficient of the sound mode, β is the reconstruction loss weight coefficient of the vibration mode, γ is the reconstruction loss weight coefficient of the temperature mode, and x si is the sound modal input feature vector of the i-th sample, x vi is the vibration mode input eigenvector of the i-th sample, x ti is the original input eigenvector of the temperature mode of the i-th sample;

[0093] The weighted reconstruction error loss function L is defined as reconThe practical significance of is to measure the overall error degree of the unsupervised learning model when reconstructing the multi-source perception data of the box-type transformer, and to give differentiated importance to the reconstruction error of each modality by setting the weight coefficients of different modalities. The weighted reconstruction error loss function plays a core role in the objective function in the model training process, guiding the model to pay more attention to the accurate reconstruction of key modal features, thereby improving the effectiveness of feature expression and the sensitivity of anomaly detection. Specifically, when a certain modality, such as sound signals, has a strong ability to distinguish anomaly detection, its weight coefficient can be increased to encourage the model to strengthen its learning and reconstruction capabilities for this modality. Compared with uniformly processing all modalities, this weighting strategy enhances the model's ability to adapt to the importance differences between multi-modal features, effectively avoids information dilution or misleading, and improves the expressiveness and stability of the overall model under multi-source perception data. Therefore, this weighted reconstruction error loss function not only improves the model's attention to key modalities.

[0094] S26, build a time context predictor, receive the current time step unified fusion vector z i Then output the predicted value of the next time step Forming a time trend modeling structure;

[0095] S27. Define the joint loss function as L total :

[0096]

[0097] Among them, μ is the prediction error weight coefficient, λ is the sparse constraint coefficient, and x i+1 is the true multimodal input data at the i+1th time step, R(z i ) The i-th unified fusion vector z generated by the encoder i Regularization term on ;

[0098] The joint loss function L is defined as totalThe practical significance of is that it integrates three components: weighted reconstruction error, temporal context prediction error, and sparsity regularization term, comprehensively constraining the performance of the unsupervised learning model during the training process, ensuring that the model can not only accurately reconstruct the multimodal input features of the current moment, but also reasonably predict the state evolution trend of the next moment, and maintain the sparse expression structure of the latent space. Specifically, the first weighted reconstruction error ensures that the model can accurately restore the original input data, the second temporal prediction error strengthens the temporal modeling capability of the model, enabling it to perceive abnormal trends, and the third sparsity regularization term restricts the model to activate only the most critical latent features and suppresses redundant feature interference. By introducing the coupled optimization of three different task objectives, the joint loss function enhances its time series modeling and anomaly detection sensitivity while maintaining the model's reconstruction capability, thereby constructing a multi-task self-supervised modeling framework with a reasonable structure and balanced performance. This joint loss function provides a key optimization target for the model to achieve high-quality feature representation and stronger anomaly recognition capabilities, and is one of the core mechanisms for unsupervised model construction.

[0099] S28, using a subset of data selected from the input data set in normal operating state as the training set, iteratively optimize the multi-channel encoder network, sparse decoder network and temporal context predictor during the training process, and jointly minimize the joint loss function L total , update the unsupervised learning model parameters;

[0100] S29, during the training process, the unified fusion vector z of each sample in the latent space i Save as feature representation samples to form feature representation set Z;

[0101] S210: Using the trained unsupervised learning model parameters and feature representation set as the output of the unsupervised learning model.

[0102] The present invention systematically improves the feature extraction and anomaly perception capabilities of unsupervised learning models in box-type transformer monitoring scenarios by introducing a multi-channel coding structure, a sparse decoder network, and a time context prediction module. The constructed three-modal encoder for sound, vibration, and temperature can independently model different modal feature distributions, enhancing the model's ability to represent multi-source heterogeneous data. The introduction of the sparse decoder structure forces the model to focus on key feature dimensions, effectively improving the reconstruction quality and feature expression efficiency; the weighted reconstruction error function and the joint loss mechanism ensure the model's optimal training performance between multi-modal balance and time trend prediction. At the same time, by saving a unified fusion vector to construct a feature representation set, an accurate and structured comparison standard is provided for subsequent anomaly identification. This claim realizes a closed-loop modeling process from multi-source data collection, feature extraction, modal reconstruction to trend prediction, improving the model's interpretability, generalization ability, and depth of characterization of the box-type transformer's operating status, laying a solid foundation for anomaly detection and voiceprint map construction.

[0103] In this embodiment, S3 specifically includes:

[0104] S31, inputting each group of samples in the input data set into the multi-channel encoder network in the trained unsupervised learning model to extract the potential feature vectors of the sound mode, vibration mode and temperature mode;

[0105] S32, inputting the sound mode, vibration mode and temperature mode potential feature vectors into the sparse decoder network and the time context predictor respectively, and generating the reconstructed modal value and the next time step prediction value respectively;

[0106] S33, respectively calculate the sound modal potential feature vector z si , vibration mode potential eigenvector z vi , temperature mode potential eigenvector z ti The minimum Euclidean distance between the corresponding feature representation set is used to obtain the anomaly scores of the three modes, forming a ternary anomaly indicator set a i ;

[0107] S34. Based on the feature representation set obtained in the training phase, a density scoring function based on Euclidean distance is used to perform an outlier factor scoring on the unified fusion vector of the current sample to generate a preliminary anomaly score value;

[0108] S35, introduce the modal attention weighting mechanism, according to the importance distribution of sound mode, vibration mode and temperature mode under different working conditions, the three-element abnormal index set a i The deviation values ​​of each modal feature are adaptively weighted, dynamically calibrated in combination with the preliminary anomaly score, and the aggregated anomaly vector is output;

[0109] S36. Based on the aggregated anomaly vector, set the sliding window range [ik,i], build a time smoothing scoring curve, and set the double threshold rule θ high ,θ low ,dynamically judge short-term sudden anomalies and continuous trend anomalies, and generate anomaly score distribution graph, where,k,represents the length of the sliding window;

[0110] S37. Use the local position attention mechanism to add position information weights to the areas with significant scores in the abnormal score distribution map to form focus area labels;

[0111] S38: Output the voiceprint abnormality label result for the sample judged to be abnormal i =1, and the modal aggregation offset vector is recorded as the feature deviation value.

[0112] This invention significantly improves the accuracy and sensitivity of identifying the operating status of box-type transformers by constructing a multimodal anomaly detection mechanism. First, a trained unsupervised learning model is used to extract the potential features of each modality. A sparse decoder and a temporal context predictor are then used to obtain modal reconstruction results and predicted trends, effectively capturing the subtle fluctuations of the operating signal. Furthermore, Euclidean distance calculations are performed on the feature representation set to form a ternary anomaly indicator set, enhancing the resolution of anomaly detection. By introducing a modal attention weighting mechanism, adaptive fusion of the deviation values ​​of the three modal anomaly features is achieved. The anomaly level is dynamically calibrated based on the outlier scoring results, enhancing robustness to anomaly types under different operating conditions. A temporal smoothing scoring curve is also constructed, supplemented by a dual-threshold rule and a local position attention mechanism, to enhance the comprehensive perception of short-term and trend anomalies. Focused region labels clarify the anomaly starting point and time range, improving the timeliness and accuracy of monitoring responses. The final output of the voiceprint anomaly labeling results and feature deviation values ​​provides highly reliable data support for subsequent map construction and visual early warning.

[0113] In this embodiment, the S4 specifically includes:

[0114] S41. Process the time series data in the input data set, and construct the multi-source perception data into multimodal time series samples in chronological order. Let the sample at each moment be in Represent the observation values ​​of acoustic signal, vibration signal and temperature signal respectively;

[0115] S42, set the sliding window length and sliding step size, perform continuous window sliding operation on the multimodal time series, and form a sliding window sample sequence W j , forming a sliding window set

[0116] S43. Introducing a multi-scale window analysis mechanism. Based on different sliding window lengths, multi-scale statistical features are extracted from the sliding window sample sequence in parallel to form a multi-scale feature vector set, thereby improving the ability to simultaneously perceive both slowly changing and sudden anomalies.

[0117] S44, calculating the difference variation between adjacent windows for the statistical feature sequence at each scale to obtain a multi-scale difference sequence, and performing normalization processing to eliminate the difference in feature dimensions between different scales;

[0118] S45. Introduce an adaptive anomaly discrimination strategy based on median absolute deviation, calculate the anomaly factor value for each scale difference sequence, mark the sliding window with anomaly factor value greater than the dynamic threshold, and output the multi-scale trend anomaly candidate interval;

[0119] S46. Using a fusion voting mechanism based on modal weights, the multi-scale trend anomaly candidate intervals are jointly judged to form a globally consistent trend anomaly interval set;

[0120] S47. Assign cluster labels to the signal modes involved in each trend anomaly interval, identify the dominant anomaly mode and mark it as an anomaly indicator type, and establish a corresponding mapping relationship between the trend anomaly interval and the anomaly indicator type.

[0121] This invention significantly enhances the ability to identify trend anomalies during the operation of box-type transformers by introducing multi-scale sliding window analysis and an adaptive anomaly detection mechanism. First, a multimodal time series sample is constructed and sliding window parameters are set to ensure dynamic slicing of continuously observed data and capture data trends. Second, the multi-scale window analysis mechanism concurrently extracts statistical features at different time scales, effectively covering both slowly varying and suddenly varying anomaly patterns, improving the comprehensiveness and sensitivity of anomaly identification. Differential sequence normalization is introduced to address interference caused by differences in feature dimensions at different scales and enhance the robustness of the detection algorithm. Furthermore, an adaptive anomaly discrimination strategy based on median absolute deviation is employed to construct a dynamic threshold system for efficient screening of sliding window anomalies. A voting mechanism that incorporates modal weights ensures consistency in judgment results across scales, resulting in stable and reliable trend anomaly interval identification results. Finally, modal clustering labels are assigned to identify the dominant mode of anomalies and establish a correspondence between trend anomaly intervals and anomaly indicator types. This further enhances the interpretability and traceability of anomaly sources, providing key support for map construction and early warning response.

[0122] In this embodiment, the S5 specifically includes:

[0123] S51. Based on the input data set, extract the time series feature vector of each monitoring variable node and construct a variable feature set;

[0124] S52. Introduce a structure-aware dynamic neighborhood construction mechanism. For each variable node, a structure neighborhood set is constructed based on the sliding correlation weight at any time step t to form a sparse and controllable time-sensitive graph structure.

[0125] S53, based on the constructed sparse controllable time-sensitive graph structure G (t) =(V,E (t) ,W (t) ), where V is the set of all monitoring variable nodes, E (t) Represents the edge set between all monitoring variable nodes at time step t, W (t) is the edge weight set at time step t, where the edge weight Indicates v i With v j The correlation at time point t is calculated by calculating the dynamic correlation index between each pair of monitoring variable nodes in the sliding window, retaining the edges with strong correlation to form a strongly coupled subgraph;

[0126] S54. Introduce a node state embedding mechanism and construct an embedding function based on the node self-attention mechanism. This function jointly encodes the change trend of the monitored variable node itself and the dynamic characteristics of the structural neighborhood. The embedding function combines the feature vector of the variable at time step t and the neighborhood information of the node, and generates an embedded representation of the node in the graph structure through the self-attention mechanism.

[0127] S55. Based on the strongly coupled subgraph, execute the abnormal sensitive path backtracking strategy. Starting from the abnormal candidate node set, use the time back propagation path to mine high-frequency interaction objects, identify the abnormal propagation chain between the abnormal candidate nodes, and mark the high-risk propagation channel.

[0128] S56. Construct variable interaction tensor Comprehensively consider node state embedding, edge weight volatility, and interaction strength, and combine the anomaly propagation chain to locate the anomaly source and propagation path:

[0129]

[0130] in, For node v i The embedded feature vector at time step t, Node v j The embedded feature vector at time step t, is the square of the Euclidean distance, Indicates that at time step t, node v i With v j The strength of the correlation between Indicates that at time step t-1 node v i With v j The strength of the correlation between

[0131] Variable interaction tensor The practical significance of this method lies in accurately characterizing the anomaly propagation relationships and potential impact paths between various monitored variables during box-type transformer operation. The variable interaction influence tensor reflects the similarity of variable states by introducing the Euclidean distance between node embedding vectors. It also incorporates the difference in edge weight strength between the current and previous moments to capture the dynamic correlation changes between monitored variables. This fused representation effectively identifies variable pairs that exhibit significant interaction anomalies within a specific time window, thereby mining high-risk anomaly propagation links. This mechanism not only reveals the starting point and diffusion path of anomalies but also quantifies the dependency strength between variables under abnormal conditions, providing structured support for variable association scoring and anomaly source location. The introduction of this tensor expands the traditional graph structure from single static edge weight modeling to a time-sensitive, dynamically adjustable interaction model, significantly enhancing the granularity and credibility of anomaly perception and demonstrating excellent practical feasibility and technological innovation.

[0132] S57, introduce a multi-perspective causal impact scoring mechanism and construct a variable correlation scoring function Φ(v i ):

[0133]

[0134] Among them, Centrality (v i ) represents the centrality index of the variable in the graph, BackPathScore(v i ) represents the abnormal path backtracking score, λ1, λ2, λ3 are weight coefficients;

[0135] Variable association score function Φ(v i The practical significance of this method lies in quantitatively assessing the causal influence of each monitored variable in the operating status of a box-type transformer, thereby identifying variable nodes that have a key impact on abnormal changes in the system. The variable association scoring function integrates two core dimensions: the centrality index of the variable node in the graph structure, which reflects the importance of the variable's position in the structure graph and the density of its connections with other variables; and the anomaly path backtracking score, which measures the strength of the variable's influence in the identified anomaly propagation path. By setting a weight coefficient to combine the two, a weighted causal influence score is output for each variable, achieving precise quantification of potential anomaly drivers. This scoring result can not only be used to rank key variables in the anomaly chain but also provide a quantitative reference for variable dependency fluctuation analysis and the formulation of operation and maintenance intervention strategies. This mechanism breaks through the limitations of traditional static causal analysis by introducing the dynamic nature of the graph structure and anomaly backtracking path information. It has excellent interpretability, real-time performance, and adaptability, effectively improving the system's perception and traceability capabilities for complex anomaly transmission mechanisms.

[0136] S58, based on the variable association scoring function Φ(v i ) output results, screen the key variables with significant fluctuations in variable dependencies, and combine the dominant propagation path with the structural embedding vector to output the final variable association information set.

[0137] By constructing a sparse, controllable, time-sensitive graph structure and combining it with a multi-perspective causal analysis method, this paper innovatively improves the modeling accuracy and interpretability of abnormal correlations between variables in box-type transformer operation monitoring. First, through a structure-aware dynamic neighborhood mechanism, the graph connection relationship between variables is dynamically constructed in the time series, achieving an effective expression of the time-varying correlation structure in multi-source perception data. Second, by introducing node state embedding and self-attention mechanisms, the node's own characteristics and neighborhood interaction features are jointly modeled to generate a graph embedding representation with dynamic context-aware capabilities, significantly enhancing the model's responsiveness to local structural changes. In terms of anomaly detection, anomaly propagation path backtracking is performed based on strongly coupled subgraphs, accurately capturing potential high-frequency abnormal interaction paths between variables, identifying high-risk propagation chains, and supporting in-depth anomaly cause analysis. At the same time, a variable interaction influence tensor is constructed, and a multi-perspective causal scoring mechanism is introduced to comprehensively score the centrality between variables, abnormal propagation paths, and edge weight volatility, ensuring the scientificity and stability of variable abnormal correlation judgment. The variable association information set finally outputted provides a solid data foundation and structural support for the subsequent voiceprint map construction and multi-dimensional fusion warning, and has significant advantages such as high efficiency, explainability and real-time performance.

[0138] In this embodiment, S6 specifically includes:

[0139] S61. Construct a voiceprint atlas library by taking the voiceprint anomaly marking results, feature deviation values, trend anomaly intervals, anomaly indicator types, and potential anomaly variables as a reference with the feature representation set;

[0140] S62. Based on the voiceprint anomaly marking results and feature deviation values, each abnormal monitoring sample and the associated feature representation vector are stored as node information in the voiceprint atlas. Referring to the feature representation set, the feature vectors in the feature representation set are compared with the feature vectors of the abnormal monitoring sample to determine the abnormality level of each sample. The samples are then stored as reference nodes in the voiceprint atlas.

[0141] S63. Based on the abnormal trend interval and the abnormal indicator type, determine whether each sample is abnormal, mark each node in the voiceprint atlas library as normal or abnormal, and clarify the status category of each node in the voiceprint atlas library;

[0142] S64. Combining the potential abnormal variables and the feature representation set, define the relative position and attributes of each node in the voiceprint atlas, and construct a relationship mapping between nodes based on the relative positions and attributes;

[0143] S65. Dynamically update the voiceprint atlas library, update the abnormal marking results and trend abnormality intervals according to the real-time monitoring data, and adjust the relationship and weight between the nodes in the voiceprint atlas library;

[0144] S66. Use the updated voiceprint atlas library for anomaly detection and real-time monitoring. Through the multi-dimensional visualization of the voiceprint atlas library, display abnormal samples, abnormal propagation paths and characteristic information, and provide them to operation and maintenance personnel to assist in decision-making.

[0145] By constructing a voiceprint atlas library, this invention enables structured storage, dynamic updating, and visual representation of multidimensional anomaly information in the operating status of box-type transformers, significantly enhancing the understanding and diagnostic efficiency of abnormal conditions. First, the voiceprint atlas library integrates multi-source heterogeneous information, including voiceprint anomaly labeling results, feature deviation values, trend anomaly intervals, anomaly indicator types, and potential anomaly variables. Using a feature representation set as a reference, it systematically stores abnormal sample nodes and their attributes, forming a structured anomaly knowledge representation framework. Second, by explicitly annotating node states and modeling inter-feature relationships, the atlas library possesses robust state differentiation and relational reasoning capabilities, supporting the identification of propagation paths for complex anomaly conditions. Furthermore, a dynamic update mechanism is introduced to adaptively adjust the atlas node states and edge weights based on real-time monitoring data, ensuring that the atlas library always reflects the latest operating status and effectively supporting online monitoring and anomaly tracing analysis. Finally, through the multi-dimensional visualization of the atlas, operation and maintenance personnel can intuitively understand the attributes, deviation features, and associated paths of each abnormal node, improving the intuitiveness and decision-making efficiency of fault diagnosis. The present invention has significant advantages in improving the depth and accuracy of perception of the operating status of box-type transformers, such as high real-time performance, strong interpretability and good scalability.

[0146] Example 1:

[0147] To verify the feasibility of this invention, it was applied to a substation in an industrial park. Sixteen box-type transformers were deployed in the park, each responsible for critical power transmission and branching tasks. This area has long suffered from complex operating environments and frequent electrical load fluctuations. Common faults include abnormal temperature rise, loose boxes, and safety hazards caused by local overloads.

[0148] Traditional monitoring systems generate alarms based on a single temperature or current acquisition rule, resulting in low recognition accuracy, delayed response, and difficulty interpreting. To verify the effectiveness of the proposed "unsupervised learning-based box-type transformer monitoring method," we deployed the system on a S11-M-630 / 10 box-type transformer for continuous monitoring between May 1 and May 3, 2025.

[0149] During the deployment phase, the system connected three modal sensors through edge nodes: an acoustic sensor, a triaxial accelerometer, and a temperature probe. The system sampled data every two minutes for 72 hours, generating 2,160 sets of raw data. By excluding sections of the operation log marked as maintenance periods and known alarm periods, 1,500 sets of normal operation data were selected as normal samples for training and constructing an unsupervised multimodal model.

[0150] During real-time operation, the system extracts the multimodal feature vectors of the current sample and compares them with the feature representation set constructed during the training phase. Compared to traditional manual inspections and threshold triggering, this system, based on multimodal fusion, unsupervised modeling, and graph reasoning, enables detection approximately 35 minutes earlier, significantly improving the proactiveness and interpretability of anomaly identification. Furthermore, the visual display of nodes in the graph library transforms fault analysis from a "point" to a "chain," helping maintenance personnel quickly identify the impacting path.

[0151] Table 1 Comparison of typical abnormality detection data of box-type transformers

[0152]

[0153]

[0154] Table 1 above demonstrates the anomaly recognition capabilities of traditional methods and the present invention in monitoring the operating status of box-type transformers, covering differences in sound scores, vibration change rates, temperature changes, and anomaly recognition results. This time-period analysis and comparison clearly demonstrates the advantages of the present invention in early anomaly warning and highly sensitive detection.

[0155] From 1:00 PM to 1:10 PM, both the traditional method and the proposed method were judged as "normal." During this period, the equipment operated smoothly, with low scores for all modal features. The sound score was in the 0.21–0.25 range, the vibration rate of change was between 0.06 and 0.08, and the temperature remained around 42°C. The identification results of the traditional and proposed models were consistent, demonstrating that both methods possess stable judgment capabilities under low-risk conditions.

[0156] Starting at 1:20 PM, our invention detected a slight deviation in voiceprint characteristics and an increasing trend in vibration. The sound score increased to 0.31, and the vibration change rate reached 0.10. Traditional methods, however, still had lower values ​​of only 0.25 and 0.08, respectively, triggering no alarm. Our invention provided an early warning at this stage, enabling earlier risk detection.

[0157] At 13:40 and 13:50, the proposed method was more sensitive to abnormal signals, identifying them as "abnormal," while the traditional method still detected them as "normal." At this time, the temperature gradually rose to 44.9°C, and the sound scores in the proposed model rose to 0.42 and 0.48, with vibration change rates of 0.15 and 0.18, respectively. Combining modal fusion with feature deviation, the proposed model successfully identified potential equipment anomalies in advance.

[0158] Starting at 2:00 PM, both methods were judged as "abnormal," but the proposed method achieved higher scores, with a sound score of 0.55 and a vibration rate of change of 0.25, demonstrating more accurate identification of multimodal trends. Furthermore, at 2:20 PM and 2:30 PM, despite a slight decrease in voiceprint and vibration, the proposed method maintained its "abnormal" judgment, demonstrating its ability to retain historical trend memory and employ time-smoothing mechanisms to prevent both misjudgments and missed detections.

[0159] In summary, the present invention demonstrates enhanced sensitivity and response speed in anomaly identification, particularly during the anomaly precursor phase (13:20–13:40), providing a significant early warning advantage. This improves the reliability and practicality of the intelligent monitoring system for box-type transformers. By combining multimodal fusion features, a contextual prediction mechanism, and a density scoring algorithm, the present invention effectively reduces false alarm and missed alarm rates, demonstrating significant engineering value.

[0160] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A box-type transformer monitoring method based on unsupervised learning, characterized in that: The steps include: S1. Collect multi-source sensory data during the operation of the box-type transformer, pre-process the multi-source sensory data, and generate a standardized input data set; S2. Select data samples in normal operating state from the input data set, build an unsupervised learning model, train the data samples, and obtain a feature representation set; S3. Input the input data set into the trained unsupervised learning model, extract the current operating status features, and compare the output of the unsupervised learning model with the feature representation set. Combined with the outlier detection algorithm, identify the abnormal state of the voiceprint, and output the voiceprint abnormality labeling result and feature deviation value; S4. Perform continuous window sliding and statistical feature analysis on the time series data in the input data set, and output the trend anomaly interval and anomaly indicator type; S5. Construct a variable association graph structure based on multi-source perception data, use graph structure analysis methods to identify correlation changes between various monitoring quantities, locate potential abnormal variables and propagation paths, and output variable association information; S6. Construct a voiceprint atlas library by taking the voiceprint anomaly marking results, feature deviation values, trend anomaly intervals, anomaly indicator types, and potential anomaly variables into consideration with reference to the feature representation set. S7. Integrate the voiceprint anomaly marking results, trend anomaly intervals, variable association information and voiceprint atlas library, perform multi-dimensional visualization through the monitoring platform, trigger abnormal alarm information to be pushed, and realize real-time monitoring and intelligent early warning of the operating status of the box-type transformer.

2. The method for monitoring a box-type transformer substation based on unsupervised learning according to claim 1 is characterized in that: The multi-source perception data specifically includes sound, vibration and temperature signals collected during the operation of the box-type transformer.

3. The method for monitoring a box-type transformer substation based on unsupervised learning according to claim 1 is characterized in that: The preprocessing of the multi-source perception data specifically includes performing denoising, normalization and time synchronization operations on the multi-source perception data.

4. The method for monitoring a box-type transformer substation based on unsupervised learning according to claim 1 is characterized in that: The S2 specifically includes: S21, extracting samples collected during the no-alarm operation period of the box-type transformer from the input data set, and constructing a data subset under normal operating conditions, the data subset including the sound modal sample sequence X s ={x s1 ,x s2 ,…,x sn }、Vibration mode sample sequence X v ={x v1 ,x v2 ,…,x vn } and the temperature modal sample sequence X t ={x t1 ,x t2 ,…,x tn }, the samples of each time step i consist of sound mode, vibration mode and temperature mode, where n is the total number of samples; S22, construct a multi-channel encoder network, set up a sound mode encoder, a vibration mode encoder, and a temperature mode encoder respectively, each encoder receives the corresponding sound mode, vibration mode, and temperature mode input, and outputs the sound mode, vibration mode, and temperature mode potential feature vector z si 、z vi and z ti ; S23, perform weighted fusion of the latent feature vectors of the sound mode, vibration mode and temperature mode in the latent space to generate a unified fusion vector z i ; S24. Build a sparse decoder network g φ , the input is the unified fusion vector z i , output the reconstructed modal value Corresponding to sound, vibration and temperature signals respectively, a sparsity constraint mechanism is introduced during the unsupervised learning model training process to restrict most dimensions in the latent feature vectors of sound mode, vibration mode and temperature mode to remain close to zero, activate a small number of key feature dimensions, and obtain encoder output with sparse representation capability; S25. Define weighted reconstruction error loss function L recon , measures the reconstruction difference between the reconstructed modal value and the multi-source perception data, and assigns different weights according to the importance of the modality: Among them, α is the reconstruction loss weight coefficient of the sound mode, β is the reconstruction loss weight coefficient of the vibration mode, γ is the reconstruction loss weight coefficient of the temperature mode, and x si is the sound modal input feature vector of the i-th sample, x vi is the vibration mode input eigenvector of the i-th sample, x ti is the original input eigenvector of the temperature mode of the i-th sample; S26, build a time context predictor, receive the current time step unified fusion vector z i Then output the predicted value of the next time step Forming a time trend modeling structure; S27. Define the joint loss function as L total : Among them, μ is the prediction error weight coefficient, λ is the sparse constraint coefficient, and x i+1 is the true multimodal input data at the i+1th time step, R(z i ) The i-th unified fusion vector z generated by the encoder i Regularization term on ; S28, using a subset of data selected from the input data set in normal operating state as the training set, iteratively optimize the multi-channel encoder network, sparse decoder network and temporal context predictor during the training process, and jointly minimize the joint loss function L total , update the unsupervised learning model parameters; S29, during the training process, the unified fusion vector z of each sample in the latent space i Save as feature representation samples to form feature representation set Z; S210: Using the trained unsupervised learning model parameters and feature representation set as the output of the unsupervised learning model.

5. The method for monitoring a box-type transformer substation based on unsupervised learning according to claim 1 is characterized in that: The S3 specifically includes: S31, inputting each group of samples in the input data set into the multi-channel encoder network in the trained unsupervised learning model to extract the potential feature vectors of the sound mode, vibration mode and temperature mode; S32, inputting the sound mode, vibration mode and temperature mode potential feature vectors into the sparse decoder network and the time context predictor respectively, and generating the reconstructed modal value and the next time step prediction value respectively; S33, respectively calculate the sound modal potential feature vector z si , vibration mode potential eigenvector z vi , temperature mode potential eigenvector z ti The minimum Euclidean distance between the corresponding feature representation set is used to obtain the anomaly scores of the three modes, forming a ternary anomaly indicator set a i ; S34. Based on the feature representation set obtained in the training phase, a density scoring function based on Euclidean distance is used to perform an outlier factor scoring on the unified fusion vector of the current sample to generate a preliminary anomaly score value; S35, introduce the modal attention weighting mechanism, according to the importance distribution of sound mode, vibration mode and temperature mode under different working conditions, the three-element abnormal index set a i The deviation values ​​of each modal feature are adaptively weighted, dynamically calibrated in combination with the preliminary anomaly score, and the aggregated anomaly vector is output; S36. Based on the aggregated anomaly vector, set the sliding window range [ik,i], build a time smoothing scoring curve, and set the double threshold rule θ high ,θ low ,dynamically judge short-term sudden anomalies and continuous trend anomalies, and generate anomaly score distribution graph, where,k,represents the length of the sliding window; S37. Use the local position attention mechanism to add position information weights to the areas with significant scores in the abnormal score distribution map to form focus area labels; S38: Output the voiceprint abnormality label result for the sample judged to be abnormal i =1, and the modal aggregation offset vector is recorded as the feature deviation value.

6. The method for monitoring a box-type transformer substation based on unsupervised learning according to claim 1 is characterized in that: The S4 specifically includes: S41. Process the time series data in the input data set, and construct the multi-source perception data into multimodal time series samples in chronological order. Let the sample at each moment be in Represent the observation values ​​of acoustic signal, vibration signal and temperature signal respectively; S42, set the sliding window length and sliding step size, perform continuous window sliding operation on the multimodal time series, and form a sliding window sample sequence W j , forming a sliding window set S43. Introducing a multi-scale window analysis mechanism. Based on different sliding window lengths, multi-scale statistical features are extracted from the sliding window sample sequence in parallel to form a multi-scale feature vector set, thereby improving the ability to simultaneously perceive both slowly changing and sudden anomalies. S44, calculating the difference variation between adjacent windows for the statistical feature sequence at each scale to obtain a multi-scale difference sequence, and performing normalization processing to eliminate the difference in feature dimensions between different scales; S45. Introduce an adaptive anomaly discrimination strategy based on median absolute deviation, calculate the anomaly factor value for each scale difference sequence, mark the sliding window with anomaly factor value greater than the dynamic threshold, and output the multi-scale trend anomaly candidate interval; S46. Using a fusion voting mechanism based on modal weights, the multi-scale trend anomaly candidate intervals are jointly judged to form a globally consistent trend anomaly interval set; S47. Assign cluster labels to the signal modes involved in each trend anomaly interval, identify the dominant anomaly mode and mark it as an anomaly indicator type, and establish a corresponding mapping relationship between the trend anomaly interval and the anomaly indicator type.

7. The method for monitoring a box-type transformer substation based on unsupervised learning according to claim 1 is characterized in that: The S5 specifically includes: S51. Based on the input data set, extract the time series feature vector of each monitoring variable node and construct a variable feature set; S52. Introduce a structure-aware dynamic neighborhood construction mechanism. For each variable node, a structure neighborhood set is constructed based on the sliding correlation weight at any time step t to form a sparse and controllable time-sensitive graph structure. S53, based on the constructed sparse controllable time-sensitive graph structure G (t) =(V,E (t) ,W (t) ), where V is the set of all monitoring variable nodes, E (t) Represents the edge set between all monitoring variable nodes at time step t, W (t) is the edge weight set at time step t, where the edge weight Indicates v i With v j The correlation at time point t is calculated by calculating the dynamic correlation index between each pair of monitoring variable nodes in the sliding window, retaining the edges with strong correlation to form a strongly coupled subgraph; S54. Introduce a node state embedding mechanism and construct an embedding function based on the node self-attention mechanism. This function jointly encodes the change trend of the monitored variable node itself and the dynamic characteristics of the structural neighborhood. The embedding function combines the feature vector of the variable at time step t and the neighborhood information of the node, and generates an embedded representation of the node in the graph structure through the self-attention mechanism. S55. Based on the strongly coupled subgraph, execute the abnormal sensitive path backtracking strategy. Starting from the abnormal candidate node set, use the time back propagation path to mine high-frequency interaction objects, identify the abnormal propagation chain between the abnormal candidate nodes, and mark the high-risk propagation channel. S56. Construct variable interaction tensor Comprehensively consider node state embedding, edge weight volatility, and interaction strength, and combine the anomaly propagation chain to locate the anomaly source and propagation path: in, For node v i The embedded feature vector at time step t, Node v j The embedded feature vector at time step t, is the square of the Euclidean distance, Indicates that at time step t, node v i With v j The strength of the correlation between Indicates that at time step t-1 node v i With v j The strength of the correlation between S57, introduce a multi-perspective causal impact scoring mechanism and construct a variable correlation scoring function Φ(v i ): Among them, Centrality (v i ) represents the centrality index of the variable in the graph, BackPathScore(v i ) represents the abnormal path backtracking score, λ1, λ2, λ3 are weight coefficients; S58, based on the variable association scoring function Φ(v i ) output results, screen the key variables with significant fluctuations in variable dependencies, and combine the dominant propagation path with the structural embedding vector to output the final variable association information set.

8. The method for monitoring a box-type transformer substation based on unsupervised learning according to claim 1 is characterized in that: The S6 specifically includes: S61. Construct a voiceprint atlas library by taking the voiceprint anomaly marking results, feature deviation values, trend anomaly intervals, anomaly indicator types, and potential anomaly variables as a reference with the feature representation set; S62. Based on the voiceprint anomaly marking results and feature deviation values, each abnormal monitoring sample and the associated feature representation vector are stored as node information in the voiceprint atlas. Referring to the feature representation set, the feature vectors in the feature representation set are compared with the feature vectors of the abnormal monitoring sample to determine the abnormality level of each sample. The samples are then stored as reference nodes in the voiceprint atlas. S63. Based on the abnormal trend interval and the abnormal indicator type, determine whether each sample is abnormal, mark each node in the voiceprint atlas library as normal or abnormal, and clarify the status category of each node in the voiceprint atlas library; S64. Combining the potential abnormal variables and the feature representation set, define the relative position and attributes of each node in the voiceprint atlas, and construct a relationship mapping between nodes based on the relative positions and attributes; S65. Dynamically update the voiceprint atlas library, update the abnormal marking results and trend abnormality intervals according to the real-time monitoring data, and adjust the relationship and weight between the nodes in the voiceprint atlas library; S66. Use the updated voiceprint atlas library for anomaly detection and real-time monitoring. Through the multi-dimensional visualization of the voiceprint atlas library, display abnormal samples, abnormal propagation paths and characteristic information, and provide them to operation and maintenance personnel to assist in decision-making.

Images