A miniature circuit breaker anomaly identification and early warning method based on deep learning

Abstract

This invention discloses a deep learning-based method for anomaly identification and early warning of miniature circuit breakers (MCBs), comprising the following steps: S1, collecting operational data and thermal imaging images of the MCBs; S2, performing time alignment and normalization processing to generate a modal alignment tensor; S3, inputting the modal alignment tensor into an improved PatchCore structure to sequentially complete deformation patch extraction, graph structure modeling, topological perturbation scoring, and information entropy compression; S4, extracting a set of deformation-stable patches; S5, constructing a graph structure and generating an embedded representation; S6, calculating perturbation scores and information entropy to construct a sparse feature memory; S7, comparing data during the inference stage to obtain a multi-dimensional anomaly score; S8, fusing the scores to generate an anomaly heatmap and risk level labels, and outputting the anomaly location and early warning information. This invention integrates thermal imaging and operational data, improving the accuracy and real-time performance of circuit breaker anomaly identification.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and power equipment monitoring technology, and in particular to a method for anomaly identification and early warning of miniature circuit breakers based on deep learning. Background Technology

[0002] In the operation and maintenance of power systems, miniature circuit breakers (MCBs) are critical low-voltage protection components, and their operating status directly affects power safety and equipment protection efficiency. Traditional circuit breaker anomaly detection methods mostly rely on single or limited sensor data such as voltage, current, and temperature, and achieve anomaly judgment by setting thresholds or constructing simple rule models. However, with the increasing complexity of power distribution systems and the diversification of fault types, the detection accuracy of single signals is no longer sufficient to meet practical needs, especially in early fault identification and high-precision early warning.

[0003] To improve the accuracy of circuit breaker anomaly identification, some studies have introduced multi-feature fusion methods based on machine learning, attempting to build diagnostic models using algorithms such as support vector machines and decision trees. These methods have some generalization ability in static scenarios, but their generalization ability and response speed remain limited when facing complex, dynamic data with multimodal inputs. Furthermore, most current mainstream methods neglect the spatial structural features of image information, especially fine-grained anomaly patterns such as hotspot distributions and overheated edges in thermal imaging images, which are often coarsely processed or ignored, resulting in insufficient model perception of minute deformations and local anomalies.

[0004] In recent years, with the development of deep learning technology, graph neural networks, attention mechanisms, and image patch modeling methods have been gradually applied to anomaly detection tasks in industrial scenarios. Anomaly detection frameworks based on patches can model local changes in images, thereby capturing more detailed anomaly features. However, existing patch modeling methods are mostly static models, lacking the ability to model the effects of deformation, and also do not fully consider the structural relationships and semantic similarities between patches, resulting in limited recognition performance when faced with complex power equipment images.

[0005] Furthermore, in terms of anomaly scoring mechanisms, traditional methods often employ linear fusion or uniform threshold judgment, lacking modeling of the synergistic effects of multiple factors such as graph structure perturbation, information entropy fluctuations, and memory offsets. This makes it difficult to achieve dynamic perception and real-time early warning of potential fault trends. For high-dimensional, highly heterogeneous, and multimodal data features, existing anomaly fusion methods lack interpretability and robustness, easily generating false alarms or missed alarms, severely restricting their promotion and application in the refined operation and maintenance of power equipment.

[0006] Therefore, how to provide a method for anomaly identification and early warning of miniature circuit breakers based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a deep learning-based method for anomaly identification and early warning of miniature circuit breakers. This invention fully integrates thermal imaging images and multi-source operational data to construct a modal alignment tensor and introduces an improved PatchCore structure. Combining a deformation-aware patch extraction mechanism, a heterogeneous graph structure modeling method, a topological perturbation scoring mechanism, and an information entropy-driven memory compression strategy, it details the entire process of deep modeling and anomaly identification of circuit breaker states in complex environments. Furthermore, it constructs a graph structure-guided scoring offset mechanism and a nonlinear anomaly fusion function to enhance the identification capability of multi-dimensional anomaly features. It possesses the advantages of fine structural modeling, robust anomaly detection, and complete expression of early warning information.

[0008] A method for anomaly identification and early warning of miniature circuit breakers based on deep learning according to an embodiment of the present invention includes the following steps: S1. Collect operating data and thermal imaging images of miniature circuit breakers; S2. Perform time alignment and normalization processing on the running data and thermal imaging image to generate a modal alignment tensor; S3. Input the modality alignment tensor into the improved PatchCore structure, which includes a deformation-aware patch extraction module, a graph structure modeling module, a topological perturbation scoring module, and an information entropy-driven memory compression module. S4. Extract a set of deformation-stable patches using the deformation-aware patch extraction module; S5. Input the Patch set into the graph structure modeling module to construct a graph structure with Patch as nodes and generate a graph structure embedding representation; S6. The graph structure is embedded to represent the input topology perturbation scoring module to calculate the topology perturbation score. At the same time, the information entropy is calculated for the embedded representation of the Patch set to construct a sparse feature memory. S7. During the inference phase, the modal alignment tensor of the test sample is sequentially input into each module to obtain the graph structure embedding representation, topological perturbation score and graph attention change score of the test sample, and compare them with the sparse feature memory to calculate the multidimensional anomaly score. S8. Integrate the multi-dimensional anomaly scores to generate an anomaly heatmap and risk level labels, and output the anomaly location and early warning information of the miniature circuit breaker.

[0009] Optionally, the operating data includes voltage signals, current signals, temperature signals, and contact resistance signals.

[0010] Optionally, S3 specifically includes: S31. The modal alignment tensor is processed by the deformation-aware Patch extraction module to extract a set of deformation-stable patches. S32. Construct a graph structure for the Patch set using the graph structure modeling module, and generate a graph structure embedding representation; S33. The graph structure embedding representation is processed by the topology perturbation scoring module to obtain the topology perturbation score; S34. The embedded representation of the Patch set is processed by the information entropy-driven memory compression module to generate a sparse feature memory library.

[0011] Optionally, S4 specifically includes: S41. Perform a sliding window operation in the modal alignment tensor to extract a set of patches with a fixed size and a fixed step size. Each patch is a sub-region in the tensor and is marked with the corresponding spatial location index. S42. Set up several control points evenly in each patch to form a control point set. Input the control point set into the control point offset prediction network and calculate the lateral and longitudinal offset of each control point. S43. Add the original coordinates of the control points to the offsets to construct a set of deformation control points, and establish a nonlinear deformation function based on the set of deformation control points: ; in, and Represents the original x and y coordinates of the pixels in the patch. This indicates the output coordinates after deformation. For affine terms, The total number of control points. For the first Deformation weight of each control point For the current pixel and the first The Euclidean distance between control points, function Denotes radial basis functions, defined as follows: ; S44. Map the pixel positions in each Patch according to the deformation function, use the interpolation method to obtain the mapped position values, and generate a set of Patches after deformation adjustment. S45. Combine the deformed patch with the original running data to form a fused patch set; S46. Select the patches that meet the deformation stability conditions from the fused patch set to form a deformation-stable patch set, and input it into the graph structure modeling module as modeling input.

[0012] Optionally, S5 specifically includes: S51. Label each patch in the deformation-stable patch set according to its source and texture features. The types include thermal feature patch, sensor feature patch, edge patch and static patch. S52. Using each Patch as a node in the graph structure, construct connecting edges based on the spatial position and type label between nodes, construct a heterogeneous adjacency matrix, use Euclidean distance to calculate the edge weights between nodes of the same type, and use semantic cosine similarity to calculate the edge weights between nodes of different types. S53. Generate a heterogeneous graph structure tensor containing node features, type labels, and connection weights; S54. Input the heterogeneous graph structure tensor into the graph neural network model, select the graph convolution kernel according to the type label, and perform weighted feature aggregation based on adjacency relationship for each type of node to obtain the updated node representation; S55. Perform global average pooling on all updated node representations to generate graph structure embedding representations.

[0013] Optionally, S6 specifically includes: S61. For each node in the graph structure embedding representation, calculate the adjacency change rate under different topological perturbations, and define the perturbation weight matrix. Each element Represents a node In a disturbed scenario, adjacent nodes Changes in connection strength; S62. Construct the tensor of node perturbation influence. The element is defined as: ; in, For nodes The perturbation score, and These represent the edge weights under normal and disturbed conditions, respectively. For node type similarity weights, This represents the total number of nodes in the graph; S63. The embedding representation of each patch in the deformation-stable patch set is input to the channel response evaluation module, and the mean response vector is calculated based on the channel dimension. With response probability distribution Construct the information entropy scoring vector: ; in, The total number of channels. Indicates channel The normalized response probability, This is the information entropy score vector corresponding to the current patch; S64. Merge the perturbation scores of each patch node. Information entropy score Construct a joint scoring tensor ,in To integrate the weighting coefficients, the proportion of structural perturbation to response entropy is controlled; S65. Based on the joint score tensor, perform sparse sampling on the Patch node embedding representation to select the top-ranked nodes by joint score. The Patch node will construct a sparse feature memory based on the corresponding embedded representation.

[0014] Optionally, S7 specifically includes: S71. Perform node-level distance comparison on the graph structure embedding representation generated for the test samples, and calculate the structural difference score based on the sparse feature memory, denoted as... ,in Number the Patch nodes; S72. Obtain the connectivity of each Patch node in the adjacency graph. Calculate the rating offset factor: ; in, Indicates the first The rating offset factor for each patch node. For adjustment coefficients, This represents the maximum connectivity among all nodes. S73. Extract the perturbation amplitude of the corresponding node from the topology perturbation score and the graph attention change score. With attention drift value Score of structural differences Together as input for fusion; S74. Calculate the final anomaly score using a nonlinear fusion function: ; in, Indicates the first Anomaly score for each patch node , , To integrate the weighting coefficients, satisfy the following conditions: ; S75, all The anomaly score tensor is reconstructed based on the spatial index of the Patch node, forming a multidimensional anomaly score.

[0015] Optionally, S8 specifically includes: S81. Input the multidimensional anomaly score tensor into the two-dimensional heat map mapping network. Based on the spatial position index of each Patch node in the modal alignment tensor, perform a weighted compression operation on the multidimensional anomaly score to obtain a two-dimensional anomaly score matrix, and project it onto the thermal imaging image dimension to generate an initial anomaly heat map. S82. Perform piecewise normalization on the initial anomaly heatmap, and normalize the score values ​​in the score tensor. Map to a fixed color gradation range and set abnormal color encoding rules; S83. Based on the set risk interval threshold, perform grading processing on the abnormal heatmap and construct a risk level label matrix. Matrix elements Indicates the risk level of the corresponding location; S84, Introduction of Position Guiding Function Modulating the abnormal heatmap to adjust the risk response intensity between the edge and center regions, the modulation function is defined as follows: ; in, Indicates the coordinate position on the abnormal heatmap. Indicates the center location of the heatmap. This is the location distribution adjustment coefficient; S85. The modulated abnormal heat map and risk level label matrix are used as output results to output the abnormal location and early warning information of the miniature circuit breaker.

[0016] The beneficial effects of this invention are: This invention constructs a deep learning-based method for anomaly identification and early warning in miniature circuit breakers, overcoming the limitations of existing technologies in terms of anomaly detection accuracy, feature fusion capability, and risk representation. By collecting operational data such as voltage, current, temperature, and contact resistance signals, and combining them with thermal imaging images, a unified modal alignment tensor is established to achieve time synchronization and numerical normalization, providing a unified representation basis for subsequent multimodal feature modeling.

[0017] This invention introduces a deformation-aware patch extraction mechanism, employing control point offset prediction and a nonlinear deformation function to achieve stable feature extraction under conditions of minute deformation. It integrates raw operational data to construct a patch-level fusion representation and filters for deformation stability to ensure that features entering the graph structure modeling stage have stable semantics and clear boundaries. The graph structure modeling stage uses a heterogeneous graph construction approach, connecting thermal features, sensory features, edge features, and static feature patch nodes using spatial relationships and type labels. Edge weights are calculated by combining Euclidean distance and semantic cosine similarity between nodes, and a graph structure embedding representation is generated through a graph neural network, enhancing the expressive power of local structures.

[0018] In the scoring and memory construction phases, this invention constructs perturbation scores and information entropy scores based on the topological perturbation sensitivity of nodes in the graph structure and the channel response information entropy of the embedded representation, respectively. A joint score tensor is generated through weighted fusion, and sparse sampling is performed to construct a sparse feature memory. In the inference phase, a graph structure-guided scoring offset mechanism is introduced. Combining structural difference scores, perturbation amplitude, and attention drift indicators, a multidimensional anomaly score tensor is generated using a nonlinear fusion function. Finally, the score tensor is mapped to a two-dimensional heatmap, and an anomaly heatmap and risk level labels are generated by combining a location-guided function and a risk grading mechanism, achieving accurate labeling and graded early warning of anomaly locations in miniature circuit breakers. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for anomaly identification and early warning of miniature circuit breakers based on deep learning proposed in this invention; Figure 2 This is a structural diagram of the deformation-sensing patch extraction process of a deep learning-based method for anomaly identification and early warning of miniature circuit breakers proposed in this invention. Figure 3 This is a schematic diagram illustrating the abnormal heatmap generation process of a deep learning-based method for anomaly identification and early warning of miniature circuit breakers proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1-3 A method for anomaly identification and early warning of miniature circuit breakers based on deep learning, comprising the following steps: S1. Collect operating data and thermal imaging images of miniature circuit breakers; S2. Perform time alignment and normalization processing on the running data and thermal imaging image to generate a modal alignment tensor; S3. Input the modality alignment tensor into the improved PatchCore structure, which includes a deformation-aware patch extraction module, a graph structure modeling module, a topological perturbation scoring module, and an information entropy-driven memory compression module. S4. Extract a set of deformation-stable patches using the deformation-aware patch extraction module; S5. Input the Patch set into the graph structure modeling module to construct a graph structure with Patch as nodes and generate a graph structure embedding representation; S6. The graph structure is embedded to represent the input topology perturbation scoring module to calculate the topology perturbation score. At the same time, the information entropy is calculated for the embedded representation of the Patch set to construct a sparse feature memory. S7. During the inference phase, the modal alignment tensor of the test sample is sequentially input into each module to obtain the graph structure embedding representation, topological perturbation score and graph attention change score of the test sample, and compare them with the sparse feature memory to calculate the multidimensional anomaly score. S8. Integrate the multi-dimensional anomaly scores to generate an anomaly heatmap and risk level labels, and output the anomaly location and early warning information of the miniature circuit breaker.

[0022] This invention achieves multi-dimensional modeling and identification of abnormal states of miniature circuit breakers by constructing a multi-module deep structure based on modal alignment tensors. The entire process incorporates patch extraction, graph structure modeling, scoring mechanisms, and anomaly visualization, effectively integrating operational and image data to improve the completeness and stability of anomaly identification. This provides early warning results with greater spatial resolution and semantic interpretability for high-risk circuit breakers.

[0023] In this embodiment, the operating data includes voltage signals, current signals, temperature signals, and contact resistance signals.

[0024] This invention integrates key indicators such as voltage, current, temperature, and contact resistance into the definition of operating data, comprehensively covering typical physical parameters that affect the operating state of circuit breakers, enhancing the ability to characterize abnormal behavior under different operating conditions, and helping to build a more accurate multimodal analysis foundation from the signal level.

[0025] In this embodiment, S3 specifically includes: S31. The modal alignment tensor is processed by the deformation-aware Patch extraction module to extract a set of deformation-stable patches. S32. Construct a graph structure for the Patch set using the graph structure modeling module, and generate a graph structure embedding representation; S33. The graph structure embedding representation is processed by the topology perturbation scoring module to obtain the topology perturbation score; S34. The embedded representation of the Patch set is processed by the information entropy-driven memory compression module to generate a sparse feature memory library.

[0026] This invention clarifies the specific functional division of Patch extraction, graph structure generation, perturbation scoring, and information entropy compression, enabling the PatchCore structure to have higher module independence and debuggability, which helps to flexibly adjust the recognition strategy for different task scenarios and achieve high-precision embedded expression.

[0027] In this embodiment, S4 specifically includes: S41. Perform a sliding window operation in the modal alignment tensor to extract a set of patches with a fixed size and a fixed step size. Each patch is a sub-region in the tensor and is marked with the corresponding spatial location index. S42. Set up several control points evenly in each patch to form a control point set. Input the control point set into the control point offset prediction network and calculate the lateral and longitudinal offset of each control point. S43. Add the original coordinates of the control points to the offsets to construct a set of deformation control points, and establish a nonlinear deformation function based on the set of deformation control points: ; in, and Represents the original x and y coordinates of the pixels in the patch. This indicates the output coordinates after deformation. For affine terms, The total number of control points. For the first Deformation weight of each control point For the current pixel and the first The Euclidean distance between control points, function Denotes radial basis functions, defined as follows: ; S44. Map the pixel positions in each Patch according to the deformation function, use the interpolation method to obtain the mapped position values, and generate a set of Patches after deformation adjustment. S45. Combine the deformed patch with the original running data to form a fused patch set; S46. Select the patches that meet the deformation stability conditions from the fused patch set to form a deformation-stable patch set, and input it into the graph structure modeling module as modeling input.

[0028] This invention introduces deformation control points and nonlinear affine deformation modeling during the patch extraction process, and achieves flexible adjustment of the patch structure through control point offset and interpolation reconstruction, which significantly improves the stability of the patch structure and its adaptability to local thermal deformation areas, thereby ensuring the physical consistency and recognizability of the input image structure.

[0029] In this embodiment, S5 specifically includes: S51. Label each patch in the deformation-stable patch set according to its source and texture features. The types include thermal feature patch, sensor feature patch, edge patch and static patch. S52. Using each Patch as a node in the graph structure, construct connecting edges based on the spatial position and type label between nodes, construct a heterogeneous adjacency matrix, use Euclidean distance to calculate the edge weights between nodes of the same type, and use semantic cosine similarity to calculate the edge weights between nodes of different types. S53. Generate a heterogeneous graph structure tensor containing node features, type labels, and connection weights; S54. Input the heterogeneous graph structure tensor into the graph neural network model, select the graph convolution kernel according to the type label, and perform weighted feature aggregation based on adjacency relationship for each type of node to obtain the updated node representation; S55. Perform global average pooling on all updated node representations to generate graph structure embedding representations.

[0030] This invention constructs a heterogeneous graph structure that integrates spatial location, type label, and multi-source features during the graph structure modeling process. It sets edge weights based on semantic similarity and spatial constraints, and completes feature aggregation and embedding learning of multiple types of nodes through graph neural networks, thereby enhancing the ability to model contextual associations in structurally abnormal regions.

[0031] In this embodiment, S6 specifically includes: S61. For each node in the graph structure embedding representation, calculate the adjacency change rate under different topological perturbations, and define the perturbation weight matrix. Each element Represents a node In a disturbed scenario, adjacent nodes Changes in connection strength; S62. Construct the tensor of node perturbation influence. The element is defined as: ; in, For nodes The perturbation score, and These represent the edge weights under normal and disturbed conditions, respectively. For node type similarity weights, This represents the total number of nodes in the graph; S63. The embedding representation of each patch in the deformation-stable patch set is input to the channel response evaluation module, and the mean response vector is calculated based on the channel dimension. With response probability distribution Construct the information entropy scoring vector: ; in, The total number of channels. Indicates channel The normalized response probability, This is the information entropy score vector corresponding to the current patch; S64. Merge the perturbation scores of each patch node. Information entropy score Construct a joint scoring tensor ,in To integrate the weighting coefficients, the proportion of structural perturbation to response entropy is controlled; S65. Based on the joint score tensor, perform sparse sampling on the Patch node embedding representation to select the top-ranked nodes by joint score. The Patch node will construct a sparse feature memory based on the corresponding embedded representation.

[0032] This invention constructs a joint scoring tensor based on perturbation scoring and channel information entropy, and forms a sparse feature memory bank through structural perturbation evaluation and response uncertainty quantification. While improving the compression effect of the memory module, it retains key discrimination information, supports an efficient abnormal sample matching process, and improves the recognition efficiency in the inference stage.

[0033] In this embodiment, S7 specifically includes: S71. Perform node-level distance comparison on the graph structure embedding representation generated for the test samples, and calculate the structural difference score based on the sparse feature memory, denoted as... ,in Number the Patch nodes; S72. Obtain the connectivity of each Patch node in the adjacency graph. Calculate the rating offset factor: ; in, Indicates the first The rating offset factor for each patch node. For adjustment coefficients, This represents the maximum connectivity among all nodes. S73. Extract the perturbation amplitude of the corresponding node from the topology perturbation score and the graph attention change score. With attention drift value Score of structural differences Together as input for fusion; S74. Calculate the final anomaly score using a nonlinear fusion function: ; in, Indicates the first Anomaly score for each patch node , , To integrate the weighting coefficients, satisfy the following conditions: ; S75, all The anomaly score tensor is reconstructed based on the spatial index of the Patch node, forming a multidimensional anomaly score.

[0034] This invention introduces multiple scoring elements such as structural differences, attention drift, and perturbation amplitude during the inference stage, and generates a patch-level multidimensional anomaly scoring tensor through a scoring offset mechanism and a nonlinear fusion function. This solves the problem of the single dimension of anomaly score determination in traditional methods and improves the interpretability and accuracy of the scoring.

[0035] In this embodiment, S8 specifically includes: S81. Input the multidimensional anomaly score tensor into the two-dimensional heat map mapping network. Based on the spatial position index of each Patch node in the modal alignment tensor, perform a weighted compression operation on the multidimensional anomaly score to obtain a two-dimensional anomaly score matrix, and project it onto the thermal imaging image dimension to generate an initial anomaly heat map. S82. Perform piecewise normalization on the initial anomaly heatmap, and normalize the score values ​​in the score tensor. Map to a fixed color gradation range and set abnormal color encoding rules; S83. Based on the set risk interval threshold, perform grading processing on the abnormal heatmap and construct a risk level label matrix. Matrix elements Indicates the risk level of the corresponding location; S84, Introduction of Position Guiding Function Modulating the abnormal heatmap to adjust the risk response intensity between the edge and center regions, the modulation function is defined as follows: ; in, Indicates the coordinate position on the abnormal heatmap. Indicates the center location of the heatmap. This is the location distribution adjustment coefficient; S85. The modulated abnormal heat map and risk level label matrix are used as output results to output the abnormal location and early warning information of the miniature circuit breaker.

[0036] This invention utilizes the spatial location of Patch nodes for weighted compression and heatmap projection during anomaly heatmap generation, and introduces a risk level label matrix and a location modulation function to classify and adjust the heatmap. Finally, it outputs an early warning map with spatial distribution perception and risk classification capabilities, achieving accurate visualization of anomaly identification results and zonal risk warnings.

[0037] Example 1: To verify the feasibility of this invention in practice, it was applied to a high-density industrial power distribution environment where a large number of miniature circuit breakers operate continuously for extended periods. Some circuit breakers are prone to latent faults such as thermal fatigue and insulation degradation due to severe current fluctuations, abnormally high contact resistance, or poor heat dissipation. Traditional maintenance methods mainly rely on manual periodic inspections and threshold triggering mechanisms, which can only identify extreme anomalies and suffer from problems such as delayed early warning, inaccurate location, and high maintenance costs.

[0038] To address this pain point, the operation and maintenance system incorporates a deep learning-based method for anomaly identification and early warning of miniature circuit breakers proposed in this invention. First, a multi-modal acquisition module is deployed at the substation site to acquire real-time circuit breaker operating data, including parameters such as voltage, current, temperature, and contact resistance. Simultaneously, thermal imaging images are acquired. After time alignment and normalization preprocessing, the acquired data generates a modal alignment tensor, which is then fed into an improved PatchCore structure for deep analysis.

[0039] This method performs structured slicing of thermal image regions through a deformation-aware patch extraction module, and removes unstable patches by combining control point deformation functions. Then, the set of deformation-stable patches is input into a graph structure modeling module to construct a heterogeneous graph. Each patch node generates a graph structure embedding representation based on its spatial and texture features, and enters a topology perturbation scoring module and an information entropy-driven compression module to generate a sparse feature memory. Subsequently, in the inference stage, new samples are sequentially subjected to embedding comparison, topology perturbation scoring, graph attention scoring, and other operations. A scoring offset mechanism and a nonlinear fusion function are used to generate multidimensional anomaly scores. Finally, an anomaly heatmap and risk level labels are output, enabling accurate identification and risk level judgment of internal anomalies of circuit breakers.

[0040] To verify the performance of the method of this invention in actual operation, a set of comparative experiments were designed. The traditional threshold triggering method and the method of this invention were deployed on circuit breaker systems in the same type of scenario. Abnormal events captured during actual operation and normal operation samples were selected for identification and comparison. The statistical identification results are shown in Table 1. Table 1 Comparison and Analysis of Anomaly Detection Performance

[0041] The comparative data shows that traditional methods have significant blind spots in anomaly detection, especially in responding to latent thermal faults and nonlinear heating processes. In contrast, the method of this invention, by integrating multimodal sensing data with a graph-driven scoring mechanism, accurately distinguishes between normal fluctuations and abnormal trends, completing risk identification and location marking before internal circuit breaker faults cause physical damage. The system can issue warnings on average more than 4 hours in advance, with an accuracy rate approaching 99%, and significantly lower false alarm and missed alarm rates than traditional methods, effectively reducing manual intervention and wasted maintenance resources.

[0042] Furthermore, in multiple test samples, it was found that the abnormal heat map constructed by this method clearly marked the abnormal temperature rise area and the structural displacement area in the thermal imaging image. Combined with the risk level label, it is convenient for inspection personnel to directly identify key risk points and shorten the response time. During the continuous operation phase, the system has good stability and the scoring mechanism has adaptive capabilities. It can maintain the identification accuracy in environments with large sample changes and is suitable for diverse electrical environment deployments.

[0043] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for anomaly identification and early warning of miniature circuit breakers based on deep learning, characterized in that, Includes the following steps: S1. Collect operating data and thermal imaging images of miniature circuit breakers; S2. Perform time alignment and normalization processing on the running data and thermal imaging image to generate a modal alignment tensor; S3. Input the modality alignment tensor into the improved PatchCore structure, which includes a deformation-aware patch extraction module, a graph structure modeling module, a topological perturbation scoring module, and an information entropy-driven memory compression module. S4. Extract a set of deformation-stable patches using the deformation-aware patch extraction module; S5. Input the Patch set into the graph structure modeling module to construct a graph structure with Patch as nodes and generate a graph structure embedding representation; S6. The graph structure is embedded to represent the input topology perturbation scoring module to calculate the topology perturbation score. At the same time, the information entropy is calculated for the embedded representation of the Patch set to construct a sparse feature memory. S7. During the inference phase, the modal alignment tensor of the test sample is sequentially input into each module to obtain the graph structure embedding representation, topological perturbation score and graph attention change score of the test sample, and compare them with the sparse feature memory to calculate the multidimensional anomaly score. S8. Integrate the multi-dimensional anomaly scores to generate an anomaly heatmap and risk level labels, and output the anomaly location and early warning information of the miniature circuit breaker.

2. The method for anomaly identification and early warning of miniature circuit breakers based on deep learning according to claim 1, characterized in that, The operating data includes voltage signals, current signals, temperature signals, and contact resistance signals.

3. The method for anomaly identification and early warning of miniature circuit breakers based on deep learning according to claim 1, characterized in that, S3 specifically includes: S31. The modal alignment tensor is processed by the deformation-aware Patch extraction module to extract a set of deformation-stable patches. S32. Construct a graph structure for the Patch set using the graph structure modeling module, and generate a graph structure embedding representation; S33. The graph structure embedding representation is processed by the topology perturbation scoring module to obtain the topology perturbation score; S34. The embedded representation of the Patch set is processed by the information entropy-driven memory compression module to generate a sparse feature memory library.

4. The method for anomaly identification and early warning of miniature circuit breakers based on deep learning according to claim 1, characterized in that, S4 specifically includes: S41. Perform a sliding window operation in the modal alignment tensor to extract a set of patches with a fixed size and a fixed step size. Each patch is a sub-region in the tensor and is marked with the corresponding spatial location index. S42. Set up several control points evenly in each patch to form a control point set. Input the control point set into the control point offset prediction network and calculate the lateral and longitudinal offset of each control point. S43. Add the original coordinates of the control points to the offsets to construct a set of deformation control points, and establish a nonlinear deformation function based on the set of deformation control points: ； in, and Represents the original x and y coordinates of the pixels in the patch. This indicates the output coordinates after deformation. For affine terms, The total number of control points. For the first Deformation weight of each control point For the current pixel and the first The Euclidean distance between control points, function Denotes radial basis functions, defined as follows: ; S44. Map the pixel positions in each Patch according to the deformation function, use the interpolation method to obtain the mapped position values, and generate a set of Patches after deformation adjustment. S45. Combine the deformed patch with the original running data to form a fused patch set; S46. Select the patches that meet the deformation stability conditions from the fused patch set to form a deformation-stable patch set, and input it into the graph structure modeling module as modeling input.

5. The method for anomaly identification and early warning of miniature circuit breakers based on deep learning according to claim 1, characterized in that, S5 specifically includes: S51. Label each patch in the deformation-stable patch set according to its source and texture features. The types include thermal feature patch, sensor feature patch, edge patch and static patch. S52. Using each Patch as a node in the graph structure, construct connecting edges based on the spatial position and type label between nodes, construct a heterogeneous adjacency matrix, use Euclidean distance to calculate the edge weights between nodes of the same type, and use semantic cosine similarity to calculate the edge weights between nodes of different types. S53. Generate a heterogeneous graph structure tensor containing node features, type labels, and connection weights; S54. Input the heterogeneous graph structure tensor into the graph neural network model, select the graph convolution kernel according to the type label, and perform weighted feature aggregation based on adjacency relationship for each type of node to obtain the updated node representation; S55. Perform global average pooling on all updated node representations to generate graph structure embedding representations.

6. The method for anomaly identification and early warning of miniature circuit breakers based on deep learning according to claim 1, characterized in that, S6 specifically includes: S61. For each node in the graph structure embedding representation, calculate the adjacency change rate under different topological perturbations, and define the perturbation weight matrix. Each element Represents a node In a disturbed scenario, adjacent nodes Changes in connection strength; S62. Construct the tensor of node perturbation influence. The element is defined as: ； in, For nodes The perturbation score, and These represent the edge weights under normal and disturbed conditions, respectively. For node type similarity weights, This represents the total number of nodes in the graph; S63. The embedding representation of each patch in the deformation-stable patch set is input to the channel response evaluation module, and the mean response vector is calculated based on the channel dimension. With response probability distribution Construct information entropy scoring vector ; S64. Merge the perturbation scores of each patch node. Information entropy score Construct a joint scoring tensor ,in To integrate the weighting coefficients, the proportion of structural perturbation to response entropy is controlled; S65. Based on the joint score tensor, perform sparse sampling on the Patch node embedding representation to select the top-ranked nodes by joint score. The Patch node will construct a sparse feature memory based on the corresponding embedded representation.

7. The method for anomaly identification and early warning of miniature circuit breakers based on deep learning according to claim 1, characterized in that, Specifically, S7 includes: S71. Perform node-level distance comparison on the graph structure embedding representation generated for the test samples, and calculate the structural difference score based on the sparse feature memory, denoted as... ,in Number the Patch nodes; S72. Obtain the connectivity of each Patch node in the adjacency graph. Calculate the rating offset factor: ； in, Indicates the first The rating offset factor for each patch node. For adjustment coefficients, This represents the maximum connectivity among all nodes. S73. Extract the perturbation amplitude of the corresponding node from the topology perturbation score and the graph attention change score. With attention drift value Score of structural differences Together as input for fusion; S74. Calculate the final anomaly score using a nonlinear fusion function: ； in, Indicates the first Anomaly score for each patch node , , To integrate the weighting coefficients, satisfy the following conditions: ; S75, all The anomaly score tensor is reconstructed based on the spatial index of the Patch node, forming a multidimensional anomaly score.

8. The method for anomaly identification and early warning of miniature circuit breakers based on deep learning according to claim 1, characterized in that, S8 specifically includes: S81. Input the multidimensional anomaly score tensor into the two-dimensional heat map mapping network. Based on the spatial position index of each Patch node in the modal alignment tensor, perform a weighted compression operation on the multidimensional anomaly score to obtain a two-dimensional anomaly score matrix, and project it onto the thermal imaging image dimension to generate an initial anomaly heat map. S82. Perform piecewise normalization on the initial anomaly heatmap, and normalize the score values ​​in the score tensor. Map to a fixed color gradation range and set abnormal color encoding rules; S83. Based on the set risk interval threshold, perform grading processing on the abnormal heatmap and construct a risk level label matrix. Matrix elements Indicates the risk level of the corresponding location; S84, Introduction of Position Guiding Function Modulating the abnormal heatmap to adjust the risk response intensity between the edge and center regions, the modulation function is defined as follows: ； in, Indicates the coordinate position on the abnormal heatmap. Indicates the center location of the heatmap. This is the location distribution adjustment coefficient; S85. The modulated abnormal heat map and risk level label matrix are used as output results to output the abnormal location and early warning information of the miniature circuit breaker.

Images