Substation isolating switch state intelligent identification and fault diagnosis method and system based on attention mechanism

Abstract

This invention discloses a method and system for intelligent identification and fault diagnosis of disconnector switch status in substations based on an attention mechanism. The method includes: acquiring current signals, vibration signals, and infrared thermal images of the disconnector switch to construct a multimodal time-series data stream; extracting temporal and spatial features using a Transformer-CNN hybrid network, and performing feature interaction through cross-modal cross-attention; constructing a graph neural network to encode physical coupling relationships as graph edges for message passing, and outputting a fault feature map; outputting diagnostic results and thermal fault segmentation masks through multi-task output using a classification head and a segmentation head; predicting current features and fault trend residuals using a multi-granularity self-attention network; and finally performing reliable decision fusion through a Venn-Abers module to output a comprehensive fault type with a confidence interval. This invention significantly improves the accuracy and robustness of disconnector switch fault diagnosis through multimodal feature fusion, graph neural network physical relationship modeling, and reliable decision fusion.

Description

Technical Field

[0001] This invention belongs to the field of power system equipment condition monitoring and fault diagnosis technology, specifically relating to a method and system for intelligent identification and fault diagnosis of substation disconnect switch status based on attention mechanism. Background Technology

[0002] Disconnecting switches are core high-voltage switchgear in substations used for electrical isolation, and their operating status directly affects the safety and reliability of the power system. Disconnecting switches are exposed to harsh environments such as wind, sun, and rain for extended periods, making them prone to mechanical failures such as jamming, loose transmissions, incomplete opening and closing, and poor contact. These failures can lead to increased contact resistance, overheating of the conductive circuit, and in severe cases, equipment burnout or even power outages. Therefore, real-time status monitoring and fault diagnosis of disconnecting switches are of great importance.

[0003] Currently, fault diagnosis methods for disconnecting switches are mainly divided into three categories: methods based on traditional signal processing, methods based on single-modal deep learning, and methods based on multi-modal fusion. Methods based on traditional signal processing extract time-frequency domain features from current, vibration, or torque signals and combine them with classifiers such as support vector machines and backpropagation neural networks for fault identification. However, these methods rely on manual feature design and struggle to fully uncover the deep features of complex faults. Methods based on single-modal deep learning use convolutional neural networks or recurrent neural networks for end-to-end feature learning of current or vibration signals. However, the information content of a single signal source is limited, resulting in a low recognition rate for easily confused faults such as jamming and normal states. In recent years, some research has begun to attempt to fuse current and vibration signals for fault diagnosis or use infrared thermal imaging for thermal fault detection, but existing methods have the following shortcomings:

[0004] First, current signals, vibration signals, and infrared thermal images have different time scales and physical properties. Existing methods lack effective time alignment and multimodal fusion mechanisms, making it difficult to achieve collaborative modeling of temporal and spatial information. Second, there are clear physical coupling relationships between the components of a disconnector switch (e.g., the motor driving force is transmitted to the moving contact via a transmission link). However, existing deep learning methods only learn statistical correlations from the data, failing to embed prior knowledge of the equipment's physical structure into the model, resulting in poor model interpretability and insufficient generalization ability. Third, existing methods typically treat fault diagnosis and thermal fault segmentation as independent tasks, lacking a multi-task joint optimization mechanism. Furthermore, the diagnostic results lack confidence assessment, making it difficult to meet the reliability requirements for decision-making in engineering applications.

[0005] Therefore, how to construct an intelligent identification and fault diagnosis method for disconnector switch status that can integrate multimodal signals, embed physical prior knowledge, and output reliable diagnostic results is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for intelligent identification and fault diagnosis of the status of substation disconnect switches based on an attention mechanism. This invention aims to solve the technical problems in the prior art, such as the difficulty in effectively fusing multimodal signals, the difficulty in embedding physical prior knowledge into deep learning models, and the lack of confidence assessment of diagnostic results, so as to achieve high-precision and high-reliability intelligent diagnosis of disconnect switch faults.

[0007] In a first aspect, embodiments of this application provide a method for intelligent identification and fault diagnosis of the status of substation disconnect switches based on an attention mechanism, the method comprising:

[0008] S1. Collect the current signal, vibration signal and infrared thermal image of the disconnecting switch, unify the timestamps and align the events of the three, and construct a multimodal time-series data stream;

[0009] S2. Input the aligned current signal and vibration signal into the Transformer branch to capture long-term dependencies, and input the aligned infrared thermal image into the CNN branch to extract local spatial features. The two branches interact through a cross-modal cross attention module to output current-vibration fusion features and infrared spatial features.

[0010] S3. Using the current-vibration fusion features and infrared spatial features as graph nodes and the physical coupling relationship between disconnector components as graph edges, a graph neural network is constructed and multiple rounds of message passing are performed through a graph convolutional network to output a fault feature map;

[0011] S4. Input the fault feature map into the classification head and the segmentation head at the same time, and output the current diagnosis result, vibration diagnosis result and thermal fault segmentation mask;

[0012] S5. Using the current-vibration fusion features within the historical time window output in step S2 as input, the predicted current features and fault trend residuals are output through a multi-granularity self-attention prediction network.

[0013] S6. Input the current diagnosis results, vibration diagnosis results, thermal fault segmentation mask and predicted current characteristics into the Venn-Abers credible decision fusion module, and output a comprehensive fault type with a confidence interval.

[0014] Secondly, embodiments of this application provide an intelligent identification and fault diagnosis system for the status of substation disconnect switches based on an attention mechanism, applied to the method described in the first aspect, the system comprising:

[0015] The multimodal data acquisition and alignment module is used to acquire the current signal, vibration signal and infrared thermal image of the disconnecting switch, unify the timestamps and align the events of the three to construct a multimodal time-series data stream;

[0016] The hybrid feature extraction module is used to input the aligned current signal and vibration signal into the Transformer branch to capture long-term dependencies, and input the aligned infrared thermal image into the CNN branch to extract local spatial features. The two branches interact through a cross-modal cross attention module to output current-vibration fusion features and infrared spatial features.

[0017] The graph neural network fusion module is used to construct a graph neural network by taking the current-vibration fusion features and infrared spatial features as graph nodes and the physical coupling relationship between disconnecting switch components as graph edges, and then using a graph convolutional network to perform multi-round message passing to output a fault feature map.

[0018] The multi-task output module is used to simultaneously input the fault feature map into the classification head and the segmentation head, and output the current diagnosis result, vibration diagnosis result and thermal fault segmentation mask.

[0019] The multi-granularity prediction module is used to take the current-vibration fusion features within the historical time window output by the hybrid feature extraction module as input, and output the predicted current features and fault trend residuals through the multi-granularity self-attention prediction network.

[0020] The trusted decision fusion module is used to input the current diagnosis results, vibration diagnosis results, thermal fault segmentation mask and predicted current characteristics into the Venn-Abers trusted decision fusion module, and output a comprehensive fault type with a confidence interval.

[0021] Thirdly, embodiments of this application provide an electronic device, including:

[0022] processor;

[0023] Memory used to store processor-executable instructions;

[0024] The processor is configured to implement the attention-based intelligent identification and fault diagnosis method for substation disconnect switch status as described in the first aspect when executing the instructions.

[0025] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program that instructs a device to execute the attention-based intelligent identification and fault diagnosis method for substation disconnect switch status as described in the first aspect.

[0026] By unifying the timestamps and aligning events for current, vibration, and infrared signals, a multimodal time-series data stream is constructed, solving the problem of difficult collaborative processing of signals with different sampling frequencies. A Transformer-CNN hybrid network and a cross-modal cross-attention mechanism are employed to achieve effective interaction and fusion of temporal and spatial features. The physical coupling relationship of the mechanical transmission chain of the disconnector switch is innovatively encoded as graph edges, and a graph neural network is constructed for message passing, enabling the model to learn fault representations that conform to physical laws, thus improving the model's interpretability and generalization ability. By sharing the classification head and segmentation head structure of the fault feature map, collaborative optimization of fault type identification and thermal fault region segmentation is achieved, allowing the two tasks to mutually promote each other and improve overall diagnostic performance. A multi-granularity self-attention prediction network is used to predict current-vibration fusion features at multiple time scales, and early fault warnings are achieved through prediction residuals, saving valuable time for equipment maintenance. The Venn-Abers calibrator is introduced to perform probability calibration and confidence assessment on the diagnostic results of each mode. Combined with the physical fault evolution mechanism, evidence weighted fusion is performed to output a comprehensive fault type and fault evidence chain with confidence intervals, which significantly improves the reliability and interpretability of the diagnostic results. Attached Figure Description

[0027] Figure 1 This is a schematic flowchart of an attention-based method for intelligent identification and fault diagnosis of substation disconnect switch status, provided in an embodiment of this application.

[0028] Figure 2 This is an architecture diagram of a substation disconnector status intelligent identification and fault diagnosis system based on an attention mechanism, provided as an embodiment of this application.

[0029] Figure 3 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0030] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0031] It should be noted that in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0032] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0033] Example 1

[0034] Figure 1 This is a schematic flowchart illustrating an intelligent identification and fault diagnosis method for the status of a substation disconnector based on an attention mechanism, provided as an embodiment of this application. Figure 1 As shown, a method for intelligent identification and fault diagnosis of substation disconnect switch status based on an attention mechanism includes:

[0035] S1. Acquire current signals, vibration signals, and infrared thermal images from the disconnector switch, unify the timestamps and align the events of these three data sources, and construct a multimodal time-series data stream. Establish a unified time reference for the multi-source signals, enabling time-aligned processing of current, vibration, and infrared data.

[0036] Specifically, in this embodiment, the method for constructing the multimodal time-series data stream in step S1 includes:

[0037] Using the moment when the disconnector switch opening and closing action command is issued as the zero point of time, the current signal and vibration signal are timestamped according to the preset sampling frequency, and the infrared thermal image is timestamped according to the acquisition frame rate. The three are unified to the same time reference through linear interpolation. According to the physical timing of the disconnector switch action process, the opening and closing process is divided into three time windows: the start-up stage, the travel stage, and the arrival stage. The current-vibration-infrared multimodal data frame sequence in each window is constructed respectively.

[0038] This embodiment takes the disconnecting switch of a 220kV substation as an example to explain in detail the implementation method of constructing a multimodal time-series data stream in step S1.

[0039] First, a current transformer is installed inside the operating mechanism box of the disconnector switch to collect the three-phase current signal of the drive motor. A vibration acceleration sensor is installed on the surface of the moving contact operating box of the disconnector switch to collect vibration signals during the opening and closing process. An infrared thermal imaging camera is mounted directly in front of the disconnector switch to collect infrared thermal images of the equipment's operating status. The sampling parameters for the three sensors are set as follows: current signal sampling frequency is 10kHz, vibration signal sampling frequency is 12.8kHz, and infrared thermal image acquisition frame rate is 25Hz.

[0040] The moment the disconnector switch's opening or closing command is issued is taken as the zero point of time. When the control system issues an opening or closing command, three sensors are simultaneously triggered to begin recording data. Current and vibration signals are timestamped according to their respective sampling frequencies, meaning each sampling point records its time offset relative to the zero point. Infrared thermal images are timestamped according to the acquisition frame rate, meaning each frame records the time offset of its exposure start time relative to the zero point.

[0041] Because the three signals have different sampling frequencies, the collected data points are unevenly distributed along the time axis. To solve this problem, a linear interpolation method is used to unify the three signals to the same time reference. Specifically, using the sampling time point of the current signal as the reference time axis, linear interpolation is used to fill in the missing points of the vibration signal and infrared thermal image on this time axis, so that the three signals have corresponding data values ​​at every time point, thus forming a time-aligned multimodal data stream.

[0042] Then, based on the physical timing of the disconnector's operation, the entire opening and closing process is divided into three consecutive time windows:

[0043] The first window is the startup phase, which starts from zero time and ends when the moving contact begins to move. This phase reflects the startup characteristics of the drive motor and the initial clearance of the transmission mechanism.

[0044] The second window is the stroke stage, which corresponds to the period from when the moving contact starts to move until it approaches the stationary contact. This stage reflects the smoothness of the transmission mechanism's operation and the movement trajectory of the moving contact.

[0045] The third window is the positioning stage, which corresponds to the moving contact contacting the stationary contact until the movement completely stops. This stage reflects the impact characteristics of the contact and the stability after positioning.

[0046] Within the three time windows mentioned above, aligned current signal segments, vibration signal segments, and infrared thermal image frame sequences are extracted to form the multimodal data frame sequence corresponding to each window. Each data frame sequence simultaneously contains the current waveform, vibration waveform, and corresponding infrared thermal image within that time window, providing multimodal input data with a physical stage division basis for subsequent steps.

[0047] The above method achieves precise temporal alignment of multi-source signals with different sampling frequencies, and divides the window based on the physical stage of the action process, laying a data foundation for subsequent feature extraction and fusion.

[0048] S2. The aligned current signal and vibration signal are input into the Transformer branch to capture long-term dependencies, and the aligned infrared thermal image is input into the CNN branch to extract local spatial features. The two branches interact through a cross-modal cross-attention module, outputting current-vibration fused features and infrared spatial features. The Transformer extracts the temporal features of current vibration, and the CNN extracts the infrared spatial features. Cross-attention enables the interaction and fusion of the two, achieving cross-modal interaction and complementary enhancement of current / vibration temporal features and infrared spatial features.

[0049] Specifically, in this embodiment, the cross-modal attention module in step S2 is implemented as follows:

[0050] The current-vibration time series features output from the Transformer branch are used as query information, and the infrared spatial features output from the CNN branch are used as key and value information. Attention weights are generated by calculating the similarity between the query information and the key information. The value information is then weighted and aggregated using these weights to obtain fused features. The fused features are added to the original Transformer branch output and CNN branch output through residual connections to obtain enhanced current-vibration fused features and infrared spatial features.

[0051] In this embodiment, the cross-modal attention module in step S2 is implemented as follows:

[0052] First, a Transformer branch is constructed to process the current and vibration signals. The current and vibration signal segments from the multimodal time-series data stream constructed in step S1 are concatenated to form a two-dimensional input matrix, where each row represents a time step and each column represents a signal channel. The Transformer branch consists of multiple stacked encoder layers, each containing a multi-head self-attention sublayer and a feedforward neural network sublayer. After multi-layer encoding, the Transformer branch outputs current-vibration time-series features, which preserve the long-range dependence of current and vibration signals in the time dimension, characterizing the torque change process of the drive motor and the temporal distribution of mechanical impact.

[0053] Simultaneously, a CNN branch is constructed to process the infrared thermal images. The sequence of infrared thermal image frames from the multimodal data stream constructed in step S1 is input into the CNN branch. The CNN branch employs a multi-layer convolutional structure, with shallow convolutional kernels used to extract low-level features such as edges and textures, and deep convolutional kernels used to extract high-level semantic features such as hot spot regions and temperature gradients. After global average pooling, the CNN branch outputs infrared spatial features, which preserve the location distribution information of the thermal fault region in the image space.

[0054] Since current-vibration temporal features and infrared spatial features reside in the time domain and spatial domain respectively, their feature dimensions and physical meanings differ, making direct fusion difficult. Therefore, this invention designs a cross-modal cross-attention module to achieve the interaction and fusion of these two heterogeneous features.

[0055] In the cross-modal attention module, the current-vibration time-series features output from the Transformer branch are used as query information, and the infrared spatial features output from the CNN branch are used as key and value information. The specific processing flow is as follows:

[0056] The first step is to calculate the similarity between the query information and the key information. For each time step vector in the current-vibration time series features, the dot product similarity is calculated with all spatial location vectors in the infrared spatial features to obtain the attention weight matrix. This weight matrix reflects the degree of correlation between the current-vibration state at each moment and each spatial location in the image.

[0057] The second step involves weighted aggregation of the value information using attention weights. The attention weights calculated in the previous step are multiplied by the infrared spatial features to obtain the weighted aggregated fused features. This fused feature allows the spatial location information associated with each moment of the current-vibration features to be retrieved from the infrared spatial features.

[0058] The third step is residual connection enhancement. The fused features are added to the original Transformer branch output and the original CNN branch output respectively through residual connections. Specifically, the fused features are added element-by-element to the current-vibration temporal features to obtain enhanced current-vibration fused features, which incorporate infrared spatial information while retaining their own temporal information; the fused features are also added element-by-element to the infrared spatial features to obtain enhanced infrared spatial features, which incorporate current-vibration temporal information while retaining their own spatial information.

[0059] Through the aforementioned cross-modal cross-attention module, bidirectional information flow is achieved between current-vibration timing features and infrared spatial features, with the output enhanced features simultaneously containing complementary information from both modes. For example, when a disconnector experiences a mechanical jamming fault, abnormal peaks appear in the current waveform. This timing anomaly information, through cross-attention, guides the model to focus more on the spatial features of the transmission linkage region in the infrared image. Conversely, when a high-temperature hotspot appears in the moving contact region of the infrared image, this spatial information guides the model to focus more on the timing segment corresponding to increased contact resistance in the current waveform. This cross-modal interaction mechanism significantly improves the model's ability to identify complex faults.

[0060] As another optional implementation, the feature interaction between the Transformer branch and the CNN branch in step S2 adopts a bidirectional cross-attention mechanism: using the infrared spatial features output by the CNN branch as query information and the current-vibration time series features output by the Transformer branch as key and value information, a first-direction cross-attention feature is calculated to achieve the attention of infrared features to the current-vibration time series features; simultaneously, using the current-vibration time series features output by the Transformer branch as query information and the infrared spatial features output by the CNN branch as key and value information, a second-direction cross-attention feature is calculated to achieve the attention of current-vibration time series features to infrared features; the cross-attention features of the first and second directions are weighted and fused to output the bidirectional enhanced current-vibration fused features and infrared spatial features.

[0061] In this embodiment, step S2 employs a bidirectional cross-attention mechanism to achieve feature interaction between the Transformer branch and the CNN branch. Unlike unidirectional cross-attention, this embodiment uses a bidirectional cross-attention mechanism, enabling the current-vibration temporal features and infrared spatial features to mutually influence and enhance each other in two directions, achieving deeper multimodal information fusion.

[0062] First, the cross-attention feature in the first direction is calculated, which realizes the attention of infrared spatial features to current-vibration timing features.

[0063] The infrared spatial features output by the CNN branch are used as query information. An infrared spatial feature is a three-dimensional feature tensor, whose height and width correspond to the spatial dimensions of the infrared image, and the number of channels corresponds to different semantic features. Each spatial location in this feature represents the temperature distribution and hotspot information of a local region in the infrared image.

[0064] The current-vibration timing features output from the Transformer branch are used as key and value information. The current-vibration timing feature is a two-dimensional feature matrix, where the time dimension corresponds to the sampling time of the opening and closing process, and the channel dimension corresponds to the waveform characteristics of the current and vibration signals. Each time step in this feature represents the motor current value and vibration amplitude at that moment.

[0065] In the first-direction cross-attention calculation, for each spatial location in the infrared spatial features, the similarity between its feature vector and the current-vibration time-series feature vectors of all time steps is calculated one by one to obtain the attention weight of that spatial location for each time step. The larger the attention weight, the closer the correlation between the thermal fault features of that spatial location and the current-vibration state at the corresponding time. Then, these attention weights are used to perform a weighted summation of the current-vibration time-series features of all time steps to obtain the first-direction cross-attention feature of that spatial location. After traversing all spatial locations, a complete first-direction cross-attention feature map is formed. This feature map enables each spatial region of the infrared image to obtain its associated time-series information. For example, when a high-temperature hotspot appears in the moving contact region of the infrared image, the features of that region will focus on the time period corresponding to the increase in contact resistance in the current waveform.

[0066] Then, the cross-attention feature in the second direction is calculated, which realizes the attention of the current-vibration time sequence feature to the infrared spatial feature.

[0067] The current-vibration time-series features output by the Transformer branch are used as query information. For each time step in the current-vibration time-series features, the similarity between its feature vector and the feature vectors of all spatial locations in the infrared spatial features is calculated one by one to obtain the attention weight of each spatial location at that time step. The larger the attention weight, the closer the correlation between the current-vibration state at that moment and the thermal fault features of the corresponding spatial location in the infrared image. Then, these attention weights are used to perform a weighted summation of the infrared spatial features of all spatial locations to obtain the second-direction cross-attention features for that time step. After traversing all time steps, a complete second-direction cross-attention feature sequence is formed. This feature sequence enables each time point of the current-vibration waveform to obtain its associated spatial information. For example, when an abnormal peak appears in the current waveform, the features at that moment will focus on whether there is a thermal anomaly in the transmission link or moving contact area in the infrared image.

[0068] Finally, the cross-attention features from the first and second directions are weighted and fused to output the bidirectional enhanced features.

[0069] The specific fusion method is as follows: The first-direction cross-attention feature map is weighted and superimposed with the original infrared spatial features by channel to obtain a bidirectional enhanced infrared spatial feature. Each spatial location in this feature retains the original temperature distribution information and incorporates time-series-related current-vibration state information. Simultaneously, the second-direction cross-attention feature sequence is weighted and superimposed with the original current-vibration time-series features by time step to obtain a bidirectional enhanced current-vibration fused feature. Each time step in this feature retains the original waveform information and incorporates spatially related thermal fault distribution information.

[0070] Through the aforementioned bidirectional cross-attention mechanism, the current-vibration temporal characteristics and infrared spatial characteristics achieve complete bidirectional information flow. Taking a mechanism jamming fault as an example, when the current waveform shows a jamming characteristic peak, the second-direction attention guides the model to focus on the transmission linkage region in the infrared image; when frictional heat generation occurs in this region in the infrared image, the first-direction attention guides the model to review the exact moment of jamming in the current waveform. The attention in the two directions mutually verifies and reinforces each other, enabling the model to simultaneously confirm the existence of the fault from both temporal and spatial dimensions, significantly improving the accuracy and robustness of the diagnosis.

[0071] S3. Using the current-vibration fusion features and infrared spatial features as graph nodes, and the physical coupling relationships between disconnector components as graph edges, a graph neural network is constructed and multi-round message passing is performed through a graph convolutional network to output a fault feature map. The physical relationship of the mechanical transmission chain of the disconnector is modeled as a graph structure, and fault information is aggregated through message passing between nodes.

[0072] Specifically, in this embodiment, the method for constructing graph nodes in step S3 includes:

[0073] The current-vibration fusion feature is decomposed into multiple time sub-nodes, each time sub-node corresponding to a timing stage in the disconnector switch operation process, the timing stage including the start-up stage, the travel stage, and the arrival stage; the infrared spatial feature is decomposed into multiple spatial sub-nodes, each spatial sub-node corresponding to a physical component area of ​​the disconnector switch, the physical component area including the drive motor area, the transmission linkage area, the moving contact area, and the stationary contact area; all time sub-nodes and spatial sub-nodes are used together as the graph node set of the graph neural network.

[0074] In this embodiment, the specific implementation of step S3, which decomposes the current-vibration fusion features and infrared spatial features into graph nodes, is as follows:

[0075] After completing cross-modal bidirectional cross-attention fusion, this embodiment obtains two types of enhanced features: one is the bidirectionally enhanced current-vibration fusion feature, and the other is the bidirectionally enhanced infrared spatial feature. In order to embed the physical structure knowledge of the disconnecting switch into the deep learning model, this invention reorganizes these two types of features into a graph structure, where graph nodes represent feature units with different physical meanings.

[0076] First, the current-vibration fusion characteristics are decomposed into time sub-nodes.

[0077] The current-vibration fusion feature is a two-dimensional feature matrix whose time dimension covers the entire opening and closing process. Based on the three time windows divided in step S1, the feature matrix is ​​divided into three continuous feature segments in the time dimension, with each feature segment corresponding to a timing stage.

[0078] The first time sub-node corresponds to the startup phase. This sub-node encapsulates the current and vibration characteristics from the issuance of the opening / closing command to the start of movement of the moving contact. During this phase, the drive motor starts and overcomes static friction, the current waveform exhibits a startup spike, and the vibration signal is weak. The characteristics of this sub-node can reflect the startup characteristics of the motor and the initial clearance state of the transmission mechanism.

[0079] The second time sub-node corresponds to the travel phase. This sub-node encapsulates the current and vibration characteristics from the start of the moving contact's movement to its approach of the stationary contact. During this phase, the motor drives the transmission linkage, the current waveform is relatively stable, and the vibration signal reflects the smoothness of the transmission mechanism's operation. The characteristics of this sub-node can reflect the friction state of the transmission mechanism and the movement trajectory of the moving contact.

[0080] The third time sub-node corresponds to the positioning stage. This sub-node encapsulates the current and vibration characteristics between the moving contact making contact with the stationary contact and the complete cessation of movement. During this stage, the moving contact impacts the stationary contact, resulting in an impact peak in the current waveform and severe vibration. The characteristics of this sub-node can reflect the impact characteristics of the contact and the stability after positioning.

[0081] Through the above decomposition, the three time sub-nodes represent the timing characteristics of three different physical stages in the process of disconnecting switch operation.

[0082] Then, the infrared spatial features are decomposed into spatial sub-nodes.

[0083] The infrared spatial feature is a three-dimensional feature tensor whose spatial extent covers the entire infrared image of the disconnector. Based on the mechanical structure of the disconnector and the locations prone to thermal failures, the infrared spatial feature is divided into four feature regions in space, with each feature region corresponding to a physical component.

[0084] The first spatial sub-node corresponds to the drive motor region. This sub-node encapsulates the spatial features of the drive motor's location in the infrared image, reflecting the temperature distribution and heat dissipation status of the motor windings. Motor overheating is often an indirect manifestation of abnormal drive current or bearing jamming.

[0085] The second spatial sub-node corresponds to the transmission link area. This sub-node encapsulates the spatial features of the transmission link mechanism's location in the infrared image, reflecting the frictional heating at the link joints. Link jamming or poor lubrication can cause the temperature in this area to rise.

[0086] The third spatial sub-node corresponds to the moving contact region. This sub-node encapsulates the spatial features of the moving contact's location in the infrared image, reflecting the temperature distribution of the moving contact's contact surface. Poor contact of the moving contact can lead to increased contact resistance and localized overheating.

[0087] The fourth spatial sub-node corresponds to the stationary contact region. This sub-node encapsulates the spatial features of the stationary contact location in the infrared image, reflecting the temperature distribution of the stationary contact surface. Oxidation or ablation of the stationary contact can also lead to localized overheating.

[0088] Through the above decomposition, the four spatial sub-nodes represent the spatial characteristics of the four key physical components of the disconnector switch.

[0089] Finally, the three temporal sub-nodes and four spatial sub-nodes are combined to form the graph node set of the graph neural network, totaling seven graph nodes. Each graph node is a high-dimensional feature vector, where the temporal sub-nodes carry fault information in the temporal dimension, and the spatial sub-nodes carry fault information in the spatial dimension. These nodes constitute the basic unit for subsequent message passing in the graph neural network, laying the foundation for establishing connections between physical components and action phases.

[0090] Specifically, in this embodiment, the construction of graph edges and message passing in step S3 includes: based on the mechanical transmission chain of the disconnector, the torque transmission direction and contact relationship between the drive motor, transmission link, moving contact, stationary contact, and insulating tie rod are used as graph edges to connect the corresponding time sub-nodes and spatial sub-nodes; in each round of message passing in the graph convolutional network, each graph node receives the feature information transmitted by its neighboring nodes, and performs weighted aggregation of the neighboring information according to the physical coupling strength to update its own node features; after multiple rounds of message passing, the features of all graph nodes are aggregated into a unified fault feature map, which simultaneously contains the temporal state information and spatial distribution information of the disconnector.

[0091] In this embodiment, the method for constructing graph edges in step S3 and the specific implementation process of multi-round message passing using the graph convolutional network are as follows:

[0092] First, construct graph edges based on the mechanical transmission chain of the disconnecting switch, and connect the time sub-nodes and the space sub-nodes.

[0093] The mechanical transmission chain of the disconnecting switch has a clearly defined torque transmission direction: the drive motor generates driving torque, which is transmitted to the moving contact through the transmission link. The moving contact contacts the stationary contact to complete the closing, or separates from the stationary contact to complete the opening. Furthermore, an insulating rod connects the transmission link and the moving contact, providing insulation. Based on this physical structure, the following connection relationship is established in this embodiment:

[0094] Connect the drive motor spatial sub-node to the startup phase time sub-node. This is because the startup characteristics of the drive motor directly affect the current waveform during startup, and motor overheating faults mainly occur during startup.

[0095] Connect the spatial sub-node of the transmission link to the time sub-node of the stroke stage. This is because the transmission link moves continuously during the stroke stage, and faults such as friction and jamming of the link will be reflected in the current and vibration signals during the stroke stage. At the same time, the frictional heat generated during this stage will also be reflected in the infrared image of the transmission link area.

[0096] Connect the moving contact space sub-node and the stationary contact space sub-node to the positioning stage time sub-node. This is because the contact between the moving contact and the stationary contact occurs during the positioning stage. Faults such as poor contact and increased contact resistance will be reflected in the current impact waveform and vibration signal during the positioning stage, and contact heating is also concentrated in this stage.

[0097] The spatial sub-nodes of the insulating tie rod are simultaneously connected to the time sub-nodes of both the travel and arrival phases. This is because the insulating tie rod is responsible for transmitting motion during the travel phase and bears the contact pressure of the contact head during the arrival phase; its mechanical state affects the characteristics of both phases.

[0098] Through the above connections, each edge in the graph neural network represents the physical coupling relationship between the disconnecting switch component and the action phase.

[0099] Then, multiple rounds of message passing are performed using the graph convolutional network.

[0100] In the first round of message passing, each graph node collects feature information from its neighboring nodes. Taking the spatial sub-node of the transmission link as an example, it is connected to the time sub-node of the stroke stage. In the first round of message passing, the spatial sub-node of the transmission link receives the current-vibration timing features from the time sub-node of the stroke stage. These features contain information on the stability of the motor current and the fluctuation of the vibration signal during the stroke stage. Simultaneously, the time sub-node of the stroke stage also receives infrared spatial features from the spatial sub-node of the transmission link. These features contain information on the temperature distribution in the transmission link region. Through this bidirectional message passing, timing and spatial features begin to flow between nodes.

[0101] The second round of message passing involves each graph node weighting and aggregating neighbor information based on physical coupling strength. Physical coupling strength is pre-defined based on the actual mechanical relationships between components: the coupling strength between the drive motor and the starting phase is high because the motor's driving characteristics completely determine the current waveform during starting; the coupling strength between the transmission link and the stroke phase is medium because the link's state influences, but does not completely determine, the characteristics of the stroke phase; the coupling strength between the moving contact, stationary contact, and the positioning phase is high because the contact state is a core feature of the positioning phase. When aggregating neighbor information, nodes assign larger aggregation weights to neighbors with high coupling strength and smaller aggregation weights to neighbors with low coupling strength, allowing the model to selectively integrate information from different sources. After aggregation, each node updates its feature vector based on the newly received information.

[0102] The third round of message transmission: After the first two rounds of transmission, the information has been sufficiently disseminated in the diagram. This embodiment performs a third round of message transmission, enabling information to propagate across multiple physical stages. For example, an abnormality in the drive motor will be transmitted from the start-up phase time sub-node to the transmission link space sub-node, then from the transmission link space sub-node to the travel phase time sub-node, ultimately affecting the arrival phase time sub-node. This cross-stage information propagation allows the model to capture the characteristics of cascading failures, such as motor jamming leading to increased stress on the transmission link, which in turn causes link overheating and poor contact at the contacts.

[0103] After three rounds of message passing, the features of all graph nodes have been fully integrated with temporal and spatial information. At this point, the features of all graph nodes are stitched together along the channel dimension to form a unified fault feature map. This fault feature map simultaneously includes the following information: current-vibration timing characteristics of the startup, travel, and arrival phases; infrared spatial characteristics of the drive motor, transmission linkage, moving contact, stationary contact, and insulating tie rod areas; and physical coupling information between each component and each phase. This fault feature map provides a comprehensive, structured, and physically consistent fault representation for subsequent classification and segmentation tasks.

[0104] S4. Simultaneously input the fault feature map into the classification head and the segmentation head, and output the current diagnosis result, vibration diagnosis result, and thermal fault segmentation mask. This achieves synchronous output and joint optimization of fault type identification and thermal fault area location.

[0105] Specifically, in this embodiment, the specific implementation of the classification head and segmentation head in step S4 includes: the classification head adopts a multi-label classification structure, simultaneously outputting a first fault category probability distribution corresponding to the current signal and a second fault category probability distribution corresponding to the vibration signal for the input fault feature map, wherein the first fault category and the second fault category are independent but share the same fault feature map; the segmentation head adopts a pixel-by-pixel classification structure, predicting a binary label for each pixel in the input fault feature map as belonging to a thermal fault region or a non-thermal fault region, and outputting a thermal fault segmentation mask with the same resolution as the input infrared thermal image; the classification head and the segmentation head share the fault feature map as a common input, and are collaboratively optimized through a multi-task joint loss function during training, so that the classification task and the segmentation task promote each other.

[0106] In this embodiment, the specific implementation of step S4, in which the fault feature map is simultaneously input into the classification head and the segmentation head, and the current diagnosis result, vibration diagnosis result, and thermal fault segmentation mask are output, is as follows:

[0107] After completing multiple rounds of message passing in the graph neural network, this embodiment obtains a unified fault feature map. This fault feature map includes the three-stage timing information of the disconnector in the time dimension, the four-component region information in the spatial dimension, and the physical coupling relationship information between each component and each stage. In order to make full use of this rich feature representation, this embodiment designs a parallel classification head and segmentation head structure to achieve synchronous output of fault type identification and thermal fault region location.

[0108] The specific implementation method of the classification header is as follows:

[0109] The classification head employs a multi-label classification structure. Its input is a one-dimensional feature vector obtained by global pooling of the fault feature map, and its output is a probability distribution of multiple independent fault categories. Unlike traditional single-label classification, multi-label classification allows the same input to belong to multiple fault categories simultaneously, which is consistent with the situation where multiple faults may occur concurrently in the actual operation of disconnect switches.

[0110] The classification head contains two independent classification branches: the first branch outputs the fault category probability distribution corresponding to the current signal, and the second branch outputs the fault category probability distribution corresponding to the vibration signal. Each branch is independent, meaning it has its own fully connected layer and output layer, but they share the same input fault feature map.

[0111] The current diagnostic results output by the first classification branch include the probabilities of the following fault categories: normal state, mechanism jamming, three-phase asynchrony, mechanism looseness, and incomplete closing. This branch mainly judges based on the timing characteristics related to the current waveform in the fault feature diagram, such as the current peak during the start-up phase, the current stability during the travel phase, and the current surge characteristics during the closing phase.

[0112] The vibration diagnosis results output by the second classification branch include the same fault category probability distribution: normal state, mechanism jamming, three-phase asynchrony, mechanism loosening, and incomplete closing. This branch mainly judges based on the temporal characteristics related to the vibration signal in the fault feature diagram, such as the vibration amplitude during the start-up phase, the vibration continuity during the stroke phase, and the vibration impact intensity during the arrival phase.

[0113] Although the two classification branches output the same set of fault categories, their diagnostic results may differ due to the varying sensitivities of current and vibration signals to the same fault. For example, a mechanism jamming fault manifests as a significant peak anomaly in the current waveform and as increased impact in the vibration signal; both are identifiable. Conversely, a mechanism loosening fault is more pronounced in the vibration signal but relatively subtle in the current waveform. After the two branches output independently, the subsequent decision fusion module will perform a comprehensive judgment based on both.

[0114] The specific implementation method of the segmentation head is as follows:

[0115] The segmentation head employs a pixel-by-pixel classification structure. Its input is the fault feature map itself, and its output is a thermal fault segmentation mask with the same resolution as the input infrared thermal image. Unlike the classification head, the segmentation head needs to independently determine each spatial location in the fault feature map, identifying whether the pixel belongs to a thermal fault region or a non-thermal fault region.

[0116] The segmentation head consists of multiple layers of transposed convolutions, progressively restoring the spatial resolution of the fault feature map to the size of the original infrared thermal image. During the upsampling process at each layer, the segmentation head fuses edge information from shallow feature maps and semantic information from deep feature maps through skip connections, ensuring clear and accurate segmentation boundaries.

[0117] The output of the segmentation head is a binary mask image, where each pixel has a value of either 0 or 1. A pixel with a value of 1 indicates that the location belongs to a thermal fault area, corresponding to the abnormally heated part of the disconnector switch; a pixel with a value of 0 indicates that the location belongs to a non-thermal fault area, corresponding to a normal temperature area or the background. This mask image has the same width and height as the original infrared thermal image, enabling precise indication of the spatial location of the thermal fault.

[0118] For example, when a fault occurs in the moving contact of a disconnecting switch, the mask output by the splitter will form a bright area at the corresponding position of the moving contact in the infrared image, clearly indicating the specific location and range of the thermal fault.

[0119] The collaborative optimization mechanism between the classification head and the segmentation head is as follows:

[0120] The classification head and segmentation head share the same fault feature map as input, and are collaboratively optimized during training using a multi-task joint loss function. The joint loss function consists of three parts: current classification loss, vibration classification loss, and segmentation loss. The weighted sum of these three losses is used as the total loss, and backpropagation simultaneously optimizes the classification head, segmentation head, and preceding Transformer, CNN, and graph neural network branches.

[0121] This collaborative optimization mechanism enables the classification and segmentation tasks to mutually reinforce each other. On one hand, the segmentation head provides precise information on the location of thermal faults, helping the classification head to more accurately determine the fault type. For example, when the segmentation head detects a high-temperature hotspot in the moving contact area, the classification head will tend to output a fault category related to overheating when determining the fault type. On the other hand, the fault type information provided by the classification head helps the segmentation head to more accurately locate the thermal fault area. For example, when the classification head determines that the mechanism is stuck, the segmentation head will pay more attention to the thermal anomalies in the transmission linkage area rather than the moving contact area. Through this two-way reinforcement, the classification and segmentation tasks jointly improve the overall diagnostic performance.

[0122] S5. Using the current-vibration fusion features within the historical time window output in step S2 as input, a multi-granularity self-attention prediction network is used to output predicted current features and fault trend residuals. Based on historical data, the future trend of current change is predicted, and early fault signs are identified through the prediction residuals.

[0123] Specifically, in this embodiment, the implementation of the multi-granularity self-attention prediction network in step S5 includes: segmenting the current-vibration fusion features within the historical time window output in step S2 according to three time scales: fine-grained, medium-grained, and coarse-grained. Fine-grained corresponds to local waveform features within a single sampling period, medium-grained corresponds to trend features within a single action phase of the disconnector switch, and coarse-grained corresponds to the overall evolution features of the complete opening or closing process. Self-attention weights are independently calculated for each of the three time scale segments, allowing the network to focus on important time step features within each scale at each granularity. The self-attention features calculated at the three granularities are then fused across scales, allowing features from different time scales to complement and enhance each other, outputting predicted current features. The difference between the predicted current features and the actual acquired current signal is used as the fault trend residual, which characterizes the degree of abnormal deviation of the current signal and serves as the basis for early fault warning.

[0124] In this embodiment, the specific implementation method of using the current-vibration fusion features within the historical time window as input and outputting the predicted current features and fault trend residuals through a multi-granularity self-attention prediction network in step S5 is as follows:

[0125] After completing the multi-task output, this embodiment further utilizes historical data to predict future current characteristics in order to achieve early fault warning. The input to the prediction network is the current-vibration fusion feature within the historical time window output in step S2, which includes the current waveform and vibration waveform information of the disconnecting switch over a past period.

[0126] First, the input current-vibration fusion features are segmented according to three different time scales to form three feature representations: fine-grained, medium-grained, and coarse-grained.

[0127] Fine-grained segmentation corresponds to local waveform features within a single sampling period. The sampling frequency of current signals is typically 10 kHz, with each sampling period lasting 0.1 milliseconds. Fine-grained segmentation divides the continuous current-vibration characteristics into segments based on single sampling periods, with each segment containing waveform details within only one period. Features at this granularity can capture instantaneous changes in the current waveform, such as the precise amplitude of a startup spike or the steepness of the rise edge of an impulse pulse. Fine-grained features are sensitive to transient faults but are susceptible to noise interference.

[0128] Medium-granularity segmentation corresponds to the trend characteristics within a single operating phase of the disconnector switch. Based on the three time windows (start-up phase, travel phase, and arrival phase) defined in step S1, medium-granularity segmentation treats all sampling points within each phase as a single segment. Each segment contains hundreds to thousands of sampling points within that phase, reflecting the overall trend of the current waveform within that phase. Examples include the overall trend of the current rising from zero to its peak and then falling back during the start-up phase, the stability trend of the current during the travel phase, and the attenuation trend of the current after its initial surge during the arrival phase. Medium-granularity characteristics are sensitive to phase-specific faults and have strong anti-interference capabilities.

[0129] Coarse-grained segmentation corresponds to the overall evolution characteristics of the complete opening or closing process. It treats all sampling points of the entire opening and closing process as a single segment without any further division. This segment includes the entire waveform from command issuance to action completion, reflecting the overall evolution of the current-vibration characteristics. Examples include the "one large, two small" waveform of the current during closing and the decreasing waveform of the current during opening. Coarse-grained segmentation is sensitive to overall faults and has the strongest anti-interference capability, but it loses local detail information.

[0130] Through the above three granularity segmentation, the same current-vibration fusion feature is represented as a feature sequence at three different time scales, which describe the operating status of the equipment from three levels: local details, stage trends, and overall evolution.

[0131] Then, the self-attention weights are calculated independently for the segmentation features at the three time scales.

[0132] For fine-grained feature sequences, the self-attention mechanism calculates the correlation between each sampling period and all other sampling periods. For example, there may be a correlation between the first and last sampling periods of the startup phase, as they correspond to the current rise and fall processes, respectively. Through self-attention weights, the fine-grained network can focus on important time-step features within this scale, such as the peak point of current spikes or the starting point of vibration shocks.

[0133] For medium-granularity feature sequences, the self-attention mechanism calculates the degree of correlation between each action phase and other action phases. For example, the peak current height in the initiation phase may be correlated with the current stability in the travel phase, and the vibration fluctuations in the travel phase may be correlated with the impact intensity in the arrival phase. Through self-attention weights, the medium-granularity network can focus on important phase features within this scale, such as the stable segment of the travel phase and the impact segment of the arrival phase.

[0134] For coarse-grained feature sequences, the self-attention mechanism calculates the long-range dependencies between different time segments throughout the entire opening and closing process. For example, the motor starting characteristics at the initial closing stage and the contact characteristics at the end of the closing stage may have a causal relationship. Through self-attention weights, the coarse-grained network can focus on important overall features within this scale, such as the position of the main peak and the proportion of secondary peaks in a "one large and two small" waveform.

[0135] The three granular networks independently compute their self-attention weights and do not share parameters with each other, allowing each granularity to focus on learning the optimal feature representation at that scale.

[0136] Next, the self-attention features calculated at the three granularities are fused across scales.

[0137] Cross-scale fusion employs a step-by-step fusion strategy: First, fine-grained features are fused with medium-grained features. Specifically, the local waveform features output by the fine-grained network are grouped according to their respective stages. The average value within each group is then added element-wise to the stage trend features output by the medium-grained network, integrating the fine-grained local details into the medium-grained stage features. Next, the fused features are fused with coarse-grained features. The global average value of the medium-grained features across the stage dimension is then concatenated with the coarse-grained features, integrating the stage trend information into the overall evolutionary features.

[0138] Through this cross-scale fusion, features from different time scales complement and enhance each other: fine-grained features provide local detail corrections for medium-grained features, medium-grained features provide stage division criteria for coarse-grained features, and coarse-grained features provide global context constraints for both fine-grained and medium-grained features. The final output is a predicted current feature, which is a prediction of the current waveform within a future time window.

[0139] Finally, the fault trend residual is calculated as a basis for early fault warning.

[0140] The predicted current characteristics are compared point-by-point with the actual acquired current signal within the same time window, and the difference between the two is calculated. This difference is the fault trend residual, which is used to characterize the degree of abnormal deviation of the current signal.

[0141] When the disconnector is in normal operation, the predicted current characteristics closely match the actual current signal, with a small and stable residual value near zero. However, when a fault begins to develop but has not yet fully manifested, the actual current signal starts to deviate from the normal pattern, while the predicted current characteristics remain based on historical normal patterns, leading to a gradual increase in the residual between the two. For example, in the early stages of a mechanism jamming fault, the friction of the transmission linkage gradually increases, but not yet to the extent that it significantly affects the current waveform. At this time, the amplitude of the actual current waveform begins to rise slowly, while the predicted current characteristics still predict a lower amplitude based on historical normal data, resulting in a positive deviation in the residual.

[0142] When the residual continues to increase and exceeds a preset threshold, the system issues an early fault warning signal. The warning signal includes the following information: the time point when the anomaly began, the trend of the residual change (increasing or decreasing), and the most significant current characteristic type of the anomaly (amplitude, phase, or waveform). Maintenance personnel can use the warning signal to schedule repairs in advance and eliminate potential problems before the fault worsens.

[0143] Taking a mechanism jamming fault as an example, the change process of the fault trend residual is as follows: In the early stage of the fault, the residual begins to show a slight positive deviation during the stroke stage; as the jamming worsens, the deviation of the residual during the stroke stage gradually increases, and at the same time, the residual during the positioning stage also begins to show abnormalities; when the residual exceeds the warning threshold, the system issues a warning message: "The transmission mechanism is suspected of jamming, and maintenance is recommended." In this way, this embodiment achieves early detection and trend prediction of disconnector switch faults.

[0144] S6. Input the current diagnosis results, vibration diagnosis results, thermal fault segmentation mask, and predicted current characteristics into the Venn-Abers reliable decision fusion module, and output a comprehensive fault type with confidence intervals. Perform probability calibration and conflict resolution on each diagnostic result, and output a comprehensive diagnostic conclusion with reliability assessment.

[0145] Specifically, in this embodiment, the Venn-Abers trusted decision fusion module in step S6 is implemented as follows:

[0146] The current diagnostic results, vibration diagnostic results, thermal fault segmentation mask, and predicted current characteristics are used as four inputs, which are fed into the Venn-Abers calibrator for probability calibration. The calibrated fault probability and its confidence interval are output for each diagnostic result. When the fault type judgments of the four diagnostic results are consistent, the fault type and the overall confidence level are directly output. When there is a conflict in the fault type judgments of the four diagnostic results, the evidence is weighted according to the physical fault evolution mechanism of the disconnecting switch: if the thermal fault segmentation mask identifies a high-temperature area and the current diagnostic result is abnormal, the overheating-related fault type is given priority; if the vibration diagnostic result identifies mechanical impact characteristics and the current diagnostic result waveform is abnormal, the mechanical jamming-related fault type is given priority; if the predicted current characteristic residual continues to increase but the current diagnostic results are all normal, an early warning signal and the predicted fault trend are output. The finally determined fault type and its confidence interval are output as the comprehensive fault type, and a fault evidence chain containing evidence from each source is generated for fault tracing and maintenance decision-making.

[0147] Example 2

[0148] like Figure 2 As shown, this application provides an architecture diagram of a substation disconnector status intelligent identification and fault diagnosis system based on an attention mechanism, applied to the method described in Embodiment 1, including:

[0149] The multimodal data acquisition and alignment module 210 is used to acquire the current signal, vibration signal and infrared thermal image of the disconnecting switch, unify the timestamps and align the events of the three to construct a multimodal time-series data stream.

[0150] The hybrid feature extraction module 220 is used to input the aligned current signal and vibration signal into the Transformer branch to capture long-term dependencies, and input the aligned infrared thermal image into the CNN branch to extract local spatial features. The two branches interact through a cross-modal cross-attention module to output current-vibration fusion features and infrared spatial features.

[0151] The graph neural network fusion module 230 is used to construct a graph neural network by taking the current-vibration fusion features and infrared spatial features as graph nodes and the physical coupling relationship between disconnecting switch components as graph edges, and then using a graph convolutional network to perform multi-round message passing to output a fault feature map.

[0152] The multi-task output module 240 is used to simultaneously input the fault feature map into the classification head and the segmentation head, and output the current diagnosis result, vibration diagnosis result and thermal fault segmentation mask.

[0153] The multi-granularity prediction module 250 is used to take the current-vibration fusion features within the historical time window output by the hybrid feature extraction module as input, and output the predicted current features and fault trend residuals through the multi-granularity self-attention prediction network.

[0154] The credible decision fusion module 260 is used to input the current diagnosis results, vibration diagnosis results, thermal fault segmentation mask and predicted current characteristics into the Venn-Abers credible decision fusion module, and output a comprehensive fault type with a confidence interval.

[0155] Figure 3 This is an electronic device provided in one embodiment of this application. For example... Figure 3 As shown, the electronic device includes at least the following components: processor 301 and memory 300, communication interface 303, and bus 302.

[0156] In this embodiment of the application, memory 300 is used to store executable instructions of processor 301, which, when configured to execute instructions, implements the method as described in the first aspect.

[0157] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0158] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Information processed by these systems is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (FlashROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0159] It should be noted that a portion of the electronic device described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0160] It should be noted that the computer mentioned here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, computer-readable recording media refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage systems such as hard drives built into the computer.

[0161] Furthermore, computer-readable recording media can include: media that dynamically stores programs for short periods of time, such as communication lines used when transmitting programs via networks like the Internet or communication lines like telephone lines; and media that store programs for fixed periods of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining them with programs already recorded in the computer.

[0162] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (system group) composed of multiple systems. Each system constituting the system group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a system group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0163] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A method for intelligent identification and fault diagnosis of the status of disengaging switches in substations based on an attention mechanism, characterized in that, Includes the following steps: S1. Collect the current signal, vibration signal and infrared thermal image of the disconnecting switch, unify the timestamps and align the events of the three, and construct a multimodal time-series data stream; S2. Input the aligned current signal and vibration signal into the Transformer branch to capture long-term dependencies, and input the aligned infrared thermal image into the CNN branch to extract local spatial features. The two branches interact through a cross-modal cross attention module to output current-vibration fusion features and infrared spatial features. S3. Using the current-vibration fusion features and infrared spatial features as graph nodes and the physical coupling relationship between disconnector components as graph edges, a graph neural network is constructed and multiple rounds of message passing are performed through a graph convolutional network to output a fault feature map; S4. Input the fault feature map into the classification head and the segmentation head at the same time, and output the current diagnosis result, vibration diagnosis result and thermal fault segmentation mask; S5. Using the current-vibration fusion features within the historical time window output in step S2 as input, the predicted current features and fault trend residuals are output through a multi-granularity self-attention prediction network. S6. Input the current diagnosis results, vibration diagnosis results, thermal fault segmentation mask and predicted current characteristics into the Venn-Abers credible decision fusion module, and output a comprehensive fault type with a confidence interval.

2. The method according to claim 1, characterized in that, The specific method for constructing the multimodal time-series data stream in step S1 includes: Using the moment when the disconnector switch opens and closes as the zero point, the current signal and vibration signal are timestamped according to the preset sampling frequency, and the infrared thermal image is timestamped according to the acquisition frame rate. The three are then unified to the same time reference through linear interpolation. Based on the physical timing of the disconnector's operation, the opening and closing process is divided into three time windows: the start-up phase, the travel phase, and the positioning phase. A multi-modal data frame sequence of current, vibration, and infrared is constructed for each window.

3. The method according to claim 1, characterized in that, The specific implementation of the cross-modal attention module in step S2 includes: The current-vibration time series features output by the Transformer branch are used as query information, and the infrared spatial features output by the CNN branch are used as key information and value information. Attention weights are generated by calculating the similarity between query information and key information, and then the value information is weighted and aggregated to obtain fused features. The fused features are added to the original Transformer branch output and CNN branch output through residual connections to obtain the enhanced current-vibration fused features and infrared spatial features.

4. The method according to claim 1, characterized in that, In step S2, the feature interaction between the Transformer branch and the CNN branch adopts a bidirectional cross-attention mechanism. Using the infrared spatial features output by the CNN branch as query information and the current-vibration time series features output by the Transformer branch as key and value information, the first direction cross-attention feature is calculated to achieve the attention of infrared features to current-vibration time series features; Meanwhile, using the current-vibration time series features output by the Transformer branch as query information and the infrared spatial features output by the CNN branch as key and value information, the second-direction cross-attention feature is calculated to achieve the attention of the current-vibration time series features to the infrared features. The cross-attention features of the first and second directions are weighted and fused to output bidirectional enhanced current-vibration fusion features and infrared spatial features.

5. The method according to claim 1, characterized in that, The methods for constructing graph nodes in step S3 include: The current-vibration fusion feature is decomposed into multiple time sub-nodes. Each time sub-node corresponds to a timing stage in the operation process of the disconnecting switch. The timing stage includes the start-up stage, the travel stage, and the arrival stage. The infrared spatial features are decomposed into multiple spatial sub-nodes, each spatial sub-node corresponding to a physical component area of ​​the disconnect switch. The physical component area includes a drive motor area, a transmission linkage area, a moving contact area, and a stationary contact area. All temporal and spatial child nodes are combined into a graph node set for the graph neural network.

6. The method according to claim 1, characterized in that, The methods for constructing graph edges and passing messages in step S3 include: Based on the mechanical transmission chain of the disconnector, the torque transmission direction and contact relationship between the drive motor, transmission link, moving contact, stationary contact, and insulating tie rod are used as graph edges to connect the corresponding time sub-nodes and space sub-nodes; In each round of message passing in a graph convolutional network, each graph node receives feature information passed by its neighbor nodes, and performs weighted aggregation of the neighbor information according to the physical coupling strength to update its own node features. After multiple rounds of message passing, the features of all graph nodes are aggregated into a unified fault feature map, which contains both the timing status information and spatial distribution information of the disconnecting switch.

7. The method according to claim 1, characterized in that, The specific implementation methods of the classification head and segmentation head in step S4 include: The classification head adopts a multi-label classification structure, which simultaneously outputs the probability distribution of the first fault category corresponding to the current signal and the probability distribution of the second fault category corresponding to the vibration signal for the input fault feature map. The first fault category and the second fault category are independent of each other but share the same fault feature map. The segmentation head adopts a pixel-by-pixel classification structure, predicts the binary label of each pixel in the input fault feature map as belonging to the thermal fault region or the non-thermal fault region, and outputs a thermal fault segmentation mask with the same resolution as the input infrared thermal image. The classification head and the segmentation head share the fault feature map as a common input. During training, they are jointly optimized through a multi-task joint loss function, so that the classification task and the segmentation task promote each other.

8. The method according to claim 1, characterized in that, The specific implementation of the multi-granularity self-attention prediction network in step S5 includes: The current-vibration fusion features within the historical time window output in step S2 are divided into three time scales: fine-grained, medium-grained, and coarse-grained. Fine-grained corresponds to the local waveform features within a single sampling period, medium-grained corresponds to the trend features within a single action phase of the disconnecting switch, and coarse-grained corresponds to the overall evolution features of the complete opening or closing process. The self-attention weights are calculated independently for the segmentation features of the three time scales, so that the network focuses on the important time step features within that scale at each granularity. The self-attention features calculated at the three granularities are fused across scales to make the features at different time scales complement and enhance each other, and output the predicted current features. The difference between the predicted current characteristics and the actual acquired current signal is used as the fault trend residual. This residual is used to characterize the degree of abnormal deviation of the current signal and serves as the basis for early fault warning.

9. The method according to claim 1, characterized in that, The specific implementation of the Venn-Abers trusted decision fusion module in step S6 includes: The current diagnostic results, vibration diagnostic results, thermal fault segmentation mask and predicted current characteristics are used as four inputs, which are respectively input to the Venn-Abers calibrator for probability calibration, and the calibrated fault probability and its confidence interval are output for each diagnostic result. When the fault type judgments of the four diagnostic results are consistent, the fault type and overall confidence level are directly output. When there is a conflict in the fault type judgment of the four diagnostic results, the evidence is weighted according to the physical fault evolution mechanism of the disconnecting switch: if the thermal fault segmentation mask identifies a high temperature area and the current diagnosis result is abnormal, the overheating-related fault type is given priority; if the vibration diagnosis result identifies mechanical impact characteristics and the current diagnosis result waveform is abnormal, the mechanical jamming-related fault type is given priority; if the predicted current characteristic residual continues to increase but the current diagnosis results are all normal, an early warning signal and a predicted fault trend are output. The final determined fault type and its confidence interval are output as a comprehensive fault type, and a fault evidence chain containing evidence from various sources is generated for fault tracing and operation and maintenance decisions.

10. A substation disconnector status intelligent identification and fault diagnosis system based on an attention mechanism, applied to the method described in any one of claims 1 to 9, characterized in that, The system includes: The multimodal data acquisition and alignment module is used to acquire the current signal, vibration signal and infrared thermal image of the disconnecting switch, unify the timestamps and align the events of the three to construct a multimodal time-series data stream; The hybrid feature extraction module is used to input the aligned current signal and vibration signal into the Transformer branch to capture long-term dependencies, and input the aligned infrared thermal image into the CNN branch to extract local spatial features. The two branches interact through a cross-modal cross attention module to output current-vibration fusion features and infrared spatial features. The graph neural network fusion module is used to construct a graph neural network by taking the current-vibration fusion features and infrared spatial features as graph nodes and the physical coupling relationship between disconnecting switch components as graph edges, and then using a graph convolutional network to perform multi-round message passing to output a fault feature map. The multi-task output module is used to simultaneously input the fault feature map into the classification head and the segmentation head, and output the current diagnosis result, vibration diagnosis result and thermal fault segmentation mask. The multi-granularity prediction module is used to take the current-vibration fusion features within the historical time window output by the hybrid feature extraction module as input, and output the predicted current features and fault trend residuals through the multi-granularity self-attention prediction network. The trusted decision fusion module is used to input the current diagnosis results, vibration diagnosis results, thermal fault segmentation mask and predicted current characteristics into the Venn-Abers trusted decision fusion module, and output a comprehensive fault type with a confidence interval.

Images