Deep learning-based high-voltage switch cabinet partial discharge fault diagnosis method and system

Abstract

The application discloses a high-voltage switch cabinet partial discharge fault diagnosis method based on deep learning, comprising the following steps: acquiring a partial discharge monitoring sequence and an operation state sequence and preprocessing to generate a partial discharge time sequence sample; embedding the sample into an initial hidden state sequence; constructing a virtual cabinet room propagation structure according to the cabinet compartment relationship to generate a virtual cabinet room state sequence; generating a coded hidden state through FEDformer decomposition and frequency enhancement coding; extracting and advancing a candidate fault development state to generate a mature fault state vector; determining an initial diagnosis category and an evidence area and performing counterfactual erasure to generate a counterfactual hidden state; and outputting a fault diagnosis result after secondary diagnosis. The application realizes structured and credible diagnosis of high-voltage switch cabinet partial discharge faults, and improves fault recognition accuracy and anti-interference capability.

Description

Technical Field

[0001] This invention relates to the field of power equipment fault diagnosis technology, and in particular to a method and system for diagnosing partial discharge faults in high-voltage switchgear based on deep learning. Background Technology

[0002] High-voltage switchgear is a critical piece of equipment in power distribution systems. During long-term operation, it is prone to partial discharge due to insulation aging, moisture, poor contact, or concentrated electric fields. Partial discharge is usually an early sign of insulation degradation. If it is not identified in time, it may further develop into insulation breakdown, phase-to-phase short circuits, or equipment shutdown accidents. Online monitoring and fault diagnosis of the partial discharge status of high-voltage switchgear are of great significance.

[0003] Existing partial discharge diagnostic methods typically collect discharge monitoring data and operational status data, and then output fault categories through threshold judgment, manual feature extraction, spectrum recognition, or deep learning models. While these methods can achieve basic identification, they often treat monitoring data as ordinary time-series data and fail to fully utilize the connection relationships between compartments within the high-voltage switchgear, making it difficult for the model to express the propagation characteristics of partial discharge signals between different compartments.

[0004] Most existing deep learning diagnostic models directly output classification results based on encoded features, lacking a step-by-step judgment process for the development state of the fault. When there are interference pulses, random noise, or non-fault anomalies in the monitoring signal, the model tends to give diagnostic conclusions directly based on local responses, making it difficult to verify whether the diagnostic category is truly supported by valid fault evidence.

[0005] Therefore, how to provide a method and system for diagnosing partial discharge faults in high-voltage switchgear based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method and system for diagnosing partial discharge faults in high-voltage switchgear based on deep learning. This invention utilizes virtual cabinet propagation, fault development state advancement, and hidden-layer counterfactual erasure techniques to perform structured diagnosis of partial discharge monitoring data, which has the advantages of high fault identification accuracy, strong anti-interference ability, and high diagnostic reliability.

[0007] The deep learning-based partial discharge fault diagnosis method for high-voltage switchgear according to an embodiment of the present invention includes the following steps: Acquire the partial discharge monitoring sequence and operating status sequence of the high-voltage switchgear, perform preprocessing on the partial discharge monitoring sequence and operating status sequence, and generate partial discharge time series samples; The partial discharge time sequence sample is input into the embedding unit to generate the initial hidden state sequence; A virtual cabinet propagation structure is constructed based on the internal compartment connection relationship of the high-voltage switchgear. The initial hidden state sequence is converted into cabinet node states. Direction-controllable state transfer and boundary blocking are performed between cabinet node states to generate a virtual cabinet state sequence. The virtual cabinet state sequence is input into the FEDformer sequence decomposition unit and the frequency enhancement coding unit to generate the coded hidden state; Extract candidate fault development states from the coded hidden state, advance the candidate fault development states step by step along the preset fault stage chain, call the coded hidden state to correct the state at each stage, and generate a mature fault state vector. An initial diagnostic category is generated based on the mature fault state vector. Evidence regions supporting the initial diagnostic category are determined in the encoded hidden state and virtual cabinet state sequence. Hidden layer counterfactual erasure is performed on the evidence regions to generate counterfactual hidden states. The counterfactual hidden state is input into the secondary diagnostic unit, and the fault diagnosis result is generated based on the change relationship between the initial diagnostic category and the secondary diagnostic category.

[0008] Optionally, the preprocessing of the partial discharge monitoring sequence and the operating status sequence specifically includes: Acquire the partial discharge monitoring sequence collected by the partial discharge monitoring device and the operating status sequence collected by the high-voltage switchgear operation monitoring device; The partial discharge monitoring sequence and the operating status sequence are synchronized in time, aligned with power frequency phase, denoised, normalized and windowed to generate partial discharge time series samples.

[0009] Optionally, the step of inputting partial discharge time-series samples into the embedding unit specifically includes: The embedding unit is set at the input of the improved FEDformer, which converts the partial discharge time series samples into sample feature vectors, and fuses the sample feature vectors with the time position vector and the channel identifier vector to generate the initial hidden state sequence. During training, the initial hidden state sequence is fed into the virtual cabinet propagation structure, and the parameters of the embedded unit are updated through backpropagation based on the error between the fault diagnosis result and the sample label.

[0010] Optionally, the construction of the virtual cabinet propagation structure based on the internal compartment connection relationship of the high-voltage switchgear specifically includes: Based on the connection relationship between the internal compartments of the high-voltage switchgear, a compartment propagation topology is established, the compartments are abstracted as compartment nodes, the connection between adjacent compartments is abstracted as a directed propagation connection, and the propagation connection type is configured for the directed propagation connection. Read the installation location of each monitoring channel in the initial hidden state sequence, merge the hidden states of the channels in the same compartment into the corresponding cabinet node, aggregate the hidden states of the channels merged into the corresponding cabinet node, and generate the cabinet node state. The state of the cabinet node is propagated edge by edge according to the direction of the directed propagation connection, and the initial state of the cabinet node is mapped to the state to be propagated. Boundary determination is performed at the boundary of the compartment corresponding to the directed propagation connection. Reserve the propagation state that meets the propagation connection type, and block the propagation state that does not meet the propagation connection type. The preserved state to be propagated is written into the termination cabinet node and updated with the original state of the termination cabinet node. The edge-by-edge propagation, boundary determination and state update are repeatedly executed according to the time sequence of the partial discharge time series samples to generate a virtual cabinet state sequence.

[0011] Optionally, the step of inputting the virtual cabinet state sequence into the FEDformer sequence decomposition unit and the frequency enhancement coding unit specifically includes: The virtual cabinet state sequence is input into the FEDformer sequence decomposition unit, and trend component separation and periodic component separation are performed on the virtual cabinet state sequence to obtain the decomposed cabinet state sequence. The decomposed cabinet state sequence is input into the frequency enhancement coding unit, and frequency domain transformation, frequency component selection and inverse transformation reconstruction are performed on the decomposed cabinet state sequence to obtain the frequency enhancement state sequence. The frequency-enhanced state sequence is input into the encoder feedforward layer and residual normalization layer for hidden state update, generating the encoded hidden state.

[0012] Optionally, the step of invoking the encoded hidden state for state correction in each stage of advancement specifically includes: Global pooling and linear mapping are performed on the encoded hidden states to obtain the candidate fault development states; The candidate fault development state is input into the state advancement unit, and multi-level state advancement is carried out according to the preset fault stage chain. Each level of state advancement reads the candidate fault development state and the encoded hidden state of the previous stage, and writes the hidden layer features corresponding to the current stage in the encoded hidden state into the candidate fault development state of the previous stage to obtain the candidate fault development state of the current stage. After each state progression, the state transition discrimination unit determines whether the candidate fault development state of the current stage meets the conditions for entering the next stage. When the candidate fault development state passes through the state transition judgment of each level in the preset fault stage chain continuously, the last level candidate fault development state is determined as the mature fault state vector. When a candidate fault development state fails to pass the state transition judgment at any level, a low-confidence label is added to the candidate fault development state, and the candidate fault development state after adding the low-confidence label is used as the mature fault state vector.

[0013] Optionally, determining the region of evidence supporting the initial diagnostic category in the sequence of encoded hidden states and virtual cabinet states specifically includes: The mature fault state vector is mapped to the category score of each preset diagnostic category, and the preset diagnostic category with the highest category score is determined as the initial diagnostic category; Generate a category retrieval vector according to the initial diagnosis category, and apply the category retrieval vector to the encoded hidden state and the virtual cabinet state sequence respectively, and perform response retrieval on the time slice features in the encoded hidden state and the cabinet state in the virtual cabinet state sequence. During the response retrieval process, the coded hidden state response and the virtual cabinet state response within the same time slice are compared for consistency. The time slice and cabinet position that both point to the initial diagnostic category are retained, and the positions that respond only in the coded hidden state or only in the virtual cabinet state sequence are removed. The retained locations are merged sequentially along the time line, and adjacent retained locations with an interval less than a preset interval threshold are merged into the same candidate evidence region; The region with the strongest response is selected from the candidate evidence region and identified as the evidence region supporting the initial diagnostic category.

[0014] Optionally, the process of performing hidden counterfactual erasure on the evidence region specifically includes: Read the time slice range, cabinet location, and hidden state index of the evidence area record, and determine the hidden state corresponding to the evidence area in the encoded hidden state and virtual cabinet state sequence according to the time slice range, cabinet location, and hidden state index; Read the category retrieval vector corresponding to the initial diagnosis category, decompose the hidden state corresponding to the evidence region into category-related components and category-independent components along the direction of the category retrieval vector, delete the category-related components, and retain the category-independent components; Select hidden states that are not covered by the evidence area from adjacent time slices and adjacent cabinet locations in the evidence area, and generate alternative hidden states based on the selected hidden states. Write the alternative hidden state into the evidence region after deleting the category-related components. The hidden state corresponding to the evidence region no longer retains the category-related information that supports the initial diagnostic category. The boundaries of the evidence region after writing the alternative hidden state are continuously updated to generate a counterfactual hidden state.

[0015] Optionally, the relationship between the initial diagnostic category and the secondary diagnostic category specifically includes: Input the counterfactual hidden state into the secondary diagnostic unit to obtain the secondary category score corresponding to each preset diagnostic category, and determine the preset diagnostic category with the largest secondary category score as the secondary diagnostic category. Compare the initial diagnosis category with the secondary diagnosis category. When the secondary diagnosis category is different from the initial diagnosis category, confirm the evidence region that the initial diagnosis category depends on and write the initial diagnosis category into the fault diagnosis result. When the secondary diagnosis category is the same as the initial diagnosis category, and the difference between the secondary category score and the initial category score is less than a preset change threshold, the fault diagnosis result is marked as a low-confidence diagnosis result. When the secondary diagnostic category is the same as the initial diagnostic category, and the difference between the secondary category score and the initial category score is greater than or equal to a preset change threshold, the initial diagnostic category and confidence decay flag are written into the fault diagnosis result.

[0016] A deep learning-based partial discharge fault diagnosis system for high-voltage switchgear according to an embodiment of the present invention includes: The data processing module is used to acquire the partial discharge monitoring sequence and the operating status sequence of the high-voltage switchgear, perform preprocessing on the partial discharge monitoring sequence and the operating status sequence, and generate partial discharge time series samples. An embedded processing module is used to convert partial discharge time series samples into an initial hidden state sequence and send the initial hidden state sequence into the improved FEDformer. The virtual cabinet propagation module is used to establish the cabinet propagation topology based on the internal compartment connection relationship of the high-voltage switchgear, convert the initial hidden state sequence into the cabinet node state, and generate the virtual cabinet state sequence through directed propagation connection, boundary determination and state update. The FEDformer encoding module is used to perform sequence decomposition and frequency enhancement encoding on the virtual cabinet state sequence to generate encoded hidden states. The fault development state advancement module is used to extract candidate fault development states from the encoded hidden states and advance along the preset fault stage chain to generate a mature fault state vector. The counterfactual diagnosis module is used to generate an initial diagnosis category based on the mature fault state vector, determine the evidence region, perform hidden counterfactual erasure on the evidence region, and generate a fault diagnosis result based on the secondary diagnosis result.

[0017] The beneficial effects of this invention are: (1) This invention introduces the internal compartment connection relationship of high voltage switch cabinet into the partial discharge fault diagnosis process. By using the virtual cabinet propagation structure, the hidden state of ordinary sensors is converted into a virtual cabinet state sequence, so that the model can combine the propagation relationship between different compartments in the cabinet during diagnosis, thereby reducing misjudgment caused by relying on only a single monitoring channel.

[0018] (2) The present invention sets up a fault development state advancement process after the FEDformer encoding hidden state, and advances the candidate fault development state step by step along the preset fault stage chain. In each advancement, the encoded hidden state is called to correct the state, so that the diagnosis process no longer relies solely on the one-time classification output, and can better express the change process of partial discharge from the occurrence of abnormality to the stabilization of fault.

[0019] (3) The present invention performs secondary diagnostic verification on the initial diagnostic category by determining the evidence region and erasing the hidden counterfactual evidence. After erasing the hidden evidence supporting the initial diagnostic category, the final fault diagnosis result is generated based on the change relationship between the secondary diagnosis result and the initial diagnosis result, thereby improving the credibility and anti-interference ability of the diagnosis result. Attached Figure Description

[0020] 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 the deep learning-based partial discharge fault diagnosis method for high-voltage switchgear proposed in this invention. Figure 2 This is a schematic diagram of the module of the high-voltage switchgear partial discharge fault diagnosis system based on deep learning proposed in this invention. Figure 3 This is a virtual cabinet propagation structure diagram of the high-voltage switchgear partial discharge fault diagnosis method and system based on deep learning proposed in this invention. Detailed Implementation

[0021] 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.

[0022] refer to Figures 1-3 A deep learning-based method for diagnosing partial discharge faults in high-voltage switchgear includes the following steps: Acquire the partial discharge monitoring sequence and operating status sequence of the high-voltage switchgear, perform preprocessing on the partial discharge monitoring sequence and operating status sequence, and generate partial discharge time series samples; The partial discharge time sequence sample is input into the embedding unit to generate the initial hidden state sequence; A virtual cabinet propagation structure is constructed based on the internal compartment connection relationship of the high-voltage switchgear. The initial hidden state sequence is converted into cabinet node states. Direction-controllable state transfer and boundary blocking are performed between cabinet node states to generate a virtual cabinet state sequence. The virtual cabinet state sequence is input into the FEDformer sequence decomposition unit and the frequency enhancement coding unit to generate the coded hidden state; Extract candidate fault development states from the coded hidden state, advance the candidate fault development states step by step along the preset fault stage chain, call the coded hidden state to correct the state at each stage, and generate a mature fault state vector. An initial diagnostic category is generated based on the mature fault state vector. Evidence regions supporting the initial diagnostic category are determined in the encoded hidden state and virtual cabinet state sequence. Hidden layer counterfactual erasure is performed on the evidence regions to generate counterfactual hidden states. The counterfactual hidden state is input into the secondary diagnostic unit, and the fault diagnosis result is generated based on the change relationship between the initial diagnostic category and the secondary diagnostic category.

[0023] In this embodiment, the preprocessing of the partial discharge monitoring sequence and the operating status sequence specifically includes: The partial discharge monitoring sequence is acquired from the partial discharge monitoring device, and the operating status sequence is acquired from the high-voltage switchgear operation monitoring device. The partial discharge monitoring sequence is acquired by the partial discharge monitoring device installed in the high-voltage switchgear cabinet, grounding circuit, or cable connection. The partial discharge monitoring sequence includes at least one of the following: discharge amplitude sequence, discharge number sequence, discharge phase sequence, and discharge energy sequence. The operating status sequence is acquired by the high-voltage switchgear operation monitoring device. The operating status sequence includes at least one of the following: electrical operating status sequence, environmental status sequence, and switch action status sequence. Both the partial discharge monitoring sequence and the operating status sequence are arranged in timestamp order. The partial discharge monitoring sequence and the operating status sequence are synchronized in time, aligned with power frequency phase, denoised, normalized, and windowed to generate partial discharge time series samples. Time synchronization uses a unified timestamp to align the partial discharge monitoring sequence and the operating status sequence, and interpolates missing timestamps using adjacent sampling points. Power frequency phase alignment uses the zero-crossing point of the high-voltage switchgear voltage signal as the phase reference, divides one power frequency cycle into several phase intervals, and maps the partial discharge monitoring sequence to the corresponding phase interval. Denoising uses wavelet denoising to remove background noise and transient interference. Normalization uses max-min normalization to convert data of different dimensions to a unified numerical range. Windowing slices the processed sequence according to a preset time length or a preset number of power frequency cycles, with each slice forming a partial discharge time series sample. The partial discharge time series sample is a two-dimensional time series matrix, where the rows of the two-dimensional time series matrix represent time slices or power frequency phase segments, and the columns represent partial discharge monitoring features and operating status features.

[0024] In this embodiment, inputting partial discharge timing samples into the embedding unit specifically includes: The embedding unit is located at the input of the improved FEDformer. It converts partial discharge time series samples into sample feature vectors and fuses these vectors with time position vectors and channel identifier vectors to generate an initial hidden state sequence. The partial discharge time series samples use a time window of length 128, with each time window containing 16-dimensional features. The embedding unit uses a one-dimensional convolutional mapping method to map the 16-dimensional features into a 64-dimensional sample feature vector. The one-dimensional convolution kernel has a length of 3, a stride of 1, and a padding length of 1. The time position vector is generated using a sinusoidal position encoding method. The channel identifier vector is generated using a trainable lookup table, and the channel identifier vector has a dimension of 64. The fusion method uses element-wise addition, adding the sample feature vector, time position vector, and channel identifier vector to obtain the initial hidden state sequence. During training, the initial hidden state sequence is fed into the virtual cabinet propagation structure. Based on the error between the fault diagnosis result and the sample label, the parameters of the embedding unit are updated through backpropagation. The sample label is represented by an integer class label, with 0 for normal state and 1 for partial discharge fault state. The fault diagnosis result is represented by a binary classification probability vector. The error is calculated using cross-entropy loss. The training optimizer is the Adam optimizer, with a learning rate of 0.001, a batch size of 32, and 100 training rounds. During backpropagation, the one-dimensional convolution parameters and channel identifier vector lookup table parameters of the embedding unit are updated synchronously with the other trainable parameters of the improved FEDformer.

[0025] In this embodiment, constructing the virtual cabinet propagation structure according to the internal compartment connection relationship of the high-voltage switchgear specifically includes: A cabinet propagation topology is established based on the connection relationships between the internal compartments of the high-voltage switchgear. The compartments are abstracted as cabinet nodes, and the connections between adjacent compartments are abstracted as directed propagation connections. Propagation connection types are configured for these directed propagation connections. The cabinet propagation topology is represented by a directed graph, with each cabinet node corresponding one-to-one with a compartment within the high-voltage switchgear. In one specific embodiment, the internal compartments of the high-voltage switchgear are divided into a busbar compartment, a circuit breaker compartment, a cable compartment, and an instrument compartment, with each of the four compartments corresponding to a different cabinet node. The directed propagation connections are set according to the actual connection direction between the internal compartments of the high-voltage switchgear, and the propagation connection type is represented by a connection type identifier, which is set to three categories: electrical connection, spatial connectivity, and grounding connection. The installation positions of each monitoring channel in the initial hidden state sequence are read, and the hidden states of channels in the same compartment are merged into the corresponding cabinet node. The hidden states of channels merged into the corresponding cabinet node are aggregated to generate the cabinet node state. The installation position of each monitoring channel is determined by the installation configuration table, which records the monitoring channel number, the installation compartment corresponding to the monitoring channel, and the connection type identifier corresponding to the monitoring channel. When the hidden states of channels are merged into the cabinet node, the hidden states of channels belonging to the same installation compartment are assigned to the same cabinet node according to the installation configuration table. The aggregation adopts average aggregation, which averages the hidden states of channels received by the same cabinet node to obtain the cabinet node state. When the same cabinet node does not receive any hidden states of channels, the cabinet node state corresponding to the previous time slice is used as the current cabinet node state. The state of the cabinet node is propagated edge by edge according to the direction of the directed propagation connection, mapping the initial cabinet node state to the state to be propagated. Boundary determination is performed at the compartment boundary corresponding to the directed propagation connection. The edge propagation is performed in the order of the directed propagation connections in the cabinet propagation topology, and each directed propagation connection is processed once in one propagation round. For the current directed propagation connection, the initial cabinet node state, the terminating cabinet node state, and the propagation connection type are read. The initial cabinet node state is input into the linear mapping layer to obtain the state to be propagated. During boundary determination, the propagation connection type of the current directed propagation connection is matched with the connection type identifier corresponding to the initial cabinet node. The propagation state that meets the propagation connection type is retained, and the propagation state that does not meet the propagation connection type is blocked. When the propagation connection type is consistent with the response source type, the propagation state is retained. When the propagation connection type is inconsistent with the response source type, the propagation state is blocked. The response source type is determined by the monitoring channel type that generates the propagation state. The response source type includes three categories: electric field response, acoustic response, and ground loop response. The retained state to be propagated is written to the termination cabinet node and updated with the original state of the termination cabinet node. The process of edge-by-edge propagation, boundary determination, and state update is repeated according to the time sequence of the partial discharge time series samples to generate a virtual cabinet state sequence. Within each time slice, edge-by-edge propagation is first performed according to the directed propagation connections in the cabinet propagation topology, mapping the state of the starting cabinet node to the state to be propagated. Boundary determination is performed at the compartment boundary corresponding to the current directed propagation connection, which is completed by comparing the propagation connection type with the connection type identifier corresponding to the starting cabinet node. When the boundary determination result is "retained," the state to be propagated is written to the receiving buffer of the termination cabinet node. When the boundary determination result is "blocked," the state to be propagated is not written to the receiving buffer of the termination cabinet node. When the same termination cabinet node receives multiple states to be propagated within the same time slice, average aggregation is performed on the multiple states to be propagated in the receiving buffer. The average aggregation result is updated with the original state of the termination cabinet node to obtain the cabinet node state of the current time slice. The above process is repeated according to the time sequence of the partial discharge time series samples, arranging the cabinet node states of each time slice in chronological order.

[0026] In this embodiment, inputting the virtual cabinet state sequence into the FEDformer sequence decomposition unit and the frequency enhancement coding unit specifically includes: The virtual counter room state sequence is input into the FEDformer sequence decomposition unit, which performs trend component separation and periodic component separation on the virtual counter room state sequence to obtain the decomposed counter room state sequence. The FEDformer sequence decomposition unit is the sequence decomposition structure in the FEDformer model. After receiving the virtual counter room state sequence, the sequence decomposition unit separates the trend component and periodic component of the virtual counter room state sequence according to the sequence decomposition method of the FEDformer model. The virtual counter room state sequence is the input object of the FEDformer sequence decomposition unit, and the sequence decomposition structure itself adopts the existing structure of the FEDformer model. The decomposed cabinet state sequence is input into the frequency enhancement coding unit. Frequency domain transformation, frequency component selection, and inverse transformation reconstruction are performed on the decomposed cabinet state sequence to obtain the frequency enhancement state sequence. The frequency enhancement coding unit is the frequency enhancement coding structure in the FEDformer model. Frequency domain transformation, frequency component selection, and inverse transformation reconstruction are all processing steps in the FEDformer frequency enhancement coding structure. Without changing the internal processing method of the FEDformer frequency enhancement coding structure, the decomposed cabinet state sequence is used as the input of the frequency enhancement coding structure, and the FEDformer frequency enhancement coding structure generates the frequency enhancement state sequence according to the preset frequency enhancement process. The frequency-enhanced state sequence is input into the encoder feedforward layer and residual normalization layer for hidden state update, generating the encoded hidden state; the encoder feedforward layer and residual normalization layer are existing structures in the FEDformer encoder; the frequency-enhanced state sequence is sequentially processed by the feedforward calculation, residual connection and normalization in the FEDformer encoder.

[0027] In this embodiment, the process of invoking the encoded hidden state for state correction at each stage of advancement specifically includes: Global pooling and linear mapping are performed on the encoded hidden state to obtain the candidate fault development state. The encoded hidden state is the time series hidden state output by the FEDformer frequency enhancement coding unit. Global pooling adopts the average pooling method to average the encoded hidden state in the time dimension to obtain the window-level coding vector. The linear mapping is completed by a fully connected layer to convert the window-level coding vector into the candidate fault development state. The candidate fault development state serves as the starting state of the preset fault stage chain. The candidate fault development state is input into the state advancement unit, and multi-level state advancement is performed according to the preset fault stage chain. Each level of state advancement reads the candidate fault development state and the encoded hidden state of the previous stage, and writes the hidden layer features corresponding to the current stage in the encoded hidden state into the candidate fault development state of the previous stage to obtain the candidate fault development state of the current stage. The preset fault stage chain is set as a three-level stage chain, which consists of the initial abnormal stage, the development discharge stage, and the stable fault stage. The state advancement unit is implemented using a gated loop unit. During each level of state advancement, the candidate fault development state of the previous stage is used as the previous state of the gated loop unit, and the stage hidden layer features obtained by passing the stage selection layer through the encoded hidden state are used as the current input of the gated loop unit, outputting the candidate fault development state of the current stage. The stage selection layer uses a fully connected layer and a Softmax function to generate stage selection coefficients, and extracts the hidden layer features corresponding to the current stage from the encoded hidden state according to the stage selection coefficients. The preset fault stage chain is set according to the number of internal inferences of the state advancement unit. The three-level stage chain corresponds to three consecutive state advancement calculations. After the candidate fault development state enters the state advancement unit, the first state advancement calculation corresponds to the initial abnormal stage, the second state advancement calculation corresponds to the development discharge stage, and the third state advancement calculation corresponds to the stable fault stage. Each level of state advancement calculation is set with independent stage parameters and a stage selection layer. The stage selection layer extracts the hidden layer features corresponding to the current level from the same encoded hidden state and writes the hidden layer features corresponding to the current level into the candidate fault development state output by the previous stage. After each state progression, a state transition discrimination unit determines whether the current candidate fault development state meets the conditions for entering the next stage. The state transition discrimination unit is implemented using a two-layer fully connected network. The state transition discrimination unit receives the current candidate fault development state and outputs a transition probability. When the transition probability is greater than or equal to 0.6, the current candidate fault development state enters the next stage; when the transition probability is less than 0.6, the current candidate fault development state fails the state transition discrimination. When the candidate fault development state passes through the state transition judgment of each level in the preset fault stage chain continuously, the last level candidate fault development state is determined as the mature fault state vector. When a candidate fault development state fails to pass any state transition judgment, a low-confidence label is added to the candidate fault development state, and the candidate fault development state with the added low-confidence label is used as the mature fault state vector. The low-confidence label is represented by a trainable label vector. When a candidate fault development state fails to pass any state transition judgment, the candidate fault development state at the current stage that fails the state transition judgment is concatenated with the low-confidence label vector and mapped to the mature fault state vector through a fully connected layer.

[0028] In this embodiment, determining the evidence region supporting the initial diagnostic category in the sequence of coded hidden states and virtual cabinet states specifically includes: The mature fault state vector is mapped to the category score of each preset diagnostic category, and the preset diagnostic category with the highest category score is determined as the initial diagnostic category. The mature fault state vector is input into a fully connected classification layer, and the fully connected classification layer outputs the category score corresponding to each preset diagnostic category. The category score is converted into a category probability through a Softmax function, and the preset diagnostic category with the highest category probability is taken as the initial diagnostic category. The preset diagnostic categories are determined during the training sample labeling stage. In one specific implementation, the preset diagnostic categories are set as six categories: normal state, corona discharge, surface discharge, internal discharge, floating potential discharge, and interference signal. A category retrieval vector is generated based on the initial diagnostic category. This category retrieval vector is then applied to the encoded hidden state and the virtual cabinet state sequence, respectively, to perform response retrieval on the time-slice features in the encoded hidden state and the cabinet states in the virtual cabinet state sequence. The category retrieval vector is generated from a category vector table, where each preset diagnostic category corresponds to a trainable category vector. After determining the initial diagnostic category, the category vector corresponding to the initial diagnostic category is read from the category vector table as the category retrieval vector. The response retrieval uses vector similarity calculation, performing dot product calculations on the category retrieval vector with each time-slice feature in the encoded hidden state and each cabinet state in the virtual cabinet state sequence to obtain the encoded hidden state response and the virtual cabinet state response. During the response retrieval process, the coded hidden state response and the virtual cabinet state response within the same time slice are compared for consistency. Time slices and cabinet locations that simultaneously point to the initial diagnostic category are retained, while locations that respond only in the coded hidden state or only in the virtual cabinet state sequence are removed. During the consistency comparison, normalization processing is performed on the coded hidden state response and the virtual cabinet state response respectively. When the coded hidden state response corresponding to the same time slice is greater than 0.6, and the virtual cabinet state response of at least one cabinet location within the time slice is greater than 0.6, the time slice and the corresponding cabinet location are marked as reserved locations. When only the coded hidden state response or only the virtual cabinet state response is greater than 0.6 in the same time slice, the time slice and the corresponding cabinet location are not marked as reserved locations. The system performs continuous merging of the retained positions in chronological order, merging adjacent retained positions with an interval less than a preset interval threshold into the same candidate evidence region. Continuous merging is performed according to the time slice order of the partial discharge time series samples. The preset interval threshold is set to 2 time slices. When the interval between two adjacent retained positions is less than 2 time slices, the two adjacent retained positions are assigned to the same candidate evidence region. When the interval between two adjacent retained positions is greater than or equal to 2 time slices, the two adjacent retained positions are assigned to different candidate evidence regions. The region with the strongest response is selected from the candidate evidence regions and determined as the evidence region supporting the initial diagnostic category. The response intensity of each candidate evidence region is obtained by adding the average value of the encoded latent state response and the average value of the virtual cabinet state response within the candidate evidence region. The response intensities of all candidate evidence regions are sorted, and the candidate evidence region with the strongest response intensity is determined as the evidence region supporting the initial diagnostic category. The evidence region records the corresponding time slice range and cabinet location.

[0029] In this embodiment, performing hidden counterfactual erasure on the evidence area specifically includes: The evidence area record is read by determining the time slice range, cabinet location, and hidden state index. The hidden state corresponding to the evidence area is determined in the coded hidden state and virtual cabinet state sequence according to the time slice range, cabinet location, and hidden state index. The evidence area record is generated by the evidence area determination process and includes the start time slice, end time slice, cabinet location number, and hidden state index. The coded hidden state is read according to the time slice index, and the virtual cabinet state sequence is read according to the time slice index and cabinet location number. The hidden state that falls within the range of the start time slice to the end time slice and the range corresponding to the cabinet location number is determined as the hidden state corresponding to the evidence area. Read the category retrieval vector corresponding to the initial diagnostic category, decompose the hidden state corresponding to the evidence region into category-related components and category-independent components along the direction of the category retrieval vector, delete the category-related components, and retain the category-independent components; for each hidden state corresponding to the evidence region, calculate the projection of the hidden state onto the direction of the category retrieval vector, determine the projection result as the category-related component, and determine the remaining part after subtracting the category-related component from the hidden state as the category-independent component; when deleting the category-related component, set the category-related component to zero, and retain the category-independent component as the basic hidden state after erasure; Hidden states not covered by the evidence area are selected from adjacent time slices and adjacent cabinet locations within the evidence area. Alternate hidden states are generated based on the selected hidden states. Adjacent time slices are the time slices one before the start time slice and one after the end time slice of the evidence area. Adjacent cabinet locations are cabinet nodes that have a directed propagation connection with cabinet locations in the evidence area. When the hidden state corresponding to the adjacent time slice and adjacent cabinet location does not fall into the evidence area, this hidden state is selected as a candidate alternative hidden state. Alternate hidden states are generated using the average value of the candidate alternative hidden states. When no candidate alternative hidden state exists, the average value of all hidden states outside the evidence area is used as the alternative hidden state. The alternative hidden state is written into the evidence region after the category-related components are deleted. The hidden state corresponding to the evidence region no longer retains the category-related information that supports the initial diagnostic category. When writing, each erased basic hidden state in the evidence region is added to the alternative hidden state to obtain the hidden state of the evidence region after writing. The hidden state of the evidence region after writing retains the category-independent components and the alternative context is provided by the adjacent time slices and adjacent cabinet locations. The boundaries of the evidence region after writing the replacement hidden state are continuously updated to generate a counterfactual hidden state. The continuous update is performed at the start and end boundaries of the evidence region. The hidden state after writing at the start boundary is averaged with the hidden state corresponding to the time slice one time slice before the start time slice. The hidden state after writing at the end boundary is averaged with the hidden state corresponding to the time slice one time slice after the end time slice. After the boundary averaging is completed, the updated encoded hidden state is used as the counterfactual hidden state.

[0030] In this embodiment, the specific aspects based on the relationship between the initial diagnostic category and the secondary diagnostic category include: The counterfactual hidden state is input into the secondary diagnostic unit to obtain the secondary category score corresponding to each preset diagnostic category. The preset diagnostic category with the largest secondary category score is determined as the secondary diagnostic category. The secondary diagnostic unit adopts the same category mapping method as the initial diagnostic category generation process. Compare the initial diagnosis category with the secondary diagnosis category. When the secondary diagnosis category is different from the initial diagnosis category, confirm the evidence region that the initial diagnosis category depends on and write the initial diagnosis category into the fault diagnosis result. The preset change threshold is set to 0.2. The initial category score is the Softmax category probability corresponding to the initial diagnosis category, and the secondary category score is the Softmax category probability corresponding to the secondary diagnosis category. The difference is the absolute value of the initial category score minus the secondary category score. When the secondary diagnostic category is the same as the initial diagnostic category, and the difference between the secondary category score and the initial category score is less than the preset change threshold, the fault diagnosis result is marked as a low-confidence diagnosis result; the confidence decay flag indicates that the initial diagnostic category remains unchanged after counterfactual erasure but the category score decreases; the confidence decay flag is 1, and is 0 when no confidence decay occurs; When the secondary diagnostic category is the same as the initial diagnostic category, and the difference between the secondary category score and the initial category score is greater than or equal to a preset change threshold, the initial diagnostic category and confidence decay flag are written into the fault diagnosis result. The fault diagnosis result is output in a triplet format, which is "diagnosis category, diagnosis confidence flag, category score". The diagnosis category is one of the initial diagnostic category, the secondary diagnostic category, or the low confidence category. The diagnosis confidence flag is one of the confirmation flag, the low confidence flag, or the confidence decay flag. The category score is the Softmax category probability corresponding to the initial diagnostic category.

[0031] A deep learning-based partial discharge fault diagnosis system for high-voltage switchgear includes: The data processing module is used to acquire the partial discharge monitoring sequence and the operating status sequence of the high-voltage switchgear, perform preprocessing on the partial discharge monitoring sequence and the operating status sequence, and generate partial discharge time series samples. An embedded processing module is used to convert partial discharge time series samples into an initial hidden state sequence and send the initial hidden state sequence into the improved FEDformer. The virtual cabinet propagation module is used to establish the cabinet propagation topology based on the internal compartment connection relationship of the high-voltage switchgear, convert the initial hidden state sequence into the cabinet node state, and generate the virtual cabinet state sequence through directed propagation connection, boundary determination and state update. The FEDformer encoding module is used to perform sequence decomposition and frequency enhancement encoding on the virtual cabinet state sequence to generate encoded hidden states. The fault development state advancement module is used to extract candidate fault development states from the encoded hidden states and advance along the preset fault stage chain to generate a mature fault state vector. The counterfactual diagnosis module is used to generate an initial diagnosis category based on the mature fault state vector, determine the evidence region, perform hidden counterfactual erasure on the evidence region, and generate a fault diagnosis result based on the secondary diagnosis result.

[0032] Example 1: To verify the feasibility of this invention in practice, it was applied to a partial discharge fault diagnosis scenario in high-voltage switchgear. During long-term operation, high-voltage switchgear experiences complex propagation characteristics of partial discharge signals between different compartments due to factors such as insulation aging, moisture, or poor contact. Traditional diagnostic methods typically classify multi-channel monitoring data directly as time-series inputs, ignoring the structural relationships within the cabinet compartments. This leads to misjudgments and missed alarms in environments with strong interference or complex signals.

[0033] In this embodiment, partial discharge monitoring data and operating status data of the high-voltage switchgear are collected and preprocessed, including time synchronization, power frequency phase alignment, noise reduction, normalization, and windowing, to generate partial discharge time series samples. These samples are then input into the embedding processing module to generate an initial hidden state sequence. A virtual cabinet propagation structure is constructed based on the internal compartment connections, mapping the initial hidden state sequence to the state of each compartment node. Directed propagation, boundary determination, and state updates are performed between nodes to form a virtual cabinet state sequence, enabling the model to simulate the propagation process of partial discharge signals between different areas within the cabinet.

[0034] The virtual cabinet state sequence is input into the FEDformer encoding module, where sequence decomposition and frequency enhancement encoding are performed to obtain coded hidden states. Subsequently, candidate fault development states are extracted from these hidden states, and the fault development state advancement module is used to advance the state step-by-step along a preset fault stage chain. At each stage, the coded hidden states are invoked for state correction, ultimately forming a mature fault state vector. This mature fault state vector not only contains current state information but also retains information about the fault evolution process, enabling a more accurate reflection of the fault development trend.

[0035] The counterfactual diagnosis module generates an initial diagnosis category based on the mature fault state vector and identifies the evidence region supporting this diagnosis category in the encoded hidden states and virtual cabinet state sequences. A hidden-layer counterfactual erasure operation is performed on the evidence region to remove category-related information, and alternative hidden states are generated using adjacent time slices and adjacent cabinet states to form counterfactual hidden states. These counterfactual hidden states are then input into the secondary diagnosis unit. By comparing the changes between the initial and secondary diagnosis categories, the final fault diagnosis result is generated, improving the model's diagnostic reliability and anti-interference capability.

[0036] To verify the performance of this invention, it was compared with the LSTM method and the original FEDformer method. Test data covered normal state, surface discharge, internal discharge, floating potential discharge, and interference state, and a unified training and test set was used for evaluation. Evaluation metrics included diagnostic accuracy, false negative rate, interference exclusion rate, counterfactual verification pass rate, and average computation time per sample.

[0037] Table 1: Performance Comparison of Different Fault Diagnosis Methods

[0038] As shown in Table 1, the present invention outperforms the comparative methods in terms of diagnostic accuracy, false negative rate, interference exclusion rate, and counterfactual verification pass rate. The diagnostic accuracy reaches 98.4%, and the false negative rate is reduced to 1.3%, indicating that the present invention can more stably identify partial discharge faults. The interference exclusion rate reaches 96.8%, corresponding to the virtual cabinet propagation structure and the counterfactual erasure process, indicating that the model can better distinguish between real partial discharge characteristics and interference responses. The counterfactual verification pass rate reaches 95.1%, indicating that the initial diagnostic category can change significantly after erasing the evidence area, and the diagnostic results are highly dependent on valid evidence. The average computation time per sample is 19.6 ms, slightly higher than the original FEDformer, mainly due to the increased computational load caused by virtual cabinet propagation, fault development state progression, and the counterfactual verification mechanism, but it still meets the requirements for online diagnosis. The results of this embodiment verify the advantages of the present invention in fault diagnosis accuracy, anti-interference ability, and diagnostic reliability.

[0039] 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 diagnosing partial discharge faults in high-voltage switchgear based on deep learning, characterized in that, Includes the following steps: Acquire the partial discharge monitoring sequence and operating status sequence of the high-voltage switchgear, perform preprocessing on the partial discharge monitoring sequence and operating status sequence, and generate partial discharge time series samples; The partial discharge time sequence sample is input into the embedding unit to generate the initial hidden state sequence; A virtual cabinet propagation structure is constructed based on the internal compartment connection relationship of the high-voltage switchgear. The initial hidden state sequence is converted into cabinet node states. Direction-controllable state transfer and boundary blocking are performed between cabinet node states to generate a virtual cabinet state sequence. The virtual cabinet state sequence is input into the FEDformer sequence decomposition unit and the frequency enhancement coding unit to generate the coded hidden state; Extract candidate fault development states from the coded hidden state, advance the candidate fault development states step by step along the preset fault stage chain, call the coded hidden state to correct the state at each stage, and generate a mature fault state vector. An initial diagnostic category is generated based on the mature fault state vector. Evidence regions supporting the initial diagnostic category are determined in the encoded hidden state and virtual cabinet state sequence. Hidden layer counterfactual erasure is performed on the evidence regions to generate counterfactual hidden states. The counterfactual hidden state is input into the secondary diagnostic unit, and the fault diagnosis result is generated based on the change relationship between the initial diagnostic category and the secondary diagnostic category.

2. The method for diagnosing partial discharge faults in high-voltage switchgear based on deep learning according to claim 1, characterized in that, The preprocessing of the partial discharge monitoring sequence and the operating status sequence specifically includes: Acquire the partial discharge monitoring sequence collected by the partial discharge monitoring device and the operating status sequence collected by the high-voltage switchgear operation monitoring device; The partial discharge monitoring sequence and the operating status sequence are synchronized in time, aligned with power frequency phase, denoised, normalized and windowed to generate partial discharge time series samples.

3. The method for diagnosing partial discharge faults in high-voltage switchgear based on deep learning according to claim 1, characterized in that, The step of inputting partial discharge time-series samples into the embedding unit specifically includes: The embedding unit is set at the input of the improved FEDformer, which converts the partial discharge time series samples into sample feature vectors, and fuses the sample feature vectors with the time position vector and the channel identifier vector to generate the initial hidden state sequence. During training, the initial hidden state sequence is fed into the virtual cabinet propagation structure, and the parameters of the embedded unit are updated through backpropagation based on the error between the fault diagnosis result and the sample label.

4. The method for diagnosing partial discharge faults in high-voltage switchgear based on deep learning according to claim 1, characterized in that, The construction of the virtual cabinet propagation structure based on the internal compartment connection relationship of the high-voltage switchgear specifically includes: Based on the connection relationship between the internal compartments of the high-voltage switchgear, a compartment propagation topology is established, the compartments are abstracted as compartment nodes, the connection between adjacent compartments is abstracted as a directed propagation connection, and the propagation connection type is configured for the directed propagation connection. Read the installation location of each monitoring channel in the initial hidden state sequence, merge the hidden states of the channels in the same compartment into the corresponding cabinet node, aggregate the hidden states of the channels merged into the corresponding cabinet node, and generate the cabinet node state. The state of the cabinet node is propagated edge by edge according to the direction of the directed propagation connection, and the initial state of the cabinet node is mapped to the state to be propagated. Boundary determination is performed at the boundary of the compartment corresponding to the directed propagation connection. Reserve the propagation state that meets the propagation connection type, and block the propagation state that does not meet the propagation connection type. The preserved state to be propagated is written into the termination cabinet node and updated with the original state of the termination cabinet node. The edge-by-edge propagation, boundary determination and state update are repeatedly executed according to the time sequence of the partial discharge time series samples to generate a virtual cabinet state sequence.

5. The method for diagnosing partial discharge faults in high-voltage switchgear based on deep learning according to claim 1, characterized in that, The step of inputting the virtual cabinet state sequence into the FEDformer sequence decomposition unit and frequency enhancement coding unit specifically includes: The virtual cabinet state sequence is input into the FEDformer sequence decomposition unit, and trend component separation and periodic component separation are performed on the virtual cabinet state sequence to obtain the decomposed cabinet state sequence. The decomposed cabinet state sequence is input into the frequency enhancement coding unit, and frequency domain transformation, frequency component selection and inverse transformation reconstruction are performed on the decomposed cabinet state sequence to obtain the frequency enhancement state sequence. The frequency-enhanced state sequence is input into the encoder feedforward layer and residual normalization layer for hidden state updates, generating coded hidden states.

6. The method for diagnosing partial discharge faults in high-voltage switchgear based on deep learning according to claim 4, characterized in that, The specific steps of invoking the encoded hidden state for state correction at each stage of advancement include: Global pooling and linear mapping are performed on the encoded hidden states to obtain the candidate fault development states; The candidate fault development state is input into the state advancement unit, and multi-level state advancement is carried out according to the preset fault stage chain. Each level of state advancement reads the candidate fault development state and the encoded hidden state of the previous stage, and writes the hidden layer features corresponding to the current stage in the encoded hidden state into the candidate fault development state of the previous stage to obtain the candidate fault development state of the current stage. After each state progression, the state transition discrimination unit determines whether the candidate fault development state of the current stage meets the conditions for entering the next stage. When the candidate fault development state passes through the state transition judgment of each level in the preset fault stage chain continuously, the last level candidate fault development state is determined as the mature fault state vector. When a candidate fault development state fails to pass the state transition judgment at any level, a low-confidence label is added to the candidate fault development state, and the candidate fault development state after adding the low-confidence label is used as the mature fault state vector.

7. The method for diagnosing partial discharge faults in high-voltage switchgear based on deep learning according to claim 6, characterized in that, The specific steps of identifying the region of evidence supporting the initial diagnostic category in the sequence of coded hidden states and virtual cabinet states include: The mature fault state vector is mapped to the category score of each preset diagnostic category, and the preset diagnostic category with the highest category score is determined as the initial diagnostic category; Generate a category retrieval vector according to the initial diagnosis category, and apply the category retrieval vector to the encoded hidden state and the virtual cabinet state sequence respectively, and perform response retrieval on the time slice features in the encoded hidden state and the cabinet state in the virtual cabinet state sequence. During the response retrieval process, the coded hidden state response and the virtual cabinet state response within the same time slice are compared for consistency. The time slice and cabinet position that both point to the initial diagnostic category are retained, and the positions that respond only in the coded hidden state or only in the virtual cabinet state sequence are removed. The retained locations are merged sequentially along the time line, and adjacent retained locations with an interval less than a preset interval threshold are merged into the same candidate evidence region; The region with the strongest response is selected from the candidate evidence region and identified as the evidence region supporting the initial diagnostic category.

8. The method for diagnosing partial discharge faults in high-voltage switchgear based on deep learning according to claim 7, characterized in that, The specific steps of performing hidden counterfactual erasure on the evidence area include: Read the time slice range, cabinet location, and hidden state index of the evidence area record, and determine the hidden state corresponding to the evidence area in the encoded hidden state and virtual cabinet state sequence according to the time slice range, cabinet location, and hidden state index; Read the category retrieval vector corresponding to the initial diagnosis category, decompose the hidden state corresponding to the evidence region into category-related components and category-independent components along the direction of the category retrieval vector, delete the category-related components, and retain the category-independent components; Select hidden states that are not covered by the evidence area from the adjacent time slices and adjacent cabinet locations of the evidence area, and generate alternative hidden states based on the selected hidden states. Write the alternative hidden state into the evidence region after deleting the category-related components. The hidden state corresponding to the evidence region no longer retains the category-related information that supports the initial diagnostic category. The boundaries of the evidence region after writing the alternative hidden state are continuously updated to generate a counterfactual hidden state.

9. The method for diagnosing partial discharge faults in high-voltage switchgear based on deep learning according to claim 8, characterized in that, The specific details of the relationship between the initial diagnostic category and the secondary diagnostic category include: Input the counterfactual hidden state into the secondary diagnostic unit to obtain the secondary category score corresponding to each preset diagnostic category, and determine the preset diagnostic category with the largest secondary category score as the secondary diagnostic category. Compare the initial diagnosis category with the secondary diagnosis category. When the secondary diagnosis category is different from the initial diagnosis category, confirm the evidence region that the initial diagnosis category depends on and write the initial diagnosis category into the fault diagnosis result. When the secondary diagnosis category is the same as the initial diagnosis category, and the difference between the secondary category score and the initial category score is less than a preset change threshold, the fault diagnosis result is marked as a low-confidence diagnosis result. When the secondary diagnostic category is the same as the initial diagnostic category, and the difference between the secondary category score and the initial category score is greater than or equal to a preset change threshold, the initial diagnostic category and confidence decay flag are written into the fault diagnosis result.

10. A deep learning-based partial discharge fault diagnosis system for high-voltage switchgear, applied to any one of the deep learning-based partial discharge fault diagnosis methods for high-voltage switchgear as described in claims 1 to 9, characterized in that, include: The data processing module is used to acquire the partial discharge monitoring sequence and the operating status sequence of the high-voltage switchgear, perform preprocessing on the partial discharge monitoring sequence and the operating status sequence, and generate partial discharge time series samples. An embedded processing module is used to convert partial discharge time series samples into an initial hidden state sequence and send the initial hidden state sequence into the improved FEDformer. The virtual cabinet propagation module is used to establish the cabinet propagation topology based on the internal compartment connection relationship of the high-voltage switchgear, convert the initial hidden state sequence into the cabinet node state, and generate the virtual cabinet state sequence through directed propagation connection, boundary determination and state update. The FEDformer encoding module is used to perform sequence decomposition and frequency enhancement encoding on the virtual cabinet state sequence to generate encoded hidden states. The fault development state advancement module is used to extract candidate fault development states from the encoded hidden states and advance along the preset fault stage chain to generate a mature fault state vector. The counterfactual diagnosis module is used to generate an initial diagnosis category based on the mature fault state vector, determine the evidence region, perform hidden counterfactual erasure on the evidence region, and generate a fault diagnosis result based on the secondary diagnosis result.

Images