Power transformer partial discharge signal extraction and diagnosis method combined with deep learning

Abstract

The invention discloses a deep learning-combined power transformer partial discharge signal extraction and diagnosis method. The method comprises the following steps of S1, setting a multi-channel synchronous acquisition system in a power transformer body area to acquire a multi-dimensional original partial discharge data set; s2, preprocessing the acquired multi-dimensional original partial discharge data set; s3, performing time alignment and amplitude matching on the processed signal, and dividing the processed signal into a sliding time window to construct a standard input tensor; s4, constructing an attention enhancement model fused by the convolutional neural network and the bidirectional gating circulation unit; s5, performing supervised training on the attention enhancement model by using the labeled sample; s6, inputting the real-time signal into the training model, and outputting a discharge type label; s7, risk grade evaluation is carried out in combination with statistical characteristics; and S8, generating a structured diagnosis report and uploading the structured diagnosis report to a monitoring platform. According to the invention, multi-source signals and a depth model are fused, and intelligent diagnosis and risk assessment of transformer partial discharge are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnostic technology for power equipment, and in particular to a method for extracting and diagnosing partial discharge signals from power transformers using deep learning. Background Technology

[0002] Partial discharge is an early sign of insulation degradation and potential fault development in power transformers, and its condition monitoring and accurate identification are crucial for the safe operation of power systems. Traditional partial discharge detection methods mainly rely on various sensing technologies such as electrical methods, ultra-high frequency methods, and acoustic emission methods. These methods collect discharge signals through sensors and perform frequency or time domain analysis to determine the presence and type of discharge. However, existing methods generally suffer from weak multi-source signal fusion capabilities, high levels of human intervention, and unstable identification accuracy. On the one hand, under field operating conditions, partial discharge signals are often overwhelmed by power frequency interference, environmental noise, and signals from other equipment, making signal extraction difficult. On the other hand, traditional feature extraction methods heavily rely on empirical rules and lack the ability to deeply model complex nonlinear time-series signals, making it difficult to accurately identify the type and development trend of partial discharge. Furthermore, many diagnostic systems still rely on manual judgment and intervention, lacking a unified and efficient closed-loop path for signal processing, identification, and evaluation.

[0003] While some studies have attempted to incorporate machine learning algorithms for partial discharge pattern recognition, current technologies still face bottlenecks such as the strong subjectivity of feature dimension selection, insufficient modeling capabilities for time series analysis, and recognition accuracy limited by noise. Particularly under conditions of collaborative acquisition by multiple types of sensors, the time synchronization, amplitude consistency, and structural uniformity of different signal sources are often difficult to guarantee, leading to challenges in subsequent model processing. Furthermore, existing systems generally fail to achieve a complete automated process from signal acquisition, preprocessing, classification and recognition to risk level assessment and visualization output, resulting in low operational efficiency and delayed diagnostic response.

[0004] To address the aforementioned limitations and shortcomings of existing technologies, this invention proposes a method for extracting and diagnosing partial discharge signals from power transformers using deep learning. This method involves constructing a multi-point sensor deployment and multi-channel synchronous acquisition system, combining wavelet packet decomposition and normalization processing to achieve high-quality signal extraction. Furthermore, a composite deep network structure integrating convolutional neural networks, bidirectional gated recurrent units, and attention mechanisms is designed to enhance the feature representation and discharge type identification capabilities of multi-source time-series signals. By introducing a standardized three-dimensional input tensor construction method and a sliding time window mechanism, model input consistency is ensured, and risk grading logic is used to determine the severity level of different types of discharge events. Finally, through the integration of structured diagnostic report generation and a remote communication module, real-time, intelligent, and closed-loop management of the insulation status of power transformers is achieved, overcoming the technical deficiencies of existing technologies in terms of signal extraction accuracy, diagnostic automation, and system response capabilities.

[0005] Therefore, how to provide a method for extracting and diagnosing partial discharge signals from power transformers using 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 for extracting and diagnosing partial discharge signals from power transformers that combines deep learning. This invention integrates the acquisition of multiple types of partial discharge signals, multi-scale wavelet processing, and deep neural network modeling to construct an end-to-end partial discharge diagnosis and risk assessment process, thereby improving the accuracy of signal extraction and recognition precision. It enables intelligent analysis, risk classification, and remote monitoring of transformer insulation status, and has good engineering practicality and promotional value.

[0007] The method for extracting and diagnosing partial discharge signals from power transformers using deep learning, according to embodiments of the present invention, includes the following steps: S1. Ultra-high frequency electromagnetic sensors, current pulse sensors, and piezoelectric acoustic emission sensors are installed in the winding joint area, insulating oil channel area, and outside the metal shell of the power transformer. A multi-channel synchronous acquisition system is used to simultaneously acquire the partial discharge signals output by the three types of sensors while the transformer is energized, forming a multi-dimensional raw partial discharge dataset containing different physical channels and spatial sources. S2. Noise filtering and baseline drift correction are performed on the acquired multi-channel raw signals. Bandpass filters are used to eliminate power frequency and higher harmonic interference. Based on wavelet packet decomposition, the signals are decomposed into multi-scale components containing the target frequency band, and then normalized to ensure uniformity across all channels. S3. Map the signal amplitude to a unified range to improve the comparability of multi-source features; S4. Perform time alignment and amplitude matching on the preprocessed signals of different channels to construct a unified multi-channel signal matrix, divide it into fixed-length sliding time window segments according to time order, form an input tensor with time dimension, channel dimension and sampling point dimension, and perform batch encapsulation processing for use in the neural network training and inference process; S5. Construct an attention enhancement model that integrates a convolutional neural network and a bidirectional gated recurrent unit, the model including: a multi-layer convolutional neural network module for extracting local spatial patterns of partial discharge signals; a bidirectional gated recurrent unit module for modeling the evolution of signals over time; and a module for applying the signal to the time-series feature map. The model includes: an attention mechanism module that weights key segments; and a fully connected module for generating the final classification output; S5, supervised training of the model using historical partial discharge data samples with manual labels, using the input tensor generated in step S3 as the training sample input, using the cross-entropy loss function to measure classification bias, and iteratively adjusting the model weights through backpropagation combined with a gradient-based parameter update method until the model reaches convergence on both the training and validation sets; S6, inputting the partial discharge signal collected in real time during the operation of the power transformer and processed by steps S2 and S3 into the model trained in step S5, and the model sequentially extracting spatial structure features, temporal features, etc. The features are weighted and aggregated to form a global embedding vector, and finally the identification result label of partial discharge type is output through the fully connected classification module; S7, based on the discharge type label and its corresponding signal statistical parameters output in step S6, including amplitude change trend, recurrence frequency, waveform envelope shape and signal duration, etc., multi-dimensional risk assessment is carried out in combination with preset classification rules to generate the corresponding partial discharge risk level label; S8, the identification label and risk level are encapsulated to generate a structured diagnostic report, which is uploaded to the power system monitoring platform through a remote communication interface, and at the same time, status alarms, data archiving and operation and maintenance suggestions are triggered to realize continuous tracking and intelligent management of the internal insulation status of power transformers.

[0008] Optionally, S1 specifically includes: S11. Set up sensor installation nodes at the location of the power transformer body, selecting areas including the high-voltage winding lead outlet, winding end, adjacent area of ​​capacitor bushing, inner wall of oil tank, and neutral grounding terminal as deployment areas; S12. Install multiple types of sensor devices, including electromagnetic sensors, current pulse sensors, and piezoelectric acoustic emission sensors, at each node to ensure that the sensors spatially cover the high-risk area of ​​transformer discharge, forming a multi-physical channel signal acquisition system; S13. Use a data acquisition device with multi-channel synchronous acquisition capability to parallelly access the output signals of multiple types of sensors, and configure independent numbers and identifiers for each channel for subsequent data alignment and correlation modeling; S14. Set the initialization parameters of the acquisition channels, including channel activation status, sampling order, buffer size, and data storage path, to ensure that the acquisition system achieves time synchronization and data isolation between different types of signals; S15. Start the acquisition program to continuously record the original partial discharge signals output by each sensor while the transformer is running, forming a time-continuous and channel-independent multi-dimensional original partial discharge dataset.

[0009] Optionally, S2 specifically includes: S21. Bandpass filtering is performed on the raw partial discharge signals acquired from each channel to remove low-frequency power frequency interference and high-frequency background noise, while retaining the frequency components related to partial discharge activity. S22. Wavelet packet decomposition is performed on the bandpass-filtered signals to divide the original signals into multiple frequency band sub-signals, and the reconstructed components containing the target frequency components are selected as the effective signals. S23. Amplitude normalization is performed on the reconstructed effective signals to map the amplitudes of each channel signal to a unified standard range, ensuring consistency of multi-channel data at the feature scale. S24. The normalized sensor signals are stored in their corresponding cache structures, maintaining the channel number and time order, providing a consistent data format for subsequent feature combination and model input.

[0010] Optionally, S3 specifically includes: S31. Perform time-base alignment on the pre-processed signals from different types of sensors. Based on the clock index information recorded by the acquisition system, perform time interpolation, cropping, or alignment buffer adjustment on each channel signal to ensure that the sampling points of all channels have consistent start and end marks on the same time axis. S32. Perform amplitude matching processing on the time-aligned signals. Based on a preset amplitude normalization function, readjust the amplitude scale of all channel signals to construct an equal expression of signal amplitude values ​​within a consistent range, eliminating the interference of differences in discharge response amplitude between channels on subsequent feature learning. S33. Divide the processed multi-channel signals into time sequences using a sliding time window. S34. A fixed window length and overlap step size are set to divide the entire signal sequence into multiple continuous signal segments, each representing the dynamic behavior of partial discharge events in different time periods; S35. The multi-channel signals in each time window segment are stacked into a unified matrix structure according to the channel order, and then combined sequentially into a three-dimensional input tensor with time dimension, channel dimension, and sampling point dimension. The input tensor is used to construct input data batches in units of windows; S36. The constructed multi-dimensional input tensors are batch-encapsulated to generate a standardized neural network input sample structure, while recording the window index, time period identifier, and sensor source label to provide a consistent input format for subsequent model training and real-time inference processes.

[0011] Optionally, the attention enhancement model includes an input layer, a multi-layer convolutional neural network module, a bidirectional gated recurrent unit module, an attention mechanism module, a fully connected module, and an output layer. The input layer receives the three-dimensional input tensor encapsulated in step S3, and the three-dimensional input tensor includes time series sampling points and corresponding amplitude information. The multi-layer convolutional neural network module includes multiple convolutional layers, normalization layers, and activation function layers. The convolutional layers extract the spatial features of the signal from the input tensor. The output of each convolutional layer is processed by a batch normalization layer and mapped by a nonlinear activation function layer to obtain a feature map. The feature map is used to represent the edge structure, frequency domain pattern, and inter-channel correlation of the partial discharge signal. The feature map is then subjected to dimensionality reduction or flattening to generate a two-dimensional feature sequence with a time dimension structure. The bidirectional gated loop unit module includes a forward gated loop unit layer and a reverse gated loop unit layer. The two layers process the flattened feature sequence from the forward and backward directions of time, respectively, to extract the change state of the signal in the bidirectional time path and output the hidden state sequence. The attention mechanism module performs weighted operations on the hidden states of all time steps to generate an aggregated feature vector to enhance the contribution of key time segments. The aggregated feature vector is then fused with the high-dimensional feature map of the last layer of the convolution module to form a comprehensive expression feature that includes local spatial patterns and global temporal dependencies. The fully connected module receives the fused comprehensive expression features. The fully connected module consists of at least one fully connected layer, and non-linear activation functions are connected between each layer to complete feature transformation and classification decision mapping. The output layer is located at the end of the fully connected module. The number of output nodes is equal to the number of partial discharge type categories, and it generates classification labels corresponding to the discharge signals, which are provided to subsequent system modules as partial discharge diagnosis results.

[0012] Optionally, S5 specifically includes: S51. Collect multi-source partial discharge signal samples recorded during historical operation, and manually label each signal data. The manual labeling is based on expert judgment or experimental verification to form a one-to-one correspondence between partial discharge signals and discharge type labels, and construct a standard supervised training sample set; S52. Bind the three-dimensional input tensor constructed in step S3 with the label to form training sample pairs, index and encode all samples, and perform stratified sampling according to the sample time distribution and category distribution to divide them into training set, validation set and test set, with each subset maintaining non-overlapping relationships; S53. Define the training parameter configuration file, and set the optimizer type and loss function. The format, learning rate adjustment strategy, regularization penalty term, training batch size, maximum number of training epochs, and early stopping monitoring indicators are used to control the convergence path and generalization ability during model training; S54, the parameters of the convolutional neural network module, bidirectional gated recurrent unit module, attention mechanism module, and fully connected output module in the attention enhancement model are initialized respectively, and a uniform random seed is used to ensure the repeatability of model training, and a complete computation graph and training dependencies are constructed; S55, during training, sample tensors are loaded from the training set in batches and input into the attention enhancement model to perform forward computation, and local spatial features are extracted sequentially through convolutional layers, and then... The bidirectional gated recurrent unit module models temporal dependencies, generates weighted aggregated features through the attention mechanism module, and outputs the predicted label distribution through the fully connected module; S56, calculates the classification error between the predicted output and the actual label, inputs the error into the cross-entropy loss function to obtain the loss value of the current training batch, and executes the backpropagation algorithm based on the constructed error propagation path to automatically solve the gradient values ​​of all trainable parameters in the attention enhancement model; S57, based on the calculated gradient information, calls the set optimizer to perform weight update operations on the parameters of each layer in the attention enhancement model, and records the changes in the loss curve and classification accuracy in each round. Changes are used for visual monitoring and dynamic adjustment during the training process; S58. After each training round, the current attention enhancement model is evaluated using the validation set. The loss value, accuracy, confusion matrix, and class balance index on the validation set are recorded to evaluate the stability and generalization performance of the attention enhancement model. The current best model snapshot is updated based on the validation set index; S59. When the attention enhancement model does not observe a significant improvement in the validation index in multiple consecutive training rounds, or when the training has reached the maximum round limit, the training process is terminated, and the final stable model structure, weight parameters, configuration file, and label mapping rules are saved for subsequent inference module calls.

[0013] Optionally, S6 specifically includes: S61, inputting the partial discharge signal collected in real time during the operation of the power transformer and processed by steps S2 and S3 into the attention enhancement model trained in step S5; S62, first passing the input tensor to the convolutional neural network module, and sequentially passing it through multiple sets of convolutional layers, normalization layers, and activation function layers to extract the feature patterns and local structural information of the partial discharge signal in the spatial domain, and outputting a spatial feature map; S63, flattening the spatial feature map and converting it into a time series format, inputting it into the bidirectional gated recurrent unit module, wherein the bidirectional gated recurrent unit module models the temporal evolution relationship within the sequence based on two paths, forward and reverse, and outputs a sequence containing... S64. The hidden state sequence with time dependency is input to the attention mechanism module, and weights are assigned to the output of all time steps to enhance the features of key time segments and form an aggregated global embedding vector; S65. The global embedding vector is input to the fully connected module, which consists of multiple linear mapping layers and nonlinear transformation layers. Finally, a probability distribution vector of partial discharge type is generated at the output layer, and the label corresponding to the maximum probability value is used as the current recognition result; S66. The identified discharge type label, the original signal timestamp, and the acquisition channel information are output together and transmitted to the risk assessment and diagnostic report generation module for result archiving, visualization, and uploading.

[0014] Optionally, S7 specifically includes: S71. Receive the discharge type tag output from step S6, and acquire the original signal segment corresponding to the discharge type tag and its acquisition data in each channel; S72. Perform amplitude statistical analysis on the original signal segment, extracting the maximum amplitude, average amplitude, fluctuation range, and amplitude change trend to measure discharge intensity and stability characteristics; S73. Perform frequency domain analysis on the original signal segment, extracting frequency-related parameters including repetition frequency, dominant frequency energy distribution, and discharge interval density to reflect the repetition and periodicity of partial discharge events; S74. Model the signal contour morphology based on the time-domain envelope extraction algorithm, acquiring the upper envelope, lower envelope, and overall morphological change trend of the waveform envelope to identify abnormal discharge morphological change characteristics; S75. Measure the signal duration... Measurement and classification are performed, and the duration, concentration, and time span of continuous signal segments under the same label are combined to assess the level of discharge persistence; S76, the extracted multidimensional statistical features are input into the risk assessment module and compared with the preset risk classification rules in the system to determine the risk level score of the current discharge event under different assessment dimensions; S77, the scores of each assessment dimension are weighted and summarized according to the comparison results to generate a comprehensive risk assessment result corresponding to the current signal segment, and divided into multiple levels, including low risk, medium risk, and high risk categories; S78, the generated partial discharge risk level identifier, its corresponding identification label, statistical features, and signal time range are packaged together to form structured risk information for subsequent report generation and remote upload modules to call.

[0015] Optionally, S8 specifically includes: S81. Receive the discharge type tag output from step S6 and the risk level identifier output from step S7, and integrate them with the corresponding original signal data, statistical characteristics, and processing timestamp; S82. Construct a diagnostic report data structure, organizing the integrated information into data entities containing equipment identifier, signal channel information, identification tag, risk level, statistical indicators, and time interval content; S83. Structure the diagnostic data into a unified message format, generate a diagnostic report document conforming to the power system data interface specification, and assign a unique diagnostic number and timestamp index; S84. Call the remote communication interface to transmit the diagnostic report document to the receiving port of the power system monitoring platform through a power-specific communication protocol or edge computing interface; S85. Upload the diagnostic report. Upon successful completion, the status alarm module is automatically triggered. Based on the identification tag and risk level, it determines whether to generate a real-time alarm event and pushes it to the relevant alarm channels. S86: The data archiving module is invoked to store the diagnostic results information in the local database and cloud server, establishing a searchable index structure sorted by device number, signal type, and diagnostic time. S87: Combining the identification tag and risk level, the operation and maintenance suggestion generation module is invoked to match corresponding inspection suggestions, handling measures, and maintenance cycle reminders from the rule base, generating action guidelines for operation and maintenance personnel. S88: The diagnostic report, alarm information, archived records, and operation and maintenance suggestions are uniformly organized into a comprehensive diagnostic output package for the monitoring system to call and display, realizing visualized monitoring of the insulation status of power transformers, fault trend tracking, and closed-loop management.

[0016] The beneficial effects of this invention are: This invention systematically reconstructs the process of acquiring, processing, identifying, and evaluating partial discharge signals from power transformers, significantly improving the automation, intelligence, and accuracy of partial discharge diagnosis. It overcomes the technical bottlenecks of traditional methods, such as low signal extraction accuracy, weak model representation ability, and high dependence on manual intervention, and has the following significant beneficial effects.

[0017] First, this invention achieves a combination of high spatial resolution and channel synchronization in sensor deployment and signal acquisition, fully covering the key discharge-sensitive areas of the transformer. It also adopts a multi-channel parallel acquisition structure to ensure that various partial discharge signals are physically synchronized and isolated at the time level, laying a high-quality data foundation for subsequent signal fusion and classification.

[0018] Secondly, in the signal processing stage, the inventors proposed a multi-scale filtering strategy and normalization process based on wavelet packet decomposition, which effectively improved the recognition of partial discharge signals in complex electromagnetic environments. By constructing a unified format and aligning the amplitudes of signals from different channels, and generating standard input tensors using a sliding time window method, the model training and inference possess high structural stability and temporal consistency, thereby improving the overall data controllability and interpretability of the model input.

[0019] Finally, the deep neural network structure constructed in this invention, which integrates convolutional neural networks, bidirectional gated recurrent units, and attention mechanisms, possesses both spatial feature extraction and temporal dependency modeling capabilities. It can perform in-depth characterization of multi-channel, multi-scale discharge features and significantly enhances the model's response sensitivity to key discharge segments through the attention mechanism, thereby greatly improving the accuracy and generalization ability of partial discharge type identification. This invention designs a multi-dimensional risk assessment mechanism based on discharge statistical features. Combining the identification results with multi-dimensional indicators such as amplitude, frequency, duration, and waveform morphology, a complete discharge level assessment system is constructed, achieving end-to-end closed-loop judgment from discharge identification to risk assessment. Simultaneously, the accompanying diagnostic report encapsulation, remote communication uploading, and alarm triggering mechanisms provide real-time, visual, and structured data support for the power system monitoring platform, effectively supporting the optimization of transformer equipment operation and maintenance strategies and online status assessment.

[0020] In summary, this invention not only achieves a fully automated processing path in terms of technical process and improves the model's ability to express complex partial discharge signals in terms of algorithm structure, but also realizes a truly deployable, traceable, and manageable intelligent diagnostic system in terms of system integration and result application, which has good engineering practical value and promotion prospects. Attached Figure Description

[0021] 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:

[0022] Figure 1 This is an overall flowchart of the power transformer partial discharge signal extraction and diagnosis method combining deep learning proposed in this invention; Figure 2 This is a structural diagram of the attention enhancement model that integrates a convolutional neural network and a bidirectional gated recurrent unit proposed in this invention. Figure 3 This is a schematic diagram of the multi-source partial discharge signal preprocessing and tensor construction process proposed in this invention. Detailed Implementation

[0023] 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.

[0024] refer to Figure 1-3 The method for extracting and diagnosing partial discharge signals from power transformers using deep learning includes the following steps: S1. Ultra-high frequency electromagnetic sensors, current pulse sensors, and piezoelectric acoustic emission sensors are installed in the winding joint area, insulating oil channel area, and outside the metal shell of the power transformer. A multi-channel synchronous acquisition system is used to simultaneously acquire the partial discharge signals output by the three types of sensors while the transformer is energized, forming a multi-dimensional raw partial discharge dataset containing different physical channels and spatial sources. S2. Noise filtering and baseline drift correction are performed on the acquired multi-channel raw signals. Bandpass filters are used to eliminate power frequency and higher harmonic interference. Based on wavelet packet decomposition, the signals are decomposed into multi-scale components containing the target frequency band, and then normalized to ensure uniformity across all channels. S3. Map the signal amplitude to a unified range to improve the comparability of multi-source features; S4. Perform time alignment and amplitude matching on the preprocessed signals of different channels to construct a unified multi-channel signal matrix, divide it into fixed-length sliding time window segments according to time order, form an input tensor with time dimension, channel dimension and sampling point dimension, and perform batch encapsulation processing for use in the neural network training and inference process; S5. Construct an attention enhancement model that integrates a convolutional neural network and a bidirectional gated recurrent unit, the model including: a multi-layer convolutional neural network module for extracting local spatial patterns of partial discharge signals; a bidirectional gated recurrent unit module for modeling the evolution of signals over time; and a module for applying the signal to the time-series feature map. The model includes: an attention mechanism module that weights key segments; and a fully connected module for generating the final classification output; S5, supervised training of the model using historical partial discharge data samples with manual labels, using the input tensor generated in step S3 as the training sample input, using the cross-entropy loss function to measure classification bias, and iteratively adjusting the model weights through backpropagation combined with a gradient-based parameter update method until the model reaches convergence on both the training and validation sets; S6, inputting the partial discharge signal collected in real time during the operation of the power transformer and processed by steps S2 and S3 into the model trained in step S5, and the model sequentially extracting spatial structure features, temporal features, etc. The features are weighted and aggregated to form a global embedding vector, and finally the identification result label of partial discharge type is output through the fully connected classification module; S7, based on the discharge type label and its corresponding signal statistical parameters output in step S6, including amplitude change trend, recurrence frequency, waveform envelope shape and signal duration, etc., multi-dimensional risk assessment is carried out in combination with preset classification rules to generate the corresponding partial discharge risk level label; S8, the identification label and risk level are encapsulated to generate a structured diagnostic report, which is uploaded to the power system monitoring platform through a remote communication interface, and at the same time, status alarms, data archiving and operation and maintenance suggestions are triggered to realize continuous tracking and intelligent management of the internal insulation status of power transformers.

[0025] This invention presents a complete closed-loop processing path from partial discharge signal acquisition to deep learning diagnosis and result visualization, effectively solving problems in existing systems such as difficulty in integrating multi-source signals, reliance on human experience in the diagnostic process, and poor response timeliness. By integrating structured input, deep model representation, and remote reporting linkage mechanisms, this method supports real-time intelligent monitoring of the insulation status of power transformers, improving identification accuracy, operation and maintenance response speed, and large-scale deployment capabilities, demonstrating strong engineering practicality and scalability.

[0026] In this embodiment, S1 specifically includes: S11. Set up sensor installation nodes at the location of the power transformer body, selecting areas including the high-voltage winding lead outlet, winding end, adjacent area of ​​capacitor bushing, inner wall of oil tank, and neutral grounding terminal as deployment areas; S12. Install multiple types of sensor devices, including electromagnetic sensors, current pulse sensors, and piezoelectric acoustic emission sensors, at each node to ensure that the sensors spatially cover the high-risk area of ​​transformer discharge, forming a multi-physical channel signal acquisition system; S13. Use a data acquisition device with multi-channel synchronous acquisition capability to parallelly access the output signals of multiple types of sensors, and configure independent numbers and identifiers for each channel for subsequent data alignment and correlation modeling; S14. Set the initialization parameters of the acquisition channels, including channel activation status, sampling order, buffer size, and data storage path, to ensure that the acquisition system achieves time synchronization and data isolation between different types of signals; S15. Start the acquisition program to continuously record the original partial discharge signals output by each sensor while the transformer is running, forming a time-continuous and channel-independent multi-dimensional original partial discharge dataset.

[0027] This invention fully considers the spatial characteristics of areas prone to partial discharge in power transformers when implementing sensor deployment strategies, proposing a nodalized arrangement method for multiple types of sensors to ensure the spatial coverage integrity and physical diversity of discharge signals. Simultaneously, through a multi-channel numbering mechanism and synchronous acquisition configuration, consistency of different signal sources on the time axis is achieved, effectively improving cross-channel data alignment efficiency and signal fusion quality, and enhancing the adaptability of subsequent time-series modeling to heterogeneous inputs.

[0028] In this embodiment, S2 specifically includes: S21. Bandpass filtering is performed on the raw partial discharge signals acquired from each channel to remove low-frequency power frequency interference and high-frequency background noise, while retaining the frequency components related to partial discharge activity. S22. Wavelet packet decomposition is performed on the bandpass-filtered signals to divide the original signals into multiple frequency band sub-signals, and the reconstructed components containing the target frequency components are selected as the effective signals. S23. Amplitude normalization is performed on the reconstructed effective signals to map the amplitudes of each channel signal to a unified standard range, ensuring consistency of multi-channel data at the feature scale. S24. The normalized sensor signals are stored in their corresponding cache structures, maintaining the channel number and time order, providing a consistent data format for subsequent feature combination and model input.

[0029] This invention employs a multi-stage processing flow combining bandpass filtering and wavelet packet decomposition, which can effectively remove the partial discharge component of the target in complex operating backgrounds, enhancing the significance and stability of the discharge characteristic signal. By normalizing the amplitude of each channel, a data input format with a unified expression scale is constructed, solving the feature expression offset problem caused by the difference in response amplitude of multiple types of sensors, and significantly improving the learning consistency and robustness of the neural network for cross-channel data.

[0030] In this embodiment, S3 specifically includes: S31. Perform time-base alignment on the pre-processed signals from different types of sensors. Based on the clock index information recorded by the acquisition system, perform time interpolation, cropping, or alignment buffer adjustment on each channel signal to ensure that the sampling points of all channels have consistent start and end marks on the same time axis. S32. Perform amplitude matching processing on the time-aligned signals. Based on a preset amplitude normalization function, readjust the amplitude scale of all channel signals to construct an equal expression of signal amplitude values ​​within a consistent range, eliminating the interference of differences in discharge response amplitude between channels on subsequent feature learning. S33. Divide the processed multi-channel signals into time sequences using a sliding time window. S34. A fixed window length and overlap step size are set to divide the entire signal sequence into multiple continuous signal segments, each representing the dynamic behavior of partial discharge events in different time periods; S35. The multi-channel signals in each time window segment are stacked into a unified matrix structure according to the channel order, and then combined sequentially into a three-dimensional input tensor with time dimension, channel dimension, and sampling point dimension. The input tensor is used to construct input data batches in units of windows; S36. The constructed multi-dimensional input tensors are batch-encapsulated to generate a standardized neural network input sample structure, while recording the window index, time period identifier, and sensor source label to provide a consistent input format for subsequent model training and real-time inference processes.

[0031] To address the issues of time axis error and amplitude inconsistency in multi-channel signal acquisition, this invention introduces an information alignment mechanism based on the acquisition system clock index and an amplitude consistency mapping method to achieve a unified representation of signals in a multi-dimensional structure. By using a sliding time window partitioning mechanism to construct continuous signals as trainable samples, the time-dependent behavior of partial discharge events is effectively extracted, providing a dynamic segment-level input foundation for the model and enhancing the subsequent modeling ability for sequence variation features.

[0032] In this embodiment, the attention enhancement model includes an input layer, a multi-layer convolutional neural network module, a bidirectional gated recurrent unit module, an attention mechanism module, a fully connected module, and an output layer: The input layer receives the three-dimensional input tensor encapsulated in step S3, and the three-dimensional input tensor includes time series sampling points and corresponding amplitude information. The multi-layer convolutional neural network module includes multiple convolutional layers, normalization layers, and activation function layers. The convolutional layers extract the spatial features of the signal from the input tensor. The output of each convolutional layer is processed by a batch normalization layer and mapped by a nonlinear activation function layer to obtain a feature map. The feature map is used to represent the edge structure, frequency domain pattern, and inter-channel correlation of the partial discharge signal. The feature map is then subjected to dimensionality reduction or flattening to generate a two-dimensional feature sequence with a time dimension structure. The bidirectional gated loop unit module includes a forward gated loop unit layer and a reverse gated loop unit layer. The two layers process the flattened feature sequence from the forward and backward directions of time, respectively, to extract the change state of the signal in the bidirectional time path and output the hidden state sequence. The attention mechanism module performs weighted operations on the hidden states of all time steps to generate an aggregated feature vector to enhance the contribution of key time segments. The aggregated feature vector is then fused with the high-dimensional feature map of the last layer of the convolution module to form a comprehensive expression feature that includes local spatial patterns and global temporal dependencies. The fully connected module receives the fused comprehensive expression features. The fully connected module consists of at least one fully connected layer, and non-linear activation functions are connected between each layer to complete feature transformation and classification decision mapping. The output layer is located at the end of the fully connected module. The number of output nodes is equal to the number of partial discharge type categories, and it generates classification labels corresponding to the discharge signals, which are provided to subsequent system modules as partial discharge diagnosis results.

[0033] This invention designs a deep neural network structure that integrates convolutional modules, bidirectional gated recurrent modules, and an attention mechanism to construct a multidimensional information extraction framework that simultaneously possesses spatial awareness and temporal dependency modeling capabilities. The convolutional modules enhance the structural recognition of local spatial patterns, the gated recurrent network tracks the changing paths of partial discharge signal sequences, and the attention mechanism dynamically focuses on key segment signals, thereby achieving accurate classification of complex discharge morphologies and improving global representation capabilities.

[0034] In this embodiment, S5 specifically includes: S51. Collect multi-source partial discharge signal samples recorded during historical operation, and manually label each signal data. The manual labeling is based on expert judgment or experimental verification to form a one-to-one correspondence between partial discharge signals and discharge type labels, and construct a standard supervised training sample set; S52. Bind the three-dimensional input tensor constructed in step S3 with the label to form training sample pairs, index and encode all samples, and perform stratified sampling according to the sample time distribution and category distribution to divide them into training set, validation set and test set, with each subset maintaining non-overlapping relationships; S53. Define the training parameter configuration file, and set the optimizer type and loss function. The format, learning rate adjustment strategy, regularization penalty term, training batch size, maximum number of training epochs, and early stopping monitoring indicators are used to control the convergence path and generalization ability during model training; S54, the parameters of the convolutional neural network module, bidirectional gated recurrent unit module, attention mechanism module, and fully connected output module in the attention enhancement model are initialized respectively, and a uniform random seed is used to ensure the repeatability of model training, and a complete computation graph and training dependencies are constructed; S55, during training, sample tensors are loaded from the training set in batches and input into the attention enhancement model to perform forward computation, and local spatial features are extracted sequentially through convolutional layers, and then... The bidirectional gated recurrent unit module models temporal dependencies, generates weighted aggregated features through the attention mechanism module, and outputs the predicted label distribution through the fully connected module; S56, calculates the classification error between the predicted output and the actual label, inputs the error into the cross-entropy loss function to obtain the loss value of the current training batch, and executes the backpropagation algorithm based on the constructed error propagation path to automatically solve the gradient values ​​of all trainable parameters in the attention enhancement model; S57, based on the calculated gradient information, calls the set optimizer to perform weight update operations on the parameters of each layer in the attention enhancement model, and records the changes in the loss curve and classification accuracy in each round. Changes are used for visual monitoring and dynamic adjustment during the training process; S58. After each training round, the current attention enhancement model is evaluated using the validation set. The loss value, accuracy, confusion matrix, and class balance index on the validation set are recorded to evaluate the stability and generalization performance of the attention enhancement model. The current best model snapshot is updated based on the validation set index; S59. When the attention enhancement model does not observe a significant improvement in the validation index in multiple consecutive training rounds, or when the training has reached the maximum round limit, the training process is terminated, and the final stable model structure, weight parameters, configuration file, and label mapping rules are saved for subsequent inference module calls.

[0035] In the supervised learning process, this invention constructs a complete training system covering sample labeling, data partitioning, training parameter tuning, gradient updating, validation and evaluation, and model saving, significantly enhancing the robustness and generalization ability of deep models under different sample distributions. By introducing stratified sampling, early stopping strategies, and model snapshot mechanisms, overfitting and training oscillations are effectively avoided, ensuring that the final model has strong engineering deployment stability and portability, adapting to various operating environments.

[0036] In this embodiment, S6 specifically includes: S61, inputting the partial discharge signal collected in real time during the operation of the power transformer and processed by steps S2 and S3 into the attention enhancement model trained in step S5; S62, first passing the input tensor to the convolutional neural network module, and sequentially passing it through multiple sets of convolutional layers, normalization layers, and activation function layers to extract the feature patterns and local structural information of the partial discharge signal in the spatial domain, and outputting a spatial feature map; S63, flattening the spatial feature map and converting it into a time series format, inputting it into the bidirectional gated loop unit module, wherein the bidirectional gated loop unit module models the time evolution relationship within the sequence based on the forward and reverse paths respectively, and outputs... S64. The hidden state sequence containing time dependencies is input into the attention mechanism module, which assigns weights to the outputs of all time steps to enhance the features of key time segments and form an aggregated global embedding vector. S65. The global embedding vector is input into the fully connected module, which consists of multiple linear mapping layers and nonlinear transformation layers. Finally, a probability distribution vector of partial discharge type is generated at the output layer, and the label corresponding to the maximum probability value is used as the current recognition result. S66. The identified discharge type label, the original signal timestamp, and the acquisition channel information are output together and transmitted to the risk assessment and diagnostic report generation module for result archiving, visualization, and uploading.

[0037] This invention achieves efficient integration of deep network structures with real-time input signals during the model deployment phase, enabling rapid identification of partial discharge types without affecting the normal operation of transformers. Through spatial structural feature extraction, temporal information modeling, and attention-weighted aggregation inference paths, it effectively enhances the sensitivity of the identification system to short-term abnormal changes and complex signal combinations, providing highly reliable criteria support for subsequent state assessment and maintenance.

[0038] In this embodiment, S7 specifically includes: S71. Receive the discharge type tag output from step S6, and acquire the original signal segment corresponding to the discharge type tag and its acquisition data in each channel; S72. Perform amplitude statistical analysis on the original signal segment, extracting the maximum amplitude, average amplitude, fluctuation range, and amplitude change trend to measure discharge intensity and stability characteristics; S73. Perform frequency domain analysis on the original signal segment, extracting frequency-related parameters including repetition frequency, dominant frequency energy distribution, and discharge interval density to reflect the repetition and periodicity of partial discharge events; S74. Model the signal contour morphology based on the time-domain envelope extraction algorithm, acquiring the upper envelope, lower envelope, and overall morphological change trend of the waveform envelope to identify abnormal discharge morphological change characteristics; S75. Measure the signal duration... Measurement and classification are performed, and the duration, concentration, and time span of continuous signal segments under the same label are combined to assess the level of discharge persistence; S76, the extracted multidimensional statistical features are input into the risk assessment module and compared with the preset risk classification rules in the system to determine the risk level score of the current discharge event under different assessment dimensions; S77, the scores of each assessment dimension are weighted and summarized according to the comparison results to generate a comprehensive risk assessment result corresponding to the current signal segment, and divided into multiple levels, including low risk, medium risk, and high risk categories; S78, the generated partial discharge risk level identifier, its corresponding identification label, statistical features, and signal time range are packaged together to form structured risk information for subsequent report generation and remote upload modules to call.

[0039] The risk assessment mechanism proposed in this invention integrates multi-dimensional signal statistical parameters and risk rule models to construct a hierarchical assessment system based on discharge characteristics. It can dynamically adapt the risk characteristic weights under different discharge modes, identify high-risk signal events with drastic amplitude changes, concentrated frequency, or continuous anomalies, thereby achieving accurate risk level classification, effectively guiding on-site maintenance and preventative repair decisions, and improving the predictability and safety of power grid operation.

[0040] In this embodiment, S8 specifically includes: S81. Receive the discharge type tag output from step S6 and the risk level identifier output from step S7, and integrate them with the corresponding original signal data, statistical characteristics, and processing timestamp; S82. Construct a diagnostic report data structure, organizing the integrated information into data entities containing equipment identifier, signal channel information, identification tag, risk level, statistical indicators, and time interval content; S83. Structure the diagnostic data into a unified message format, generate a diagnostic report document conforming to the power system data interface specification, and assign a unique diagnostic number and timestamp index; S84. Call the remote communication interface to transmit the diagnostic report document to the receiving port of the power system monitoring platform through a power-specific communication protocol or edge computing interface; S85. Upload the diagnostic report. Upon successful completion, the status alarm module is automatically triggered. Based on the identification tag and risk level, it determines whether to generate a real-time alarm event and pushes it to the relevant alarm channels. S86: The data archiving module is invoked to store the diagnostic results information in the local database and cloud server, establishing a searchable index structure sorted by device number, signal type, and diagnostic time. S87: Combining the identification tag and risk level, the operation and maintenance suggestion generation module is invoked to match corresponding inspection suggestions, handling measures, and maintenance cycle reminders from the rule base, generating action guidelines for operation and maintenance personnel. S88: The diagnostic report, alarm information, archived records, and operation and maintenance suggestions are uniformly organized into a comprehensive diagnostic output package for the monitoring system to call and display, realizing visualized monitoring of the insulation status of power transformers, fault trend tracking, and closed-loop management.

[0041] This invention constructs a structured diagnostic report data structure and integrates remote communication, alarm push, and intelligent suggestion linkage modules to complete the information loop of diagnostic results throughout the "identification-reporting-feedback" chain. Through standardized message format and platform access interface adaptation, it can quickly connect with mainstream monitoring platforms to achieve visualized management, historical tracing, and intelligent intervention of partial discharge diagnostic information, thereby enhancing the system's digital operation and maintenance capabilities and fault response efficiency.

[0042] Example 1: To verify the feasibility of this invention in practice, it was deployed in the field operation environment of a 220kV main substation. This substation, located in East China, has equipment that has been in service for over 15 years and has experienced multiple sudden trips and insulation oil deterioration in recent years. Traditional partial discharge detection relies on only a single type of sensor, such as ultrasonic or high-frequency current transformers, lacking the ability to fuse multiple physical signals. This makes it difficult to locate the partial discharge source and results in a high false alarm rate, severely impacting the judgment efficiency and diagnostic accuracy of maintenance personnel.

[0043] To verify the feasibility and effectiveness of the method of the present invention in a real power system, 12 sensing nodes were deployed on the main transformer body of the substation. Electromagnetic UHF sensors, current pulse HFCT sensors and piezoelectric ceramic acoustic emission sensors were respectively deployed. The various sensors were deployed at the winding lead outlet, bushing root, inner wall of the oil tank and neutral point lead-out end, etc., to ensure that partial discharge signals with different physical characteristics can be collected synchronously from multiple directions.

[0044] The acquisition system used a high-speed synchronous acquisition platform supporting 24 parallel channels, with a sampling frequency of 200 MHz and a data buffer of 2 GB per channel. The sampling duration was set to 30 minutes, the sampling window interval was 10 seconds, and each trigger continuously recorded a raw signal sequence of 300 ms in length. The sampled data ultimately formed a raw dataset of approximately 82 GB.

[0045] The acquired signals were first subjected to bandpass filtering (10–80 MHz) and wavelet packet decomposition to remove power frequency noise and background interference. Then, normalization was performed on the signals from different channels, mapping all channel signals to the amplitude range of [-1, 1]. Signals with a time-domain synchronization error within ±20 ns underwent time interpolation alignment, and a sliding window mechanism was used to construct sliding window samples with a window length of 10 ms and a step size of 5 ms.

[0046] After the input tensor is constructed, it is fed into the deep neural network model proposed in this invention. The model consists of a three-layer convolutional structure (each layer contains 64 5×1 convolutional kernels), a bidirectional GRU structure (with 128 hidden units), and an attention weighting module. The entire network contains approximately 490,000 trainable parameters and is trained on an NVIDIA RTX 3090 GPU platform. The training set and validation set account for 70% and 30% of the total, respectively, and a total of 4,000 sets of real historical labeled samples are used for training.

[0047] During training, the model achieved stable convergence in the 28th round, with the cross-entropy loss decreasing to 0.031 and the validation accuracy remaining stable at 97.3%. Test set results show that the model's recognition accuracy for three typical partial discharge types (floating potential discharge, creepage discharge, and local internal air gap discharge) is 98.5%, 96.8%, and 97.6%, respectively, with an average accuracy of 97.6%, significantly outperforming the recognition performance of the traditional SVM model (87.2%) and KNN model (81.5%) in the same scenario.

[0048] During the diagnostic phase, the model successfully identified several frequent high-amplitude, short-cycle internal partial discharge events in a certain channel, with a duration of up to 12 minutes. The risk assessment module, considering parameters such as the average discharge amplitude (0.67 V), repetition frequency (52 times / second), and duration (average 28 ms), determined it to be of "medium-high risk level." The system automatically generated a diagnostic report and pushed it to the remote monitoring master station. The system completed the upload of the diagnostic results and risk alarm response within 3 seconds. This report was confirmed as valid during the team's routine inspection the following day and served as supporting evidence for the subsequent inspection of the local oil circuit insulation.

[0049] After the system was deployed, it identified a total of 193 partial discharge events within a week of continuous operation, of which 186 were effectively identified and accurately classified. The false alarm rate was controlled within 3.6%, and the false negative rate was less than 1%. Through the risk classification suggestion document generated by this system, the on-site maintenance plan was dynamically adjusted, and the operation and maintenance cost decreased by approximately 18.2%. This marked the first time that a graded maintenance mechanism driven by partial discharge signals was realized.

[0050] In summary, this embodiment fully demonstrates the high adaptability and excellent performance of the present invention in a real power grid environment, proving that the present invention can effectively solve the three key problems of "difficulty in fusion of multi-source data, weak feature expression, and high false alarm rate" in traditional partial discharge detection, and has significant technical advantages and engineering value in deep fusion signal processing and intelligent diagnosis.

[0051] 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 extracting and diagnosing partial discharge signals from power transformers using deep learning, characterized in that: Includes the following steps: S1. Set up a multi-channel synchronous acquisition system in the power transformer body area to acquire multi-dimensional raw partial discharge datasets; S2. Preprocess the acquired multi-dimensional raw partial discharge datasets; S3. Time-align and amplitude-match the preprocessed signals from different channels to construct a unified multi-channel signal matrix, which is then divided into fixed-length sliding time window segments according to time order to form the input tensor. S4. Construct an attention enhancement model that integrates a convolutional neural network and a bidirectional gated recurrent unit; S5. Supervise the training of the attention enhancement model using historical partial discharge data samples with manual labels, use the cross-entropy loss function to measure classification bias, and iteratively adjust the weights through backpropagation combined with a gradient-based parameter update method; S6. Input the partial discharge signal into the trained attention enhancement model, extract spatial structure features and temporal behavior features in sequence, and aggregate them with weights to form a global embedding vector. Finally, output the identification result label of the partial discharge type through a fully connected classification module; S7. Based on the discharge type label and the statistical parameters of its corresponding signal, perform multi-dimensional risk assessment in combination with preset grading rules to generate the corresponding partial discharge risk level label; S8. The identification tags and risk levels are packaged to generate a structured diagnostic report. The structured diagnostic report is uploaded to the power system monitoring platform through a remote communication interface, and status alarms, data archiving and operation and maintenance suggestions are generated at the same time.

2. The method for extracting and diagnosing partial discharge signals from power transformers using deep learning as described in claim 1, characterized in that, S1 specifically includes: S11. Set up sensor installation nodes at the locations of the power transformer body, selecting areas including the high-voltage winding lead outlet, winding end, adjacent area of ​​capacitor bushing, inner wall of oil tank, and neutral grounding terminal as deployment areas; S12. Install multiple types of sensor devices, including electromagnetic sensors, current pulse sensors, and piezoelectric acoustic emission sensors, at each node; S13. Use a data acquisition device with multi-channel synchronous acquisition capability to parallelly access the output signals of multiple types of sensors, and assign independent numbers and identifiers to each channel; S14. Set the initialization parameters of the acquisition channels, including channel activation status, sampling order, buffer size, and data storage path; S15. Start the acquisition program and continuously record the original partial discharge signals output by each sensor while the transformer is running, forming a time-continuous, channel-independent multidimensional original partial discharge dataset.

3. The method for extracting and diagnosing partial discharge signals from power transformers using deep learning as described in claim 1, characterized in that, S2 specifically includes: S21. Bandpass filtering is performed on the raw partial discharge signals acquired from each channel to remove low-frequency power frequency interference and high-frequency background noise, while retaining the frequency components related to partial discharge activity. S22. Wavelet packet decomposition is performed on the bandpass filtered signals to divide the original signals into multiple frequency band sub-signals, and the reconstructed components containing the target frequency components are selected as the effective signals. S23. Amplitude normalization is performed on the reconstructed effective signals to map the amplitude of each channel signal to a unified standard range. S24. The normalized sensor signals are stored in the corresponding buffer structures, keeping the channel number and time order unchanged.

4. The method for extracting and diagnosing partial discharge signals from power transformers using deep learning as described in claim 1, characterized in that, S3 specifically includes: S31. Perform unified time reference alignment on the pre-processed signals from different types of sensors. Based on the clock index information recorded by the acquisition system, perform time interpolation, cropping, or alignment buffer adjustment on each channel signal. S32. Perform amplitude matching processing on the time-aligned signals. Based on a preset amplitude normalization function, readjust the amplitude scale of all channel signals to construct an equal expression of signal amplitude values ​​within a consistent range. S33. Divide the processed multi-channel signals into sliding time windows according to time order. Set a fixed window length and overlap step size to split the entire signal sequence into multiple continuous signal segments, each representing the dynamic behavior of partial discharge events in different time periods. S34. Stack the multi-channel signals in each time window segment into a unified matrix structure according to channel order, and sequentially combine them into a three-dimensional input tensor with time dimension, channel dimension, and sampling point dimension. The input tensor constructs input data batches in units of windows. S35. Perform batch encapsulation processing on the constructed multi-dimensional input tensors to generate a standardized neural network input sample structure, while recording the window index, time period identifier, and sensor source label.

5. The method for extracting and diagnosing partial discharge signals from power transformers using deep learning as described in claim 1, characterized in that, The attention enhancement model includes an input layer, a multi-layer convolutional neural network module, a bidirectional gated recurrent unit module, an attention mechanism module, a fully connected module, and an output layer. The input layer receives the three-dimensional input tensor encapsulated in step S3, and the three-dimensional input tensor includes time series sampling points and corresponding amplitude information. The multi-layer convolutional neural network module includes multiple convolutional layers, normalization layers, and activation function layers. The convolutional layers extract the spatial features of the signal from the input tensor. Each convolutional layer is followed by a batch normalization layer and a nonlinear activation function layer in sequence to output a feature map. The feature map is then subjected to dimensionality reduction or flattening to generate a two-dimensional feature sequence with a time dimension structure. The bidirectional gated loop unit module includes a forward gated loop unit layer and a reverse gated loop unit layer. The two layers process the flattened feature sequence from the forward and backward directions of time, respectively, to extract the change state of the signal in the bidirectional time path and output the hidden state sequence. The attention mechanism module performs weighted operations on the hidden states at all time steps to generate an aggregated feature vector. The aggregated feature vector is then fused with the high-dimensional feature map of the last layer of the convolution module to form a comprehensive expressive feature that includes local spatial patterns and global temporal dependencies. The fully connected module receives the fused comprehensive representation features. The fully connected module consists of at least one fully connected layer, and nonlinear activation functions are connected between the layers. The output layer is located at the end of the fully connected module. The number of output nodes is equal to the number of partial discharge type categories, and it generates classification labels corresponding to the discharge signals.

6. The method for extracting and diagnosing partial discharge signals from power transformers using deep learning as described in claim 1, characterized in that, S5 specifically includes: S51. Collect multi-source partial discharge signal samples recorded during historical operation, and manually label each signal data. The manual labeling is based on expert judgment or experimental verification to form a one-to-one correspondence between partial discharge signals and discharge type labels, and construct a standard supervised training sample set; S52. Bind the three-dimensional input tensor constructed in step S3 with the label to form training sample pairs, index and encode all samples, and perform stratified sampling according to the sample time distribution and category distribution to divide them into training set, validation set and test set, with each subset maintaining non-overlapping relationships; S53. Define training parameters... Set the configuration file to specify the optimizer type, loss function form, learning rate adjustment strategy, regularization penalty term, training batch size, maximum number of training epochs, and early stopping monitoring indicators; S54, initialize the parameters of the convolutional neural network module, bidirectional gated recurrent unit module, attention mechanism module, and fully connected output module in the attention enhancement model, using a unified random seed to construct a complete computation graph and training dependencies; S55, during training, load sample tensors from the training set in batches, input them into the attention enhancement model to perform forward computation, extract local spatial features sequentially through convolutional layers, and through bidirectional... The gated recurrent unit module models temporal dependencies, generates weighted aggregated features through the attention mechanism module, and outputs the predicted label distribution through the fully connected module; S56, calculates the classification error between the predicted output and the actual label, inputs the error into the cross-entropy loss function to obtain the loss value of the current training batch, and executes the backpropagation algorithm based on the constructed error propagation path to automatically solve for the gradient values ​​of all trainable parameters in the attention enhancement model; S57, based on the calculated gradient information, calls the set optimizer to perform weight update operations on the parameters of each layer in the attention enhancement model, and records each round. The changes in the loss curve and classification accuracy are recorded in S58. After each training round, the current attention enhancement model is evaluated using the validation set. The loss value, accuracy, confusion matrix, and class balance index on the validation set are recorded to evaluate the stability and generalization performance of the attention enhancement model. The current best model snapshot is updated based on the validation set index. When the attention enhancement model does not observe a significant improvement in the validation index in multiple consecutive training rounds, or when the training has reached the maximum round limit, the training process is terminated, and the final stable model structure, weight parameters, configuration file, and label mapping rules are saved.

7. The method for extracting and diagnosing partial discharge signals from power transformers using deep learning as described in claim 1, characterized in that, S6 specifically includes: S61, inputting the partial discharge signal collected in real time during the operation of the power transformer and processed by steps S2 and S3 into the attention enhancement model trained in step S5; S62, first passing the input tensor to the convolutional neural network module, and then sequentially passing it through multiple sets of convolutional layers, normalization layers, and activation function layers to extract the feature patterns and local structure information of the partial discharge signal in the spatial domain, and outputting a spatial feature map; S63, flattening the spatial feature map and converting it into a time series format, inputting it into the bidirectional gated recurrent unit module, which models the temporal evolution relationship within the sequence based on both forward and reverse paths, and outputs a time-series-inclusive model. S64. The hidden state sequence of dependencies is input into the attention mechanism module, and weights are assigned to the output of all time steps to enhance the features of key time segments and form an aggregated global embedding vector; S65. The global embedding vector is input into the fully connected module, which consists of multiple linear mapping layers and nonlinear transformation layers. Finally, a probability distribution vector of partial discharge type is generated in the output layer, and the label corresponding to the maximum probability value is used as the current recognition result; S66. The identified discharge type label, the original signal timestamp, and the acquisition channel information are output together and transmitted to the risk assessment and diagnostic report generation module for result archiving, visualization, and uploading.

8. The method for extracting and diagnosing partial discharge signals from power transformers using deep learning as described in claim 1, characterized in that, Specifically, S7 includes: S71. Receive the discharge type tag output from step S6, and acquire the original signal segment corresponding to the discharge type tag and its acquisition data in each channel; S72. Perform amplitude statistical analysis on the original signal segment, extracting the maximum amplitude, average amplitude, fluctuation range, and amplitude change trend; S73. Perform frequency domain analysis on the original signal segment, extracting frequency-related parameters including repetition frequency, dominant frequency energy distribution, and discharge interval density; S74. Model the signal contour morphology based on the time domain envelope extraction algorithm, and acquire the upper envelope, lower envelope, and overall morphological change trend of the waveform envelope; S75. Measure and classify the signal duration, combining the data under the same tag... The duration, concentration, and time span of the continuous signal segment are used to assess its discharge persistence level; S76, the extracted multidimensional statistical features are input into the risk assessment module and compared with the preset risk grading rules in the system to determine the risk level score of the current discharge event under different assessment dimensions; S77, the scores of each assessment dimension are weighted and summarized according to the comparison results to generate the comprehensive risk assessment result corresponding to the current signal segment, and divided into multiple levels, including low risk, medium risk, and high risk categories; S78, the generated partial discharge risk level identifier, its corresponding identification label, statistical features, and signal time range are packaged together to form structured risk information.

9. The method for extracting and diagnosing partial discharge signals from power transformers using deep learning as described in claim 1, characterized in that, S8 specifically includes: S81. Receive the discharge type tag output from step S6 and the risk level identifier output from step S7, and integrate them with the corresponding original signal data, statistical characteristics, and processing timestamp; S82. Construct a diagnostic report data structure, organizing the integrated information into data entities containing equipment identifier, signal channel information, identification tag, risk level, statistical indicators, and time interval content; S83. Structure the diagnostic data into a unified message format, generate a diagnostic report document conforming to the power system data interface specification, and assign a unique diagnostic number and timestamp index; S84. Call the remote communication interface to transmit the diagnostic report document to the power system monitoring system via a dedicated power communication protocol or edge computing interface. The control platform's receiving port; S85. After the diagnostic report is successfully uploaded, the status alarm module is automatically triggered to determine whether to generate a real-time alarm event based on the identification tag and risk level, and push it to the relevant alarm channel; S86. The data archiving module is called to store the diagnostic result information in the local database and cloud server, and establish a searchable index structure sorted by device number, signal type and diagnostic time; S87. Combining the identification tag and risk level, the operation and maintenance suggestion generation module is called to match the corresponding inspection suggestions, handling measures and maintenance cycle reminders from the rule base, and generate action guidelines for operation and maintenance personnel; S88. The diagnostic report, alarm information, archived records and operation and maintenance suggestions are uniformly organized into a comprehensive diagnostic output package.

Images