PRPD quality driven converter transformer partial discharge characteristic decoupling identification method and system

Abstract

The invention discloses a PRPD quality-driven converter transformer partial discharge feature decoupling identification method and system, and the method comprises the steps: carrying out the data preprocessing and feature extraction of historical partial discharge feature PRPD data, obtaining three types of statistical features of the historical partial discharge feature PRPD data, training a three-branch convolutional neural network model, and deploying an inference engine; and performing data preprocessing and feature extraction on the real-time partial discharge feature PRPD data to obtain three types of statistical features, inputting the three types of statistical features into the inference engine, and performing fault type identification on the real-time partial discharge feature PRPD data. According to the method, the problems of time sequence information loss and high-dimensional data redundancy in a traditional PRPD identification method are solved, and the method is suitable for partial discharge monitoring and fault early warning application of a converter transformer of a power system.

Description

Technical Field

[0001] This invention relates to the field of partial discharge fault monitoring of converter transformers, and in particular to a method and system for decoupling and identifying partial discharge characteristics of converter transformers driven by PRPD quality. Background Technology

[0002] Converter transformers (CVTs), as core equipment in high-voltage direct current (HVDC) transmission systems, undertake the critical task of AC / DC power conversion, and their operating status directly affects the safety and stability of the power grid. Partial discharge (PD) is a significant indicator of insulation degradation in converter transformers. Prolonged partial discharge gradually erodes the insulating medium, ultimately leading to equipment failure or even system paralysis. Different types of partial discharge (such as air gap discharge, floating discharge, surface discharge, and pin-plate discharge) reflect different insulation defect mechanisms and degrees of severity. Accurately identifying the type of partial discharge fault is crucial for developing targeted maintenance strategies and preventing major accidents.

[0003] Existing methods for identifying partial discharge fault types in converter transformers mainly face the following problems: First, methods based on PRPD image recognition require converting phase-amplitude data into two-dimensional spectral images. This conversion process loses the precise numerical information and temporal correlation characteristics of the original signal. Furthermore, the parameter settings for image generation (such as resolution, color mapping, coordinate scaling, etc.) have a significant impact on the classification results, making it difficult to guarantee the consistency of feature representation. Second, while existing deep learning methods based on original time-series signals retain the original form of the data, they fail to fully utilize the periodic distribution pattern of partial discharge signals in the phase dimension, resulting in insufficient mining of the physical meaning of PRPD features, leading to poor model interpretability and limited generalization ability. Third, in practical applications, data collected at different time periods often exhibit amplitude distribution differences and baseline drift. Existing methods lack an effective cross-time adaptive standardization mechanism, causing a significant decrease in the model's recognition accuracy on newly collected data. Fourth, partial discharge signals are often mixed with a large amount of low-quality data and noise interference, severely affecting the reliability and real-time performance of fault identification. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a PRPD-driven decoupling identification method and system for partial discharge characteristics of converter transformers, which enables high-precision and robust identification of typical partial discharge faults, effectively improving the intelligence level and practicality of online monitoring of partial discharge in converter transformers.

[0005] Technical solution: The present invention provides a PRPD-driven method for decoupling and identifying partial discharge characteristics of converter transformers, comprising the following steps:

[0006] Historical partial discharge characteristics (PRPD) data of the converter transformer were collected;

[0007] The historical partial discharge characteristic PRPD data is preprocessed and feature extracted to obtain the first feature of the historical partial discharge characteristic PRPD data. The first feature includes the maximum amplitude feature, the average amplitude feature, and the discharge count feature.

[0008] A three-branch convolutional neural network model is constructed for fault type identification. The three-branch convolutional neural network model is trained using the first feature. The three-branch convolutional neural network model uses three branches to extract features from the maximum amplitude feature, the average amplitude feature, and the discharge count feature, respectively.

[0009] The trained three-branch convolutional neural network model is converted into an inference engine adapted to the hardware acceleration platform.

[0010] Real-time partial discharge characteristic (PRPD) data of the converter transformer is collected. The real-time PRPD data is preprocessed and feature extracted to obtain a second feature, which includes the maximum amplitude feature, the average amplitude feature, and the discharge count feature. The second feature is input into the inference engine to identify the fault type of the real-time PRPD data.

[0011] Furthermore, data preprocessing and feature extraction include the following steps:

[0012] The partial discharge characteristic PRPD data is processed by time-series binning, and the partial discharge characteristic PRPD data is divided into several data segments according to a fixed time window;

[0013] Evaluate the discharge event density of each data segment and select high-quality data segments;

[0014] For high-quality data segments, perform 360-degree fine-grained binning statistics according to the phase dimension, and calculate the maximum amplitude characteristics, average amplitude characteristics, and discharge count characteristics within each phase interval.

[0015] Furthermore, the discharge event density of each data segment was evaluated, and high-quality data segments were selected, including:

[0016] The total number of discharges within each data segment is used as the discharge event density.

[0017] The continuous data segment is divided into several non-overlapping time windows with a fixed window length. Within each time window, the number of high-quality data segments is counted according to a first threshold of discharge event density. The quality level of the time window is calculated based on the number of high-quality data segments.

[0018] Based on a second threshold for the number of high-quality data segments, the time series window is divided into high-quality active mode, normal quality mode and low-quality mode.

[0019] Filter high-quality data segments in each mode.

[0020] Furthermore, filtering high-quality data segments within each mode includes:

[0021] For high-quality active mode, high-quality data segments are retained on a time-series window basis until the quality level of the current time-series window is lower than the third threshold.

[0022] For the standard quality mode, all high-quality data segments are retained;

[0023] For low-quality modes, a certain number of consecutive timelines are merged into a high-quality data segment;

[0024] The quality level of the time series window is calculated based on the number of high-quality data segments within the time series window and the length of the time series window.

[0025] Furthermore, for high-quality data segments, a 360-degree fine-grained binning statistical analysis is performed according to the phase dimension, calculating the maximum amplitude characteristic, average amplitude characteristic, and discharge count characteristic within each phase interval, including:

[0026] The 0-360 degree phase space is uniformly divided into 360 phase intervals.

[0027] The phase interval to which the discharge event belongs is determined based on the phase angle of the discharge event corresponding to the partial discharge characteristic PRPD data;

[0028] For each phase interval, calculate the maximum amplitude characteristic, average amplitude characteristic, and discharge count characteristic of the partial discharge characteristic PRPD data.

[0029] Furthermore, training the three-branch convolutional neural network model using the first feature includes:

[0030] The first feature is reshaped into a three-dimensional matrix. In the three-dimensional matrix, if there is no discharge event in a certain phase interval, the maximum amplitude feature, the average amplitude feature, and the discharge count feature are all set to 0.

[0031] Furthermore, the three-branch convolutional neural network model includes a feature decoupling layer, a multi-branch feature learning layer, a cross-branch interaction layer, an adaptive fusion layer, a fully connected classification layer, and an output layer;

[0032] The feature decoupling layer receives the first feature in the form of a three-dimensional matrix and decomposes it into three independent channels;

[0033] The multi-branch feature learning layer performs independent deep feature extraction on each of the three channels;

[0034] The cross-branch interaction layer employs a lightweight attention mechanism for information exchange and feature complementarity between channels;

[0035] The adaptive fusion layer weighted and fused the deep features of the three channels to obtain high-level features;

[0036] The fully connected classification layer and output layer map high-level features to the fault type label space to achieve fault type classification decision.

[0037] Furthermore, real-time partial discharge characteristic (PRPD) data of the converter transformer is collected. After the real-time PRPD data accumulates to a certain time window length, the real-time PRPD data is preprocessed and feature extracted to obtain a second feature. The second feature is then input into the inference engine to identify the fault type of the real-time PRPD data.

[0038] The present invention discloses a PRPD-driven converter transformer partial discharge characteristic decoupling and identification system, comprising:

[0039] The historical data acquisition module is used to collect historical partial discharge characteristic (PRPD) data of the converter transformer.

[0040] The data preprocessing and feature extraction module is used to perform data preprocessing and feature extraction on the historical partial discharge feature PRPD data to obtain the first feature of the historical partial discharge feature PRPD data. The first feature includes the maximum amplitude feature, the average amplitude feature, and the discharge count feature.

[0041] The model design and training inference module is used to construct a three-branch convolutional neural network model for fault type identification. The three-branch convolutional neural network model is trained using the first feature. The three-branch convolutional neural network model uses three branches to extract features from the maximum amplitude feature, the average amplitude feature, and the discharge count feature, respectively.

[0042] The model compilation and deployment module is used to convert the trained three-branch convolutional neural network model into an inference engine adapted to the hardware acceleration platform.

[0043] The real-time data acquisition and online inference module is used to acquire real-time partial discharge characteristic (PRPD) data of the converter transformer, perform data preprocessing and feature extraction on the real-time PRPD data to obtain a second feature, which includes the maximum amplitude feature, the average amplitude feature, and the discharge count feature; input the second feature into the inference engine to identify the fault type of the real-time PRPD data.

[0044] The electronic device of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements the PRPD quality-driven decoupling identification method for partial discharge characteristics of converter transformers.

[0045] The computer-readable storage medium of the present invention stores a computer program, which, when executed by a processor, implements the PRPD quality-driven decoupling and identification method for partial discharge characteristics of converter transformers.

[0046] The computer program product of the present invention includes a computer program that, when executed by a processor, implements a converter transformer partial discharge characteristic decoupling and identification method driven by the PRPD quality.

[0047] Beneficial effects: Compared with the prior art, the advantages of the present invention are as follows:

[0048] 1. This invention employs a full-phase fine-grained binning statistical method, dividing the phase space into 360 one-degree intervals, which can more precisely characterize the distribution pattern of partial discharge in the phase dimension. Simultaneously, it extracts three types of statistical features—maximum amplitude, average amplitude, and discharge frequency—for each phase interval, providing multi-dimensional characterization of discharge properties and fully mining the deep-level pattern information in the PRPD data. This feature extraction method preserves the accurate numerical information of the original data while avoiding the information loss caused by image conversion methods, providing high-quality input features for subsequent deep learning models.

[0049] 2. The three-branch convolutional neural network architecture designed in this invention constructs independent learning branches for three types of statistical features. Each branch focuses on extracting local patterns and global distribution features of a specific statistic in the phase dimension, avoiding feature interference between different statistics. The outputs of the three branches are organically combined through a feature fusion layer, achieving collaborative learning and complementary enhancement of multi-statistic information. Compared to a single network structure, this multi-branch architecture can more effectively capture the differences in statistical features among different discharge types, significantly improving the accuracy and robustness of fault identification.

[0050] 3. Traditional methods are greatly affected by time and have poor generalization ability. The cross-time adaptive normalization strategy proposed in this invention effectively solves the problem of data distribution drift over time in practical applications by calculating global statistics on the training set and applying them to data from different time periods. This strategy eliminates the differences in data amplitude across different time periods while preserving the sparsity features of zero-padding positions, ensuring the consistency of feature representation.

[0051] 4. The dynamic quality monitoring mechanism designed in this invention can adaptively adjust the data filtering strategy according to the discharge event density. This mechanism effectively suppresses the interference of noise on model training and inference, and improves the reliability of fault identification. Attached Figure Description

[0052] Figure 1 This is a flowchart of the decoupling and identification method for partial discharge characteristics of converter transformers according to an embodiment of the present invention.

[0053] Figure 2 This is a schematic diagram of the data preprocessing and feature extraction process according to an embodiment of the present invention.

[0054] Figure 3 This is a schematic diagram of the sliding window filtering process of the dynamic quality monitoring mechanism in an embodiment of the present invention.

[0055] Figure 4 This is a schematic diagram of a multi-branch convolutional neural network architecture according to an embodiment of the present invention.

[0056] Figure 5 This is a schematic diagram of the real-time data acquisition and online inference process according to an embodiment of the present invention. Detailed Implementation

[0057] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0058] like Figure 1 As shown, the PRPD quality-driven converter transformer partial discharge characteristic decoupling identification method includes the following steps.

[0059] Step 1: Collect historical partial discharge characteristic (PRPD) data of the converter transformer;

[0060] Step 2: Perform data preprocessing and feature extraction on the historical partial discharge characteristic PRPD data to obtain the first feature of the historical partial discharge characteristic PRPD data. The first feature includes the maximum amplitude feature, the average amplitude feature, and the discharge count feature.

[0061] Step 3: Construct a three-branch convolutional neural network model for fault type identification. Train the three-branch convolutional neural network model using the first feature. The three-branch convolutional neural network model uses three branches to extract features from the maximum amplitude feature, average amplitude feature, and discharge count feature, respectively.

[0062] Step 4: Convert the trained three-branch convolutional neural network model into an inference engine adapted to the hardware acceleration platform.

[0063] Step 5: Collect real-time partial discharge characteristic (PRPD) data of the converter transformer; perform data preprocessing and feature extraction on the real-time PRPD data to obtain a second feature, which includes the maximum amplitude feature, the average amplitude feature, and the discharge count feature; input the second feature into the inference engine to identify the fault type of the real-time PRPD data.

[0064] Further, step 1 is the historical data acquisition phase, which involves obtaining partial discharge characteristic value information from the converter transformer active defense device or sensors. The sources of this characteristic value information include, but are not limited to, ultra-high frequency signals, high-frequency current transformer signals, or acoustic signals. Each data entry contains key parameters such as the amplitude, phase, and timestamp of the discharge event. In practice, partial discharge characteristic value information can be obtained from a CSV file.

[0065] Furthermore, such as Figure 2 As shown, step 2 is the data preprocessing and feature extraction stage. The collected PRPD feature value data is processed by time-series binning, and the continuous discharge events are divided into multiple data segments according to a fixed 1-second time window. The discharge event density of each data segment is evaluated through a dynamic quality monitoring mechanism to select high-quality data segments. The selected data segments are then subjected to 360-degree fine binning statistics according to the phase dimension to calculate three statistical features: maximum amplitude, average amplitude, and discharge count within each phase interval. The feature data is normalized using a cross-time adaptive standardization strategy.

[0066] Specifically, the time-series binning process in step 2 includes binning the continuously acquired PRPD feature value data stream according to the time dimension. Specifically, using a fixed 1-second window as the unit, the continuous discharge event sequence is divided into multiple independent data segments, each corresponding to a time bin. Each time bin contains phase-amplitude pairs of all discharge events within that time window. It is important to note that, to control computational resources, regardless of whether the data is high-frequency, ultra-high-frequency, or acoustic, each bin can represent a maximum of 1250 discharge events. Bins with fewer than 1250 events are padded with zeros to 1250, and bins with more than 1250 events are truncated to the first 1250, ensuring that all samples within a batch have a consistent dimension. After storing the data in HDF5 format, all time bins are arranged in timestamp order to form a time-series dataset. The specific steps of the time-series binning process are as follows:

[0067] Step 2.1.1: Time window binning. The continuously acquired PRPD feature value data stream is divided into fixed time windows. The preferred time window size is 1 second, meaning each bin contains all discharge events acquired within 1 second. For each time window, phase-amplitude pairs of all discharge events within that time period are collected, forming an independent data segment. The number of discharge events in each data segment is not fixed and depends on the actual discharge activity within that time period.

[0068] Step 2.1.2: For all timeboxes within the same HDF5 file, sort them according to their timestamps to ensure the temporal continuity of the data. Extract the index number of each timebox by parsing the timestamp information in the group name, and sort the timeboxes using a sorting algorithm (such as quicksort) to generate an ordered list of timeboxes.

[0069] Specifically, the dynamic quality monitoring mechanism in step 2 includes: using an adaptive filtering strategy with a sliding window to assess and classify the time-bin data, effectively suppressing noise and low-quality data interference. First, the total number of discharges in each time-bin is counted as a density index, reflecting the discharge activity within that time period. High and low event thresholds are set, optionally using an adaptive threshold setting method, dynamically calculating the thresholds based on the event density statistical characteristics of the dataset to improve the method's adaptability to different datasets. Continuous time-bins are divided into multiple non-overlapping time-series windows with a fixed window length. The number of high-quality and medium-quality time-bins within each window is counted, and a window quality index is defined to determine the overall quality level of the window. For example... Figure 3 As shown, the specific steps of the dynamic quality monitoring mechanism are as follows:

[0070] Step 2.2.1: For each timebox (k is the time bin index), count the total number of discharges. As a density indicator, it reflects the discharge activity during that time period.

[0071] Step 2.2.2: Quality level classification, setting high event thresholds and low event threshold Optionally, an adaptive threshold setting method can be used, based on the event density statistics of the dataset (mean μ and standard deviation). Dynamic calculation , This improves the method's adaptability to different datasets.

[0072] Step 2.2.3: Time window construction and feature extraction. The continuous time bins are divided into multiple non-overlapping time windows according to a fixed window length M. For each window, statistically analyze its quality distribution characteristics:

[0073]

[0074]

[0075] These represent the number of high-quality and medium-quality timeboxes within the window, respectively. Define the window quality metrics:

[0076]

[0077] This indicates the percentage of high-quality timeboxes within the window, used to determine the overall quality level of the window.

[0078] Step 2.2.4, Data Quality Pattern Recognition, based on the initial window The quality characteristics are used to identify the quality mode of the data file. There are three modes:

[0079] High-quality activity mode: Judgment criteria , indicates that the data shows high discharge activity from the very beginning.

[0080] Standard Quality Model: Judgment Criteria The data is mostly of medium quality.

[0081] The remaining cases are classified as low-quality modes, indicating sparse discharge or high noise levels. This is an adjustable parameter.

[0082] Step 2.2.5: Adaptive filtering strategy execution, including three strategies:

[0083] Strategy 1, a strict screening strategy, is applicable to high-quality active modes. Perform the following actions:

[0084] 1) Filter window by window, for each window Preserve high-quality timeboxes:

[0085]

[0086] 2) Quality Monitoring: Calculate the quality indicators for the current window. ,like If the quality is significantly reduced, processing of subsequent windows will be stopped.

[0087] 3) Output Results ,in The index of the window that triggers a stop due to quality degradation.

[0088] Strategy 2, a basic screening strategy, is applicable to the standard quality model. It removes low-quality data from all time bins.

[0089]

[0090] This strategy retains data of medium to high quality, maximizing the utilization of effective samples.

[0091] Strategy 3: Use a time accumulation strategy for low-quality data.

[0092] 1) Constructing the cumulative window: Set the cumulative window length L, and merge the discharge events of L consecutive time bins into one cumulative sample. .

[0093] 2) Cumulative event density calculation:

[0094]

[0095] 3) Cumulative sample screening:

[0096]

[0097] in This is the cumulative threshold.

[0098] Step 2.2.6: Output and Evaluation of Screening Results. Output the set of time bins S after screening, as input for subsequent feature extraction. Simultaneously, calculate the screening effectiveness evaluation index:

[0099]

[0100]

[0101] in For retention rate, This represents the average discharge density after screening.

[0102] Specifically, step 2, the 360-degree fine-grained binning statistics, includes: extracting refined phase statistical features from each timebox data that has undergone quality screening. The phase space from 0 to 360 degrees is uniformly divided into 360 phase intervals, each interval spanning 1 degree. For each phase interval, all discharge events within that timebox are traversed, and all events whose phase angle falls within that interval are selected. The maximum amplitude feature, average amplitude feature, and discharge frequency feature are calculated. For phase intervals without discharge events, all three feature values ​​are set to 0. Finally, each timebox is converted into a feature matrix of shape (360, 3), where 360 ​​represents the phase dimension and 3 represents the three statistical feature dimensions. This feature representation method not only preserves the fine distribution information of the PRPD spectrum in the phase dimension but also enhances the expressive power of the features through the fusion of multiple statistics. The specific steps of the 360-degree fine-grained binning statistics are as follows:

[0103] Step 2.3.1: Fine-grained full-phase interval division. The complete phase space from 0 to 360 degrees is uniformly divided into 360 phase intervals, with each interval spanning 1 degree. Each phase interval is defined as:

[0104] ,

[0105] in, , , .

[0106] Step 2.3.2: Determine the phase attribution of discharge events. For the set of all discharge events within a given time bin... ,in, This represents the amplitude of the j-th discharge event. This represents the phase angle, where N is the total number of discharge events within the timeframe. For each discharge event, based on its phase angle... Determine its corresponding phase interval:

[0107]

[0108] in This represents the function for rounding up.

[0109] Construct a mapping relationship from phase intervals to the set of discharge events:

[0110] ,

[0111] in, This represents the set of all discharge events belonging to the i-th phase interval.

[0112] Step 2.3.3: Statistical characteristic calculation, for each phase interval The corresponding set of discharge events Calculate the three types of statistical values ​​respectively. Initialize the feature array by creating three arrays of length 360 to store the statistical features of the 360 ​​phase intervals: `max` represents the maximum amplitude feature, `mean` represents the average amplitude feature, and `count` represents the number of discharges feature. Then, all phase intervals are traversed to determine if a discharge event occurs within that interval. If so, all amplitude values ​​within the interval are extracted to form an amplitude set.

[0113]

[0114] calculate Maximum value:

[0115]

[0116] calculate Arithmetic mean in:

[0117]

[0118] Calculate the characteristics of the number of discharges:

[0119]

[0120] Finally, the calculated three types of feature values ​​are stored in the corresponding feature arrays. Each array contains 360 elements, which correspond to the statistical features of 360 phase intervals.

[0121] Specifically, the cross-time adaptive standardization strategy in step 2 includes: to address the degradation in model generalization performance caused by differences in data distribution across different time periods, a cross-time adaptive standardization strategy based on training set statistics is adopted. First, global statistics for the three types of features are calculated on the training set: for the feature matrices of all training samples, the mean and standard deviation of the maximum amplitude feature, average amplitude feature, and discharge frequency feature are calculated respectively. During standardization, zero-padding positions in the feature matrix are identified (i.e., phase intervals where all three types of feature values ​​are 0), and standardization is performed only on feature values ​​at non-zero-padding positions: for each feature dimension, normalization is performed using the formula (x - mean) / std, where mean and std are the global statistics for that feature dimension. After standardization, the feature values ​​at zero-padding positions are restored to 0, preserving the sparsity of the data. For the validation and test sets, the statistics calculated from the training set are directly used for standardization, ensuring that data from different time periods are used for model inference at the same numerical scale. This standardization strategy eliminates differences in data distribution while retaining meaningful zero-value information in the original data, improving the model's cross-time adaptability. The specific steps of the cross-time adaptive normalization strategy are as follows:

[0122] Step 2.4.1: Feature matrix reshaping.

[0123] The feature vectors extracted through 360-degree full-phase binning statistical feature extraction are reshaped into a three-dimensional matrix. For a dataset containing N samples, the feature vector dimension of each sample is 360×3=1080, which is then reshaped as follows:

[0124]

[0125] Among them, the three channels of the third dimension correspond to the maximum amplitude feature, the average amplitude feature, and the discharge number feature, respectively.

[0126] Step 2.4.2: Zero padding position identification.

[0127] Identify the zero-padding positions in the feature matrix. If there is no discharge event in a certain phase interval, all three types of feature values ​​are 0. Define a zero-padding mask. When all three features of the i-th phase interval of the n-th sample are 0, ,otherwise .

[0128] Step 2.4.3: Calculate global statistics for the training set.

[0129] Global statistics for all three features are computed only on the training set. The mean is calculated for each feature channel. and standard deviation .

[0130] Further, step 3 is the model design and training inference stage, which constructs a three-branch convolutional neural network model. The three branches perform independent convolutional learning on the maximum value feature, the average value feature, and the count feature, respectively, to extract the deep patterns of each statistic. The outputs of the three branches are concatenated and dimensionality reduced through a feature fusion layer. A fully connected classifier is used to output the recognition results. The model is trained and optimized using labeled data, and the model performance is evaluated on the validation set.

[0131] Specifically, such as Figure 4 As shown, step 3 constructs a network architecture combining a multi-branch one-dimensional convolutional neural network with adaptive feature fusion, including a feature decoupling layer, a multi-branch feature learning layer, a cross-branch interaction layer, an adaptive fusion layer, and a fully connected classification layer. The feature decoupling layer decomposes the input tensor along the statistical feature dimension into three independent channels: a peak feature channel, a mean feature channel, and a density feature channel, corresponding to the maximum amplitude, average amplitude, and number of discharges, respectively. The multi-branch feature learning layer contains three parallel branches with identical structures but independent parameters, each using a multi-layer one-dimensional convolutional neural network. The cross-branch interaction layer employs a lightweight attention mechanism to achieve information interaction and feature complementarity between branches, concatenating the feature maps of the three branches along the channel dimension to form a joint feature representation. Channel attention weights are calculated using global average pooling and a fully connected network to achieve adaptive enhancement of different statistical features. The adaptive fusion layer employs a learnable weighted fusion strategy, calculating the dynamic weight coefficients of the three branches through a fusion weight network and fusing the features of the three branches using a weighted summation method. This allows for dynamic adjustment of the contribution ratio of the three statistical features based on the feature differences of different discharge types. The fully connected classification layer maps the fused high-level features to the fault type label space, realizes classification decision and outputs fault identification results.

[0132] Specifically, different training models can be set for different sensor data, for example, deploying separate model weights for high-frequency and ultra-high-frequency data.

[0133] The following sections introduce the feature decoupling layer, multi-branch feature learning layer, cross-branch interaction layer, adaptive fusion layer, and fully connected classification layer.

[0134] (a) A feature decoupling layer receives the feature matrix after 360-degree full-phase binning statistics and standardization, and decouples the channels according to the statistical dimension. More preferably, the input data is a three-dimensional tensor of shape (N, 360, 3), where N is the batch size (preferably 32 or 64), 360 is the phase dimension, and 3 is the statistical feature dimension. The feature decoupling layer decomposes the input tensor along the third dimension into three independent channels: a peak feature channel (N, 360, 1), a mean feature channel (N, 360, 1), and a density feature channel (N, 360, 1), corresponding to the maximum amplitude, average amplitude, and number of discharges, respectively. This decoupling operation is based on the physical differences of the three statistical features, laying the foundation for subsequent independent learning.

[0135] (b) A multi-branch feature learning layer is used to extract deep features independently from the three feature channels. More preferably, the network contains three parallel branches with identical structures but independent parameters, each employing a multi-layer one-dimensional convolutional neural network structure. Taking the peak feature branch as an example, the first convolutional kernel size is 5, the number of output channels is 32, and the output feature map shape is (N, 32, 360); the second convolutional kernel size is 3, the number of output channels is 64, and the output shape after pooling is (N, 64, 180); the third convolutional kernel size is 3, the number of output channels is 64, and the output shape after pooling is (N, 64, 90). Each convolutional layer is followed by a batch normalization layer, a ReLU activation function, and a Dropout regularization layer (preferably with a dropout rate of 0.3) to improve the model's training stability and generalization ability. The output feature map shapes of the three branches are all (N, 64, 90), representing the deep patterns of the three statistical features in the phase dimension.

[0136] (c) A cross-branch interaction layer is used to achieve information exchange and feature complementarity between branches based on independent learning. More preferably, a lightweight attention mechanism is used to implement cross-branch interaction: First, the feature maps of the three branches are concatenated along the channel dimension to form a joint feature representation of (N, 192, 90); then, the channel attention weight vector (N, 192) is calculated through global average pooling and two fully connected layers; next, the attention weights are reshaped and applied to the joint feature, and then the enhanced joint feature is separated back into three independent branch features (N, 64, 90)×3 along the channel dimension. This interaction mechanism allows the three branches to maintain independence while dynamically adjusting their respective contributions according to the characteristics of the input data, improving the flexibility of feature representation. Optionally, the cross-branch interaction layer can also be implemented using gating mechanisms or residual connections.

[0137] (d) An adaptive fusion layer is used to intelligently fuse the features of the three branches enhanced by interaction to generate a unified high-level feature representation. More preferably, a learnable weighted fusion strategy is adopted: First, the feature maps of the three branches output by the interaction layer are flattened to obtain three feature vectors, each with dimensions (N, 5760); then, the dynamic weight coefficients of the three branches are calculated through a fusion weight network (containing a fully connected layer and Softmax activation). , , Finally, the features of the three branches are fused using a weighted summation method to obtain a fused feature vector (N, 5760). This adaptive fusion mechanism can dynamically adjust the contribution ratio of the three statistical features according to the feature differences of different discharge types, and has stronger expressive power compared to fixed splicing or averaging fusion methods. Optionally, the adaptive fusion layer can also be implemented using attention fusion, gated fusion, or multi-scale fusion.

[0138] (e) A fully connected classification layer is used to map the fused high-level features to the fault type label space to achieve classification decision. More preferably, the network contains two fully connected layers. The first layer maps the fused features from 5760 dimensions to 128 dimensions, and performs batch normalization, ReLU activation and Dropout. The second layer maps the 128-dimensional features to the output.

[0139] During model training, model performance is evaluated on a validation set, and the parameters of the model with the highest accuracy on the validation set and their corresponding standardized statistics are saved. During model inference, the data to be identified undergoes the same data processing procedures as in the training phase, including temporal binning, dynamic quality monitoring, and 360-degree feature extraction. The standardized statistics saved in the training phase are used for feature standardization to ensure consistency in data processing.

[0140] In some optional embodiments, the above neural network architecture exhibits good scalability and flexibility. For different application scenarios, the following parameters and structures can be adjusted: 1. The number of branches can be flexibly set according to the statistical feature dimension, supporting configurations of 2, 3, or more branches; 2. Parameters such as the number of convolutional layers, kernel size, and number of output channels in each branch can be adjusted according to the data scale and computing resources; 3. Cross-branch interaction layers and adaptive fusion layers can adopt various implementation methods, including but not limited to attention mechanisms, gating mechanisms, and residual connections; 4. The number of layers and the dimension of hidden layers in the fully connected classification layer can be configured according to the complexity of the classification task.

[0141] Further, step 4 is the model compilation and deployment stage, which converts the trained deep learning model into a format adapted to the Lingxi APU hardware acceleration platform, and performs operator optimization, quantization compression and memory layout adjustment through model compilation tools to generate an inference engine that can run efficiently on edge computing devices; and completes the deployment and performance verification of the model on the target hardware platform.

[0142] Furthermore, step 5 is the real-time data acquisition and online inference stage. In the actual operating environment, the system receives the partial discharge monitoring data of the converter transformer in real time, performs feature extraction and preprocessing according to the method in step 2, calls the deployed inference engine to identify the fault type of the real-time data, and outputs the identification results and confidence information to the monitoring interface to provide decision support for operation and maintenance personnel.

[0143] Specifically, such as Figure 5 As shown, step 5 involves receiving the PRPD feature value data stream transmitted by the converter transformer partial discharge monitoring device in real time, buffering the data in chronological order, and performing feature extraction and standardization operations on the received real-time data when the accumulated data volume reaches a complete time bin. This includes 360-degree phase bin statistics, dynamic quality assessment, and cross-time standardization to generate feature data that meets the model input requirements. The preprocessed feature data is converted to float16 format, and the deployed APU inference engine is invoked to perform fault identification. The identification results (fault category, confidence level) are output to the monitoring interface to provide maintenance personnel with real-time equipment status information and fault warnings.

[0144] In some optional embodiments, step 5 specifically includes the following steps:

[0145] Step 5.1: Multi-threaded file monitoring and intelligent scheduling.

[0146] When the system starts, it initializes multiple processing threads according to the configuration. Each thread corresponds to a specific combination of sensor type and model type. The main monitoring thread periodically scans the data directory and distributes newly added CSV files to the corresponding processing queues based on file path keywords. A file modification time and persistent recording mechanism is used to avoid duplicate processing.

[0147] Step 5.2: Screening of rapid discharge events.

[0148] Prioritize reading the "maximum waveform value" and "minimum waveform value" columns from the CSV file, use vectorized operations to batch calculate the absolute values ​​and compare them with the discharge threshold, and only filter discharge events that exceed the threshold for data parsing.

[0149] Step 5.3: Data preprocessing and batch inference.

[0150] A fixed number of samples are randomly selected from the selected discharge events. An appropriate preprocessing strategy is chosen, sensor gain configuration is applied, and the data is converted into the data format required by the model. The APU inference engine is selected according to the model type for batch inference. The probability distribution of all samples is averaged to obtain the confidence voting result. The category with the highest average confidence is selected as the final prediction.

[0151] Step 5.4: Multi-channel comprehensive analysis.

[0152] The single-file inference results include information such as the number of discharges, prediction categories, and confidence distribution, and are saved as a CSV file. The comprehensive judgment module collects the inference results of different channels of the same converter transformer according to the time window. If multiple channels or different sensor types predict the same result, the confidence level is increased. If the predictions conflict, a weighted decision is made based on sensor priority and confidence level.

[0153] Step 5.5: Exception handling and system monitoring.

[0154] When file processing fails, the number of retries is recorded and the file is re-added to the queue. Once the maximum number of retries is reached, the file is marked as processed to avoid infinite retries. The processing records of deleted files are cleaned up periodically. Thread exceptions are automatically captured so as not to affect other threads. Performance metrics such as throughput and average processing time are calculated in real time.

[0155] The PRPD-driven converter transformer partial discharge characteristic decoupling identification system of the present invention includes:

[0156] The historical data acquisition module is used to collect historical partial discharge characteristic (PRPD) data of the converter transformer.

[0157] The data preprocessing and feature extraction module is used to perform data preprocessing and feature extraction on the historical partial discharge feature PRPD data to obtain the first feature of the historical partial discharge feature PRPD data. The first feature includes the maximum amplitude feature, the average amplitude feature, and the discharge count feature.

[0158] The model design and training inference module is used to construct a three-branch convolutional neural network model for fault type identification. The three-branch convolutional neural network model is trained using the first feature. The three-branch convolutional neural network model uses three branches to extract features from the maximum amplitude feature, the average amplitude feature, and the discharge count feature, respectively.

[0159] The model compilation and deployment module is used to convert the trained three-branch convolutional neural network model into an inference engine adapted to the hardware acceleration platform.

[0160] The real-time data acquisition and online inference module is used to acquire real-time partial discharge characteristic (PRPD) data of the converter transformer, perform data preprocessing and feature extraction on the real-time PRPD data to obtain a second feature, which includes the maximum amplitude feature, the average amplitude feature, and the discharge count feature; input the second feature into the inference engine to identify the fault type of the real-time PRPD data.

[0161] Furthermore, the data acquisition module is connected to the converter transformer current transformer or active defense device to obtain the original characteristic value information of partial discharge faults. The types of characteristic value data include, but are not limited to: high-frequency partial discharge data, ultra-high frequency partial discharge data, ultrasonic partial discharge data, etc.

[0162] Furthermore, the data preprocessing and feature extraction module is responsible for converting the raw PRPD feature value data into a high-quality feature representation suitable for deep learning model input. This module receives the phase-amplitude pair data stream from the data acquisition module, and outputs a standardized full-phase binned statistical feature matrix after systematic processing.

[0163] Furthermore, the model design and training inference module is responsible for constructing a deep learning model and achieving automatic classification of partial discharge faults. This module receives a standardized feature matrix from the data preprocessing and feature extraction module, performs deep feature learning and pattern recognition through a three-branch convolutional neural network, and finally outputs the fault identification result and confidence level.

[0164] Furthermore, the model compilation and deployment module is responsible for converting the trained deep learning model into a high-efficiency inference engine adapted to the hardware acceleration platform. This module receives the optimal model parameter file output from the model design, training, and inference module, and after format conversion, compilation optimization, and deployment verification, it finally generates an inference engine that can run efficiently on the Lingxi APU hardware platform.

[0165] Furthermore, the real-time data acquisition and online inference module is responsible for real-time monitoring and fault identification of partial discharge data generated during the operation of the converter transformer. This module receives real-time PRPD data streams from the converter transformer monitoring device, performs online feature extraction and APU inference, and outputs fault type identification results and confidence information, providing maintenance personnel with timely equipment status warnings and decision support.

[0166] The electronic device of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements the PRPD quality-driven decoupling identification method for partial discharge characteristics of converter transformers.

[0167] The computer-readable storage medium of the present invention stores a computer program, which, when executed by a processor, implements the PRPD quality-driven decoupling and identification method for partial discharge characteristics of converter transformers.

[0168] The computer program product of the present invention includes a computer program that, when executed by a processor, implements a converter transformer partial discharge characteristic decoupling and identification method driven by the PRPD quality.

[0169] The computer-readable storage medium may include RAM, ROM, EEPROM, CD-ROM or other optical disc storage devices, magnetic disk storage devices or other magnetic storage devices, flash memory, or any other media that can be used to store program code in the form of instructions or data structures and is accessible by a computer.

[0170] The processor is used to execute computer programs stored in memory to implement the various steps in the methods described in the above embodiments.

Claims

1. A PRPD-driven method for decoupling and identifying partial discharge characteristics of converter transformers, characterized in that, Includes the following steps: Historical partial discharge characteristics (PRPD) data of the converter transformer were collected; The historical partial discharge characteristic PRPD data is preprocessed and feature extracted to obtain the first feature of the historical partial discharge characteristic PRPD data. The first feature includes the maximum amplitude feature, the average amplitude feature, and the discharge count feature. A three-branch convolutional neural network model is constructed for fault type identification. The three-branch convolutional neural network model is trained using the first feature. The three-branch convolutional neural network model uses three branches to extract features from the maximum amplitude feature, the average amplitude feature, and the discharge count feature, respectively. The trained three-branch convolutional neural network model is converted into an inference engine adapted to the hardware acceleration platform. Real-time partial discharge characteristic (PRPD) data of the converter transformer is collected. The real-time PRPD data is preprocessed and feature extracted to obtain a second feature, which includes the maximum amplitude feature, the average amplitude feature, and the discharge count feature. The second feature is input into the inference engine to identify the fault type of the real-time PRPD data.

2. The PRPD-driven decoupling and identification method for partial discharge characteristics of converter transformers according to claim 1, characterized in that, Data preprocessing and feature extraction include the following steps: The partial discharge characteristic PRPD data is processed by time-series binning, and the partial discharge characteristic PRPD data is divided into several data segments according to a fixed time window; Evaluate the discharge event density of each data segment and select high-quality data segments; For high-quality data segments, perform 360-degree fine-grained binning statistics according to the phase dimension, and calculate the maximum amplitude characteristics, average amplitude characteristics, and discharge count characteristics within each phase interval.

3. The PRPD-driven decoupling and identification method for partial discharge characteristics of converter transformers according to claim 2, characterized in that, The discharge event density of each data segment was evaluated, and high-quality data segments were selected, including: The total number of discharges within each data segment is used as the discharge event density. The continuous data segment is divided into several non-overlapping time windows with a fixed window length. Within each time window, the number of high-quality data segments is counted according to a first threshold of discharge event density. The quality level of the time window is calculated based on the number of high-quality data segments. Based on a second threshold for the number of high-quality data segments, the time series window is divided into high-quality active mode, normal quality mode and low-quality mode. Filter high-quality data segments in each mode.

4. The PRPD-driven decoupling and identification method for partial discharge characteristics of converter transformers according to claim 3, characterized in that, Filtering high-quality data segments in each mode includes: For high-quality active mode, high-quality data segments are retained on a time-series window basis until the quality level of the current time-series window is lower than the third threshold. For the standard quality mode, all high-quality data segments are retained; For low-quality modes, a certain number of consecutive timelines are merged into a high-quality data segment; The quality level of the time series window is calculated based on the number of high-quality data segments within the time series window and the length of the time series window.

5. The PRPD-driven decoupling and identification method for partial discharge characteristics of converter transformers according to claim 1, characterized in that, For high-quality data segments, a 360-degree fine-grained binning statistical analysis is performed according to the phase dimension. The maximum amplitude characteristic, average amplitude characteristic, and discharge count characteristic within each phase interval are calculated, including: The 0-360 degree phase space is uniformly divided into 360 phase intervals. The phase interval to which the discharge event belongs is determined based on the phase angle of the discharge event corresponding to the partial discharge characteristic PRPD data; For each phase interval, calculate the maximum amplitude characteristic, average amplitude characteristic, and discharge count characteristic of the partial discharge characteristic PRPD data.

6. The PRPD-driven decoupling and identification method for partial discharge characteristics of converter transformers according to claim 1, characterized in that, Training the three-branch convolutional neural network model using the first feature includes: The first feature is reshaped into a three-dimensional matrix. In the three-dimensional matrix, if there is no discharge event in a certain phase interval, the maximum amplitude feature, the average amplitude feature, and the discharge count feature are all set to 0.

7. The PRPD-driven decoupling and identification method for partial discharge characteristics of converter transformers according to claim 6, characterized in that, The three-branch convolutional neural network model includes a feature decoupling layer, a multi-branch feature learning layer, a cross-branch interaction layer, an adaptive fusion layer, a fully connected classification layer, and an output layer. The feature decoupling layer receives the first feature in the form of a three-dimensional matrix and decomposes it into three independent channels; The multi-branch feature learning layer performs independent deep feature extraction on each of the three channels; The cross-branch interaction layer employs a lightweight attention mechanism for information exchange and feature complementarity between channels; The adaptive fusion layer weighted and fused the deep features of the three channels to obtain high-level features; The fully connected classification layer and output layer map high-level features to the fault type label space to achieve fault type classification decision.

8. The PRPD-driven decoupling and identification method for partial discharge characteristics of converter transformers according to claim 2, characterized in that, Real-time partial discharge characteristic (PRPD) data of the converter transformer is collected. After the real-time PRPD data accumulates to a certain time window length, the real-time PRPD data is preprocessed and feature extracted to obtain a second feature. The second feature is then input into the inference engine to identify the fault type of the real-time PRPD data.

9. A PRPD-driven decoupling identification system for partial discharge characteristics of converter transformers, characterized in that, include: The historical data acquisition module is used to collect historical partial discharge characteristic (PRPD) data of the converter transformer. The data preprocessing and feature extraction module is used to perform data preprocessing and feature extraction on the historical partial discharge feature PRPD data to obtain the first feature of the historical partial discharge feature PRPD data. The first feature includes the maximum amplitude feature, the average amplitude feature, and the discharge count feature. The model design and training inference module is used to construct a three-branch convolutional neural network model for fault type identification. The three-branch convolutional neural network model is trained using the first feature. The three-branch convolutional neural network model uses three branches to extract features from the maximum amplitude feature, the average amplitude feature, and the discharge count feature, respectively. The model compilation and deployment module is used to convert the trained three-branch convolutional neural network model into an inference engine adapted to the hardware acceleration platform. The real-time data acquisition and online inference module is used to acquire real-time partial discharge characteristic (PRPD) data of the converter transformer, perform data preprocessing and feature extraction on the real-time PRPD data to obtain a second feature, which includes the maximum amplitude feature, the average amplitude feature, and the discharge count feature; input the second feature into the inference engine to identify the fault type of the real-time PRPD data.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is loaded into the processor, it implements the PRPD quality-driven decoupling identification method for partial discharge characteristics of converter transformers according to any one of claims 1-8.

11. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the PRPD quality-driven decoupling identification method for partial discharge characteristics of converter transformers according to any one of claims 1-8.

12. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the PRPD quality-driven decoupling identification method for partial discharge characteristics of converter transformers according to any one of claims 1-8.

Images