Partial discharge defect type identification method based on dual-stage feature fusion

Abstract

The invention discloses a partial discharge defect type identification method based on dual-stage feature fusion, and belongs to the technical field of power equipment state monitoring and image identification, and the method comprises the steps: obtaining original discharge waveform data, and drawing a PRPD map; partial discharge statistical features in the PRPD atlas are extracted, and RFE is used for feature selection; the selected statistical characteristics are learned by using CatBoost, and a primary recognition model is constructed; performing type identification on the collected data sample by using a primary identification model, and identifying and screening out a primary discharge type; inputting the PRPD atlas of the screened samples of the primary discharge type into a convolutional neural network (CNN), and extracting image features; inputting the image features of the sample and the selected statistical features into Mamba for fusion to form a new feature vector; according to the method, the problem of discharge mode recognition confusion in the partial discharge process can be solved, the advantages of double models can be effectively exerted by adopting staged recognition, and the recognition efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment status monitoring and image recognition, and in particular to a partial discharge defect type recognition method based on double-stage feature fusion. Background Art

[0002] During the operation of power equipment, partial discharge (PD), a significant indicator of insulation degradation, has become a major cause of power grid equipment failure. General electrical equipment typically has complex internal structures and contains a wide variety of insulation types. Therefore, impurities and foreign matter inevitably enter during the manufacturing and transmission processes. This, combined with long-term operation, changes the surrounding electric field distribution. When the electric field from these impurities reaches a certain intensity, PD can be induced. Sustained discharge can damage the equipment's insulation. For example, power transformers contain insulation made of oil and paper. When the transformer is exposed to high humidity, the oil's insulating properties degrade, while the water gradually vaporizes due to rising temperatures, triggering PD. Continuous discharge damages the equipment's insulation. Furthermore, PD is accompanied by a significant amount of background noise, requiring careful signal denoising and extraction of key information during processing.

[0003] With the expansion of smart grid architectures and the iterative upgrades in power electronics research, intelligent analysis methods based on feature recognition are demonstrating significant application value in equipment condition monitoring. Research has shown that the diversity of insulation defects can trigger discharge behaviors with distinct patterns. This strong correlation between discharge characteristics and defect types provides a new technical path for fault diagnosis. By quantitatively analyzing key parameters such as the phase distribution, pulse waveform, and spectral characteristics of discharge signals, a multidimensional assessment system for insulation defect type, development cycle, and damage severity can be constructed, enabling dynamic early warning of equipment insulation status. Current research challenges in this field primarily focus on the highly complex analysis of discharge patterns. Traditional manual interpretation methods, due to their high subjectivity and low efficiency, are unable to meet engineering requirements. Although hybrid models based on statistical feature dimensionality reduction and support vector machine algorithms have shown adaptability in simple scenarios, their ability to decouple features from complex discharge patterns remains significantly insufficient, creating an urgent need for the development of new intelligent algorithms. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for identifying partial discharge defect types based on two-stage feature fusion, aiming to solve the safety hazards caused by partial discharge and thus ensure the stable operation of power equipment, and solve the technical problem that the judgment of partial discharge type mostly depends on manual and intelligent algorithms and the identification of partial discharge defect types is inaccurate. By fusing statistical features and image features, the partial discharge pattern is identified in stages, which further improves the recognition robustness and accuracy.

[0005] To solve the above technical problems, the technical solution adopted by the present invention is: a method for identifying partial discharge defect types based on two-stage feature fusion, comprising the following steps:

[0006] Step 1: Obtain original discharge waveform data;

[0007] Step 2: Preprocess the discharge waveform data and draw the PRPD spectrum;

[0008] Step 3: Extract the statistical features of partial discharge in the PRPD spectrum and perform feature selection using RFE;

[0009] Step 4: Use CatBoost to learn the selected statistical features and build a recognition model. Use the recognition model to identify the type of the collected data samples and identify and filter out the discharge type.

[0010] Step 5: Input the PRPD map of the screened single-discharge type sample into the convolutional neural network (CNN) to extract image features;

[0011] Step 6: Input the sample image features and the selected statistical features into Mamba for fusion to form a new feature vector;

[0012] Step 7: Use the YOLO model to improve and build a secondary recognition model, use the secondary recognition model to perform secondary recognition on the new feature vector, output the recognition results, and evaluate the overall model.

[0013] The further improvement of the technical solution of the present invention is that: Step 2 specifically comprises the following steps: performing noise reduction processing on the discharge waveform data, and drawing a PRPD spectrum of the partial discharge waveform data signal according to the relationship between phase and amplitude.

[0014] The further improvement of the technical solution of the present invention is that the specific steps of step 3 are as follows:

[0015] Step 3.1: Extract the statistical features of partial discharge from the PRPD spectrum, including maximum value, minimum value, mean value, maximum power, discharge factor, PRPD skewness, PRPD kurtosis, phase distribution and amplitude distribution characteristics;

[0016] Step 3.2: Use the recursive feature elimination (RFE) algorithm to perform feature space dimensionality reduction. RFE completes feature selection by recursively training the model and removing the least important features until the preset number of features is reached.

[0017] A further improvement to the technical solution of the present invention is that the phase distribution and amplitude distribution characteristics in step 3.1 are used to describe the intensity or distribution of partial discharge at different phases of the signal and are discrete values. The phase range is divided into 120 windows, each with a phase range of 3°.

[0018] A further improvement of the technical solution of the present invention is that the specific steps of constructing a recognition model in step 4 are as follows:

[0019] Step 4.1: Encode the categorical features. Randomly sort the training data, then process each sample in turn, encoding the categorical features of the current sample based on the statistics of the previous sample.

[0020] Step 4.2: Replace each value of the categorical feature with the statistic of the target value;

[0021] Step 4.3: Normalize the numerical features to make them have similar scales and then handle missing values;

[0022] Step 4.4: Initialize the CatBoost classifier and predefine the number of iterations, learning rate, tree depth, and loss function.

[0023] Step 4.5: At the beginning of training, for each sample, calculate the difference between its true label and the predicted label to obtain the gradient information;

[0024] Step 4.6: Based on the gradient information, use the gradient boosting algorithm to construct multiple decision trees. Each tree is constructed by selecting a feature and its corresponding split point at each node to minimize the loss function. Then, the samples are divided into left and right child nodes according to the selected feature and split point. This step is repeated until the decision tree reaches the preset maximum depth or the number of samples in the node falls below a set threshold.

[0025] Step 4.7: During the model training phase, use the validation set to monitor model performance. Evaluation metrics include recognition accuracy, recall, and F1 score.

[0026] A further improvement of the technical solution of the present invention is that the specific evaluation method of each evaluation indicator in step 4.7 is:

[0027] The recognition accuracy rate (Accuracy) indicates the ratio of the number of samples correctly recognized by the model to the total number of samples, reflecting the overall prediction accuracy. The formula is:

[0028]

[0029] Recall refers to the ratio of the number of positive samples correctly identified by the model to the actual number of positive samples, reflecting the ability to capture positive examples. The formula is:

[0030]

[0031] The F1 score (F1-Score) represents the harmonic mean of precision and recall, balancing the performance of the two. Its formula is:

[0032] F1-Score=2×Accuracy×Recall÷(Accuracy+Recall)

[0033] A further improvement of the technical solution of the present invention is that: the primary discharge type screened out in step 4 is specifically an air gap discharge type or a suspension discharge type, and samples of surface discharge and tip discharge types are directly output.

[0034] A further improvement of the technical solution of the present invention is that the specific steps of step 6 are as follows:

[0035] Step 6.1: Normalize the image features and statistical features to ensure that the feature values ​​are in a similar range;

[0036] Step 6.2: The input layer of the Mamba module performs image feature input and statistical feature input;

[0037] Step 6.3: Receive the image feature vector extracted by CNN and set it as Where N is the batch size, C image is the number of channels of the image feature, H and W are the height and width of the feature map;

[0038] Step 6.4: Receive the statistical feature vector after dimensionality reduction and set it as Among them C stat is the dimension of statistical features;

[0039] Step 6.5: Align the image features and statistical features to the same dimension, and use global average pooling to reduce the dimension of the image features to

[0040] Step 6.6: If the dimensions of the statistical features do not match, use a linear layer to map them to the same dimension.

[0041] Step 6.7: Input the aligned image features and statistical features into the Mamba module for fusion, and output the fused feature vector X fused .

[0042] A further improvement of the technical solution of the present invention is that the improvement of the YOLO model in step 7 is specifically as follows: MambaLayer is introduced in the Backbone part of the YOLO model, the traditional C2F block is improved in the Neck part, and the Mamba attention mechanism is added.

[0043] Due to the adoption of the above-mentioned technical solution, the technical progress achieved by the present invention is as follows: through two-stage feature extraction, selection, fusion and recognition, the accuracy and efficiency of partial discharge defect type identification are improved. CatBoost is used for primary recognition, which effectively improves the performance of the model; CNN is used to extract image features and combine them with statistical features to fuse information of different features. On this basis, an improved YOLO model is used for secondary recognition, thereby further improving the recognition ability and stability of the model. The present invention has high innovation and practicality, and can better solve the problem of confusion in discharge pattern recognition during partial discharge. The use of staged recognition can effectively give play to the advantages of the dual model, improve recognition efficiency, and provide strong support for power equipment fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention, and those skilled in the art can derive other drawings based on these drawings without inventive effort.

[0045] Figure 1 It is a flow chart of the identification method of the present invention. DETAILED DESCRIPTION

[0046] The present invention is described in further detail below in conjunction with the embodiments:

[0047] like Figure 1 The figure shows a flow chart of a method for identifying partial discharge defect types based on dual-stage feature fusion. The specific steps are as follows:

[0048] Step 1: Obtain raw discharge waveform data. Collect discharge data multiple times to obtain raw discharge waveform data, including samples of four discharge types: tip, air gap, suspension, and surface. Build a discharge database.

[0049] Step 2: Preprocess the discharge waveform data and plot the PRPD spectrum. Data preprocessing begins with denoising to reduce noise interference. Wavelet transforms or filter banks are used, but not limited to, to denoise the raw discharge waveform data, removing white noise and narrowband interference. This improves data quality and reduces the impact of noise on feature extraction. The data is then standardized to eliminate potential order-of-magnitude differences between features, ensuring equivalence and comparability in subsequent analysis. The PD waveform data signals are plotted as PRPD spectra based on the relationship between phase and amplitude to visually demonstrate discharge characteristics. The different types of collected discharges are organized and categorized to establish a PRPD spectrum database.

[0050] Step 3: Extract the statistical features of partial discharge in the PRPD spectrum,

[0051] Step 3.1: Extract statistical features of partial discharge from the PRPD spectrum, including maximum, minimum, mean, maximum power, discharge factor, PRPD skewness, PRPD kurtosis, phase distribution, and amplitude distribution characteristics; comprehensively quantify the statistical characteristics of partial discharge. The phase distribution and amplitude distribution characteristics are discrete values ​​used to describe the intensity or distribution of partial discharge at different phases. The phase range is divided into 120 windows, each with a phase range of 3°.

[0052] Step 3.2: Use the recursive feature elimination (RFE) algorithm to implement feature space dimensionality reduction. RFE recursively trains the model and removes the least important features until the preset number of features is reached. It gradually discards unimportant features and retains the main features to complete feature selection. In this embodiment, the number of features selected is 10.

[0053] Step 4: Use CatBoost to learn the selected statistical features and build a recognition model. Use the recognition model to identify the type of the collected data samples and identify and filter out the discharge type.

[0054] The specific steps to build a recognition model are as follows:

[0055] Step 4.1: Encode the categorical features. Randomly sort the training data, then process each sample in turn, encoding the categorical features of the current sample based on the statistics of the previous sample.

[0056] Step 4.2: Replace each value of the categorical feature with the statistic of the target value;

[0057] Step 4.3: Normalize the numerical features to make them have similar scales and then handle missing values;

[0058] Step 4.4: Initialize the CatBoost classifier and predefine the number of iterations, learning rate, tree depth, and loss function.

[0059] Step 4.5: At the beginning of training, for each sample, calculate the difference between its true label and the predicted label to obtain the gradient information;

[0060] Step 4.6: Based on the gradient information, use the gradient boosting algorithm to construct multiple decision trees. Each tree is constructed by selecting a feature and its corresponding split point at each node to minimize the loss function. Then, the samples are divided into left and right child nodes according to the selected feature and split point. This step is repeated until the decision tree reaches the preset maximum depth or the number of samples in the node falls below a set threshold.

[0061] Step 4.7: During the model training phase, the validation set is used to monitor model performance. Evaluation metrics include recognition accuracy, recall, and F1 score. CatBoost effectively learns the characteristic representations of partial discharge defect types from the training data and accurately classifies and identifies new samples. Recognition accuracy (Accuracy) represents the proportion of samples correctly identified by the model to the total number of samples, reflecting overall prediction accuracy. The formula is:

[0062]

[0063] Recall refers to the ratio of the number of positive samples correctly identified by the model to the actual number of positive samples, reflecting the ability to capture positive examples. The formula is:

[0064]

[0065] The F1-Score represents the harmonic mean of precision and recall, balancing the performance of the two. Its formula is:

[0066] F1-Score=2×Accuracy×Recall÷(Accuracy+Recall)

[0067] The statistical features after feature selection are used as input feature vectors to the CatBoost recognition model. If tip discharge or surface discharge is identified, the results are directly output. Meanwhile, if suspended discharge or tip discharge samples are identified, they are filtered out. In other words, the primary discharge recognition model performs an initial screening, selecting suspended discharge or tip discharge samples as primary discharge types.

[0068] Step 5: Input the PRPD maps of the selected single-discharge samples into a convolutional neural network (CNN) to extract image features. Specifically, input the PRPD maps of the selected suspended discharge or tip discharge samples into the CNN to extract image features. The CNN can automatically extract deep-level image features from the PRPD maps. Select ResNet-50 as the CNN model to extract image features from the PRPD maps.

[0069] Step 6: Input the sample image features and the selected statistical features into Mamba for fusion to form a new feature vector;

[0070] Step 6.1: Normalize the image features and statistical features to ensure that the feature values ​​are in a similar range;

[0071] Step 6.2: Design the dual input interfaces of the Mamba module to receive image features and statistical features respectively. The input layer includes a feature conversion component to adapt the input features of different formats to the Mamba processing dimensions. The image feature input path preserves the spatial structure information, while the statistical feature input path maintains its sequential nature. Through independent processing channels, the unique characteristics of the two feature types are preserved.

[0072] Step 6.3: Receive the image feature vector extracted by CNN and set it as Where N is the batch size, C image is the number of channels of the image feature, H and W are the height and width of the feature map. The image features are spatially compressed and the spatial dimension of each channel is compressed to 1 through global average pooling, resulting in a dimension of [N, C image ] feature vector, retaining the global semantic information of the image;

[0073] Step 6.4: Receive the statistical feature vector after dimensionality reduction and set it as Among them C stat is the dimension of statistical features;

[0074] Step 6.5: Align the image features and statistical features to the same dimension, and use global average pooling to reduce the dimension of the image features to

[0075] Step 6.6: If the dimensions of the statistical features do not match, use a linear layer to map them to the same dimension. The weight matrix W of the linear layer is initialized to random values ​​and its dimension is C stat×D (D is the target dimension) During the training process, the back propagation algorithm is used to calculate the gradient according to the loss function and continuously adjust the parameter values ​​of the weight matrix W to learn the optimal mapping relationship. In the mapping operation, the batch dimension remains unchanged and only the feature dimension is transformed, that is, the input statistical feature vector X is transformed. stat Perform matrix multiplication X′ stat =X stat W, get the dimension-adapted statistical feature vector X′ stat ;

[0076] Step 6.7: Input the aligned image features and statistical features into the Mamba module for fusion, and output the fused feature vector X fused The Mamba module is built on a state-space model and implements sequence modeling through three core learnable parameter matrices: The parameter matrix A (shape [D, d_state]) is parameterized as A_tilde = Sigmoid(A) via the Sigmoid activation function, controlling the strength of dependencies between different time steps during state transitions; the parameter matrix B (shape [D, d_state]) is parameterized as B_tilde = Tanh(B) via the Tanh activation function, mapping input features to the state space; and the parameter matrix C (shape [D, d_state]) is also parameterized as C_tilde = Tanh(C) via the Tanh activation function, mapping the state space representation back to the feature space. Furthermore, a time step parameter dt (shape [D]) is introduced and scaled to dt_scaled = Sigmoid(dt) via the Sigmoid function to dynamically adjust the state update frequency. This design enables Mamba to adaptively capture long-distance dependencies in sequences while maintaining linear computational complexity. It is particularly suitable for processing multimodal data with different characteristics, such as the spatial dependence of image features and the global correlation of statistical features.

[0077] Mamba employs a dual gating mechanism to achieve precise control over information flow: the input gate, generated by Sigmoid(Linear(X)), controls the influx of input information; the reset gate, also generated by Sigmoid(Linear(X)), determines the degree to which the previous state influences the current state. A gating value close to 1 allows full information flow, while a gating value close to 0 inhibits the corresponding information flow. During processing, input features are first gated to generate value = SiLU(W_value*X+b_value). This is then combined with state-space parameters to calculate: state = B_tilde*value*dt_scaled, output = Einsum(′bld,ds->bld′,state,C_tilde). This combination of gating and state-space modeling enables Mamba to selectively integrate key information from multimodal features.

[0078] The processed features are further optimized through the residual structure and linear projection: first, the Mamba processed output is added to the original input through the residual connection, that is, X_residual = X+output, to ensure that the early feature information is losslessly transmitted in the deep network; then the features are mapped back to the original dimension through the linear projection layer, that is, X fused =W_proj*X_residual+b_proj;

[0079] Step 7: Use the YOLO model to improve and build a secondary recognition model, use the secondary recognition model to perform secondary recognition on the new feature vector, output the recognition results, and evaluate the overall model.

[0080] The improved YOLO model architecture is as follows: The input layer first receives the fused feature vector. The convolutional layers of the backbone network extract low-level semantic information, such as local features like edges and textures, through 3×3 and 1×1 convolutions. These convolution operations maintain sensitivity to image details. The Mamba Layer is introduced to enhance the model's global understanding of image features. Based on the state-space model (SSM), the Mamba Layer uses learnable parameters A, B, C, and a time step parameter dt to effectively capture long-range dependencies while maintaining low computational complexity. The improved C2F block in the Neck (feature fusion layer) further fuses and enhances features from the backbone. Building on the original CSP architecture, it adds cross-stage connections and feature recombination mechanisms. This multi-branch structure processes features at different scales in parallel, improving feature representation. Each C2F block contains multiple convolutional layers and residual connections to ensure that key information is not lost during feature transfer. The Mamba attention mechanism is then added to better fuse features from different layers and highlight feature channels that are more important for classification and detection tasks. Finally, the Head (detection head) continues YOLO's decoupling design and generates the final recognition result based on the fused multi-scale features.

[0081] The embodiments described above are merely descriptions of preferred implementations of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A method for identifying partial discharge defect types based on two-stage feature fusion, characterized by: The steps are as follows: Step 1: Obtain original discharge waveform data; Step 2: Preprocess the discharge waveform data and draw the PRPD spectrum; Step 3: Extract the statistical features of partial discharge in the PRPD spectrum and perform feature selection using RFE; Step 4: Use CatBoost to learn the selected statistical features and build a recognition model. Use the recognition model to identify the type of the collected data samples and identify and filter out the discharge type. Step 5: Input the PRPD map of the screened single-discharge type sample into the convolutional neural network (CNN) to extract image features; Step 6: Input the sample image features and the selected statistical features into Mamba for fusion to form a new feature vector; Step 7: Use the YOLO model to improve and build a secondary recognition model, use the secondary recognition model to perform secondary recognition on the new feature vector, output the recognition results, and evaluate the overall model.

2. The method for identifying partial discharge defect types based on dual-stage feature fusion according to claim 1, characterized in that: The specific steps of step 2 are: performing noise reduction processing on the discharge waveform data, and drawing a PRPD spectrum of the partial discharge waveform data signal according to the relationship between phase and amplitude.

3. The method for identifying partial discharge defect types based on dual-stage feature fusion according to claim 1, characterized in that: Step 3: Step 3.1: Extract the statistical features of partial discharge from the PRPD spectrum, including maximum value, minimum value, mean value, maximum power, discharge factor, PRPD skewness, PRPD kurtosis, phase distribution and amplitude distribution characteristics; Step 3.2: Use the recursive feature elimination (RFE) algorithm to perform feature space dimensionality reduction. RFE completes feature selection by recursively training the model and removing the least important features until the preset number of features is reached.

4. The method for identifying partial discharge defect types based on dual-stage feature fusion according to claim 3 is characterized in that: The phase distribution and amplitude distribution features in step 3.1 are used to describe the intensity or distribution of partial discharge at different phases of the signal and are discrete values. The phase range is divided into 120 windows, and the phase range of each window is 3°.

5. The method for identifying partial discharge defect types based on dual-stage feature fusion according to claim 1, characterized in that: The specific steps for building a recognition model in step 4 are as follows: Step 4.1: Encode the categorical features. Randomly sort the training data, then process each sample in turn, encoding the categorical features of the current sample based on the statistics of the previous sample. Step 4.2: Replace each value of the categorical feature with the statistic of the target value; Step 4.3: Normalize the numerical features to make them have similar scales and then handle missing values; Step 4.4: Initialize the CatBoost classifier and predefine the number of iterations, learning rate, tree depth, and loss function. Step 4.5: At the beginning of training, for each sample, calculate the difference between its true label and the predicted label to obtain the gradient information; Step 4.6: Based on the gradient information, use the gradient boosting algorithm to construct multiple decision trees. Each tree is constructed by selecting a feature and its corresponding split point at each node to minimize the loss function. Then, the samples are divided into left and right child nodes according to the selected feature and split point. This step is repeated until the decision tree reaches the preset maximum depth or the number of samples in the node falls below a set threshold. Step 4.7: During the model training phase, use the validation set to monitor model performance. Evaluation metrics include recognition accuracy, recall, and F1 score.

6. The method for identifying partial discharge defect types based on dual-stage feature fusion according to claim 5, characterized in that: The specific evaluation method for each evaluation indicator in step 4.7 is: The recognition accuracy rate (Accuracy) indicates the ratio of the number of samples correctly recognized by the model to the total number of samples, reflecting the overall prediction accuracy. The formula is: Recall refers to the ratio of the number of positive samples correctly identified by the model to the actual number of positive samples, reflecting the ability to capture positive examples. The formula is: The F1-Score represents the harmonic mean of precision and recall, balancing the performance of the two. The formula is: F1-Score=2×Accuracy×Recall÷(Accuracy+Recall).

7. The method for identifying partial discharge defect types based on dual-stage feature fusion according to claim 5, characterized in that: The primary discharge type screened out in step 4 is specifically an air gap discharge type or a suspension discharge type, and samples of surface discharge and tip discharge types identified as the primary discharge type are directly output.

8. The method for identifying partial discharge defect types based on dual-stage feature fusion according to claim 1, characterized in that: Step 6 The specific steps are as follows: Step 6.1: Normalize the image features and statistical features to ensure that the feature values ​​are in a similar range; Step 6.2: The input layer of the Mamba module performs image feature input and statistical feature input; Step 6.3: Receive the image feature vector extracted by CNN and set it as Where N is the batch size, C image is the number of channels of the image feature, H and W are the height and width of the feature map; Step 6.4: Receive the statistical feature vector after dimensionality reduction and set it as Among them C stat is the dimension of statistical features; Step 6.5: Align the image features and statistical features to the same dimension, and use global average pooling to reduce the dimension of the image features to Step 6.6: If the dimensions of the statistical features do not match, use a linear layer to map them to the same dimension. Step 6.7: Input the aligned image features and statistical features into the Mamba module for fusion, and output the fused feature vector X fused .

9. The method for identifying partial discharge defect types based on dual-stage feature fusion according to claim 1, characterized in that: The improvements to the YOLO model in step 7 are as follows: introducing the Mamba Layer in the Backbone part of the YOLO model, improving the traditional C2F block in the Neck part, and adding the Mamba attention mechanism.

Images