Partial discharge detection method based on adaptive feature extraction and cross-modal feature fusion

Abstract

The invention discloses a partial discharge detection method based on adaptive feature extraction and cross-modal feature fusion, and relates to the technical field of electrical equipment insulation detection, and the method comprises the steps: collecting partial discharge data; preprocessing the data; building a deep learning model; according to the method, a deep learning model is adopted for building, processed data serves as input of the model, the network is divided into three parts, namely, a double-branch 1DCNN network, a cross-modal feature fusion module and a multi-branch decision fusion module based on an attention mechanism, and while high detection precision is guaranteed, the detection accuracy is greatly improved. The method effectively improves the combined representation capability of the time domain and frequency domain signals, remarkably improves the detection efficiency, and can achieve the high-accuracy and high-stability partial discharge detection in a complex power equipment environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power equipment insulation detection, and particularly relates to a partial discharge detection method based on adaptive feature extraction and cross-modal feature fusion, which is suitable for online monitoring and fault early warning of insulated overhead conductors in urban distribution networks. BACKGROUND

[0002] With the acceleration of urbanization and the growth of distribution network load, 10kV insulated overhead conductors are widely used due to the limited adaptation space of urban environment. However, such conductors are easily affected by insulation aging, extreme weather and other factors, causing partial discharge, accelerating insulation deterioration and possibly leading to short circuit power failure.

[0003] The existing partial discharge detection methods generally use deep learning-based detection methods, and the conventional detection methods have the following obvious defects:

[0004] The deep learning-based detection methods rely on single-modal features. Time-domain signals can reflect transient characteristics but lack frequency distribution information, and frequency-domain features are difficult to capture transient changes, resulting in insufficient utilization of multi-dimensional information.

[0005] The partial discharge signals collected on site often contain a large amount of background noise, such as power frequency noise generated by line load current and electromagnetic interference noise generated by surrounding electronic equipment, which can mask the effective features of the discharge signals.

[0006] The existing deep learning models use simple filtering methods such as mean filtering and Gaussian filtering for noise reduction, which cannot effectively separate noise and effective signals. At the same time, single-modal features are more sensitive to noise, and feature extraction deviation is easy to occur under noise interference, resulting in a significant decrease in detection accuracy.

[0007] In addition, some studies attempt to combine time-domain and frequency-domain features, but mostly use static fusion methods such as simple feature splicing and weighted averaging, which cannot dynamically mine the correlation between different modal features and cannot fully utilize the advantages of multi-modal data. Therefore, it is of great significance to develop a partial discharge detection method that can realize adaptive extraction and dynamic fusion of multi-modal features and has strong anti-interference ability, in order to ensure the safe and stable operation of 10kV insulated overhead conductors and improve the power supply reliability of urban distribution networks. SUMMARY

[0008] In view of the core pain points of insufficient information utilization, weak anti-interference ability and low detection accuracy in the existing partial discharge detection technology, the present application provides a partial discharge detection method based on adaptive feature extraction and cross-modal feature fusion, which can significantly improve the detection efficiency while ensuring high detection accuracy, and can realize high-accuracy and high-stability partial discharge detection in complex power equipment environments.

[0009] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0010] A partial discharge detection method based on adaptive feature extraction and cross-modal feature fusion includes the following steps: (1) partial discharge data acquisition; (2) data preprocessing; (3) deep learning model building; and (4) result prediction and evaluation.

[0011] The specific steps are as follows:

[0012] 1) Partial Discharge Data Acquisition: The dataset used was the partial discharge dataset for overhead insulated conductors published by the ENET Center of the Technical University of Ostrava (VSB) in the Czech Republic. This dataset is derived from actual operating conditions, ensuring both realism and scenario adaptability. The sampling frequency was set to 40MHz, and each signal consisted of 800,000 sampling points.

[0013] 2) Data preprocessing: For the original signal of partial discharge detection, a parallel processing strategy is adopted to generate input data adapted for dual-modal feature extraction: On the one hand, the original time domain signal is processed by max pooling with a step size of 100 to retain transient pulse features and compress dimensions; on the other hand, the original time domain signal is simultaneously processed by FFT transformation, normalization, and low-pass filtering of 1-20MHz, and then the filtered spectrum signal is processed by max pooling with a step size of 40.

[0014] The two paths respectively enhance the core information in the time and frequency domains, providing a complementary and dimensionally adapted dual-modal data foundation for subsequent cross-modal fusion. After data preprocessing, the feature data of the two modalities are divided into training and test sets, and then randomly divided into four folds.

[0015] 3) Deep Learning Model Construction: The processed data is used as the model input. The network consists of three parts: a two-branch 1DCNN network, a cross-modal feature fusion module, and a multi-branch decision fusion module based on an attention mechanism. Its overall structure is as follows:

[0016] (1) The backbone network is a dual-branch structure consisting of two parallel 1DCNN networks, used to extract features from the time-domain and frequency-domain signals, respectively. Each branch has the same structure, consisting of five convolutional blocks, each containing a convolutional layer (Conv1d), a batch normalization layer (BatchNorm1d), an activation function (ReLU), and a pooling layer (MaxPool1d). The kernel size in the convolutional blocks is progressively increased layer by layer, gradually extracting different levels of time-domain and frequency-domain features. Through this design of progressively increasing the kernel size and expanding the feature channels, the network can effectively capture multi-scale and multi-level time-domain features in the signal. Moreover, the pooling operation (MaxPool1d) in each layer gradually reduces the spatial dimension of the data while retaining the most important feature information, thereby reducing computational cost. Finally, the adaptive pooling (AdaptiveAvgPool1d) in each branch compresses the feature map into a fixed dimension, enabling the subsequent feature fusion module to process features from the two branches more efficiently.

[0017] (2) The feature fusion module is used in two different locations. The two feature fusions complementarily fuse the output features of the third and fourth layers from the two branches (the time-domain signal branch and the additional signal branch), effectively enhancing the richness and diversity of feature expression. This cross-modal fusion strategy enables the network to make full use of the advantages of each modality when processing different modal signals, reducing the limitations and noise effects of a single modality. After each feature fusion, the fused feature is concatenated with the output feature of the next layer of the backbone network branch, thus forming a new feature vector that flows to the subsequent layers of the network. This feature concatenation strategy not only enhances the network's ability to integrate multimodal information, but also effectively supplements and compensates for the content or details that may be missing in a single modality feature. The two-level fusion module corresponds to the mid-term and deep features of the dual-branch network, forming a progressive cross-modal interaction: first, local details are fused, and then global semantics are integrated, avoiding information loss caused by a single fusion step, while allowing the fused features to directly participate in the feature extraction of subsequent branches, realizing a cyclical enhancement of fusion-re-extraction.

[0018] In the specific architecture of the cross-modal feature fusion module, the outputs of the time-domain and frequency-domain branches are first compressed to a specified output length using adaptive average pooling. Then, corresponding adaptive feature extraction methods are applied to each branch. A fusion matrix is ​​constructed and further processed through a Swing Transformer network to complete the cross-modal feature fusion. The specific steps are as follows:

[0019] In the time-domain signal branch, the difference sequence between signal feature points is first calculated. This effectively enhances the features of abrupt changes or edges in the signal. This process allows the model to focus more on rapidly changing parts of the signal during subsequent learning. The formula is:

[0020]

[0021] Where b is the batch size and c is the channel. and These are the values ​​of the i-th and j-th feature points in each channel c, respectively. Next, the difference sequence is normalized to scale the range of difference values ​​to [-1, 1] for subsequent calculation of attention weights. The normalization process is as follows:

[0022]

[0023] ε is set to 10⁻⁶ to prevent division by zero. The normalized difference sequences are then processed using a softmax function to enhance the differences between the data, thereby further highlighting significant variations. A trainable temperature-controlled parameter T is used to adjust the smoothness of the softmax function. The calculation formula is as follows:

[0024]

[0025] The sign function is used to preserve the sign information of the differential sequence, indicating that the direction of the signal abrupt change remains unchanged.

[0026] In the frequency domain signal branch, the eigenvalues ​​of signal x2 are first constrained to [-1, 1] using the tanh activation function. tanh is a smooth nonlinear activation function; compared to linear transformations, it can capture complex nonlinear patterns in the frequency domain features, avoiding bias in model weights towards larger values ​​due to differences in the magnitude of eigenvalues, thus further highlighting the overall characteristics. After activation, an inverse cosine operation is performed to calculate the angle value φ of the feature points in each channel:

[0027]

[0028] The `clamp` operation ensures that the input is within the valid range of [-1, 1], avoiding errors during calculation. Subsequently, the angular differences between frequency domain feature points in each channel `c` are calculated. These angular differences reflect the structured relationships between different components in the frequency domain, forming a frequency domain difference sequence. The specific angular difference calculation is as follows:

[0029] GADF(i,j)=sin(φ i -φ j (5)

[0030] Where φ i and φ j These are the angle values ​​of two feature points in the signal.

[0031] After completing the above steps, the softmax attention difference sequence of the temporal branch and the angular difference sequence of the frequency domain branch are fused through recombination. Specifically, the output of the temporal branch is filled into the lower triangular part of the matrix, and the output of the frequency domain branch is filled into the upper triangular part. This fills different regions of the matrix with temporal and frequency domain information respectively, effectively reducing the influence of temporal dependencies and focusing more on the independence and complementarity between modal mutations. This fusion matrix, while preserving temporal and frequency domain information, enhances the focus on key features. Finally, the Swin Transformer is used for further feature extraction on the recombined modality fusion matrix, aiming to achieve efficient fusion of cross-modal features using its window attention mechanism. Specifically, the window mechanism of the Swin Transformer can effectively capture the correlation of local regions. Windows near the diagonals contain information from both temporal and frequency domain modalities, thereby promoting interaction and fusion between modalities. In regions far from the two diagonals, the windows mainly focus on information from their respective modalities, thus ensuring the integrity and expression of the original modal features. At the same time, this fusion approach not only preserves the characteristics of a single modality, but also enhances the connection between modalities through attention mechanisms, fully exploring the interactive information between modalities.

[0032] (3) At this stage, the model has processed signals from two modalities through two backbone networks, and the fusion of the two feature fusion blocks includes fusion features from the time domain and frequency domain. The outputs of these four branches are max-pooled and concatenated in the channel dimension to become the input of the multi-branch decision fusion module; the channel attention mechanism is used for dynamic weighting adjustment, thereby enhancing the expressive power of key information and suppressing irrelevant information; the features adjusted by the attention mechanism are passed to the fully connected layer for final classification. In this way, the model can effectively integrate features from different modalities and processing stages, and promote information exchange and fusion between different modalities.

[0033] 4) Result Prediction and Evaluation: This experiment uses Adaptive Moment Estimation (Adam) as the optimizer to accelerate the training process. The initial learning rate is set to 10⁻⁴. The ReduceLROnPlateau adaptive learning rate scheduling mechanism is used to automatically decay the learning rate when the validation set performance does not improve within a certain number of epochs, achieving a more stable convergence process. The loss function is cross-entropy loss with class weights, where the positive and negative sample weight ratio is 1:10. During training, the batch size is fixed at 64, and four-fold cross-validation is used, with each fold cross-validation training for a maximum of 70 epochs. Training is performed on a designated GPU, and the number of data loading processes is set to 0.

[0034] Due to class imbalance in the dataset, conventional precision metrics alone are not representative. Therefore, we selected MCC (Mean Cross-Sectional Control), recall, and precision, with MCC as the primary evaluation criterion. The formulas for calculating these three metrics are shown in (6-8):

[0035]

[0036]

[0037] Here, TP (True Positive) refers to the number of samples that are actually positive and correctly predicted as positive, TN (True Negative) refers to the number of samples that are actually negative and correctly predicted as negative, FP (False Positive) refers to the number of samples that are actually negative but incorrectly predicted as positive, and FN (False Negative) refers to the number of samples that are actually positive but incorrectly predicted as negative.

[0038] The beneficial effects of this invention are as follows:

[0039] The method of this invention effectively improves the joint representation capability of time-domain and frequency-domain signals by constructing a multi-branch network structure, introducing an attention mechanism, and designing a cross-modal feature fusion strategy. This invention achieves multiple breakthroughs in information utilization, anti-interference capability, detection accuracy and efficiency. While ensuring high detection accuracy, it significantly improves detection efficiency and can achieve high accuracy and high stability partial discharge detection in complex power equipment environments. Attached Figure Description

[0040] Figure 1 This is a diagram of the deep learning model structure.

[0041] Figure 2 This is the overall flowchart of the present invention. Detailed Implementation

[0042] The invention will now be further described with reference to the accompanying drawings:

[0043] The flowchart of the method of the present invention is as follows Figure 2 As shown, in simple terms, it includes the following steps:

[0044] (1) Partial discharge data acquisition; (2) Data preprocessing; (3) Deep learning model building; (4) Result prediction and evaluation.

[0045] The specific implementation process of this method is as follows:

[0046] Step 1: Obtain electrical signals during equipment operation by deploying UHF and ultrasonic sensors near key parts prone to discharge, such as conductor insulators and clamps.

[0047] Step 2: Generate dual-modal input data using a parallel strategy:

[0048] Max pooling is used to reduce the dimensionality of the original time-domain signal. Simultaneously, the original time-domain signal is transformed by FFT, normalized, and low-pass filtered from 1 to 20 MHz. The spectrum signal is then max pooled and randomly divided into training and test sets in an 8:2 ratio.

[0049] Step 3: Building a deep learning model: Input the training and test sets into an adaptive feature extraction and cross-modal feature fusion model built on PyTorch, and train the model using four-fold cross-validation.

[0050] Step 4: Result Prediction and Evaluation: To verify the impact of time-domain and frequency-domain modal feature fusion on model performance, especially to verify the effectiveness of cross-modal feature fusion and the necessity of using feature fusion blocks, we evaluate the complementarity of time-domain and frequency-domain information by comparing different input modalities and fusion strategies, and analyze their synergistic effect after feature fusion.

[0051] This experiment uses Adaptive Moment Estimation (Adam) as the optimizer to accelerate the training process. The initial learning rate is set to 10. -4 The ReduceLROnPlateau adaptive learning rate scheduling mechanism is used, which automatically decays the learning rate when the validation set performance does not improve within a certain number of epochs, thus achieving a more stable convergence process. The loss function is cross-entropy loss with class weights, where the ratio of positive to negative sample weights is 1:10. During training, the batch size is fixed at 64, and four-fold cross-validation is used, with each fold cross-validation training for a maximum of 70 epochs. Training is performed on a specified GPU, and the number of data loading processes is set to 0.

[0052] Because the dataset suffers from class imbalance, conventional accuracy metrics alone are not representative. Therefore, we selected MCC (Mean Cross-Sectional Classification), recall, and precision, with MCC as the primary evaluation criterion.

[0053] The experimental results are shown in Table 1.

[0054] Table 1 Comparison of experimental results for cross-modal feature fusion

[0055]

[0056] The comparison between Experiment 1 and Experiments 2, 3, and 6 shows that the model's performance (especially MCC and Recall) is significantly lower without the cross-modal feature fusion module. Introducing the fusion module allows the model to better integrate information from the time and frequency domains, significantly improving classification performance. Experiments 2 and 3 demonstrate that even using only a single fusion module significantly improves model performance. Experiment 6 further shows that using two fusion modules simultaneously (processing frequency and time domain modalities respectively) further enhances model performance. This result highlights the improvement in model classification ability brought about by the cross-modal feature fusion block and verifies the necessity of using two feature fusion modules, helping the model to more comprehensively understand and utilize multimodal data.

[0057] The comparison results of Experiments 4, 5, and 6 show that although the model's performance improved when using a single modality and applying the fusion module, it still fell short of the effect of combining both time-domain and frequency-domain modalities. This indicates that time-domain and frequency-domain modalities are significantly complementary, and using either modality alone cannot fully leverage its advantages. Therefore, comprehensively utilizing time-domain and frequency-domain modalities, combined with an appropriate fusion module, can significantly improve the model's classification performance and fully realize the potential of multimodal data.

[0058] This experiment verifies the effectiveness of the partial discharge detection method based on adaptive feature extraction and cross-modal feature fusion, providing an efficient and reliable solution for partial discharge detection of 10kV insulated overhead conductors, and has significant engineering application value.

[0059] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements can be made without departing from the principle of the present invention, and these improvements should also be considered within the scope of protection of the present invention without creative effort.

Claims

1. A partial discharge detection method based on adaptive feature extraction and cross-modal feature fusion, characterized in that: The steps include (1) partial discharge data acquisition; (2) data preprocessing; (3) deep learning model building; (4) result prediction and evaluation; Step (1) partial discharge data acquisition: the data set used is the public overhead insulated conductor partial discharge data set, and the sampling frequency is set to 40MHz, each signal consists of 800,000 sampling points; Step (2) data preprocessing: for the original signal of partial discharge detection, a parallel processing strategy is adopted to generate input data suitable for dual-modal feature extraction: on the one hand, the original time domain signal is processed by step 100 maximum pooling to retain transient pulse features and compress dimensions; on the other hand, the original time domain signal is converted by FFT, normalized, and low-pass filtered at 1-20MHz, and then the filtered frequency spectrum signal is processed by step 40 maximum pooling; after data preprocessing, the feature data of the above two modalities is divided into training set and test set, and randomly divided into four folds; Step (3) deep learning model building: the processed data is used as the input of the model, and the network is divided into three parts: a dual-branch 1DCNN network, a cross-modal feature fusion module, and a multi-branch decision fusion module based on an attention mechanism.

2. The partial discharge detection method based on adaptive feature extraction and cross-modal feature fusion according to claim 1, characterized in that: In step (3), the backbone network is a dual-branch 1DCNN network composed of two parallel 1DCNN networks, which are used to extract features of time domain signals and frequency domain signals respectively; The structure of each branch is the same, which is composed of 5 convolution blocks, each of which contains a convolution layer, a batch normalization layer, an activation function and a pooling layer; the convolution kernel size in the convolution block adopts the mode of layer-by-layer enlargement, which gradually extracts time domain and frequency domain features at different levels; through this design of gradually increasing the convolution kernel and gradually expanding the feature channel, the network can effectively capture the multi-scale and multi-level time domain features in the signal; moreover, the pooling operation at each layer can gradually reduce the spatial dimension of the data while retaining the most important feature information, thereby reducing the computational load; finally, the adaptive pooling of each branch compresses the feature map to a fixed dimension, so that the subsequent feature fusion module can more efficiently process the features from the two branches.

3. The partial discharge detection method based on adaptive feature extraction and cross-modal feature fusion according to claim 1, characterized in that: In step (3), the feature fusion module is used in two different positions, and the two times of feature fusion respectively complement the output features of the third layer and the fourth layer from the two branches: the time domain signal branch and the additional signal branch; after each feature fusion, the fused features are spliced with the output features of the next layer of the main network branch, thereby forming a new feature vector flowing to the subsequent layers of the network; the two-level fusion modules correspond to the middle and deep features of the dual-branch network, forming a progressive cross-modal interaction: first fuse local details, then integrate global semantics, avoid information loss caused by a single fusion step, and at the same time let the fused features directly participate in the feature extraction of the subsequent branch, realizing the cycle enhancement of fusion-re-extraction.

4. The partial discharge detection method based on adaptive feature extraction and cross-modal feature fusion according to claim 3, characterized in that: In the specific architecture of the cross-modal feature fusion module, first, the outputs of the time domain and frequency domain branches are compressed to a specified output length through adaptive average pooling; then, the adaptive feature extraction method corresponding to each branch is used, a fusion matrix is constructed, and further processing is performed through the Swin Transformer network to complete cross-modal feature fusion.

5. The partial discharge detection method based on adaptive feature extraction and cross-modal feature fusion according to claim 4, characterized in that: In the time domain signal branch, first, the difference sequence between signal feature points is calculated to effectively enhance the features of mutation points or edges in the signal, and the formula is: where b is batch size, c is channel, and are the values of the i-th and j-th feature points in each channel c, respectively. Next, the difference sequence is normalized so that the range of difference values is scaled to [-1, 1] in order to calculate the attention weight later. The normalization process is as follows: where ε is set to 10-6 to prevent division by zero, and the normalized difference sequence is passed through a softmax function to enhance the differences between data, thereby further highlighting the features of significant changes. A trainable temperature control parameter T is applied to adjust the smoothness of the softmax function. The smoothness of the softmax is adjusted, and the calculation formula is: The sign function is used to maintain the sign information of the difference sequence, indicating that the direction of signal mutation does not change.

6. The partial discharge detection method based on adaptive feature extraction and cross-modal feature fusion according to claim 4, characterized in that: In the frequency domain signal branch, the signal x2 is first passed through a tanh activation function to constrain the feature values to [-1, 1]. The tanh function is a smooth nonlinear activation function that can capture complex nonlinear patterns in frequency domain features, avoiding the bias of model weights towards large values due to the magnitude difference of feature values, thereby further highlighting the overall features. After activation, the inverse cosine operation is performed to calculate the angle value φ of each channel feature point: The clamp operation ensures that the input is within the valid range of [-1, 1] to avoid errors during calculation. Subsequently, the angle difference between frequency domain feature points in each channel c is calculated. This angle difference can reflect the structured relationship between different components in the frequency domain, forming a frequency domain difference sequence. The specific angle difference calculation is as follows: GADF(i,j) = sin(φ i -φ j ) (5) where φ i and φ j are the angle values of two feature points in the signal, respectively; After completing the above steps, the softmax attention difference sequence of the time domain branch and the angle difference sequence of the frequency domain branch are fused through reorganization.

7. The partial discharge detection method based on adaptive feature extraction and cross-modal feature fusion according to claim 6, characterized in that: The time domain branch output is filled into the lower triangular part of the matrix, and the frequency domain branch output is filled into the upper triangular part. The time domain and frequency domain information are filled into different areas of the matrix, effectively reducing the influence of time sequence dependence and focusing more on the independence and complementarity between mutations in the modal; the Swin Transformer is used to further extract features from the reorganized modal fusion matrix. The window mechanism of the Swin Transformer can effectively capture the relevance of local regions, where the windows near the diagonals will contain information from both time domain and frequency domain modalities, thereby promoting interaction and fusion between modalities; while in the areas far from the two diagonals, the windows mainly focus on the information of each modality, thereby ensuring the integrity and expression of the original modal features.

8. The partial discharge detection method based on adaptive feature extraction and cross-modal feature fusion according to claim 3, characterized in that: At step (3), the model processes the signals of the two modalities through two backbone networks, and the fusion of the two feature fusion blocks contains the fusion features in the time domain and the frequency domain. The four branch outputs are maximum-pooled in the channel dimension and spliced to become the input of the multi-branch decision fusion module; the channel attention mechanism is used for dynamic weighting adjustment to enhance the expression ability of key information and suppress irrelevant information. The features adjusted by the attention mechanism are passed to the fully connected layer for final classification. Through this method, the model can effectively integrate features from different modalities and processing stages, promoting information exchange and fusion between different modalities. 9.The partial discharge detection method based on adaptive feature extraction and cross-modal feature fusion according to claim 1, characterized in that: Step (4) result prediction and evaluation: adaptive moment estimation is used as the optimizer to speed up the training process: the initial learning rate is set to 10-4; and the ReduceLROnPlateau adaptive learning rate scheduling mechanism is used to automatically decay the learning rate when the validation set performance does not improve for several epochs, to achieve a more stable convergence process; the loss function is cross-entropy loss with class weights, where the positive and negative sample weight ratio is 1:10; the batch size is fixed at 64 during training, and four-fold cross-validation is used with a maximum of 70 epochs for each fold; training is performed on the specified GPU, and the data loading process number is set to 0.

10. The partial discharge detection method based on adaptive feature extraction and cross-modal feature fusion according to claim 9, characterized in that: MCC value, recall rate and accuracy are selected, and MCC value is used as the main evaluation basis; the calculation formulas of the three evaluation indexes are shown in (6-8): Here, TP refers to the number of samples that are actually positive and correctly predicted as positive, refers to the number of samples that are actually negative and correctly predicted as negative, FP refers to the number of samples that are actually negative but incorrectly predicted as positive, and FN refers to the number of samples that are actually positive but incorrectly predicted as negative.

Images