A partial discharge detection method based on multi-modal feature and multi-model ensemble learning

Abstract

The application discloses a partial discharge detection method based on multi-modal feature and multi-model ensemble learning, relates to the technical field of power equipment detection, and comprises the following steps: collecting partial discharge data; constructing a multi-modal feature data set; building a deep learning and machine learning model; building a multi-model ensemble learning framework CMELF; and predicting and evaluating results. Compared with the prior art, the method breaks through the limitation of traditional single modal feature extraction, comprehensively captures partial discharge signal features by constructing three types of complementary multi-modal feature systems, namely time-frequency images, spectrum statistics and down-sampled spectrum, avoids detection deviation caused by information loss or noise interference under a single mode, realizes adaptive fusion of multi-scale features through the designed MSA-FuseNet network, efficiently processes high-dimensional text features through the LightGBM model, and fully utilizes the prediction confidence information of each model through the confidence weighted ensemble learning framework, thereby avoiding the defects of weak generalization ability and poor anti-interference of a single model, significantly improving the generalization ability and detection precision of the model, and improving the accuracy and reliability of partial discharge detection of power grid equipment.

Description

Technical Field

[0001] This invention relates to the field of power equipment fault detection technology, specifically to a partial discharge detection method based on multimodal features and multi-model integrated learning. Background Technology

[0002] Partial discharge (PD) is a typical indicator of the early stages of insulation aging and degradation in power equipment insulation systems. The persistent presence of PD leads to local electric field distortion, gradually forming discharge channels, which in turn cause insulation breakdown and equipment failure, posing a serious threat to the safe and reliable operation of power systems. Therefore, conducting partial discharge detection and diagnosis is of great significance for assessing the operational status of power equipment, providing early warning of faults, and predicting its lifespan.

[0003] Existing partial discharge detection methods primarily rely on single-modal feature extraction in the time, frequency, or time-frequency domains, combined with machine learning or deep learning algorithms for classification. However, partial discharge signals exhibit complex time-frequency characteristics, making it difficult for single-modal features to comprehensively represent discriminative information, thus limiting detection accuracy. Furthermore, the scarcity of fault samples in actual detection leads to class imbalance, reducing the model's learning effectiveness on fault samples. In addition, single detection models suffer from insufficient generalization ability and weak robustness to noise, making them unsuitable for complex and varied real-world detection scenarios. In contrast, multimodal feature extraction avoids the impact of noise interference on feature effectiveness under single-modal conditions and enhances the representation of unique patterns in fault signals through information complementarity between different modalities. Ensemble learning provides an efficient and accurate detection scheme. By constructing diverse base models, it fully leverages the advantages of individual models and then integrates their performance through scientific fusion strategies, correcting the inherent defects of individual models, reducing detection errors caused by model limitations, effectively addressing challenges such as noise interference in complex scenarios in practical applications, and significantly improving detection performance. Summary of the Invention

[0004] To address the technical problems of error-prone and insufficient anti-interference capabilities in partial discharge detection methods for 10kV insulated overhead conductors in complex noise environments, this invention provides a partial discharge detection method based on multimodal features and multi-model integrated learning. This method improves detection stability while ensuring high detection accuracy, enabling accurate and reliable partial discharge detection in complex power equipment environments.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0006] A partial discharge detection method based on multimodal features and multi-model ensemble learning, comprising the following steps: (1) partial discharge data acquisition; (2) construction of multimodal feature dataset; (3) deep learning and machine learning model building; (4) construction of multi-model ensemble learning framework CMELF; and (5) result prediction and evaluation.

[0007] The specific steps are as follows:

[0008] 1) Partial discharge data acquisition: The dataset used is the partial discharge signal dataset released by the ENET Center on the Kaggle platform. Data acquisition uses a single-layer inductor directly mounted on the surface of the line to measure voltage through capacitive coupling. The sampling frequency is set to 40MHz, and each signal consists of 800,000 sampling points.

[0009] 2) Construct a multimodal feature dataset: Normalize all signals to the range of [-100, 100], and determine the 500kHz-20MHz key frequency range through spectral centroid analysis. Only the key spectral signals in this range are retained for subsequent processing. In order to fully capture the discrimination information of partial discharge signals, a complementary multimodal feature dataset is constructed from three dimensions for the key spectral signals to avoid the problem of missing information in a single mode.

[0010] Specific methods include:

[0011] (1) Constructing a time-frequency graph dataset: Perform Discrete Short-Time Fourier Transform (DSTFT) on the preprocessed signal. The DSTFT formula is as follows:

[0012]

[0013] Where w[n] represents the window function. This represents the floor function, where H and L represent the window step size and window length, respectively, and m is the window index of the m-th window. This method uses a Hamming window with a window step size H = 7000 and a window length L = 8000, and generates a spectrogram as image modal data.

[0014] (2) Constructing a spectrum feature dataset: Perform Fast Fourier Transform (FFT) on the signal in the key frequency range, divide the spectrum into several segments, and extract 21 statistical indicators such as mean, standard deviation, peak factor, percentile features, etc. of each segment as shown in Table 1 to quantify the global energy distribution law of the signal; In this invention, the key spectrum signal is divided into 400 segments, and feature extraction is performed on each segment as shown in Table 1.

[0015] Table 1. Feature values ​​extracted from each key spectral signal segment.

[0016]

[0017] Where n represents the number of points in each segment, x i This represents the i-th point in the segment. Percentile characteristics refer to the values ​​of a specified percentile after sorting the magnitudes in ascending order within each segment. For example, P0 represents the minimum value, P... 50 This represents the median.

[0018] (3) Construct a spectrum downsampling dataset: Non-overlapping sliding window scanning with a length of 40 is used for key spectrum signals, and the maximum value in each window is extracted as the core feature. This method can accurately retain the frequency information reflecting the key characteristics of the partial discharge signal in the spectrum while reducing the complexity of subsequent model processing, avoiding redundant data interference and ensuring that effective discrimination information is not lost.

[0019] The feature data of the above three modalities are divided into training set, validation set and test set respectively.

[0020] 3) Deep Learning and Machine Learning Model Construction: To improve the ability to extract features from a single modality, a corresponding feature extraction model was designed. For the three constructed modality datasets, the core of the deep learning and machine learning models designed revolves around two directions: data augmentation and network design.

[0021] (1) Building a deep learning model based on time-frequency graph dataset: The class imbalance problem is solved by using image data augmentation network and MSA-FuseNet network to enhance feature extraction capability.

[0022] The DeepSMOTE-based image data augmentation network addresses the class imbalance problem caused by the scarcity of fault samples in time-frequency image data. It constructs an encoder and decoder based on a deep convolutional generative adversarial network (DCGAN), training the network only with normal samples (NPD) to learn the image feature distribution. In the low-dimensional feature space of the encoder output, fault sample feature vectors are synthesized using the SMOTE algorithm. The final layer employs the Tanh activation function to constrain the output within the range [-1, 1] for image reconstruction. To further improve image sharpness, the UnsharpMask method of the Python Image Library (PIL) is used to sharpen the reconstructed image, thereby enhancing contrast and edge detail. Its network structure is as follows: Figure 1 As shown.

[0023] A multi-scale feature fusion image classification network, MSA-FuseNet, is designed for augmented image data to fully exploit time-frequency correlations and multi-scale features in time-frequency maps. This network integrates a hybrid model of Swin Transformer, Feature Pyramid (FPN) module enhanced with Spatiotemporal Attention (CBAM), and residual blocks. It leverages the powerful feature extraction capabilities of Swin Transformer to acquire multi-level features, which are then fused through FPN to capture richer multi-scale information. Furthermore, four layers of residual blocks are used to extract deeper representations. Finally, an adaptive weighting strategy is employed to fuse the predicted outputs of the Swin Transformer and FPN branches, improving model stability. The network structure is as follows: Figure 2 As shown.

[0024] Swin Transformer is a model based on the Transformer architecture. Compared to traditional Convolutional Neural Networks (CNNs), which extract features only from local receptive fields, Swin improves the model's ability to model long-distance dependencies through a window self-attention mechanism. Secondly, this network employs a multi-scale feature extraction approach, allowing features extracted at different levels to be better adapted to downstream classification tasks.

[0025] FPN primarily addresses the multi-scale problem in object detection, typically used for multi-scale feature fusion to enhance the model's feature extraction capabilities at different scales. FPN fuses features from different scales to obtain deeper semantics. After Swin Transformer extracts feature maps with 192 and 384 output channels for stage 2 and stage 3 respectively, FPN is used for multi-scale feature fusion. FPN first uses 1×1 convolutions to unify the number of channels in each layer to 128, then uses nearest-neighbor interpolation to upsample and double the width and height of the higher-level feature maps, fusing them with the lower-level feature maps to enhance the semantic information of the lower-level features. Assuming the output feature map of stage 3 is P3 and the output of stage 2 is P2, the formula for the feature fusion mechanism is:

[0026] P'3=F.interpolate[Conv2d(P3)] (2)

[0027] P'2=Conv2d(P2)+P'3 (3)

[0028] Next, FPN optimizes the fused features by introducing CBAM, and finally uses adaptive pooling to unify the spatial size of the feature map to 8×8. Finally, feature maps of all scales are concatenated and subjected to 1x1 convolution to obtain the final output feature map, which is then passed to the Residual Block for further feature extraction.

[0029] The dual-branch adaptive weighted fusion strategy calculates weights for the outputs of the Swin Transformer and the Residual Block module separately, and then fuses their prediction results using these two weights. Specifically, the feature input of the Swin Transformer's Stage 4 output is a multilayer perceptron (MLP) consisting of a LayerNorm layer and a linear layer, and the output is a prediction score {a1, a2, a3, ..., a...}. i The Residual Block finally outputs another batch prediction score {b1,b2,b3,...,b} through the Linear layer. i}. Where i is the batch size, and each a i and b i The representation is [a i1 ,a i2 ] and [b i1 ,b i2 The vector is denoted by ]. Finally, the network performs a weighted fusion of the outputs of the two classifiers using an adaptive weighting strategy, with the weights calculated using two trainable weight parameters ω. a ω b After control and normalization via the Sigmoid activation function, the weighted fusion formula is:

[0030] c = {a1, a2, a3, ..., a i}×ω a +{b1,b2,b3,...,b i}×ω b (4)

[0031] Where c represents the final decision result obtained through additive fusion. This weighted strategy allows the model to automatically adjust its decisions based on the scores of different branches, thereby enhancing the model's accuracy and stability.

[0032] (2) Building a machine learning model based on spectral features and spectral downsampling dataset: This section uses Adaptive Synthetic Sampling (ADASYN) and a boosting tree-based classification model (LightGBM) to balance the class distribution of spectral features and spectral downsampling dataset and improve classification performance.

[0033] ADASYN addresses the class imbalance problem by adjusting the number of synthesized samples based on the distribution and synthesis difficulty of minority class samples, thereby enhancing the model's learning ability for these more difficult samples. Taking a minority class sample u as an example, the number N of synthesized samples to be generated is:

[0034]

[0035] Where G represents the total number of oversampled samples, r u The proportion of majority class samples in the neighborhood of sample u indicates the difficulty of synthesis, Δ u It is the number of majority class samples in the K nearest neighbors of sample u, where K is the total number of neighborhoods considered, and m is the total number of minority class samples.

[0036] Because a single model may exhibit varying predictive capabilities across different modality datasets, two independent LightGBM models, LGBM-SF and LGBM-DS, were trained on spectral feature datasets and spectral downsampling datasets, respectively. LightGBM is a gradient boosting decision tree (GBDT) algorithm that integrates multiple weak classifiers for learning. This algorithm employs a histogram-based decision tree to optimize the training process, efficiently transforming data distribution into discrete histogram information, thereby improving computational efficiency. LightGBM performs particularly well when processing structured data, exhibiting strong feature interpretability, fast training speed, and excellent anti-overfitting performance. It can efficiently model spectral features and feature information in spectral downsampling datasets, while balancing the computational demands of deep learning models and optimizing overall efficiency.

[0037] 4) Construction of the Multi-Model Ensemble Learning Framework CMELF: At the decision level, the confidence scores of three pre-trained models—LGBM-SF, LGBM-DS, and MSA-FuseNet—for their respective modal outputs are integrated. Then, a Support Vector Machine (SVM) is used as a meta-learner to fuse the prediction results of these models. Stratified K-Fold cross-validation is employed to divide the validation set, ensuring a balanced sample distribution. Grid search is used to optimize the hyperparameters of the SVM (regularization strength, kernel function type, and parameters, etc.) to optimally integrate the outputs of each model. This method leverages the strengths of each model in different feature spaces to enhance the final prediction, automatically balances the contribution weights of each base model, and mitigates the prediction bias of a single model.

[0038] 5) Result Prediction and Evaluation: In this experiment, the deep learning models DeepSMOTE and MSA-FuseNet used Adaptive Moment Estimation (Adam) as optimizers to accelerate the training process. MSA-FuseNet used the default momentum parameter for Adam with no weight decay; it used the ReduceLROnPlateau strategy, monitored the validation set loss (mode='min'), with a decay factor of 0.5 and a patience value of 10; the loss function used was CrossEntropy Loss, and the class weights were set to [1, 3]. The input images were uniformly adjusted to 128×128 pixels and standardized (mean [0.485, 0.456, 0.406], standard deviation [0.229, 0.224, 0.225]); the batch size was set to 64, and the number of training epochs was 100. Training was performed on a specified GPU, and the number of data loading processes was set to 0. The augmented network learning rate was set to 0.0001, the total number of training epochs was 100, the input images were uniformly adjusted to 124×124 pixels, the batch size was set to 64, the hidden layer dimension factor was set to 64, the latent space dimension was set to 1024, and the loss function was mean squared error loss (MSELoss). The LightGBM parameter settings are shown in Table 2.

[0039] Table 2 LightGBM Parameter Settings

[0040]

[0041]

[0042] 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):

[0043]

[0044] 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.

[0045] The beneficial effects of this invention are as follows:

[0046] The partial discharge detection method disclosed in this invention breaks through the limitations of traditional single-modal feature extraction compared to previous technologies. By constructing a multi-modal feature system with three complementary types—time-frequency image, spectral statistics, and downsampled spectrum—it comprehensively captures the features of partial discharge signals, avoiding detection bias caused by information loss or noise interference under single-modal conditions. The designed MSA-FuseNet network realizes adaptive fusion of multi-scale features, and the LightGBM model efficiently processes high-dimensional text features. The confidence-weighted ensemble learning framework makes full use of the prediction confidence information of each model, avoiding the defects of weak generalization ability and poor anti-interference of single models, significantly improving the generalization ability and detection accuracy of the model, and achieving accurate and reliable partial discharge detection in complex power equipment environments. Attached Figure Description

[0047] Figure 1 Image augmentation network model structure diagram.

[0048] Figure 2 This is a diagram of the MSA-FuseNet model structure.

[0049] Figure 3 This is the overall flowchart of the present invention. Detailed Implementation

[0050] The flowchart of the method of the present invention is as follows Figure 3 As shown, it includes the following steps:

[0051] (1) Acquisition of partial discharge data; (2) Construction of multimodal feature dataset; (3) Construction of deep learning and machine learning models; (4) Construction of multi-model integrated learning framework CMELF; (5) Result prediction and evaluation.

[0052] Step 1: Partial discharge acquisition; this data is acquired by collecting signals using ultrasonic sensors, ultra-high frequency sensors, etc.

[0053] Step 2: Construct a multimodal feature dataset; after filtering the key frequency bands of the collected signals, extract features from the three modes to construct a time-frequency graph dataset, a spectrum feature dataset, and a spectrum downsampling dataset; and randomly divide the training set, validation set, and test set in a ratio of 6:2:2.

[0054] Step 3: Building Deep Learning and Machine Learning Models; First, the partial discharge data is augmented using corresponding augmentation methods for different modalities; the training, validation, and test sets of the augmented time-frequency graph dataset are input into the MSA-FuseNet model built based on PyTorch; the augmented spectral feature dataset and spectral downsampling dataset are input into two different LightGBM models respectively. This step yields the pre-trained MSA-FuseNet, LGBM-SF, and LGBM-DS models.

[0055] Step 4: Building the multi-model ensemble learning framework CMELF; using the data from the three modal validation sets, input them into the corresponding pre-trained models to obtain the confidence data output by each model; use five-fold cross-validation to divide this data into training and validation sets, input it into the parameter-optimizable meta-learner SVM, and perform threshold optimization; combine the pre-trained SVM with the three pre-trained models to form the ensemble learning framework CMELF.

[0056] Step 5: Result Prediction and Evaluation; The performance of other high-performing detection algorithms designed for MSA-FuseNet on the ENET dataset was compared and analyzed. Different model combinations, different fusion strategies, and the CMELF framework were compared and analyzed. Table 3 shows the experimental results of CMELF and different model fusion strategies. Dual-model combinations such as LGBM-SF and LGBM-DS, and LGBM-DS and MSA-FuseNet, all outperformed their respective single models, indicating that model fusion can improve overall performance, classification ability, and generalization ability. CMELF showed the best performance, with all core metrics higher than single models and traditional three-model ensemble strategies. The meta-learner SVM achieves optimal decision-making by learning the confidence scores of each base model, balancing the contributions of each model. Compared to hard voting and soft voting, its recall rate reached 0.9143, significantly higher than other methods, demonstrating excellent ability in detecting minority classes. This reflects that the method enhances detection ability while maintaining high classification accuracy and overall robustness. Table 4 shows the performance comparison between the MSA-FuseNet designed in this invention and other deep learning models. The method achieves the optimal MCC value before and after data augmentation, proving that it can effectively balance classification under class imbalance, take into account both local details and global feature capture, improve classification accuracy and stability, and thus improve overall classification performance.

[0057] Table 3 Experimental results of CMELF and different model fusion strategies

[0058]

[0059]

[0060] Table 4 Performance comparison of the method of the present invention with different methods using the dataset.

[0061]

[0062] The partial discharge detection method disclosed in this invention breaks through the limitations of traditional single-modal feature extraction compared to previous technologies. By constructing a complementary multimodal feature system of three types—time-frequency image, spectral statistics, and downsampled spectrum—it comprehensively captures the features of partial discharge signals, avoiding detection bias caused by information loss or noise interference under single-modal conditions. The designed MSA-FuseNet network realizes adaptive fusion of multi-scale features, and the LightGBM model efficiently processes high-dimensional text features. The confidence-weighted ensemble learning framework makes full use of the prediction confidence information of each model, avoiding the defects of weak generalization ability and poor anti-interference of single models, significantly improving the generalization ability and detection accuracy of the model, and providing reliable technical support for partial discharge detection of power grid equipment.

[0063] 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 multimodal features and multi-model ensemble learning, characterized in that: The steps are as follows: (1) partial discharge data acquisition; (2) construction of multimodal feature dataset; (3) deep learning and machine learning model building; (4) construction of multi-model ensemble learning framework CMELF; (5) result prediction and evaluation. The specific steps are as follows: (1) Partial discharge data acquisition: Data acquisition uses a single-layer inductor directly mounted on the surface of the line to measure the voltage through capacitive coupling. The sampling frequency is set to 40MHz, and each signal consists of 800,000 sampling points. (2) Construct a multimodal feature dataset: normalize all signals to the range of [-100, 100], determine 500kHz-20MHz as the key frequency range through spectral centroid analysis, retain only the key spectral signals in this range for subsequent processing, and construct a time-frequency graph dataset, a spectral feature dataset, and a spectral downsampling dataset for the key spectral signals to avoid the problem of missing information in a single mode. The feature data of the above three modes are divided into training set, validation set, and test set respectively. (3) Building deep learning and machine learning models: Design feature extraction models. For the three modal datasets constructed, the core of building the corresponding deep learning and machine learning models revolves around two directions: data expansion and network design. (4) Construction of the multi-model ensemble learning framework CMELF: At the decision level, the confidence scores of three pre-trained models—LGBM-SF, LGBM-DS, and MSA-FuseNet—for their respective modal outputs are integrated. Then, the prediction results of these models are fused using a support vector machine as a meta-learner. The validation set is divided using hierarchical K-fold cross-validation to ensure a balanced distribution of samples. The hyperparameters of the SVM are optimized through grid search to integrate the outputs of each model in the optimal way. (5) Result prediction and evaluation: Deep learning models DeepSMOTE and MSA-FuseNet use adaptive moment estimation as the optimizer to accelerate the training process; MCC value, recall and accuracy are selected, with MCC value as the main evaluation criterion; In step (3), a deep learning model is built based on the time-frequency graph dataset: an image data augmentation network and an MSA-FuseNet network are designed to solve the class imbalance problem and enhance the feature extraction capability; The DeepSMOTE-based image data augmentation network addresses the class imbalance problem caused by the scarcity of fault samples in time-frequency image data. An encoder and decoder are constructed based on a deep convolutional generative adversarial network. The network is trained using only normal samples to learn the distribution of image features. In the low-dimensional feature space of the encoder output, the feature vector of fault samples is synthesized by the SMOTE algorithm. The final layer uses the Tanh activation function to restrict the output to the range of [-1,1] to achieve image reconstruction. To further improve image clarity, the reconstructed image is sharpened using the UnsharpMask method from the Python image library, thereby enhancing contrast and edge detail. For augmented image data, a multi-scale feature fusion image classification network, MSA-FuseNet, is designed to fully exploit the time-frequency correlations and multi-scale features in the time-frequency maps. This network integrates a hybrid model of Swing Transformer, a feature pyramid module enhanced with a spatiotemporal attention mechanism, and residual blocks. It leverages the powerful feature extraction capabilities of Swing Transformer to obtain multi-level features, which are then fused using FPN to capture richer multi-scale information. Furthermore, a four-layer residual block is used to extract deeper representations. Finally, an adaptive weighting strategy is employed to fuse the prediction outputs of the Swing Transformer and FPN branches, improving the model's stability. The Swin Transformer extracts feature maps with 192 and 384 output channels for stages 2 and 3, respectively. Multi-scale feature fusion is performed using FPN. FPN first uses 1×1 convolutions to unify the number of channels in each layer to 128, then uses nearest-neighbor interpolation to upsample and double the width and height of the high-level feature maps, fusing them with the low-level feature maps to enhance the semantic information of the low-level features. Assuming the output feature map of stage 3 is... The output of stag2 is The formula for the feature fusion mechanism is: Next, FPN optimizes the fused features by introducing CBAM, and finally uses adaptive pooling to unify the spatial size of the feature map to 8×8. Finally, feature maps of all scales are concatenated and the final output feature map is obtained by 1x1 convolution, which is then passed to ResidualBlock for further feature extraction. The dual-branch adaptive weighted fusion strategy calculates weights for the outputs of the Swin Transformer and the Residual Block module separately, and then fuses their prediction results using these two weights. Specifically, the feature input from the Stage 4 output of the Swin Transformer is a multilayer perceptron consisting of a LayerNorm layer and a linear layer, which outputs a prediction score. The Residual Block finally outputs another batch of predicted scores through the Linear layer. ;in It is the batch size, each and The representation is as follows and The vector; finally, the network weights and fuses the outputs of the two classifiers using an adaptive weighting strategy, with the weights calculated using two trainable weight parameters. , After control and normalization via the Sigmoid activation function, the weighted fusion formula is: in To obtain the final decision result through additive fusion, this weighted strategy allows the model to automatically adjust the decision based on the scores of different branches, thereby enhancing the model's accuracy and stability. In step (3), a machine learning model is built based on spectral features and spectral downsampling dataset: This part uses adaptive synthetic sampling and a boosting tree-based classification model to balance the class distribution of spectral features and spectral downsampling dataset and improve classification performance.

2. The partial discharge detection method based on multimodal features and multi-model ensemble learning according to claim 1, characterized in that: In step (2), a time-frequency graph dataset is constructed: the preprocessed signal is subjected to a Discrete Short-Time Fourier Transform (DSTFT), and the DSTFT formula is as follows: in Represents the window function. This indicates the rounding up operation. H and L represent the window step size and window length, respectively, and m is the window index of the m-th window. In this method, a Hamming window is used with a window step size H=7000 and a window length L=8000. The spectrogram is used as the image modal data during generation.

3. The partial discharge detection method based on multimodal features and multi-model ensemble learning according to claim 1, characterized in that: In step (2), a spectral feature dataset is constructed: the signal in the key frequency range is subjected to fast Fourier transform, the spectrum is divided into several segments, and the mean, standard deviation, peak factor and percentile features of each segment are extracted to quantify the global energy distribution law of the signal.

4. The partial discharge detection method based on multimodal features and multi-model ensemble learning according to claim 1, characterized in that: In step (2), a spectrum downsampling dataset is constructed: a non-overlapping sliding window with a length of 40 is used to scan the key spectrum signal, and the maximum value in each window is extracted as the core feature. This method can accurately retain the frequency information reflecting the key characteristics of the partial discharge signal in the spectrum while reducing the complexity of subsequent model processing, and ensure that effective discrimination information is not lost while avoiding redundant data interference.

5. The partial discharge detection method based on multimodal features and multi-model ensemble learning according to claim 1, characterized in that: Synthetic sampling is used to address class imbalance. By adjusting the number of synthetic samples based on the distribution and difficulty of the minority class samples, it enhances the model's learning ability for these more difficult samples. For example, the number of synthetic samples to be generated for: in This represents the total number of oversampled samples. Indicates sample The proportion of majority class samples in the neighborhood indicates the difficulty of synthesis. It is a sample of The number of majority class samples in the nearest neighbor. It is the total number of neighborhoods considered. It is the total number of minority class samples.

Images