Partial discharge unsupervised classification method and system based on cross-modal depth clustering

Abstract

The invention discloses a partial discharge unsupervised classification method and system based on cross-modal depth clustering, and the method comprises the steps: obtaining the partial discharge monitoring data of a gas insulated switchgear, carrying out the preprocessing and enhancement, converting the data into time series data, frequency domain data and image data, and carrying out the recognition of the time series data, the frequency domain data and the image data; carrying out dimension alignment on the multi-view data of different modalities; constructing a multi-view feature extraction network, extracting deep sequence features and spatial features of the aligned multi-view data, and integrating the deep sequence features and the spatial features to obtain multi-view depth features; constructing a cross-modal comparison loss function, and performing optimization training on multi-view depth features in combination with instance-level and cluster-level comparison constraints; and based on the optimized depth features, unsupervised classification of partial discharge types is realized through a clustering algorithm, and a clustering result is corrected by using a label alignment algorithm. The method does not need to depend on a prior sample label, and achieves the precise unsupervised classification of partial discharge types through multi-view feature extraction, cross-modal fusion and deep clustering.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for high-voltage power equipment, and in particular to an unsupervised classification method and system for partial discharge based on cross-modal deep clustering. Background Technology

[0002] In high-voltage power transmission systems, gas-insulated switchgear (GIS) serves as a critical infrastructure component, playing a vital role in ensuring the safety of electrical equipment and the stable operation of the system. However, with long-term operation and the influence of external environmental factors, minute physical defects may develop within the GIS, leading to localized electric field distortion and partial discharge. Partial discharge is a significant precursor to insulation degradation; if not detected and diagnosed in time, it can further cause corrosion of the insulation material and equipment breakdown, severely impacting the reliability of the power grid. Therefore, developing efficient GIS fault detection and diagnosis methods is of great importance for achieving precise maintenance and preventing major accidents.

[0003] In partial discharge (PD) identification research, early methods primarily relied on feature extraction techniques. For example, wavelet scaling maps were used to visualize the frequency domain features of PD signals; time-series-based topological data analysis was employed to extract topological invariants from the signals; and other studies attempted to introduce discriminative feature learning modules into deep learning frameworks to optimize the inter-class differentiation capabilities of phase-resolved PD maps. While these methods improved the model's accuracy in identifying PD types, their implementation heavily depended on expert experience and domain knowledge, resulting in complex feature extraction processes and limitations in practical engineering applications.

[0004] In recent years, with the development of deep learning technology, researchers have gradually adopted deep networks to replace traditional manual feature extraction methods to simplify the algorithm process and improve performance. For example, the impact of sampling windows on classification results was explored based on deep belief networks, achieving high recognition accuracy; the MixNet model was proposed and combined with generative adversarial networks for data augmentation, significantly improving the robustness of pattern recognition; other studies have constructed multi-task learning networks to simultaneously complete PD type recognition and state assessment. However, in real-world engineering scenarios, the occurrence of partial discharges is random and sparse, leading to a scarcity of high-quality labeled samples, which poses challenges to supervised learning methods due to overfitting and insufficient generalization ability.

[0005] To address the problem of scarce labeled data, some researchers have explored few-shot learning and meta-learning methods. For example, a Siamese network structure incorporating an attention mechanism employs model-independent meta-learning for parameter optimization; and a few-shot learning method based on approximate space models the correlation between attributes to enhance the model's discriminative ability. However, these methods remain sensitive to sample quality and are susceptible to noise and neighborhood bias, limiting their adaptability in complex field environments.

[0006] Meanwhile, the volume of unlabeled monitoring data generated during the operation of high-voltage equipment is enormous, but its informational value has not been fully explored using traditional methods. Although some studies have attempted to introduce unsupervised clustering methods into PD analysis, such as proposing a density-based nonparametric clustering algorithm that constructs neighborhood sets through multivariate kernel density estimation to identify discharge patterns and using clustering methods to achieve effective grouping of PD features, these methods often lack the ability to model the deep structure of the data and multimodal correlations, making it difficult to adapt to the complex and ever-changing distribution of high-dimensional PD data. Furthermore, existing methods generally lack the ability to learn collaborative features from multimodal data, failing to fully leverage the complementary advantages of cross-modal information, resulting in insufficient classification accuracy and generalization ability. Summary of the Invention

[0007] This invention addresses the limitations of existing technologies in feature extraction, reliance on labeled data, and insufficient utilization of multimodal information. It proposes an unsupervised classification method and system for partial discharge based on cross-modal deep clustering. This method does not rely on prior sample labels and achieves accurate unsupervised classification of partial discharge types through multi-view feature extraction, cross-modal fusion, and deep clustering.

[0008] To achieve the above objectives, the present invention provides the following solution: An unsupervised classification method for partial discharge based on cross-modal deep clustering, comprising the following steps: S1, acquire partial discharge monitoring data of gas-insulated switchgear, preprocess and enhance the monitoring data to obtain multimodal raw dataset and enhanced dataset; S2, the original multimodal dataset and the augmented dataset are converted into three view formats: time series data, frequency domain data and image data. The dimensions of the multimodal view data are aligned to obtain the aligned multimodal view data. S3, construct a multi-view feature extraction network to extract deep sequence features and spatial features of the aligned multi-view data, and integrate them to obtain multi-view depth features; S4, construct a cross-modal contrast loss function, and optimize the training of the multi-view depth features by combining instance-level and cluster-level contrast constraints; S5, based on the optimized deep features, achieves unsupervised classification of partial discharge types through clustering algorithm, uses label alignment algorithm to correct clustering results, and outputs final classification results.

[0009] Further, in step S1, the partial discharge monitoring data includes phase-resolved partial discharge (PRPD) data and phase-resolved pulse sequence (PRPS) data, the preprocessing includes noise filtering, and the enhancement includes data augmentation operations such as random perturbation, flipping, and scaling.

[0010] Further, in step S2, the time series data This includes PRPD and PRPS organized according to each power frequency period; the frequency domain data is obtained by applying a fast Fourier transform to the time series data, satisfying the formula... ,in, For time series data, The image data is in the frequency domain; It is obtained by converting the original data matrix into a grayscale image.

[0011] Furthermore, in step S2, the specific formula for dimension alignment is: , , In the formula, Represents the original multimodal dataset and augmented datasets Time series or frequency domain data in the middle, Represents the original multimodal dataset and augmented datasets In the image data, Linear represents a fully connected layer, Upsample represents an upsampling operation, and Cov represents a two-dimensional convolution operation.

[0012] Further, in step S3, the multi-view feature extraction network includes a shared sequence encoder and a shared image encoder. The shared sequence encoder is a three-layer stacked bidirectional long short-term memory network (BiLSTM), with 512 hidden units in each layer. The feature extraction formula is: in, This represents a three-layer BiLSTM operation; The shared image encoder is an 18-layer ResNet-18 residual network, and the feature extraction formula is: For deep feature extraction of image data, the ResNet network extracts deep spatial features layer by layer through multiple convolutional layers and residual modules. The residual modules alleviate the gradient vanishing problem by introducing shortcut connections. The multi-view depth features extracted by BiLSTM and ResNet-18 networks are uniformly represented as .

[0013] Further, in step S3, the formula for mapping the multi-view depth features to the instance-level and cluster-level feature spaces is: , , in, For a two-layer perceptron with a mapping dimension equal to the number of samples, normalize represents the normalization operation; For a two-layer perceptron with the mapping dimension aligned to the number of partial discharge types, softmax represents the softmax normalization operation.

[0014] Furthermore, in step S4, the instance-level contrast constraint adopts the InfoNCE loss function, the formula of which is: in, For instance-level temperature parameters, s(·) represents the cosine similarity calculation function between two features. It is a set The elements represent instance-level features of sample i. Enhance instance-level features of sample i's data; and The instance-level contrastive loss formula is expressed as: Where n represents the total number of samples in the dataset, , These represent the instance-level contrast loss for the original data and the augmented data, respectively. The calculation formula and same.

[0015] Further, in step S4, the formula for the cluster-level comparison constraint is: in, It is a set The elements represent the cluster-level features of sample i. The temperature parameter represents the cluster level, and s(·) represents the cosine similarity calculation function between two features. yes The augmented data is represented as the cluster-level features of the augmented data for sample i; The cross-entropy loss is expressed as: in, , k is the number of PD types; The cluster-level loss, which includes contrast constraints and cross-entropy constraints, is defined as follows: in, , These represent cluster-level comparison constraints for the original data and the augmented data, respectively.

[0016] Further, in step S4, the cross-modal contrast loss function is: Where α is the weighting coefficient. This represents the loss function for single-view cross-modal contrastive learning. This represents the cross-modal, cross-view contrastive learning loss function; The loss function of the single-view cross-modal contrastive learning Represented as: in, This indicates aligned PRPS data. Represents aligned PRPD data; The cross-modal, cross-view contrastive learning loss function for: in, These represent aligned PRPS data in time series (t), frequency domain (f), and image (p) formats, respectively. This indicates aligned PRPD data in the same format.

[0017] This invention also provides an unsupervised partial discharge classification system based on cross-modal deep clustering, applied to the above-described unsupervised partial discharge classification method based on cross-modal deep clustering, comprising: Data acquisition and preprocessing module: used to acquire partial discharge monitoring data of gas-insulated switchgear, preprocess and enhance the monitoring data to obtain multimodal raw dataset and enhanced dataset; Multi-view transformation and alignment module: used to convert the original multimodal dataset and augmented dataset into three view forms: time series data, frequency domain data and image data, and to perform dimensional alignment on the multi-view data of different modalities to obtain aligned multi-view data; Multi-view feature extraction module: used to construct a multi-view feature extraction network, extract deep sequence features and spatial features of the aligned multi-view data, and integrate them to obtain multi-view depth features; Loss function and optimization module: used to construct a cross-modal contrastive loss function, and combine instance-level and cluster-level contrastive constraints to optimize and train the multi-view depth features; Clustering and Classification Module: Based on optimized deep features, this module performs unsupervised classification of partial discharge types using a clustering algorithm, corrects the clustering results using a label alignment algorithm, and outputs the final classification result.

[0018] According to specific embodiments provided by the present invention, the unsupervised classification method and system for partial discharge based on cross-modal deep clustering disclosed in the present invention have the following technical effects: (1) A multi-view feature extraction strategy is adopted, combining ResNet-18 and BiLSTM networks to extract deep features from three views: time series, frequency domain and image. This fully explores the spatiotemporal and frequency features of partial discharge data, improves feature discriminability, and overcomes the problem of limited feature extraction capability of traditional methods.

[0019] (2) A cross-modal multi-view deep clustering model is proposed, which does not rely on labeled samples and uses unlabeled monitoring data to achieve unsupervised classification, reducing the dependence on data labeling and solving the problem of scarce labeled data in real-world scenarios.

[0020] (3) Design intra-modal and cross-modal contrast loss functions, combined with instance-level and cluster-level constraints, to promote the learning of intra-modal invariant features and cross-modal similarity features, give full play to the complementary advantages of multimodal data, and improve the generalization ability and classification accuracy of the model.

[0021] (4) By using optimization strategies such as dimension alignment, feature fusion and label alignment, the classification stability in complex partial discharge data scenarios is ensured. Experiments show that the overall classification accuracy reaches 98.32%, and all performance indicators are better than existing methods, which has high engineering application value. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the unsupervised classification method for partial discharge based on cross-modal deep clustering of the present invention; Figure 2 This is a schematic diagram of a fault generation and signal acquisition system for a GIS system according to an embodiment of the present invention; wherein, 1, transformer; 2, oscilloscope; 3, fault generation component; 4, ultra-high frequency sensor; 5, gas-insulated switchgear; 6, potential insulator; 7, high-voltage conductor; Figure 3 This is a schematic diagram of the original data feature distribution in an embodiment of the present invention, where A represents PRPD and B represents PRPS; Figure 4 This is a schematic diagram of the feature representation distribution of the CMDC model according to an embodiment of the present invention; Figure 5This is a schematic diagram of the partial discharge fault classification confusion matrix according to an embodiment of the present invention; Figure 6 This is a schematic diagram comparing the diagnostic and classification effects of different technical solutions in embodiments of the present invention. Detailed Implementation

[0024] The embodiments of the present invention are described in detail below. These embodiments are intended to explain the present invention and should not be construed as limiting the present invention. Where specific techniques or conditions are not specified in the embodiments, they are performed according to the techniques or conditions described in the literature in the art or according to the product instructions. Materials or instruments whose manufacturers are not specified are all conventional products that can be obtained commercially.

[0025] This invention provides an unsupervised classification method and system for partial discharge based on cross-modal deep clustering, which mainly achieves the following innovations: 1. A multi-view feature extraction method using ResNet-18 and BiLSTM was developed to capture deep spatial and sequence representations of partial discharge patterns. This method effectively extracts key information even with limited datasets by utilizing time series, frequency domain, and image data.

[0026] 2. A multimodal feature fusion method is proposed. First, consistent deep representations are extracted from various modalities using raw and augmented data. Then, a cross-modal feature interaction mechanism is used to fuse multi-view information. By learning similarity features and integrating complementary information from different data sources, the feature extraction capability of the model is enhanced.

[0027] 3. An innovative cross-modal multi-view deep clustering (CMDC) model was designed. This model does not rely on prior sample label knowledge and significantly improves the clustering accuracy of PD types by promoting intramodal invariant feature learning and cross-modal similarity feature learning.

[0028] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] Example 1 like Figure 1 As shown, the unsupervised partial discharge classification method based on cross-modal deep clustering provided in this embodiment of the invention includes the following steps: S1, acquire partial discharge monitoring data of gas-insulated switchgear, preprocess and enhance the monitoring data to obtain multimodal raw dataset and enhanced dataset; S2, the original multimodal dataset and the augmented dataset are converted into three view formats: time series data, frequency domain data and image data. The dimensions of the multimodal view data are aligned to obtain the aligned multimodal view data. S3, construct a multi-view feature extraction network to extract deep sequence features and spatial features of the aligned multi-view data, and integrate them to obtain multi-view depth features; S4, construct a cross-modal contrast loss function, and optimize the training of the multi-view depth features by combining instance-level and cluster-level contrast constraints; S5, based on the optimized deep features, achieves unsupervised classification of partial discharge types through clustering algorithm, uses label alignment algorithm to correct clustering results, and outputs final classification results.

[0030] Specifically, the steps of the unsupervised partial discharge classification method based on cross-modal deep clustering provided in this embodiment of the invention are described below: 1. Data Collection In step S1, the partial discharge monitoring data includes phase-resolved partial discharge (PRPD) data and phase-resolved pulse sequence (PRPS) data. The preprocessing includes operations such as noise filtering, and the enhancement includes operations such as random perturbation, flipping, and scaling data augmentation.

[0031] 2. Multi-view data acquisition The partial discharge dataset collected by GIS monitoring can be represented as follows: To extract depth features from multiple views, X is used in three forms: time series, frequency domain, and image data. The augmented dataset is represented as... .

[0032] In step S2, the time series data This includes PRPD and PRPS organized according to each power frequency period; the frequency domain data is obtained by applying Fast Fourier Transform (FFT) to the time series data, satisfying the formula... ,in, For time series data, The image data is in the frequency domain; It is obtained by converting the original data matrix into a grayscale image.

[0033] Because PRPS and PRPD have different dimensions, dimension alignment is crucial. This operation maps raw data with different dimensions to a common dimension space. The specific formula for dimension alignment is: , , In the formula, Represents the original multimodal dataset and augmented datasets Time series or frequency domain data in the middle, Represents the original multimodal dataset and augmented datasets In the image data, Linear represents a fully connected layer, Upsample represents an upsampling operation, and Cov represents a two-dimensional convolution operation.

[0034] 3. Multi-view feature extraction In step S3, the multi-view feature extraction network includes a shared sequence encoder and a shared image encoder. The shared sequence encoder is a three-layer stacked bidirectional long short-term memory network BiLSTM, and the shared image encoder is an 18-layer residual network ResNet-18.

[0035] For deep feature extraction of sequence data, LSTM can effectively capture time dependencies. When processing frequency domain data, it can also identify the frequency components of the signal and their periodicity. Therefore, a bidirectional LSTM (BiLSTM) network was designed, consisting of three stacked BiLSTM layers, each with 512 hidden units. Since the frequency domain data originates from the Fourier transform of time series data, their data lengths are consistent. Therefore, a shared BiLSTM network is used. Features are obtained from the original and augmented data. Described as: in, This represents a three-layer BiLSTM operation; For deep feature extraction from image data, the ResNet network extracts deep spatial features layer by layer through multiple convolutional layers and residual modules. The residual modules alleviate the vanishing gradient problem by introducing shortcut connections, ensuring that complex features can still be effectively learned in deep networks. It can be represented as: The multi-view depth features extracted by BiLSTM and ResNet-18 networks are uniformly represented as .

[0036] To enhance clustering performance, this invention employs a joint optimization strategy, integrating instance-level and cluster-level clustering. Instance-level clustering focuses on refining the extraction of patterns from individual samples, while cluster-level clustering promotes the aggregation of similar samples. This dual optimization method effectively improves the learning of pattern features in the dataset.

[0037] The extracted deep features H are mapped to a fixed-dimensional feature space Z. Its representation at the instance and cluster levels is defined as follows: , , in, For a two-layer perceptron with a mapping dimension equal to the number of samples, normalize represents the normalization operation; For a two-layer perceptron with the mapping dimension aligned to the number of partial discharge types, softmax represents the softmax normalization operation.

[0038] 4. Comparison Constraints In unsupervised learning, the lack of labels poses a challenge to direct model optimization, making the formulation of an effective loss function crucial. A significant advantage of contrastive learning is that it allows the model to learn representations that highlight the similarities and differences between samples. In the field of self-supervised learning, the InfoNCE loss function has demonstrated superior performance. Therefore, this invention employs InfoNCE to construct a contrastive loss function.

[0039] For instance-level contrastive learning, it is applied to each reconstructed feature. The InfoNCE loss is defined as: in, For instance-level temperature parameters, s(·) represents the cosine similarity calculation function between two features. It is a set The elements represent instance-level features of sample i. Enhance instance-level features of sample i's data; and The instance-level contrastive loss formula is expressed as: Where n represents the total number of samples in the dataset, , These represent the instance-level contrast loss for the original data and the augmented data, respectively. The calculation formula and same.

[0040] Similarly, in step S4, the formula for the cluster-level comparison constraint is: in, It is a set The elements represent the cluster-level features of sample i. The temperature parameter represents the cluster level, and s(·) represents the cosine similarity calculation function between two features. yes The augmented data is represented as the cluster-level features of the augmented data for sample i; Furthermore, cross-entropy loss can increase the distance between different types of features, thereby enhancing the model's generalization ability. Cross-entropy loss can be expressed as: in, , k is the number of PD types; The cluster-level loss, which includes contrast constraints and cross-entropy constraints, is defined as follows: in, , These represent cluster-level comparison constraints for the original data and the augmented data, respectively.

[0041] PRPD and PRPS represent two different modalities of PD data. Traditional contrastive learning methods often lack intermodal information interaction, which may hinder comprehensive pattern feature learning. To address this limitation, this invention proposes a segmented cross-modal information interaction mechanism. Initially, invariant features within each modality are learned independently by comparing the original and augmented data. This process effectively reduces interference from irrelevant features, enabling the model to focus on meaningful pattern features. Subsequently, cross-modal contrastive learning is applied to the original data to further explore consistent features across modalities.

[0042] For unimodal data, such as PRPS or PRPD, multi-view contrastive learning is employed between the raw and augmented data to identify key invariant features within the unimodality. The intramodal loss is expressed as: in, This represents the alignment of the three views of a PRPS or PRPD.

[0043] In step S4, the cross-modal contrast loss function is: Where α is the weighting coefficient, This represents the loss function for single-view cross-modal contrastive learning. This represents the cross-modal, cross-view contrastive learning loss function; The loss function of the single-view cross-modal contrastive learning Represented as: in, This indicates aligned PRPS data. Represents aligned PRPD data; The cross-modal, cross-view contrastive learning loss function for: in, These represent aligned PRPS data in time series (t), frequency domain (f), and image (p) formats, respectively. This indicates aligned PRPD data in the same format.

[0044] 5. Cross-modal multi-view deep clustering (CMDC) Based on detailed deep feature extraction and loss function construction, this invention designs a cross-modal multi-view deep clustering method. First, PRPD and PRPS are preprocessed and transformed into time-series and image formats. For time-series data, FFT is applied to obtain the corresponding frequency domain data. To achieve efficient cross-modal contrastive learning, a temporal and spatial alignment strategy is adopted for multi-view data from different modalities, laying a solid foundation for subsequent joint multi-view learning. Next, a single-modal deep learning model is developed for PRPD and PRPS. Time-series and frequency domain data are input to a shared sequence encoder to learn global patterns and time-frequency features, while image data is processed through a shared image encoder to extract spatial features. Multi-view contrastive learning within each modality captures consistent depth representations from the raw and augmented data.

[0045] To further explore intermodal relationships, a cross-modal interaction mechanism was introduced. This facilitated multi-view information exchange between PRPD and PRPS, enabling the model to learn more discriminative features from complementary information, thereby improving its ability to identify PD types. After completing multi-view cross-modal learning, deep features from various modalities and perspectives were obtained. Subsequently, t-SNE and K-means algorithms were applied to effectively cluster these PD patterns, accurately distinguishing different types. Since K-means labels are arbitrary, the Hungarian algorithm was used for label alignment to evaluate the results. This framework achieved efficient clustering and significantly improved clustering accuracy.

[0046] To validate the proposed cross-modal deep clustering model Figure 2 This paper demonstrates the use of UHF sensors to acquire PRPD and PRPS data from GIS equipment. PRPD represents the two-dimensional phase distribution by capturing the relationship between the phase, number of discharges, and discharge amplitude of partial discharge pulse signals within a single power frequency cycle (typically 50 Hz). PRPS, on the other hand, records discharge pulse signals over multiple power frequency cycles, thus providing a three-dimensional representation of time, phase, and discharge quantity. The final acquired data samples are summarized in Table 1.

[0047] Table 1. Types and number of partial discharges collected Discharge type quantity PRPD PRPS Label normal 562 √ √ 0 interference 573 √ √ 1 Solid insulation discharge 570 √ √ 2 Tip corona discharge 581 √ √ 3 Metal particle discharge 584 √ √ 4 Floating potential discharge 562 √ √ 5 During the experiment, the computer configuration included an NVIDIA GeForce GTX 3060 Ti GPU, an Intel Core i5-12600KF processor, and 16 GB of RAM. The model was run in a Python 3.8.19 environment using Torch 1.10.2. The optimizer used for different architectures was Adam, with a learning rate of 0.001 and a batch size of 256. The temperature coefficients in the instance-level and cluster-level contrastive losses were set to 1.0 and 0.5, respectively. After feature representation and clustering using CMDC, clustering results for the PD type were obtained. To effectively evaluate the model's performance, several performance metrics related to classification and clustering were introduced: (1) Accuracy (ACC): The proportion of correctly classified samples out of the total number of samples.

[0048] (2) Precision (PR): The proportion of truly positive samples among those predicted as positive by the model.

[0049] (3) F1 score: This important evaluation metric combines precision and recall and can comprehensively measure the classification ability of the model.

[0050] (4) Normalized mutual information (NMI): This indicator is based on the principle of information theory and quantifies the amount of information shared between the clustering results and the real labels.

[0051] (5) Adjusted Random Number Index (ARI): This index assesses the consistency between clustering results and true labels by evaluating the consistency of all possible sample pairs.

[0052] After using the CMDC model to perform unsupervised classification of different types of PD data, the original data was analyzed to verify the model's ability to accurately extract key pattern features. For example... Figure 3 As shown, t-SNE is applied to visualize the original high-dimensional PRPS and PRPD. It can be observed that in the PRPS and PRPD visualizations, most categories exhibit significant overlap and lack clear boundaries, making effective category separation difficult. In contrast, Figure 4 The clustering results obtained from the features learned by the CMDC model are presented. Different types of PD data are clearly separated with well-defined decision boundaries. This improvement can be attributed to the model's multi-stage feature learning, which makes it possible to extract highly discriminative representations. Therefore, even using a relatively simple t-SNE configuration, the streamlined features exhibit clear class separability, further demonstrating the effectiveness of the proposed method.

[0053] Based on the deep feature representation extracted from multimodal data by the model, the diagnostic confusion matrix for each partial discharge fault category is as follows: Figure 5As shown, the model maintains a high diagnostic accuracy for different fault types. The lowest diagnostic accuracy is for metal particle discharge, but it still reaches 96.99%. The diagnostic accuracy for other categories is higher than 97%. The model can effectively perform high-performance diagnostic classification through feature learning between different modes.

[0054] To demonstrate the effectiveness of the technical solution proposed in this invention, the performance of several unsupervised and supervised methods was compared. Due to the inherent limitations of unsupervised learning caused by the lack of labels, CVCL and CDCC, which employ unsupervised methods, performed worse than MixNet and DBN models in the final performance. Although CVCL uses multi-view data as input, the lack of an effective cross-view interaction mechanism leads to poor performance. CVCL's accuracy is only 0.7389, with NMI and ARI values ​​of 0.6394 and 0.5458, respectively. As a time-frequency data analysis method, the CDCC model showed higher classification performance, with NMI and ARI scores of 0.6683 and 0.6065, respectively, and other indicators around 0.80. This performance improvement may be attributed to the inclusion of frequency domain data, which significantly enhances the feature representation of partial discharge modes.

[0055] In contrast, MixNet and DBN demonstrated significant training performance, achieving an accuracy exceeding 0.9 and showing significant improvements across all other metrics. This improvement is primarily due to the targeted training on labeled data, which facilitates the extraction of key features. However, in industrial scenarios, acquiring data labels can be extremely complex and challenging. Utilizing large amounts of unprocessed monitoring data is a key area of ​​recent development.

[0056] The CMDC model overcomes the limitations of current partial discharge data analysis methods, skillfully extracting complex and deep features from multiple modes, thus accurately distinguishing between normal and abnormal signals. It can also accurately differentiate various anomalous partial discharge types, achieving an overall classification accuracy of 0.9832, an NMI of 0.9477, and other indicators exceeding 0.96. This capability enhances the accurate classification of partial discharge modes under real-world conditions. (See Table 2 and...) Figure 6 The results clearly show that, compared with existing research methods, the CMDC model demonstrates higher application potential in terms of performance indicators.

[0057] Table 2 Diagnostic performance of different technical solutions Solution ACC Precision F1 NMI ARI CVCL 0.7389 0.7738 0.7507 0.6394 0.5458 CDCC 0.7861 0.8037 0.7929 0.6683 0.6065 MixNet 0.9000 0.9097 0.9018 0.8147 0.7963 DBN 0.9506 0.9511 0.9503 0.8799 0.8892 CMDC 0.9832 0.9834 0.9832 0.9477 0.9600 Example 2 This invention also provides an unsupervised partial discharge classification system based on cross-modal deep clustering, applied to the above-described unsupervised partial discharge classification method based on cross-modal deep clustering, comprising: Data acquisition and preprocessing module: used to acquire partial discharge monitoring data of gas-insulated switchgear, preprocess and enhance the monitoring data to obtain multimodal raw dataset and enhanced dataset; Multi-view transformation and alignment module: used to convert the original multimodal dataset and augmented dataset into three view forms: time series data, frequency domain data and image data, and to perform dimensional alignment on the multi-view data of different modalities to obtain aligned multi-view data; Multi-view feature extraction module: used to construct a multi-view feature extraction network, extract deep sequence features and spatial features of the aligned multi-view data, and integrate them to obtain multi-view depth features; Loss function and optimization module: used to construct a cross-modal contrastive loss function, and combine instance-level and cluster-level contrastive constraints to optimize and train the multi-view depth features; Clustering and Classification Module: Based on optimized deep features, this module performs unsupervised classification of partial discharge types using a clustering algorithm, corrects the clustering results using a label alignment algorithm, and outputs the final classification result. Example 3 The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the unsupervised classification method for partial discharge based on cross-modal deep clustering as described above.

[0058] Matters not covered in this invention are common knowledge.

[0059] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0060] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. An unsupervised classification method for partial discharge based on cross-modal deep clustering, characterized in that, Includes the following steps: S1, acquire partial discharge monitoring data of gas-insulated switchgear, preprocess and enhance the monitoring data to obtain multimodal raw dataset and enhanced dataset; S2, the original multimodal dataset and the augmented dataset are converted into three view formats: time series data, frequency domain data and image data. The dimensions of the multimodal view data are aligned to obtain the aligned multimodal view data. S3, construct a multi-view feature extraction network to extract deep sequence features and spatial features of the aligned multi-view data, and integrate them to obtain multi-view depth features; S4, construct a cross-modal contrast loss function, and optimize the training of the multi-view depth features by combining instance-level and cluster-level contrast constraints; S5, based on the optimized deep features, achieves unsupervised classification of partial discharge types through clustering algorithm, uses label alignment algorithm to correct clustering results, and outputs final classification results.

2. The unsupervised classification method for partial discharge based on cross-modal deep clustering according to claim 1, characterized in that, In step S1, the partial discharge monitoring data includes phase-resolved partial discharge (PRPD) data and phase-resolved pulse sequence (PRPS) data. The preprocessing includes noise filtering operations, and the enhancement includes data augmentation operations such as random perturbation, flipping, and scaling.

3. The unsupervised classification method for partial discharge based on cross-modal deep clustering according to claim 1, characterized in that, In step S2, the time series data This includes PRPD and PRPS organized according to each power frequency period; the frequency domain data is obtained by applying a fast Fourier transform to the time series data, satisfying the formula... ,in, For time series data, The image data is in the frequency domain; It is obtained by converting the original data matrix into a grayscale image.

4. The unsupervised classification method for partial discharge based on cross-modal deep clustering according to claim 1, characterized in that, In step S2, the specific formula for dimension alignment is: ， ， In the formula, Represents the original multimodal dataset and augmented datasets Time series or frequency domain data in the middle, Represents the original multimodal dataset and augmented datasets In the image data, Linear represents a fully connected layer, Upsample represents an upsampling operation, and Cov represents a two-dimensional convolution operation.

5. The unsupervised classification method for partial discharge based on cross-modal deep clustering according to claim 4, characterized in that, In step S3, the multi-view feature extraction network includes a shared sequence encoder and a shared image encoder. The shared sequence encoder is a three-layer stacked bidirectional long short-term memory network (BiLSTM), with 512 hidden units in each layer. The feature extraction formula is: in, This represents a three-layer BiLSTM operation; The shared image encoder is an 18-layer ResNet-18 residual network, and the feature extraction formula is: For deep feature extraction of image data, the ResNet network extracts deep spatial features layer by layer through multiple convolutional layers and residual modules. The residual modules alleviate the gradient vanishing problem by introducing shortcut connections. The multi-view depth features extracted by BiLSTM and ResNet-18 networks are uniformly represented as .

6. The unsupervised classification method for partial discharge based on cross-modal deep clustering according to claim 5, characterized in that, In step S3, the formula for mapping the multi-view depth features to the instance-level and cluster-level feature spaces is as follows: ， ， in, For a two-layer perceptron with a mapping dimension equal to the number of samples, normalize represents the normalization operation; For a two-layer perceptron with the mapping dimension aligned to the number of partial discharge types, softmax represents the softmax normalization operation.

7. The unsupervised classification method for partial discharge based on cross-modal deep clustering according to claim 6, characterized in that, In step S4, the instance-level contrast constraint adopts the InfoNCE loss function, the formula of which is: in, For instance-level temperature parameters, s(·) represents the cosine similarity calculation function between two features. It is a set The elements represent instance-level features of sample i. Enhance instance-level features of sample i's data; and The instance-level contrastive loss formula is expressed as: Where n represents the total number of samples in the dataset, , These represent the instance-level contrast loss for the original data and the augmented data, respectively. The calculation formula and same.

8. The unsupervised classification method for partial discharge based on cross-modal deep clustering according to claim 7, characterized in that, In step S4, the formula for the cluster-level comparison constraint is: in, It is a set The elements represent the cluster-level features of sample i. The temperature parameter represents the cluster level, and s(·) represents the cosine similarity calculation function between two features. yes The augmented data is represented as the cluster-level features of the augmented data for sample i; The cross-entropy loss is expressed as: in, , k is the number of PD types; The cluster-level loss, which includes contrast constraints and cross-entropy constraints, is defined as follows: in, , These represent cluster-level comparison constraints for the original data and the augmented data, respectively.

9. The unsupervised classification method for partial discharge based on cross-modal deep clustering according to claim 8, characterized in that, In step S4, the cross-modal contrast loss function is: Where α is the weighting coefficient. This represents the loss function for single-view cross-modal contrastive learning. This represents the cross-modal, cross-view contrastive learning loss function; The loss function of the single-view cross-modal contrastive learning Represented as: in, This indicates aligned PRPS data. Represents aligned PRPD data; The cross-modal, cross-view contrastive learning loss function for: in, These represent aligned PRPS data in time series (t), frequency domain (f), and image (p) formats, respectively. This indicates aligned PRPD data in the same format.

10. A partial discharge unsupervised classification system based on cross-modal deep clustering, characterized in that, An application to the unsupervised partial discharge classification method based on cross-modal deep clustering as described in any one of claims 1-9, comprising: Data acquisition and preprocessing module: used to acquire partial discharge monitoring data of gas-insulated switchgear, preprocess and enhance the monitoring data to obtain multimodal raw dataset and enhanced dataset; Multi-view transformation and alignment module: used to convert the original multimodal dataset and augmented dataset into three view forms: time series data, frequency domain data and image data, and to perform dimensional alignment on the multi-view data of different modalities to obtain aligned multi-view data; Multi-view feature extraction module: used to construct a multi-view feature extraction network, extract deep sequence features and spatial features of the aligned multi-view data, and integrate them to obtain multi-view depth features; Loss function and optimization module: used to construct a cross-modal contrastive loss function, and combine instance-level and cluster-level contrastive constraints to optimize and train the multi-view depth features; Clustering and Classification Module: Based on optimized deep features, this module performs unsupervised classification of partial discharge types using a clustering algorithm, corrects the clustering results using a label alignment algorithm, and outputs the final classification result.

Images