Intelligent partial discharge on-line monitoring and fault diagnosis system based on multi-sensor fusion

Through multi-sensor fusion and lightweight capsule network combined with Hyperband algorithm to optimize hyperparameters, the problems of insufficient feature extraction and complex hyperparameter optimization in local discharge monitoring are solved, high-precision and real-time fault diagnosis are achieved, and the operation and maintenance efficiency of power equipment and system stability are improved.

CN120493064AInactive Publication Date: 2025-08-15CHONGKE INTELLIGENT TECH (ZHEJIANG) CO LTD
View PDF 0 Cites 11 Cited by

Patent Information

Application Number
CN202510565681.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-04-30
Publication Date
2025-08-15
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing multi-sensor fusion system has problems such as insufficient feature extraction, insufficient sensor signal weight adjustment, and high complexity of hyperparameter optimization in local discharge monitoring, resulting in low accuracy and reliability of local discharge fault diagnosis.

Method used

Multi-sensors are used to synchronize local discharge signals, combine wavelet noise reduction and time-frequency feature extraction, and build a weighted feature matrix, use a lightweight capsule network to perform local pulse mode feature extraction, and optimize hyperparameters through the Hyperband algorithm to achieve efficient local discharge type classification.

Benefits of technology

It improves the accuracy and real-time nature of partial discharge fault diagnosis, reduces the workload of manual inspection, optimizes operation and maintenance management, and improves the stability and safety of the power system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120493064A_ABST
    Figure CN120493064A_ABST
Patent Text Reader

Abstract

The invention discloses an intelligent partial discharge online monitoring and fault diagnosis system based on multi-sensor fusion, and the system comprises a multi-sensor collection module which is used for synchronously collecting data in a partial discharge process; the data preprocessing module is used for preprocessing the partial discharge signal data; the feature fusion module is used for constructing a fusion weighted feature matrix; the feature dimension reduction module is used for constructing a fusion feature matrix after dimension reduction; the partial discharge classification model training module is used for constructing a partial discharge type classification model by adopting a lightweight capsule network; the hyper-parameter search optimization module is used for optimizing the partial discharge type classification model; the classification model deployment module is used for deploying the optimized partial discharge type classification model; and the online reasoning and fault diagnosis module is used for receiving data in real time, generating a fault alarm signal and recording and returning fault event information. According to the invention, a real-time partial discharge on-line monitoring and intelligent fault diagnosis scheme is provided for equipment.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment fault diagnosis, and in particular to an intelligent partial discharge online monitoring and fault diagnosis system based on multi-sensor fusion. Background Art

[0002] As the scale of power systems continues to expand and their intelligence level increases, partial discharge (PD), an electrical defect that occurs inside or on the surface of electrical equipment, has become a common type of fault in the operation of power equipment. The partial discharge phenomenon not only reduces the working efficiency of power equipment, but may also cause more serious equipment damage or even system failures. Therefore, the monitoring and fault diagnosis of partial discharge are crucial to the safe and stable operation of power systems. Traditional partial discharge detection methods mostly rely on manual inspections, regular inspections, and partial discharge test equipment. These methods usually require a lot of manual intervention, which not only has limited detection accuracy, but is also time-consuming and labor-intensive, and is easily affected by human factors. The accuracy of the test results is affected. In addition, traditional equipment mostly uses a single sensor for partial discharge signal acquisition, which cannot fully and accurately capture the various partial discharge signal characteristics that may appear in the equipment under different working conditions, resulting in low diagnostic accuracy and reliability.

[0003] With the continuous advancement of sensing and data processing technologies, partial discharge monitoring systems have gradually evolved into intelligent monitoring systems based on multi-sensor fusion in recent years. These systems utilize the collaborative work of multiple sensors to simultaneously collect multiple signals, including electrical, acoustic, vibration, and electromagnetic signals. Fusion algorithms are then used to comprehensively analyze these signals, thereby improving the accuracy of partial discharge detection and diagnosis. Although multi-sensor fusion technology can improve the accuracy of partial discharge monitoring to a certain extent, existing technologies still face numerous challenges in their application.

[0004] First, existing multi-sensor fusion systems mostly rely on traditional feature extraction and classification methods, such as analysis methods based on frequency domain features, time domain features, or statistical features. These methods often fail to fully consider the complexity and diversity of the signals when extracting features from partial discharge signals, resulting in insufficient accuracy in feature extraction results. Furthermore, existing classification models often utilize traditional machine learning algorithms, such as support vector machines (SVMs) and decision trees (DTs). These algorithms suffer from high computational complexity and long training times when processing high-dimensional data, and their ability to process nonlinear data features is limited. Therefore, existing technologies for partial discharge classification and fault diagnosis still face high computational burdens and low accuracy.

[0005] Secondly, existing multi-sensor fusion systems typically use a fixed-weight fusion method to process partial discharge signals. This method ignores the differences in the weights of individual sensor signals under different circumstances. For example, under certain operating conditions, the signals of some sensors may be more representative, while the signals of other sensors may be significantly affected by noise or interference. Therefore, existing fusion methods are generally unable to dynamically adjust the weights of individual sensor signals based on different scenarios. This results in significant differences in the performance of the fused feature matrix under different fault conditions, affecting the accuracy of partial discharge fault diagnosis.

[0006] Furthermore, existing hyperparameter optimization methods have limited application in multi-sensor fusion systems. While some research has begun to introduce hyperparameter optimization techniques to improve the performance of partial discharge classification models, most existing methods use grid search or random search-based hyperparameter adjustment. These methods are not only computationally complex but also often struggle to find the optimal hyperparameter combination. When processing complex partial discharge signals, model hyperparameters significantly impact classification performance. Traditional optimization methods often fail to fully exploit the deep features in partial discharge data, limiting the effectiveness of classification models in practical applications.

[0007] In response to the above problems, some improvement measures have been proposed in the existing technology. In recent years, deep learning technology, especially convolutional neural networks (CNN) and capsule networks (CapsNet), has been applied to the field of partial discharge fault diagnosis. Deep learning models can automatically extract high-level features of multi-dimensional signals, reducing the complexity and limitations of manual feature extraction, and have good nonlinear modeling capabilities and strong generalization capabilities. Capsule networks have become a powerful tool for processing partial discharge signals due to their powerful dynamic routing mechanism and low computing requirements. However, in the existing technology, the hyperparameter adjustment of capsule networks still relies on traditional optimization methods, making it difficult to find the optimal hyperparameter combination in the complex scenarios of partial discharge signals.

[0008] In addition, some studies have introduced hyperparameter optimization methods based on evolutionary algorithms and Bayesian optimization, which can improve the efficiency and accuracy of hyperparameter optimization to a certain extent. However, currently, deep learning-based partial discharge fault diagnosis systems have not yet fully achieved automated hyperparameter optimization, and most existing optimization methods have high computational complexity, making them difficult to apply in real-time monitoring systems.

[0009] Therefore, how to provide an intelligent partial discharge online monitoring and fault diagnosis system based on multi-sensor fusion is a problem that technical personnel in this field urgently need to solve. Summary of the Invention

[0010] One objective of this invention is to propose an intelligent online partial discharge monitoring and fault diagnosis system based on multi-sensor fusion. This system leverages multi-sensor technology, deep learning algorithms, and hyperparameter optimization methods. It describes in detail the steps involved in partial discharge signal acquisition, preprocessing, feature fusion, feature dimensionality reduction, and classification model training. It also incorporates the Hyperband algorithm for hyperparameter optimization, ultimately achieving efficient and accurate identification of partial discharge fault types. This system boasts high precision, real-time performance, and reliability, effectively improving the fault diagnosis capabilities of power equipment, reducing the workload of manual inspections, optimizing operations and maintenance, and ensuring the stability and safety of power systems.

[0011] An intelligent partial discharge online monitoring and fault diagnosis system based on multi-sensor fusion according to an embodiment of the present invention includes:

[0012] Multi-sensor acquisition module, used to synchronously collect data during partial discharge;

[0013] A data preprocessing module, used for preprocessing the original partial discharge signal data;

[0014] Feature fusion module, used to unify the features after partial discharge signal processing and construct a fusion weighted feature matrix;

[0015] The feature dimension reduction module is used to perform feature selection and dimension reduction on the fused weighted feature matrix and output the fused feature matrix after dimension reduction;

[0016] The partial discharge classification model training module is used to build a partial discharge type classification model using a lightweight capsule network, and perform convolutional feature extraction, capsule encoding, dynamic routing aggregation, and fault classification decisions;

[0017] Hyperparameter search and optimization module, used to search for hyperparameters based on the Hyperband algorithm, screen the optimal hyperparameter combination, and optimize the partial discharge type classification model;

[0018] The classification model deployment module is used to deploy the optimized partial discharge type classification model to edge computing nodes or remote servers for real-time reasoning;

[0019] The online reasoning and fault diagnosis module is used to receive partial discharge feature data that has undergone preprocessing and feature fusion in real time, output partial discharge fault category labels, generate fault alarm signals, and record fault events in the database and transmit them back to the remote monitoring center.

[0020] Optionally, modules can be connected using the following methods:

[0021] S1. Synchronously collect partial discharge signal data through multiple sensors to construct a partial discharge signal data set;

[0022] S2. Preprocessing the partial discharge signal data set to obtain a partial discharge feature data set;

[0023] S3. Based on the partial discharge feature dataset, the attention mechanism is used to dynamically weight the features of each sensor and construct a fusion weighted feature matrix;

[0024] S4, performing feature dimensionality reduction processing on the fused feature matrix to obtain a fused feature matrix after dimensionality reduction;

[0025] S5. Input the fusion feature matrix after dimension reduction into the lightweight capsule network, use two-dimensional convolution operation to extract the local pulse pattern features of the partial discharge signal, and build a partial discharge type classification model;

[0026] S6. Introduce the Hyperband algorithm to optimize the hyperparameters of the lightweight capsule network, randomly initialize the hyperparameter combination, define a comprehensive evaluation function, determine the optimal hyperparameter combination, and obtain the optimized partial discharge type classification model;

[0027] S7. Deploy the optimized partial discharge type classification model, receive pre-processed and feature-fused partial discharge feature data in real time to identify the fault type, infer and predict the partial discharge fault category and generate an alarm signal, and record and transmit fault event information.

[0028] Optionally, the multiple sensors include a high-frequency current sensor, an ultrasonic sensor, and an ultra-high-frequency electromagnetic sensor.

[0029] Optionally, the preprocessing includes time synchronization alignment, wavelet noise reduction processing, wavelet packet energy decomposition, short-time Fourier transform to extract time-frequency features and generate phase-resolved partial discharge images;

[0030] The generation of phase-resolved partial discharge images includes synchronously collecting the applied voltage signal and matching it with the corresponding time axis of the synchronized electrical pulse signal; normalizing the applied voltage signal and extracting the periodic phase information. The applied voltage period is T V , calculate the phase angle; according to the moment when the discharge pulse appears in each pulse signal, extract the corresponding phase angle and pulse amplitude; with the phase angle as the horizontal coordinate and the pulse amplitude as the vertical coordinate, construct a scatter plot to form a phase-resolved partial discharge map.

[0031] Optionally, the S3 specifically includes:

[0032] S31, the synchronous electrical pulse signal characteristic data F obtained by preprocessing I , Synchronous acoustic emission signal characteristic data F A and synchronous electromagnetic radiation signal characteristic data F EPerform feature normalization processing respectively to generate the normalized electrical pulse signal feature vector F′ I , normalized acoustic emission signal feature vector F′ A and normalized electromagnetic radiation signal eigenvector F′ E ;

[0033] S32, normalizing the feature vector F′ I , F′ A , F′ E Apply feature reweighting processing to generate reweighted electrical pulse signal feature vectors Reweighted acoustic emission signal eigenvector and reweighted electromagnetic radiation signal eigenvector

[0034] S33, the reweighted feature vector Splice by feature dimension to generate the initial fusion feature matrix F stack ;

[0035] S34, based on the preliminary fusion feature matrix F stack , the attention mechanism is used to calculate the reweighted feature vectors The weight coefficients of , and the fusion weighted feature matrix is obtained:

[0036] α I +α A +α E =1;

[0037]

[0038] Among them, α I , α A , α E is the weight coefficient of the reweighted eigenvector, F attn is the fusion weighted feature matrix, is the eigenvector of the reweighted electrical pulse signal, is the eigenvector of the reweighted acoustic emission signal, is the heavily weighted electromagnetic radiation signal eigenvector.

[0039] Optionally, the S4 specifically includes:

[0040] S41, based on the fusion weighted feature matrix F attn , calculate the feature covariance matrix:

[0041]

[0042] Among them, C F is the covariance matrix between the dimensions of the fusion feature, n is the number of samples, F attnis the fusion weighted feature matrix, is the mean vector of the fused weighted feature matrix, and T is the matrix transpose operation;

[0043] S42, the covariance matrix C between the dimensions of the fusion feature F Perform feature decomposition to make the covariance matrix C between the dimensions of the fusion feature F is decomposed into the product of the eigenvector matrix and the eigenvalue matrix;

[0044] S43, according to the size of the eigenvalues in the eigenvalue matrix, in descending order, select the eigenvectors corresponding to the first k eigenvalues whose cumulative characteristic contribution rate is not less than the set threshold η, and form the dimension reduction mapping matrix V k ;

[0045] S44. Using the dimensionality reduction mapping matrix V k The fusion weighted feature matrix F attn Perform linear transformation to obtain the fusion feature matrix F after dimension reduction dim .

[0046] Optionally, the S5 specifically includes:

[0047] S51, based on the fusion feature matrix F after dimensionality reduction dim , a two-dimensional convolution operation is used to extract the local pulse pattern features of the partial discharge signal and generate a convolution feature matrix:

[0048] F conv =max(0,BN(Conv2D(F dim ,W conv )+b conv ));

[0049] Among them, F conv is the convolution feature matrix, BN(·) is the batch normalization function, Conv2D(·) is the two-dimensional convolution operation function, F dim is the fusion feature matrix after dimension reduction, W conv is the convolution kernel parameter matrix, b conv is the convolution bias term;

[0050] S52, convolution feature matrix F conv As input, for each local feature region, the linearly transformed feature vector is obtained by multiplying it by the capsule encoding weight matrix and adding the bias vector. Then, the vector compression function is applied for nonlinear normalization to generate the primary local discharge pattern capsule set U = u1, u2, u m}, where each primary capsule vector u i Express the intensity and direction information of the local pulse characteristics of partial discharge;

[0051] S53. Based on the set of primary partial discharge mode capsules U, by multiplying each primary capsule vector u i by the corresponding partial discharge category mapping weight matrix, a set of prediction vectors is generated for each primary capsule pointing to capsules of various categories;

[0052] S54. Based on the set of prediction vectors, through a dynamic routing mechanism, according to the weighted manner of the current coupling coefficient, the prediction vectors of each primary capsule are weighted and summed to obtain the input vector of each category capsule. Subsequently, the input vector is normalized by applying a vector compression function to obtain the category capsule output vector, and a set of category capsules is constructed:

[0053]

[0054] where, V output is the set of output vectors of category capsules, f is the normalization function, w ij is the coupling coefficient between the i-th category capsule and the j-th primary capsule, i is the index of the category capsule, j is the index of the primary capsule, k is the index of the prediction vector, v k is the prediction vector of the k-th primary capsule, and M is the total number of category capsules;

[0055] S55. Based on the set of category capsules, by calculating the norm of each category capsule vector, the category with the largest norm is selected as the final local discharge fault prediction result, and the predicted category label y[[ID=二十一]] pred [[ID=二十二]]is output;[[ID=二十三]] [[ID=二十四]]

[0056] [[ID=二十五]]S56. By reducing the number of primary capsules, reducing the dimension of capsule vectors, reducing the number of dynamic routing iterations, and implementing a parameter sharing strategy for the partial discharge category mapping weight matrix, the network structure is optimized and the capsule network is lightweighted;[[ID=二十六]] [[ID=二十七]]

[0057] [[ID=二十八]]S57. Through the operations in steps S51 to S56, a local discharge type classification model is output.[[ID=二十九]] [[ID=三十]]

[0058] [[ID=三十一]]Optionally, the S56 specifically includes:[[ID=三十二]] [[ID=三十三]]

[0059] [[ID=三十四]]S561. Set the number of primary capsules as N, reduce the number of primary capsules to obtain the optimized number of primary capsules N1. Control the computational complexity of the capsule network through the optimized number of primary capsules N1. By reducing the number of primary capsules, the total computational amount is reduced, and the consumption of computing resources is optimized;[[ID=三十五]] [[ID=三十六]]

[0060] [[ID=三十七]]S562. Set the original dimension of each capsule vector as D1, adjust the original dimension D1 of the capsule vector to D2, where D2 < D1. Use a dimensionality reduction mapping matrix to map the original dimension D1 to the new dimension D2 to obtain a new capsule vector. By reducing the dimension, the network computing and storage requirements are reduced; [[ID=三十八]]

[0061] S563. Set the maximum number of iterations of the original dynamic routing to K1, and reduce the original maximum number of iterations K1 to K2, where K2 < K1. By reducing the maximum number of iterations, the overall routing calculation amount is reduced, thereby improving the inference speed of the network.

[0062] S564. Perform parameter sharing processing on the game release category mapping weight matrix, and set the sharing coefficient to λ. By controlling the degree of parameter sharing, the number of shared parameters is the ratio of the original number of parameters to the sharing coefficient. By sharing weights, the calculation and storage requirements of the model are reduced.

[0063] S565. Based on the optimized number of primary capsules N1, capsule vector dimension D2, dynamic routing iteration number K2, and sharing coefficient λ, finally construct a lightweight capsule network. The computational complexity, storage requirements, and real-time inference ability of the lightweight capsule network are effectively improved, so as to meet the requirements of the partial discharge type classification task.

[0064] Optionally, the S6 specifically includes:

[0065] S61. Based on the Hyperband algorithm, set the hyperparameter search space of the partial discharge type classification model H = {(m, d, r, k, η, λ)|m ∈ M, d ∈ D, r ∈ R, k ∈ K, η ∈ L, λ ∈ Λ}, where m is the number of primary capsules, d is the capsule vector dimension, r is the number of dynamic routing iterations, k is the convolution kernel size, η is the learning rate, λ is the regularization coefficient, and M, D, R, K, L, Λ respectively represent the value sets of each hyperparameter.

[0066] S62. Based on the partial discharge type classification model, define the comprehensive evaluation function of the Hyperband algorithm:

[0067]

[0068] where E(h) is the comprehensive evaluation function, α is the classification accuracy weight coefficient, h is the set of hyperparameter combinations to be evaluated currently, N correct (h) is the number of samples correctly classified in the validation set, N total is the total number of samples in the validation set, β is the inference delay penalty weight coefficient, T inference (h) is the total time taken to process a batch of data in the inference stage, N batch is the number of samples in the inference batch, γ is the model size penalty weight coefficient, P l (h) is the number of parameters of the l-th layer network, S lis the data storage size of a single parameter in the lth layer, L is the total number of network layers, l is the network layer index, δ is the hyperparameter regularization penalty weight coefficient, p is a single hyperparameter element in the hyperparameter combination, ∈ is the single inference energy consumption penalty weight coefficient, E total (h) is the total energy consumed in the inference phase, N inference is the total number of samples involved in the reasoning process;

[0069] S63. Perform a hyperparameter search based on the hyperparameter search space and the comprehensive evaluation function. In each round of hyperparameter search, the current round is numbered s according to the preset total resource budget. The parameter reduction factor is used to control the resource allocation ratio. Specifically, in the sth round, the number of hyperparameter combinations to be sampled is obtained by multiplying the total budget resources by the reduction factor raised to the negative power of s and then dividing by the minimum resource unit, rounding up. At the same time, the amount of resources allocated to each hyperparameter combination in this round is equal to the minimum resource unit multiplied by the reduction factor raised to the power of s:

[0070]

[0071] in, is the number of hyperparameter combinations sampled in the sth round, R total is the total resource budget, s is the current round number, α is the parameter reduction factor, R min is the minimum resource unit for each set of hyperparameter combinations, For the rounding operation, The amount of resources allocated to each set of hyperparameter combinations in round s;

[0072] S64. After each round of hyperparameter search is completed, the hyperparameter combinations with lower performance evaluation scores are eliminated based on the retention rate corresponding to the current round, and the remaining hyperparameter combination set is screened. Among the remaining sets, the hyperparameter combination with the highest score is selected based on the comprehensive evaluation function score as the optimal hyperparameter combination for the current stage;

[0073] S65. Based on the optimal hyperparameter combination obtained by screening, the network structure and training parameters of the partial discharge type classification model are configured and updated according to the optimal hyperparameter combination to form an optimized partial discharge type classification model.

[0074] Optionally, the S7 specifically includes:

[0075] S71. Deploy the optimized partial discharge type classification model to the edge computing node or remote server of the online monitoring system to implement real-time reasoning services and form a classification reasoning system.

[0076] S72, collecting partial discharge signal data in real time and completing preprocessing, feature fusion and dimensionality reduction, generating a real-time fusion feature matrix, and inputting it into a classification inference system;

[0077] S73. Based on the classification reasoning system, the input real-time fusion feature matrix is input into the partial discharge type classification model, and a real-time predicted partial discharge fault category label is output;

[0078] S74. Based on the real-time predicted partial discharge fault category label and the preset fault judgment logic, a partial discharge fault diagnosis result is generated, and corresponding alarm strategies are triggered according to different fault levels to form a partial discharge fault alarm signal.

[0079] S75. Combine the partial discharge fault category label and the fault alarm signal to form a fault event record, write it into the online monitoring system database, and transmit it back to the remote monitoring center periodically or on demand.

[0080] The beneficial effects of the present invention are:

[0081] By constructing an intelligent online partial discharge monitoring and fault diagnosis system based on multi-sensor fusion, the present invention establishes a complete closed-loop mechanism from signal acquisition, data preprocessing, feature fusion, to classification model training and hyperparameter optimization, achieving efficient real-time detection of partial discharge faults and fault type identification. The system uses multiple sensors to synchronously collect partial discharge signal data, combined with advanced data processing technologies such as wavelet noise reduction and time-frequency feature extraction. By fusing the signal characteristics of different sensors, a weighted feature matrix is constructed, and a lightweight capsule network is used to extract local pulse pattern features and classify faults. In particular, the present invention introduces the Hyperband algorithm for hyperparameter optimization, further improving the accuracy and efficiency of the classification model, enabling the partial discharge type classification model to be continuously optimized in complex environments and adapt to the operating conditions of different power equipment.

[0082] In terms of system architecture, this invention uses automated feature fusion and dimensionality reduction to avoid the tedious manual design and selection of features required in traditional methods, thereby improving the accuracy and automation of feature extraction. Through a dynamic weighting mechanism, this invention automatically adjusts the weights of each sensor based on its actual signal characteristics, enabling the fused feature matrix to more accurately reflect the true characteristics of partial discharge, enhancing the system's ability to identify complex fault modes.

[0083] Furthermore, this invention utilizes a hyperparameter optimization module and the Hyperband algorithm to automatically adjust the model's hyperparameters, avoiding the inefficiency of traditional hyperparameter adjustment, which requires extensive manual intervention. The Hyperband algorithm, through its efficient resource allocation strategy, can quickly find the optimal hyperparameter combination, significantly improving the system's training efficiency and prediction accuracy. This not only enhances the overall performance of the partial discharge classification model but also enables the system to automatically adapt to different devices and operating environments, ensuring its wide applicability and stability.

[0084] Furthermore, the system of the present invention achieves a complete closed-loop process, from partial discharge fault data collection to final intelligent diagnosis and alarm. By receiving and analyzing multi-sensor fusion features in real time, the system can instantly identify the type of partial discharge fault and generate an alarm signal, providing real-time fault diagnosis support for the operation and maintenance of power equipment. This effectively reduces the workload of manual inspections and improves the accuracy of equipment fault prediction and prevention. Through the system's automated fault classification and alarm functions, the present invention significantly reduces reliance on manual intervention and improves the efficiency and reliability of operation and maintenance.

[0085] In summary, this invention, through its intelligent online partial discharge monitoring and fault diagnosis system based on multi-sensor fusion, achieves automated, efficient, and intelligent partial discharge detection, offering high precision, high real-time performance, and low manual dependency. This system not only enhances the fault diagnosis capabilities of power equipment but also optimizes operation and maintenance management, reduces maintenance costs, and improves the safety and stability of power systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0087] Figure 1 This is a flow chart of the method for the intelligent partial discharge online monitoring and fault diagnosis system based on multi-sensor fusion proposed by the present invention;

[0088] Figure 2 This is a system diagram of the intelligent partial discharge online monitoring and fault diagnosis system based on multi-sensor fusion proposed by the present invention;

[0089] Figure 3 This is a flowchart of the partial discharge classification model training and optimization of the lightweight capsule network in the intelligent partial discharge online monitoring and fault diagnosis system based on multi-sensor fusion proposed in the present invention. DETAILED DESCRIPTION

[0090] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0091] refer to Figure 1-3 , an intelligent partial discharge online monitoring and fault diagnosis system based on multi-sensor fusion, including:

[0092] Multi-sensor acquisition module, used to synchronously collect data during partial discharge;

[0093] A data preprocessing module, used for preprocessing the original partial discharge signal data;

[0094] Feature fusion module, used to unify the features after partial discharge signal processing and construct a fusion weighted feature matrix;

[0095] The feature dimension reduction module is used to perform feature selection and dimension reduction on the fused weighted feature matrix and output the fused feature matrix after dimension reduction;

[0096] The partial discharge classification model training module is used to build a partial discharge type classification model using a lightweight capsule network, and perform convolutional feature extraction, capsule encoding, dynamic routing aggregation, and fault classification decisions;

[0097] Hyperparameter search and optimization module, used to search for hyperparameters based on the Hyperband algorithm, screen the optimal hyperparameter combination, and optimize the partial discharge type classification model;

[0098] The classification model deployment module is used to deploy the optimized partial discharge type classification model to edge computing nodes or remote servers for real-time reasoning;

[0099] The online reasoning and fault diagnosis module is used to receive partial discharge feature data that has undergone preprocessing and feature fusion in real time, output partial discharge fault category labels, generate fault alarm signals, and record fault events in the database and transmit them back to the remote monitoring center.

[0100] The intelligent online monitoring and fault diagnosis system for partial discharge based on multi-sensor fusion provided by the present invention integrates multiple functional modules such as data acquisition and preprocessing, feature fusion, feature dimensionality reduction, classification model training, hyperparameter optimization and model deployment, realizing the intelligent and automated diagnosis of partial discharge faults. The system synchronously collects electrical, acoustic and electromagnetic signals during the partial discharge process through multiple sensors, constructs a weighted fusion feature matrix after preprocessing and feature fusion, and classifies the partial discharge type through a lightweight capsule network. The Hyperband algorithm is introduced to optimize hyperparameters to ensure that the model has optimal performance under different working conditions. The optimized classification model can be deployed to edge computing nodes or remote servers to perform fault diagnosis and alarm generation in real time. The system has the advantages of high precision, high real-time performance and low manual dependence, which can effectively improve the fault detection capability of power equipment, reduce the workload of manual inspections, and improve the operational safety and stability of the power system.

[0101] In this embodiment, the modules are connected through the following methods:

[0102] S1. Synchronously collect partial discharge signal data through multiple sensors to construct a partial discharge signal data set;

[0103] S2. Preprocessing the partial discharge signal data set to obtain a partial discharge feature data set;

[0104] S3. Based on the partial discharge feature dataset, the attention mechanism is used to dynamically weight the features of each sensor and construct a fusion weighted feature matrix;

[0105] S4, performing feature dimensionality reduction processing on the fused feature matrix to obtain a fused feature matrix after dimensionality reduction;

[0106] S5. Input the fusion feature matrix after dimension reduction into the lightweight capsule network, use two-dimensional convolution operation to extract the local pulse pattern features of the partial discharge signal, and build a partial discharge type classification model;

[0107] S6. Introduce the Hyperband algorithm to optimize the hyperparameters of the lightweight capsule network, randomly initialize the hyperparameter combination, define a comprehensive evaluation function, determine the optimal hyperparameter combination, and obtain the optimized partial discharge type classification model;

[0108] S7. Deploy the optimized partial discharge type classification model, receive pre-processed and feature-fused partial discharge feature data in real time to identify the fault type, infer and predict the partial discharge fault category and generate an alarm signal, and record and transmit fault event information.

[0109] The intelligent online partial discharge monitoring and fault diagnosis method based on multi-sensor fusion provided by the present invention combines a lightweight capsule network with the Hyperband algorithm. Through multiple steps including data acquisition, preprocessing, feature fusion, feature dimensionality reduction, classification model training, and hyperparameter optimization, it achieves intelligent and automated partial discharge fault diagnosis. The method uses multiple sensors to synchronously acquire partial discharge signals, constructs a signal dataset, performs preprocessing and feature fusion, generates a weighted feature matrix, and then optimizes the feature space through dimensionality reduction. A lightweight capsule network is used to extract local pulse patterns from the reduced feature matrix, constructs a classification model, and utilizes the Hyperband algorithm to efficiently optimize the model's hyperparameters to ensure optimal classification performance. Finally, the optimized model is deployed to edge computing nodes for real-time fault diagnosis and alarm generation. This method has the advantages of high precision, low computational requirements, and high real-time performance. It can significantly improve the accuracy and automation level of power equipment fault detection, reduce manual inspections, and enhance the stability and safety of power systems. It is suitable for online monitoring and fault warning of complex power equipment.

[0110] In this embodiment, the multiple sensors include a high-frequency current sensor, an ultrasonic sensor, and an ultra-high-frequency electromagnetic sensor.

[0111] The system uses high-frequency current sensors, ultrasonic sensors, and ultra-high-frequency electromagnetic sensors to collect partial discharge signals in real time and conduct comprehensive analysis. The system utilizes multi-sensor fusion technology to improve the detection accuracy and diagnostic efficiency of partial discharge faults.

[0112] In this embodiment, the preprocessing includes time synchronization alignment, wavelet noise reduction processing, wavelet packet energy decomposition, short-time Fourier transform to extract time-frequency features and generate phase-resolved partial discharge images;

[0113] The generation of phase-resolved partial discharge images includes synchronously collecting the applied voltage signal and matching it with the corresponding time axis of the synchronized electrical pulse signal; normalizing the applied voltage signal and extracting the periodic phase information. The applied voltage period is T V , calculate the phase angle; according to the moment when the discharge pulse appears in each pulse signal, extract the corresponding phase angle and pulse amplitude; with the phase angle as the horizontal coordinate and the pulse amplitude as the vertical coordinate, construct a scatter plot to form a phase-resolved partial discharge map.

[0114] The present invention effectively improves the quality and availability of partial discharge signals by performing multi-step preprocessing on partial discharge signal data, including time synchronization alignment, wavelet noise reduction, wavelet packet energy decomposition, short-time Fourier transform to extract time-frequency features and generate phase-resolved partial discharge maps. By synchronously collecting applied voltage signals and electrical pulse signals, normalizing them and extracting periodic phase information, and then calculating the phase angle and pulse amplitude, a phase-resolved partial discharge map is constructed, thereby enhancing the timing recognition capability and feature resolution capability of the partial discharge process. This preprocessing method makes the fused feature data more accurate and reliable, providing high-quality data support for subsequent fault diagnosis and classification. By optimizing the data input quality, the accuracy and stability of the subsequent model in partial discharge fault diagnosis are significantly improved. It is suitable for intelligent monitoring and fault warning of power equipment, and has the advantages of high efficiency, precision and low error.

[0115] In this embodiment, S3 specifically includes:

[0116] S31, the synchronous electrical pulse signal characteristic data F obtained by preprocessing I , Synchronous acoustic emission signal characteristic data F A and synchronous electromagnetic radiation signal characteristic data F E Perform feature normalization processing respectively to generate the normalized electrical pulse signal feature vector F′ I , normalized acoustic emission signal feature vector F′ A and normalized electromagnetic radiation signal eigenvector F′ E ;

[0117] S32, normalizing the feature vector F′ I , F′ A , F′ E Apply feature reweighting processing to generate reweighted electrical pulse signal feature vectors Reweighted acoustic emission signal eigenvector and reweighted electromagnetic radiation signal eigenvector

[0118] S33, the reweighted feature vector Splice by feature dimension to generate the initial fusion feature matrix F stack ;

[0119] S34, based on the preliminary fusion feature matrix F stack , the attention mechanism is used to calculate the reweighted feature vectors The weight coefficients of , and the fusion weighted feature matrix is obtained:

[0120] α I +α A +α E =1;

[0121]

[0122] Among them, α I , α A , α E is the weight coefficient of the reweighted eigenvector, F attn is the fusion weighted feature matrix, is the eigenvector of the reweighted electrical pulse signal, is the eigenvector of the reweighted acoustic emission signal, is the heavily weighted electromagnetic radiation signal eigenvector.

[0123] The present invention significantly improves the comparability and accuracy of signal features by performing refined processing on multi-sensor data of partial discharge signals, including feature normalization, feature reweighting, and the generation of a fused weighted feature matrix. First, by normalizing the synchronous features of electrical pulse signals, acoustic emission signals, and electromagnetic radiation signals, the uniformity between different sensor signals is ensured. Then, through feature reweighting, the key role of each sensor signal in partial discharge fault identification is enhanced, ensuring the representativeness and accuracy of the fused features. Finally, an attention mechanism is used to calculate the feature weight coefficient, further optimizing the fusion weighting effect of each signal feature and generating a more accurate fused weighted feature matrix, providing high-quality data input for subsequent fault diagnosis. This preprocessing method improves the feature expression capability of partial discharge signals and enhances the accuracy and stability of fault classification models. It is suitable for intelligent fault diagnosis and online monitoring of power equipment and has the advantages of high precision, strong adaptability, and low computational burden.

[0124] In this embodiment, the S4 specifically includes:

[0125] S41, based on the fusion weighted feature matrix F attn , calculate the feature covariance matrix:

[0126]

[0127] Among them, C F is the covariance matrix between the dimensions of the fusion feature, n is the number of samples, F attn is the fusion weighted feature matrix, is the mean vector of the fused weighted feature matrix, and T is the matrix transpose operation;

[0128] S42, the covariance matrix C between the dimensions of the fusion feature F Perform feature decomposition to make the covariance matrix C between the dimensions of the fusion feature F is decomposed into the product of the eigenvector matrix and the eigenvalue matrix;

[0129] S43, according to the size of the eigenvalues in the eigenvalue matrix, in descending order, select the eigenvectors corresponding to the first k eigenvalues whose cumulative characteristic contribution rate is not less than the set threshold η, and form the dimension reduction mapping matrix V k ;

[0130] S44. Using the dimensionality reduction mapping matrix V k The fusion weighted feature matrix F attn Perform linear transformation to obtain the fusion feature matrix F after dimension reduction dim .

[0131] The present invention optimizes the dimensionality reduction process of the fused weighted feature matrix through the calculation of the feature covariance matrix and feature decomposition. First, based on the fused weighted feature matrix, the covariance matrix between each dimension is calculated to ensure that the correlation between different features is fully considered; then, through feature decomposition, the covariance matrix is decomposed into an eigenvector matrix and an eigenvalue matrix, and then the appropriate eigenvector is selected according to the contribution rate of the eigenvalue to complete the construction of the dimensionality reduction mapping matrix. Finally, the dimensionality reduction mapping matrix is applied to the fused weighted feature matrix through linear transformation to generate a feature matrix after dimensionality reduction. This dimensionality reduction method effectively reduces the feature dimension, retains key feature information, improves computational efficiency and optimizes model performance. Through this series of processing, the system can efficiently extract the core features of the partial discharge signal, improve the accuracy and stability of the fault classification model, and is suitable for intelligent monitoring and fault diagnosis of power equipment with high accuracy and low computational complexity.

[0132] In this embodiment, the S5 specifically includes:

[0133] S51, based on the fusion feature matrix F after dimensionality reduction dim , a two-dimensional convolution operation is used to extract the local pulse pattern features of the partial discharge signal and generate a convolution feature matrix:

[0134] F conv =max(0,BN(Conv2D(F dim ,W conv )+b conv ));

[0135] Among them, F conv is the convolution feature matrix, BN(·) is the batch normalization function, Conv2D(·) is the two-dimensional convolution operation function, F dim is the fusion feature matrix after dimension reduction, W conv is the convolution kernel parameter matrix, b conv is the convolution bias term;

[0136] S52, convolution feature matrix F convAs input, for each local feature region, the linearly transformed feature vector is obtained by multiplying it by the capsule encoding weight matrix and adding the bias vector. Then, the vector compression function is applied for nonlinear normalization to generate the primary local discharge pattern capsule set U = u1, u2, u m}, where each primary capsule vector u i Express the intensity and direction information of the local pulse characteristics of partial discharge;

[0137] S53, based on the primary partial discharge pattern capsule set U, by i Multiply by the corresponding partial discharge category mapping weight matrix to generate a set of prediction vectors for each primary capsule pointing to each category capsule;

[0138] S54. Based on the prediction vector set, the prediction vectors of each primary capsule are weighted and summed according to the weighting method of the current coupling coefficient through the dynamic routing mechanism to obtain the input vector of each class capsule. The input vector is then normalized by applying a vector compression function to obtain the class capsule output vector, and the class capsule set is constructed:

[0139]

[0140] Among them, V output is the output vector set of each capsule, f is the normalization function, w ij is the coupling coefficient between the i-th category capsule and the j-th primary capsule, i is the index of the category capsule, j is the index of the primary capsule, k is the index of the prediction vector, and v k is the prediction vector of the kth primary capsule, and M is the total number of category capsules;

[0141] S55. Based on the category capsule set, by calculating the modulus length of each category capsule vector, the category with the largest modulus length is selected as the final partial discharge fault prediction result, and the predicted category label y is output. pred ;

[0142] S56. By reducing the number of primary capsules, reducing the dimension of the capsule vector, reducing the number of dynamic routing iterations, and implementing a parameter sharing strategy for the partial discharge category mapping weight matrix, the network structure is optimized to lightweight the capsule network;

[0143] S57 . Through the operations of steps S51 to S56 , a partial discharge type classification model is output.

[0144] Through the construction of a partial discharge signal feature extraction and classification model based on a lightweight capsule network, the present invention realizes efficient and accurate detection and classification of partial discharge faults. First, two-dimensional convolution operations are used to extract the features of local pulse patterns and generate a convolutional feature matrix, ensuring the full expression of signal features. Through capsule encoding and non-linear normalization processing, a set of primary partial discharge pattern capsules is generated, thereby enhancing the expression ability of partial discharge features. Combining the dynamic routing mechanism and weight optimization, each partial discharge category is accurately predicted, improving the accuracy of the classification results. Further optimizing the network structure, the lightweight capsule network reduces the computational complexity and improves the real-time inference ability of the model. Finally, through the operations of steps S51 to S56, a partial discharge type classification model is constructed, which has high precision, high efficiency and low computational burden, can effectively improve the intelligent level of power equipment fault diagnosis, is widely applicable to the online monitoring and fault warning of power systems, and has significant application prospects.

[0145] In this embodiment, S56 specifically includes:

[0146] S561. Set the number of primary capsules as N, reduce the number of primary capsules to obtain an optimized number of primary capsules N1, control the computational complexity of the capsule network through the optimized number of primary capsules N1, and reduce the total computational amount by reducing the number of primary capsules, optimizing the consumption of computing resources;

[0147] S562. Set the original dimension of each capsule vector as D1, adjust the original dimension D1 of the capsule vector to D2, where D2 < D1, use a dimensionality reduction mapping matrix to map the original dimension D1 to the new dimension D2 to obtain a new capsule vector, and reduce the network computing and storage requirements by reducing the dimension;

[0148] S563. Set the maximum number of iterations of the original dynamic routing as K1, reduce the original maximum number of iterations K1 to K2, where K2 < K1, and reduce the overall routing computational amount by reducing the maximum number of iterations, thereby improving the inference speed of the network;

[0149] S564. Perform parameter sharing processing on the partial discharge category mapping weight matrix, set the sharing coefficient as λ, and control the degree of parameter sharing. The number of shared parameters is the ratio of the original number of parameters to the sharing coefficient. By sharing weights, the computing and storage requirements of the model are reduced;

[0150] S565. Based on the optimized number of primary capsules N1, the capsule vector dimension D2, the dynamic routing iteration number K2, and the sharing coefficient λ, finally construct a lightweight capsule network. The computational complexity, storage requirements, and real-time inference ability of the lightweight capsule network are effectively improved, thus meeting the requirements of the partial discharge type classification task.

[0151] The present invention effectively reduces the computational complexity and storage requirements by optimizing multiple key parameters of the capsule network, thereby improving the real-time reasoning capability and having significant technical benefits. First, by optimizing the number of primary capsules, the consumption of computing resources is reduced, the overall computational load is significantly reduced, and efficient network computing performance is ensured. Secondly, by reducing the dimension of the capsule vector, the use of a dimensionality reduction mapping matrix effectively reduces the network computing and storage requirements, further improving the computing efficiency. Thirdly, by optimizing the maximum number of iterations of dynamic routing, the routing calculation amount is reduced, thereby accelerating the reasoning speed and improving real-time performance. Finally, by performing parameter sharing processing on the partial discharge category mapping weight matrix, the computational and storage burden of the model is further reduced. The comprehensively optimized lightweight capsule network not only effectively improves the computational complexity, storage requirements and reasoning capability, but also can meet the efficient computing requirements of the partial discharge type classification task.

[0152] In this embodiment, S6 specifically includes:

[0153] S61. Based on the Hyperband algorithm, the hyperparameter search space H = {(m, d, r, k, η, λ) | m∈M, d∈D, r∈R, k∈K, η∈L, λ∈Λ} of the partial discharge type classification model is set, where m is the number of primary capsules, d is the capsule vector dimension, r is the number of dynamic routing iterations, k is the convolution kernel size, η is the learning rate, λ is the regularization coefficient, and M, D, R, K, L, and Λ represent the value sets of each hyperparameter respectively;

[0154] S62. Based on the partial discharge type classification model, define the comprehensive evaluation function of the Hyperband algorithm:

[0155]

[0156] Among them, E(h) is the comprehensive evaluation function, α is the classification accuracy weight coefficient, h is the current hyperparameter combination set to be evaluated, N correct (h) is the number of correctly classified samples in the validation set, N total is the total number of samples in the verification set, β is the inference delay penalty weight coefficient, T inference (h) is the total time taken to process a batch of data in the inference phase, N batch is the number of samples in the inference batch, γ is the model size penalty weight coefficient, P l (h) is the number of parameters of the l-th layer network, S l is the data storage size of a single parameter in the lth layer, L is the total number of network layers, l is the network layer index, δ is the hyperparameter regularization penalty weight coefficient, p is a single hyperparameter element in the hyperparameter combination, ∈ is the single inference energy consumption penalty weight coefficient, E total(h) is the total energy consumed in the inference phase, N inference is the total number of samples involved in the reasoning process;

[0157] S63. Perform a hyperparameter search based on the hyperparameter search space and the comprehensive evaluation function. In each round of hyperparameter search, the current round is numbered s according to the preset total resource budget. The parameter reduction factor is used to control the resource allocation ratio. Specifically, in the sth round, the number of hyperparameter combinations to be sampled is obtained by multiplying the total budget resources by the reduction factor raised to the negative power of s and then dividing by the minimum resource unit, rounding up. At the same time, the amount of resources allocated to each hyperparameter combination in this round is equal to the minimum resource unit multiplied by the reduction factor raised to the power of s:

[0158]

[0159] in, is the number of hyperparameter combinations sampled in the sth round, R total is the total resource budget, s is the current round number, α is the parameter reduction factor, R min is the minimum resource unit for each set of hyperparameter combinations, For the rounding operation, The amount of resources allocated to each set of hyperparameter combinations in round s;

[0160] S64. After each round of hyperparameter search is completed, the hyperparameter combinations with lower performance evaluation scores are eliminated based on the retention rate corresponding to the current round, and the remaining hyperparameter combination set is screened. Among the remaining sets, the hyperparameter combination with the highest score is selected based on the comprehensive evaluation function score as the optimal hyperparameter combination for the current stage;

[0161] S65. Based on the optimal hyperparameter combination obtained by screening, the network structure and training parameters of the partial discharge type classification model are configured and updated according to the optimal hyperparameter combination to form an optimized partial discharge type classification model.

[0162] The present invention optimizes the hyperparameters of the partial discharge type classification model by introducing the Hyperband algorithm, thereby achieving efficient hyperparameter search and model optimization. First, based on the Hyperband algorithm, a hyperparameter search space is set, and a comprehensive evaluation function is defined, covering multiple key indicators such as classification accuracy, inference delay, model scale and energy consumption. Through precise resource allocation and round-by-round screening, the system can efficiently screen out the optimal hyperparameter combination in each round, thereby ensuring the optimal performance of the classification model under different power equipment and working environments. Finally, through hyperparameter optimization, the network structure and training parameters of the partial discharge type classification model are effectively configured and updated, significantly improving the classification accuracy and real-time reasoning capability. This method has high computational efficiency and accuracy, can significantly reduce manual intervention in hyperparameter adjustment, and improve the automation level of power equipment fault diagnosis. It is widely applicable to intelligent monitoring and fault warning of power systems, and has great application potential and practical value.

[0163] In this embodiment, the S7 specifically includes:

[0164] S71. Deploy the optimized partial discharge type classification model to the edge computing node or remote server of the online monitoring system to implement real-time reasoning services and form a classification reasoning system.

[0165] S72, collecting partial discharge signal data in real time and completing preprocessing, feature fusion and dimensionality reduction, generating a real-time fusion feature matrix, and inputting it into a classification inference system;

[0166] S73. Based on the classification reasoning system, the input real-time fusion feature matrix is input into the partial discharge type classification model, and a real-time predicted partial discharge fault category label is output;

[0167] S74. Based on the real-time predicted partial discharge fault category label and the preset fault judgment logic, a partial discharge fault diagnosis result is generated, and corresponding alarm strategies are triggered according to different fault levels to form a partial discharge fault alarm signal.

[0168] S75. Combine the partial discharge fault category label and the fault alarm signal to form a fault event record, write it into the online monitoring system database, and transmit it back to the remote monitoring center periodically or on demand.

[0169] The present invention implements real-time reasoning services by deploying the optimized partial discharge type classification model to the edge computing node or remote server of the online monitoring system, thereby improving the response speed and real-time performance of fault diagnosis. The system collects and processes partial discharge signal data in real time, completes feature fusion and dimensionality reduction, generates a real-time fusion feature matrix, and inputs it into the classification reasoning system to predict the type of partial discharge fault. Through the classification reasoning system, the partial discharge fault category label is quickly output, and the diagnosis result is generated in combination with the preset fault judgment logic. According to different fault levels, the system automatically triggers the corresponding alarm strategy and generates a fault alarm signal to ensure timely response to potential fault hazards. Finally, the fault event record is written into the online monitoring system database and periodically transmitted back to the remote monitoring center to facilitate subsequent operation and maintenance and fault tracing. This method effectively improves the intelligence level, accuracy and real-time performance of power equipment fault diagnosis, reduces the workload of manual inspections, optimizes operation and maintenance management, and has broad application prospects.

[0170] Example 1:

[0171] In order to verify the feasibility of the present invention in implementation, the present invention was applied to an intelligent partial discharge fault diagnosis system optimization project jointly carried out by a power company and a well-known domestic power equipment monitoring institute. The main goal of this project is to solve the core problems faced by power companies in the equipment monitoring process, such as inaccurate partial discharge fault detection, delayed fault alarm, and frequent manual intervention. The test period is from June 2024 to October 2024, covering periodic inspections and sudden failure scenarios of multiple power equipment. The test system is deployed in the power monitoring center in Jiangsu Province, and is linked to the data of multiple substations across the country through a remote cloud server. It adopts a high-performance computing platform with a hardware configuration of Intel Xeon Gold6240R+2×A10040G heterogeneous computing architecture. The platform data comes from historical power equipment operation data from Q3 2023 to Q2 2024.

[0172] Traditional partial discharge monitoring systems primarily rely on single-sensor data acquisition combined with simple time-domain analysis, resulting in low fault detection accuracy in the presence of significant signal interference. Furthermore, traditional systems also experience significant latency in real-time fault warnings, often requiring manual verification before issuing an alarm signal, significantly delaying fault handling response time. While existing multi-sensor fusion technology has improved monitoring accuracy, there is still room for improvement in fusion algorithms and hyperparameter optimization, particularly in efficiently combining data from different sensor types and dynamically adjusting hyperparameters based on varying environmental conditions to enhance classification model accuracy and inference speed.

[0173] To effectively address these issues, the research team adopted the "intelligent partial discharge monitoring and fault diagnosis system based on multi-sensor fusion and lightweight capsule network" proposed in this invention. The specific implementation process is as follows: First, the system uses multiple sensors, including high-frequency current sensors, ultrasonic sensors, and ultra-high-frequency electromagnetic sensors, to synchronously collect partial discharge signals from power equipment to generate a raw data set. Subsequently, data preprocessing is performed, including signal time synchronization, noise reduction, feature extraction, and wavelet packet energy decomposition, to ensure that the data quality meets the model input requirements. Next, the feature matrix after dimensionality reduction is classified and trained using a lightweight capsule network model, and hyperparameter optimization is performed in combination with the Hyperband algorithm to improve the accuracy of the classification model and its real-time reasoning capability. Finally, the optimized partial discharge classification model is deployed to the edge computing node of the online monitoring system, which receives the partial discharge signal data after preprocessing and feature fusion in real time, identifies the fault type, and generates an alarm signal.

[0174] During the two-month testing phase, the system successfully monitored and diagnosed partial discharge faults on five different types of power equipment. During system operation, the generated partial discharge fault category labels, generated through real-time inference, matched on-site inspection results, and the fault alarm response time was significantly lower than that of traditional systems. During this period, the system collected over 5,000 hours of equipment operation data, detected 12 potential partial discharge faults, and successfully diagnosed nine of them, achieving a diagnostic accuracy rate of 75%. Compared to traditional manual inspections and single-sensor systems, this system not only significantly improves fault diagnosis accuracy but also significantly shortens fault response time, effectively enhancing the operational safety of power equipment.

[0175] The following is a table of specific data results during the test period of this project:

[0176] Table 1 Comparison of test results of partial discharge fault detection and diagnosis system

[0177] Test items Test data Traditional systems System of the present invention Number of test equipment 5 units 5 units 5 units Total data collection time (hours) 5000 5000 5000 Partial discharge fault detection times 12 times 10 times 12 times Fault diagnosis accuracy - 60% 75% Fault alarm response time (average value) 15 minutes 20 minutes 5 minutes System real-time inference time (each time) - 25 seconds 12 seconds False alarm rate - 20% 5% False negative rate - 30% 10%

[0178] As can be seen from the above table, the present invention has achieved significant improvements in multiple core performance indicators for partial discharge fault monitoring of power equipment. Specifically, in terms of fault diagnosis accuracy, the present invention's intelligent partial discharge monitoring and fault diagnosis system improved fault diagnosis accuracy by 15% compared to traditional systems. During testing, the present invention's system accurately identified nine partial discharge faults in five devices, while the traditional system only identified seven, demonstrating superior fault identification capabilities.

[0179] The system of the present invention demonstrates significant advantages in fault alarm response time. Compared to conventional systems, this system reduces fault alarm response time from 20 minutes to 5 minutes, a 75% improvement in response speed. This rapid response capability significantly reduces the impact of power equipment failures on system safety and stability, minimizing potential risks and losses.

[0180] The present invention also significantly improves false alarm and missed alarm rates. While the conventional system has a false alarm rate of 20%, the present invention reduces this to 5% and the missed alarm rate from 30% to 10%. These results demonstrate that the present invention effectively reduces unnecessary alarms and improves the accuracy and reliability of partial discharge fault diagnosis, avoiding the false alarm and missed detection problems common in conventional systems.

[0181] From the overall results analysis, it can be seen that the present invention optimizes the accuracy and real-time performance of partial discharge fault diagnosis by adopting multi-sensor data fusion and a lightweight capsule network classification model. In particular, during the power equipment monitoring process, the present invention can quickly adapt to partial discharge fault modes in different types of equipment and different working environments through hyperparameter optimization and real-time inference mechanisms, thereby improving the adaptability and accuracy of the system. Compared with traditional solutions that rely on a single sensor and a fixed model structure, the method proposed in this invention is superior to existing technologies in terms of fault diagnosis accuracy, response speed, and system stability, verifying its strong adaptability and high practicality in actual online monitoring and fault diagnosis of power equipment.

[0182] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. Intelligent partial discharge online monitoring and fault diagnosis system based on multi-sensor fusion, characterized by: include: Multi-sensor acquisition module, used to synchronously collect data during partial discharge; A data preprocessing module, used for preprocessing the original partial discharge signal data; Feature fusion module, used to unify the features after partial discharge signal processing and construct a fusion weighted feature matrix; The feature dimension reduction module is used to perform feature selection and dimension reduction on the fused weighted feature matrix and output the fused feature matrix after dimension reduction; The partial discharge classification model training module is used to build a partial discharge type classification model using a lightweight capsule network, and perform convolutional feature extraction, capsule encoding, dynamic routing aggregation, and fault classification decisions; Hyperparameter search and optimization module, used to search for hyperparameters based on the Hyperband algorithm, screen the optimal hyperparameter combination, and optimize the partial discharge type classification model; The classification model deployment module is used to deploy the optimized partial discharge type classification model to edge computing nodes or remote servers for real-time reasoning; The online reasoning and fault diagnosis module is used to receive partial discharge feature data that has undergone preprocessing and feature fusion in real time, output partial discharge fault category labels, generate fault alarm signals, and record fault events in the database and transmit them back to the remote monitoring center.

2. The intelligent partial discharge online monitoring and fault diagnosis system based on multi-sensor fusion according to claim 1 is characterized in that: The modules are implemented as follows: S1. Synchronously collect partial discharge signal data through multiple sensors to construct a partial discharge signal data set; S2. Preprocessing the partial discharge signal data set to obtain a partial discharge feature data set; S3. Based on the partial discharge feature dataset, the attention mechanism is used to dynamically weight the features of each sensor and construct a fusion weighted feature matrix; S4, performing feature dimensionality reduction processing on the fused feature matrix to obtain a fused feature matrix after dimensionality reduction; S5. Input the fusion feature matrix after dimension reduction into the lightweight capsule network, use two-dimensional convolution operation to extract the local pulse pattern features of the partial discharge signal, and build a partial discharge type classification model; S6. Introduce the Hyperband algorithm to optimize the hyperparameters of the lightweight capsule network, randomly initialize the hyperparameter combination, define a comprehensive evaluation function, determine the optimal hyperparameter combination, and obtain the optimized partial discharge type classification model; S7. Deploy the optimized partial discharge type classification model, receive pre-processed and feature-fused partial discharge feature data in real time to identify the fault type, infer and predict the partial discharge fault category and generate an alarm signal, and record and transmit fault event information.

3. The intelligent partial discharge online monitoring and fault diagnosis system based on multi-sensor fusion according to claim 2 is characterized in that The multiple sensors include a high-frequency current sensor, an ultrasonic sensor and an ultra-high-frequency electromagnetic sensor.

4. The intelligent partial discharge online monitoring and fault diagnosis system based on multi-sensor fusion according to claim 2 is characterized in that: The preprocessing includes time synchronization alignment, wavelet noise reduction processing, wavelet packet energy decomposition, short-time Fourier transform to extract time-frequency features and generate phase-resolved partial discharge images; The generation of phase-resolved partial discharge images includes synchronously collecting the applied voltage signal and matching it with the corresponding time axis of the synchronized electrical pulse signal; normalizing the applied voltage signal and extracting the periodic phase information. The applied voltage period is T V , calculate the phase angle; According to the moment when the discharge pulse appears in each pulse signal, the corresponding phase angle and pulse amplitude are extracted; A scatter plot is constructed with the phase angle as the horizontal axis and the pulse amplitude as the vertical axis to form a phase-resolved partial discharge map.

5. The intelligent partial discharge online monitoring and fault diagnosis system based on multi-sensor fusion according to claim 2 is characterized in that: The S3 specifically includes: S31, the synchronous electrical pulse signal characteristic data F obtained by preprocessing I , Synchronous acoustic emission signal characteristic data F A and synchronous electromagnetic radiation signal characteristic data F E Perform feature normalization processing respectively to generate the normalized electrical pulse signal feature vector F′ I , normalized acoustic emission signal feature vector F′ A and normalized electromagnetic radiation signal eigenvector F′ E ; S32, normalizing the feature vector F′ I , F′ A , F′ E Apply feature reweighting processing to generate reweighted electrical pulse signal feature vectors Reweighted acoustic emission signal eigenvector and reweighted electromagnetic radiation signal eigenvector S33, the reweighted feature vector Splice by feature dimension to generate the initial fusion feature matrix F stack ; S34, based on the preliminary fusion feature matrix F stack , the attention mechanism is used to calculate the reweighted feature vectors The weight coefficients of , and the fusion weighted feature matrix is obtained: α I +α A +α E =1; Among them, α I , α A , α E is the weight coefficient of the reweighted eigenvector, F attn is the fusion weighted feature matrix, is the eigenvector of the reweighted electrical pulse signal, is the eigenvector of the reweighted acoustic emission signal, is the heavily weighted electromagnetic radiation signal eigenvector.

6. The intelligent partial discharge online monitoring and fault diagnosis system based on multi-sensor fusion according to claim 2 is characterized in that: The S4 specifically includes: S41, based on the fusion weighted feature matrix F attn , calculate the feature covariance matrix: Among them, C F is the covariance matrix between the dimensions of the fusion feature, n is the number of samples, F attn is the fusion weighted feature matrix, is the mean vector of the fused weighted feature matrix, and T is the matrix transpose operation; S42, the covariance matrix C between the dimensions of the fusion feature F Perform feature decomposition to make the covariance matrix C between the dimensions of the fusion feature F is decomposed into the product of the eigenvector matrix and the eigenvalue matrix; S43, according to the size of the eigenvalues in the eigenvalue matrix, in descending order, select the eigenvectors corresponding to the first k eigenvalues whose cumulative characteristic contribution rate is not less than the set threshold η, and form the dimension reduction mapping matrix V k ; S44. Using the dimensionality reduction mapping matrix V k The fusion weighted feature matrix F attn Perform linear transformation to obtain the fusion feature matrix F after dimension reduction dim .

7. The intelligent partial discharge online monitoring and fault diagnosis system based on multi-sensor fusion according to claim 2 is characterized in that: The S5 specifically includes: S51, based on the fusion feature matrix F after dimensionality reduction dim , a two-dimensional convolution operation is used to extract the local pulse pattern features of the partial discharge signal and generate a convolution feature matrix: F conv =max(0,BN(Conv2D(F dim ,W conv )+b conv )); Among them, F conv is the convolution feature matrix, BN(·) is the batch normalization function, Conv2D(·) is the two-dimensional convolution operation function, F dim is the fusion feature matrix after dimension reduction, W conv is the convolution kernel parameter matrix, b conv is the convolution bias term; S52, convolution feature matrix F conv As input, for each local feature region, the linearly transformed feature vector is obtained by multiplying it by the capsule encoding weight matrix and adding the bias vector. Then, the vector compression function is applied for nonlinear normalization to generate the primary local discharge pattern capsule set U = u1, u2, u m }, where each primary capsule vector u i Express the intensity and direction information of the local pulse characteristics of partial discharge; S53, based on the primary partial discharge pattern capsule set U, by i Multiply by the corresponding partial discharge category mapping weight matrix to generate a set of prediction vectors for each primary capsule pointing to each category capsule; S54. Based on the set of prediction vectors, through the dynamic routing mechanism, according to the weighted manner of the current coupling coefficient, perform weighted summation on the prediction vectors of each primary capsule to obtain the input vector of each category capsule. Subsequently, apply a vector compression function to the input vector for normalization processing to obtain the category capsule output vector, and construct a set of category capsules; S55. Based on the category capsule set, by calculating the modulus length of each category capsule vector, the category with the largest modulus length is selected as the final partial discharge fault prediction result, and the predicted category label y is output. pred ; S56. Optimize the network structure and lightweight the capsule network by reducing the number of primary capsules, reducing the dimension of the capsule vector, reducing the number of dynamic routing iterations, and implementing a parameter sharing strategy for the partial discharge category mapping weight matrix; S57. Through the operations of steps S51 to S56, output a partial discharge type classification model.

8. The intelligent partial discharge online monitoring and fault diagnosis system based on multi-sensor fusion according to claim 7 is characterized in that: The specific content of S56 includes: S561. Set the number of primary capsules as N, reduce the number of primary capsules to obtain the optimized number of primary capsules N1, and control the computational complexity of the capsule network through the optimized number of primary capsules N1; S562. Set the original dimension of each capsule vector as D1, adjust the original dimension D1 of the capsule vector to D2, where D2 < D1. Use a dimensionality reduction mapping matrix to map the original dimension D1 to the new dimension D2 to obtain a new capsule vector; S563. Set the maximum number of iterations of the original dynamic routing as K1, and reduce the original maximum number of iterations K1 to K2, where K2 < K1; S564. Perform parameter sharing processing on the partial discharge category mapping weight matrix, set the sharing coefficient as λ, and control the degree of parameter sharing. The number of parameters after sharing is the ratio of the original number of parameters to the sharing coefficient; S565. Based on the optimized number of primary capsules N1, the capsule vector dimension D2, the dynamic routing iteration number K2, and the sharing coefficient λ, finally construct a lightweight capsule network.

9. The intelligent partial discharge online monitoring and fault diagnosis system based on multi-sensor fusion according to claim 2 is characterized in that: The specific content of S6 includes S61. Based on the Hyperband algorithm, set the hyperparameter search space H of the partial discharge type classification model as H = {(m, d, r, k, η, λ)|m ∈ M, d ∈ D, r ∈ R, k ∈ K, η ∈ L, λ ∈ Λ}, where m is the number of primary capsules, d is the capsule vector dimension, r is the number of dynamic routing iterations, k is the convolution kernel size, η is the learning rate, λ is the regularization coefficient, and M, D, R, K, L, Λ respectively represent the value sets of each hyperparameter; S62. Based on the partial discharge type classification model, define the comprehensive evaluation function of the Hyperband algorithm: Among them, E(h) is the comprehensive evaluation function, α is the classification accuracy weight coefficient, h is the current hyperparameter combination set to be evaluated, N correct (h) is the number of correctly classified samples in the validation set, N total is the total number of samples in the verification set, β is the inference delay penalty weight coefficient, T inference (h) is the total time taken to process a batch of data in the inference phase, N batch is the number of samples in the inference batch, γ is the model size penalty weight coefficient, P l (h) is the number of parameters of the l-th layer network, S l is the data storage size of a single parameter in the lth layer, L is the total number of network layers, l is the network layer index, δ is the hyperparameter regularization penalty weight coefficient, p is a single hyperparameter element in the hyperparameter combination, ∈ is the single inference energy consumption penalty weight coefficient, E total (h) is the total energy consumed in the inference phase, N inference is the total number of samples involved in the reasoning process; S63. Based on the hyperparameter search space and the comprehensive evaluation function, perform hyperparameter search. In each round of hyperparameter search process, according to the preset total resource budget, record the current round number as s, and use a parameter reduction factor to control the resource allocation ratio. Specifically, in the s-th round, the number of hyperparameter combinations to be sampled is obtained by multiplying the total budget resource by the negative s-th power of the reduction factor and then dividing by the minimum resource unit, and rounding up. At the same time, the amount of resources allocated to each group of hyperparameter combinations in this round is equal to the minimum resource unit multiplied by the s-th power of the reduction factor; S64. After each round of hyperparameter search is completed, the hyperparameter combinations with lower performance evaluation scores are eliminated based on the retention rate corresponding to the current round, and the remaining hyperparameter combination set is screened. Among the remaining sets, the hyperparameter combination with the highest score is selected based on the comprehensive evaluation function score as the optimal hyperparameter combination for the current stage; S65. Based on the optimal hyperparameter combination obtained by screening, the network structure and training parameters of the partial discharge type classification model are configured and updated according to the optimal hyperparameter combination to form an optimized partial discharge type classification model.

10. The intelligent partial discharge online monitoring and fault diagnosis system based on multi-sensor fusion according to claim 2 is characterized in that: The S7 specifically includes: S71. Deploy the optimized partial discharge type classification model to the edge computing node or remote server of the online monitoring system to implement real-time reasoning services and form a classification reasoning system. S72, collecting partial discharge signal data in real time and completing preprocessing, feature fusion and dimensionality reduction, generating a real-time fusion feature matrix, and inputting it into a classification inference system; S73. Based on the classification reasoning system, the input real-time fusion feature matrix is input into the partial discharge type classification model, and a real-time predicted partial discharge fault category label is output; S74. Based on the real-time predicted partial discharge fault category label and the preset fault judgment logic, a partial discharge fault diagnosis result is generated, and corresponding alarm strategies are triggered according to different fault levels to form a partial discharge fault alarm signal. S75. Combine the partial discharge fault category label and the fault alarm signal to form a fault event record, write it into the online monitoring system database, and transmit it back to the remote monitoring center periodically or on demand.

Citation Information

Cited By

  • Power transformer partial discharge positioning method based on multi-sensor array fusion

    CN120742041A

  • Multi-mode sensing partial discharge monitoring device and method

    CN120847574A

  • GIS partial discharge fault simulation experiment system based on real defects

    CN121093789A

  • High-voltage cable partial discharge monitoring and early warning method and system based on multi-dimensional data perception

    CN121211230A

  • High-voltage cable partial discharge monitoring and early warning method and system based on multi-dimensional data perception

    CN121211230B