A high-voltage motor main insulation multi-source partial discharge interpretable pattern recognition method
Patent Information
- Application Number
- CN202610856235.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-15
- Publication Date
- 2026-10-30
AI Technical Summary
现有的基于深度学习的模式识别方法多集中在单一放电源的分类，对于多源局部放电的研究较少，且基于深度学习的局部放电识别方法多有“黑箱”特性，决策过程缺乏可解释性，难以为电动机绝缘状态评估和运维决策提供可靠依据
[0017]与现有技术相比，本发明的有益效果是：通过引入深度学习方法实现高压电机主绝缘多源局部放电的智能识别，能够在复杂工况下对不同放电类型进行自动特征提取与判别，提升局部放电识别的准确性与稳定性；同时，本发明支持多源及复合局部放电状态的识别，适应多种放电机理并存的实际运行环境；此外，本发明在识别过程中引入可解释性分析机制，对深度学习模型的关注区域进行可视化表征，使识别结果具有明确的物理意义和可理解性。
Smart Images
Figure CN123016148A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of the power industry, specifically to an interpretable pattern recognition method for multi-source partial discharge in the main insulation of a high-voltage motor. Background Technology
[0002] With the widespread application of rotating electric machines in industrial fields, the operating condition of their main insulation system is crucial to their safe operation. Many manufacturers, for economic reasons, frequently operate electric motors under full load or even overload conditions, and the operating environment varies depending on the application. The main insulation is subjected to various aging stresses such as high temperature, humidity, and dust. Insulation defects caused by these aging stresses can trigger partial discharges during motor operation. Continuous partial discharges over long-term operation can lead to insulation cracking, ultimately causing insulation failure and malfunction. Partial discharge (PD), as an important indicator of insulation aging, is of great significance for assessing the condition of main insulation. However, because multiple partial discharge sources are easily interfered with each other, traditional identification methods have limited capabilities under complex multi-source operating conditions.
[0003] In recent years, the rapid development of deep learning technology has provided new solutions for partial discharge pattern recognition. Convolutional neural networks (CNNs) can automatically extract high-level features directly from PRPD spectra or time-frequency images, avoiding the subjectivity and limitations of manual feature extraction. Existing deep learning-based pattern recognition methods mostly focus on the classification of single discharge sources, with limited research on multi-source partial discharges. Furthermore, deep learning-based partial discharge recognition methods often exhibit "black box" characteristics, lacking interpretability in the decision-making process and making it difficult to provide reliable basis for motor insulation condition assessment and operation and maintenance decisions.
[0004] Therefore, this paper proposes a multi-source partial discharge (PRPD) identification method for the main insulation of rotating electrical machines based on convolutional neural networks. A dataset of multi-source PRPDs from slot-to-slot, surface, and corona defects is constructed, generating PRPD spectra and establishing a multi-label image library. On the network model side, a residual network adapted to the PRPD spectra is designed, and an attention mechanism is combined to enhance the model's classification ability. On the interpretation side, gradient weighted analysis is introduced to visualize the convolutional results, determining the consistency between the model's region of interest and the prior PRPD spectra, thereby improving the model's credibility. Summary of the Invention
[0005] In view of the above situation, the purpose of this invention is to provide a method for partial discharge pattern recognition of the main insulation of high-voltage motors. The method obtains PRPD spectrum through experimental data and establishes a database. It leverages the excellent performance of convolutional neural networks in image recognition to form a method applicable to the field of main insulation pattern recognition of high-voltage motors.
[0006] The purpose of this invention, a method for identifying partial discharge modes in the main insulation of a high-voltage motor, is achieved through the following technical solution: A method for interpretable pattern recognition of multi-source partial discharge in the main insulation of a high-voltage motor, the specific implementation steps of which are as follows: S1: Based on the structure, operating environment, and potential aging stress of the main insulation of the high-voltage motor, the characteristics of insulation defects that cause discharge are obtained, and seven partial discharge models are designed according to the defect characteristics: corona discharge, slot discharge, surface discharge, corona-slot discharge, corona-surface discharge, slot-surface discharge, and corona-slot-surface discharge. S2: A partial discharge detection platform is built under laboratory conditions. The pulse current method is used to detect the partial discharge signal of the model and obtain the corresponding phase information and discharge amplitude information. S3: Based on the phase information and discharge amplitude information obtained in S2, preprocess them first, specifically by normalizing the amplitude and removing background noise; and map the preprocessed pulse signal into a two-dimensional PRPD spectrum. S4: Assign corresponding discharge type labels to the two-dimensional PRPD spectrum generated in S3, establish an image database from the labeled PRPD spectrum, input it into a deep convolutional neural network model, and output the recognition results of multi-source partial discharge. S5: Perform gradient weighted analysis on the intermediate features of the deep convolutional neural network model to generate a visualization result that reflects the region of interest of the model, and perform interpretability analysis on the multi-source partial discharge identification process.
[0007] Further, step S1 includes the following sub-steps: S11: Collect typical structural information of the main insulation of the high-voltage motor, including but not limited to the external dimensions, structural layout, insulation medium type and parameters, and geometry of conductive components of the main insulation. S12: Obtain the insulation defect characteristics of the main insulation, including but not limited to operating voltage, current, temperature, humidity, frequency, and the location of partial discharge; S13: Take several F-grade insulated coils and first observe their appearance to ensure that there are no obvious insulation damages such as cracks or wear; S14: Design models for seven types of partial discharge based on the characteristics of the defects: corona discharge, slot discharge, surface discharge, corona-slot discharge, corona-surface discharge, slot-surface discharge, and corona-slot-surface discharge.
[0008] Further, step S2 includes the following sub-steps: S21: Set up a partial discharge detection platform under laboratory conditions and use the pulse current method to detect partial discharge signals; S22: The MPD600 partial discharge detector is used to detect the reaction of the pulse on the coupling capacitor, save the data to the PC, and obtain information such as phase information and discharge amplitude.
[0009] Further, step S3 includes the following sub-steps: S31: The obtained partial discharge data is preprocessed, the discharge amplitude is linearly normalized, and mapped to the numerical range of [0,1]. S32: Generate a PRPD spectrum from the normalized data. First, divide one power frequency cycle into multiple phase intervals, and then divide the amplitude into multiple amplitude intervals. Generate the two-dimensional PRPD spectrum by color mapping to reflect the number of pulses.
[0010] Further, step S4 includes the following sub-steps: S41: Assign the corresponding discharge type label to the generated PRPD spectrum, and use multi-label annotation to characterize the composite discharge state where a single discharge type or multiple discharge types coexist. S42: The database that uses the labeled PRPD spectra as model input; the database is made of an Excel spreadsheet, the first column is the location where the PRPD spectra are stored, and the last three columns are the discharge types in each PRPD spectra; S43: The deep convolutional neural network model is a multi-layer convolutional neural network with residual connection structure. Too deep a network structure can easily lead to a waste of computing resources, while too shallow a network structure cannot extract deep features. Therefore, a residual 34 structure is adopted to enhance the model's feature extraction capability for complex partial discharge patterns.
[0011] Furthermore, the deep convolutional neural network model includes an input layer, a first convolutional layer, a main residual structure, a global average pooling layer, and a fully connected layer. The first convolutional layer uses a 7×7 convolutional kernel with a stride of 2, and passes through a 2×2 pooling window, resulting in 64 feature maps for rapid extraction of low-level features. BN+ReLU is used to introduce non-linearity, improving convergence speed and suppressing overfitting. Finally, max pooling is used for further downsampling to reduce computation. The residual network consists of four layers, each with several residual blocks. The number of residual blocks and the layer group can be represented by a vector [3,4,6,3]. There are a total of 16 residual blocks, each with two convolutional layers, for a total of 32 convolutional layers.
[0012] When extracting spatial features, convolutional neural networks use convolutional kernels to extract data features by sliding across the feature vector, as shown in the following formula: ; ; in, Represents the k-th output channel in Position output, This represents the convolution kernel weights from the m-th input channel to the k-th output channel. This represents the offset of the m-th output channel. The value at that location.
[0013] S44: Divide the input database into three parts: 70% for training, 20% for validation, and 10% for testing. Use a systematic hyperparameter optimization method to tune hyperparameters of the model, such as learning rate and batch size, until the training results of the model can meet the expected accuracy and robustness requirements.
[0014] Further, step S5 includes the following sub-steps: S51: Obtain the predicted values of the last feature layer and classification type of the trained model; S52: Backpropagate the predicted values to obtain the gradient information of the backpropagated feature layer, and calculate the importance of each channel for the last feature layer. S53: Perform a weighted summation using the ReLU function to obtain a thermal distribution map of the region of focus in the model.
[0015] Furthermore, the calculation method for step S52 is as follows: ; in, For feature layer A k The spatial location of each channel Activation value at that location, Indicate category c The predicted score, The gradient of each pixel in feature map A. The weight coefficients for each feature channel.
[0016] Furthermore, the formula for the weighted summation in step S53 is as follows: ; Where 1 represents the class activation heatmap of category c.
[0017] Compared with existing technologies, the beneficial effects of this invention are as follows: By introducing deep learning methods, intelligent identification of multi-source partial discharge in the main insulation of high-voltage motors is achieved, enabling automatic feature extraction and discrimination of different discharge types under complex operating conditions, thus improving the accuracy and stability of partial discharge identification; at the same time, this invention supports the identification of multi-source and composite partial discharge states, adapting to actual operating environments where multiple discharge mechanisms coexist; furthermore, this invention introduces an interpretability analysis mechanism during the identification process, visually representing the region of interest of the deep learning model, making the identification results have clear physical meaning and understandability. Attached Figure Description
[0018] Figure 1 This is a flowchart of the method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the main insulation model of the motor according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the spatial feature extraction process according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the visualization results of an embodiment of the present invention. Specific implementation methods
[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Example 1
[0020] The flowchart of the solution is as follows Figure 1 As shown.
[0021] Taking a typical high-voltage motor as the research object, this study collects relevant structural and operational data of its main insulation system. Specifically, it first obtains the rated voltage level, stator structure, and winding arrangement of the high-voltage motor; then, it acquires the geometric dimensions, insulation material type, and dielectric properties of the stator winding main insulation, while also recording the core structure, electromagnetic environment, and spatial distribution relationship between the windings and the casing. Furthermore, it collects key operating parameters of the high-voltage motor under normal operating conditions, including but not limited to: phase voltage, phase current, operating frequency, winding temperature, and ambient temperature and humidity. The collection of this structural information and operating parameters provides fundamental data support for subsequent partial discharge signal modeling and analysis.
[0022] This study analyzes and models different types of partial discharge conditions, targeting the common physical mechanisms of partial discharge in the main insulation of high-voltage motors. These partial discharge types include, but are not limited to, internal insulation discharge, surface discharge, and corona discharge. In an experimental platform, different insulation defect forms are designed to simulate the generation process of these partial discharges. Partial discharge signals generated under various discharge conditions are collected, and corresponding discharge type labels are recorded. For some special or complex conditions, various combinations of partial discharge conditions can be constructed by changing local insulation parameters, electric field intensity distribution, or discharge location. Through repeated experiments, multi-source partial discharge data containing different discharge modes and their combinations are obtained. The collected data is systematically organized and labeled to form a structured partial discharge sample dataset.
[0023] The acquired multi-source partial discharge data underwent preprocessing and feature construction. First, the original partial discharge signals were denoised, filtered, and invalid signals were removed to improve signal quality. Then, the data was normalized to ensure that signals of different magnitudes and sources could be modeled uniformly. Based on this, the preprocessed partial discharge signals were converted into PRPD spectra, a feature representation suitable for deep learning model input. The constructed features not only reflect the amplitude and phase distribution characteristics of partial discharges but also their spatial features.
[0024] The constructed partial discharge sample dataset was divided into training, validation, and test sets. A 70% allocation was used for training, 20% for validation, and 10% for testing. The training set was used to learn the parameters of the deep learning model, the validation set was used to optimize the model structure and training process, and the test set was used to evaluate the model's final recognition performance. Based on the training set, a deep convolutional neural network model with residual connections was constructed, and end-to-end training was performed on the partial discharge features. During model training, the network parameters were iteratively updated, enabling the model to automatically learn the discrimination features for different partial discharge types. The feature extraction process is as follows: Figure 3 As shown; simultaneously, the model training process is monitored using a validation set, and the training strategy is adjusted based on changes in model performance to improve the model's generalization ability. Through the above training process, a model for identifying multi-source partial discharge in the main insulation of high-voltage motors is obtained.
[0025] After obtaining the trained partial discharge recognition model, it is applied to the partial discharge recognition and diagnosis process. Partial discharge data from the test set is input into the model, and the corresponding partial discharge recognition results are output. Furthermore, gradient weighted analysis is performed on the intermediate feature layers of the model to generate visualization results reflecting the regions of interest during the recognition process, used to characterize the key response regions of different partial discharge types in the feature images. Combining the partial discharge recognition results and their interpretability analysis results, a comprehensive evaluation of the main insulation status of the high-voltage motor is conducted.
[0026] The above description is merely a specific embodiment of the present invention, used to illustrate the technical solution of the present invention, and is not intended to limit the scope of protection of the present invention. Any equivalent changes, substitutions, or improvements made within the scope of the technical concept and claims disclosed in the present invention should be included within the scope of protection of the present invention. The scope of protection of the present invention shall be determined by the claims.
Claims
1. A method for interpretable pattern recognition of multi-source partial discharge in the main insulation of a high-voltage motor, characterized in that, include: S1: Based on the structure, operating environment, and potential aging stress of the main insulation of the high-voltage motor, the characteristics of insulation defects that cause discharge are obtained, and seven partial discharge models are designed according to the defect characteristics: corona discharge, slot discharge, surface discharge, corona-slot discharge, corona-surface discharge, slot-surface discharge, and corona-slot-surface discharge. S2: A partial discharge detection platform is built under laboratory conditions. The pulse current method is used to detect the partial discharge signal of the model and obtain the corresponding phase information and discharge amplitude information. S3: Based on the phase information and discharge amplitude information obtained in S2, preprocess them first, and map the preprocessed pulse signal into a two-dimensional PRPD spectrum. S4: Assign corresponding discharge type labels to the two-dimensional PRPD spectrum generated in S3, establish an image database from the labeled PRPD spectrum, input it into a deep convolutional neural network model, and output the recognition results of multi-source partial discharge. S5: Perform gradient weighted analysis on the intermediate features of the deep convolutional neural network model to generate a visualization result that reflects the region of interest of the model, and perform interpretability analysis on the multi-source partial discharge identification process.
2. The method for interpretable pattern recognition of multi-source partial discharge in the main insulation of a high-voltage motor according to claim 1, characterized in that: In step S1, typical structural information of the main insulation of the high-voltage motor is collected, including the external dimensions, structural layout, insulation medium type and parameters, and geometric shape of conductive components of the main insulation; and the insulation defect characteristics of corona discharge, slot discharge, and surface discharge of the main insulation are obtained, specifically including the operating voltage, current, temperature, humidity, frequency, and location of partial discharge of the main insulation of the high-voltage motor.
3. The method for interpretable pattern recognition of multi-source partial discharge in the main insulation of a high-voltage motor according to claim 1, characterized in that, In step S2: The MPD600 partial discharge detector is used to detect the reaction of the discharge pulse on the coupling capacitor, save the data to the PC, and obtain phase information, discharge amplitude information, and the duration of the acquired signal.
4. The method for interpretable pattern recognition of multi-source partial discharge in the main insulation of a high-voltage motor according to claim 1, characterized in that, In step S3, the discharge amplitude is linearly normalized and mapped to the numerical range of [0,1]. The normalized data is then used to generate a PRPD spectrum. First, one power frequency cycle is divided into multiple phase intervals, and then the discharge amplitude is divided into multiple amplitude intervals. A two-dimensional PRPD spectrum is generated by color mapping to reflect the number of pulses.
5. The method for interpretable pattern recognition of multi-source partial discharge in the main insulation of a high-voltage motor according to claim 1, characterized in that, In step S4, the generated two-dimensional PRPD spectrum is assigned a corresponding discharge type label, and multiple labels are used to characterize a single discharge type or a composite discharge state in which multiple discharge types coexist.
6. The method for interpretable pattern recognition of multi-source partial discharge in the main insulation of a high-voltage motor according to claim 1, characterized in that, In step S4, a deep convolutional neural network model is used to enhance the model's feature extraction capability for complex partial discharge patterns; the database of input deep convolutional neural network model is divided, and a systematic hyperparameter optimization method is used to fine-tune the hyperparameters of the model until the training results of the model can meet the expected accuracy and robustness requirements.
7. The method for interpretable pattern recognition of multi-source partial discharge in the main insulation of a high-voltage motor according to claim 1, characterized in that, In step S5, gradient weighted analysis is performed on the intermediate features of the model. Specifically, backpropagation is performed on the predicted values of the last feature layer and the classification type of the trained model to obtain the gradient information of the backpropagated feature layer. The importance of each channel for the last feature layer is calculated. Weighted summation is performed using the ReLU function to obtain the heat map of the model's key focus area.
8. The method for interpretable pattern recognition of multi-source partial discharge in the main insulation of a high-voltage motor according to claim 4, characterized in that, The linear normalization method is calculated as follows: ; in and These represent the normalized value and the original value, respectively. and These represent the minimum and maximum values of each variable in the data, respectively.
9. The method for interpretable pattern recognition of multi-source partial discharge in the main insulation of a high-voltage motor according to claim 5, characterized in that: The deep convolutional neural network model comprises a multi-layer convolutional neural network with residual connections. This multi-layer convolutional neural network model includes an input layer, a first convolutional layer, a main residual structure, a global average pooling layer, and a fully connected layer. The first convolutional layer uses a 7×7 kernel with a stride of 2, and a 2×2 pooling window to obtain 64 feature maps for rapid low-level feature extraction. BN+ReLU is used to introduce non-linearity and suppress overfitting. Finally, max pooling is used for further downsampling to reduce computational cost. The main residual structure consists of four layers, each layer having several residual blocks. The relationship between the number of residual blocks and the layer group can be represented by a vector [3,4,6,3]. The total number of residual blocks is 16, and each residual block has two convolutional layers, for a total of 32 convolutional layers.
10. The method for interpretable pattern recognition of multi-source partial discharge in the main insulation of a high-voltage motor according to claim 7, characterized in that: The importance of each channel in the last feature layer is calculated as follows: ; in For feature layer A k The spatial location of each channel Activation value at that location, Indicate category The predicted score, The gradient of each pixel in feature map A. The weight coefficients for each feature channel.