Multi-source partial discharge intelligent identification method and system based on YOLOv7 algorithm

Abstract

The invention relates to a multi-source partial discharge intelligent identification method and system based on a YOLOv7 algorithm, and the method comprises the steps: obtaining a to-be-identified multi-source partial discharge PRPD spectrogram, and carrying out the data preprocessing; inputting the preprocessed spectrogram into a YOLOv7 target detection model to obtain an identification result; the YOLOv7 target detection model is optimized by the following steps: acquiring a single-source partial discharge PRPD spectrogram as a training set, acquiring a multi-source partial discharge PRPD spectrogram as a test set, and performing data preprocessing; using the training set to train a YOLOv7 target detection model, and learning and identifying spectrogram features of different types of defects; a Soft-NMS algorithm and a performance evaluation index GIoU are applied to the model, iterative training is carried out, and the model is further optimized; and performing identification performance evaluation on the trained model on the test set, and adjusting and optimizing the model according to a test result. Compared with the prior art, the accuracy of partial discharge detection of the GIS equipment is improved, and the stability and the real-time performance of a target detection system are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of partial discharge identification, and in particular to a multi-source partial discharge intelligent identification method and system based on a YOLOv7 algorithm. Background Art

[0002] With the rapid development of urban power grids, Gas Insulated Switchgear (GIS), as a key device in the power grid, plays a vital role in ensuring the reliability of power supply. However, GIS equipment may have potential insulation defects during the manufacturing and construction process. These defects may develop into discharge channels as the operating time increases, causing GIS breakdown failures, resulting in serious economic losses and personal safety risks.

[0003] In order to timely detect and avoid insulation defects in GIS equipment, partial discharge (PD) detection technology came into being. Ultra High Frequency (UHF) partial discharge detection technology is valued because it can detect internal insulation defects when GIS equipment is operating normally. Through this technology, the health status of GIS equipment can be evaluated in real time without affecting the operation of the power grid, so as to timely detect hidden dangers and avoid failures. There are free metal particles inside the GIS equipment, which will adhere to the surface of the insulator under the action of the electric field. During the operation of the power grid, partial discharge is easily generated. The discharge of metal particles has the characteristics of "small pulses and few pulses", which brings challenges to the accurate detection and identification of partial discharge.

[0004] Traditional partial discharge type recognition methods are mainly aimed at single-source partial discharge pattern recognition. These methods are based on image classification algorithms and cannot effectively handle the stacking and coverage problem of multi-source partial discharge (Partial Discharge Pattern Decomposition, PRPD) spectra. It is difficult to classify a PRPD spectrum into multiple partial discharge types. In addition, about 30% of partial discharge signals caused by insulation defects in GIS equipment are intermittent and sporadic signals. These signals pose a great threat to GIS equipment and are difficult to detect based on conventional inspection technology. They need to be identified through continuous monitoring and accumulation of long-term PRPD spectra. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a multi-source partial discharge intelligent identification method and system based on the YOLOv7 algorithm, so as to realize efficient identification of multi-source partial discharge PRPD spectra and improve the performance and accuracy of the algorithm.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] A multi-source partial discharge intelligent identification method based on the YOLOv7 algorithm comprises the following steps:

[0008] Obtain the multi-source partial discharge PRPD spectrum to be identified and perform data preprocessing;

[0009] The multi-source partial discharge PRPD spectra after data preprocessing are used as the input of the YOLOv7 target detection model to obtain the partial discharge recognition results;

[0010] Among them, the optimization steps of the YOLOv7 target detection model are as follows:

[0011] The single-source partial discharge PRPD spectrum is obtained as a training set, and the multi-source partial discharge PRPD spectrum is obtained as a test set for data preprocessing;

[0012] Using the training set to train the YOLOv7 target detection model, learning to identify spectral features of different types of defects;

[0013] Apply the Soft-NMS algorithm and the performance evaluation index GIoU to the YOLOv7 target detection model, perform iterative training, and further optimize the model;

[0014] The recognition performance of the trained YOLOv7 target detection model is evaluated on the test set, and the model is adjusted and optimized according to the test results.

[0015] Furthermore, the data preprocessing includes multiple processing such as grayscale processing, normalization, dimensionality reduction processing, and noise reduction processing.

[0016] Furthermore, the main network structure of the YOLOv7 target detection model includes Input, Backbone network, Neck network and Head network.

[0017] Furthermore, the Input includes Mosaic data enhancement and adaptive anchor box calculation.

[0018] Furthermore, when training the YOLOv7 target detection model, the configuration file of the YOLOv7 target detection model is adjusted according to the characteristics of the partial discharge PRPD spectrum.

[0019] Furthermore, the configuration file of the YOLOv7 target detection model includes an anchor box size and a number of categories.

[0020] Furthermore, the Soft-NMS algorithm is used to separate and reconstruct the multi-source partial discharge PRPD spectrum.

[0021] Furthermore, the GIoU as a performance evaluation indicator includes multiple ones of precision, recall and F1 score.

[0022] Furthermore, the adjustment and optimization include adjusting the parameters of the YOLOv7 target detection model, improving the data preprocessing steps and further optimizing the NMS algorithm.

[0023] According to another aspect of the present invention, a multi-source partial discharge intelligent identification system based on the YOLOv7 algorithm is provided, comprising:

[0024] A data processing module, used to obtain the multi-source partial discharge PRPD spectrum to be identified and perform pre-processing;

[0025] The intelligent recognition module uses the multi-source partial discharge PRPD spectrum after data preprocessing as the input of the YOLOv7 target detection model to obtain the partial discharge recognition result;

[0026] Among them, the optimization steps of the YOLOv7 target detection model are as follows:

[0027] The single-source partial discharge PRPD spectrum is obtained as a training set, and the multi-source partial discharge PRPD spectrum is obtained as a test set for data preprocessing;

[0028] Using the training set to train the YOLOv7 target detection model, learning to identify spectral features of different types of defects;

[0029] Apply the Soft-NMS algorithm and the performance evaluation index GIoU to the YOLOv7 target detection model, perform iterative training, and further optimize the model;

[0030] The recognition performance of the trained YOLOv7 target detection model is evaluated on the test set, and the model is adjusted and optimized according to the test results.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] 1. The present invention uses the YOLOv7 algorithm instead of the traditional image classification algorithm to automatically detect the overlapping of multi-source partial discharge in the PRPD spectrum, detect the spectrum features therein, and give their category information, thereby improving the model's recognition accuracy of the partial discharge spectrum features in the multi-source partial discharge PRPD spectrum.

[0033] 2. The present invention adopts the Soft-NMS algorithm to deal with the problem of overlapping bounding boxes, separates and reconstructs the multi-source partial discharge PRPD maps, better solves the problem of target overlap, and improves the accuracy of the model in partial discharge detection of GIS equipment.

[0034] 3. The present invention adopts GIoU as the performance evaluation indicator, takes into account the alignment of the bounding box, and improves the stability and accuracy of the detection model. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flow chart of the YOLOv7-based multi-source partial discharge PRPD spectrum recognition algorithm;

[0036] Figure 2 It is the PRPD spectrum of solid insulation discharge;

[0037] Figure 3 is the PRPD spectrum of suspended discharge;

[0038] Figure 4 is the PRPD spectrum of free particle discharge;

[0039] Figure 5 PRPD spectrum of tip discharge. DETAILED DESCRIPTION

[0040] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0041] Abbreviations involved:

[0042] Gas Insulated Switchgear, GIS

[0043] Partial Discharge Pattern Decomposition, PRPD

[0044] Embodiment 1:

[0045] This embodiment provides a multi-source partial discharge intelligent identification method based on the YOLOv7 algorithm. Figure 1 As shown, the following steps are included:

[0046] S1. Obtain the multi-source partial discharge PRPD spectrum to be identified and perform data preprocessing.

[0047] Collect the multi-source partial discharge PRPD spectra on site and perform data preprocessing on the collected spectrum images, including the following steps:

[0048] Grayscale processing to reduce the amount of calculation and highlight the key information in the image;

[0049] Normalizing the spectrogram image to make its pixel value range uniform helps to stabilize the model training;

[0050] Perform dimensionality reduction on the spectrogram image to reduce the complexity of the data and improve the computational efficiency of the model;

[0051] Apply noise reduction processing to remove noise from the spectrogram image and improve image quality.

[0052] S2. The multi-source partial discharge PRPD spectrum after data preprocessing is used as the input of the YOLOv7 target detection model to obtain the partial discharge recognition result.

[0053] The PRPD spectrum obtained in this embodiment is different from the previous partial discharge PRPD spectrum. The PRPD spectrum trained by the YOLOv7 algorithm has a horizontal axis representing period and a vertical axis representing phase.

[0054] The PRPD spectra of partial discharge of different types of defects obtained by the YOLOv7 target detection model are as follows: Figure 2 , Figure 3 , Figure 4 and Figure 5 As shown, Figure 2 is the PRPD spectrum of solid insulation discharge, Figure 3 is the PRPD spectrum of suspended discharge, Figure 4 is the PRPD spectrum of free particle discharge, Figure 5 is the PRPD spectrum of the tip discharge. Figure 2 , Figure 3 , Figure 4 and Figure 5 It can be seen that the multi-source partial discharge intelligent identification method based on the YOLOv7 algorithm proposed in this embodiment can accurately classify and identify the multi-source partial discharge PRPD spectra into multiple partial discharge types.

[0055] The optimization steps of the YOLOv7 target detection model are as follows:

[0056] The single-source partial discharge PRPD spectra collected on-site are used as training sets to train the model to identify different types of defect characteristics; the multi-source partial discharge PRPD spectra collected are used as test sets to verify the generalization ability and recognition effect of the model.

[0057] For data preprocessing, the YOLOv7 algorithm is used to replace the traditional image classification algorithm, and its target detection ability is used to identify the spectral features in the single-source partial discharge PRPD spectrogram; YOLOv7 uses an improved version called Anchor-Scaled YOLOv7, where Anchor refers to the prediction box. Prediction boxes are some rectangular boxes defined on the spectrogram, which cover the position and size of the possible spectral features. YOLOv7 regards each prediction box as an anchor point and uses these anchor points to predict the position and size of the partial discharge features in the spectrogram. The main network structure of the YOLOv7 algorithm includes Input, Backbone network, Neck network and Head network, among which Input includes Mosaic data enhancement and adaptive anchor box calculation. Mosaic data enhancement: Randomly select 4 PRPD spectrograms for random cropping, scaling, and splicing to ensure that small targets are evenly distributed, enrich the data set, and enhance robustness. Adaptive anchor frame calculation: The size of the initial anchor frame needs to be set before network training. During training, a predicted frame will be generated based on the initial anchor frame, and compared with the manually annotated real frame to calculate the loss function. The YOLOv7 target detection model is trained using the annotated single-source partial discharge PRPD spectrum training set to enable the model to learn to identify the spectrum features of different types of defects.

[0058] The improved Soft-NMS algorithm replaces the traditional NMS algorithm. The Soft-NMS algorithm is used to deal with the problem of bounding box overlap, better solve the problem of target overlap, and improve the accuracy of target detection. The Generalized Intersection over Union (GIoU) ​​is used instead of the traditional Intersection over Union (IoU) as the performance evaluation indicator. GIoU takes into account the alignment of the bounding box, which helps to improve the stability and accuracy of the model performance.

[0059] The Soft-NMS algorithm and the performance evaluation index GIoU were applied to the YOLOv7 target detection model for iterative training to further optimize the model. The improved YOLOv7 target detection model and the trained weights were used to infer the multi-source partial discharge PRPD spectrum test set to verify the effectiveness of the algorithm.

[0060] The size of the output tensor of the YOLOv7 object detection model depends on the network architecture and the size of the input image. Typically, the output of the model includes multiple feature maps, each corresponding to a detection of a different scale. The output tensor usually contains the following dimensions:

[0061] Batch size: The number of images processed at a time.

[0062] Grid size: The number of grid cells in each feature map. For example, if the input image is divided into a 13x13 grid, then the grid size is 13x13.

[0063] Number of predictions per grid cell: Each grid cell predicts multiple bounding boxes, and the information of each bounding box includes the coordinates of the bounding box, the confidence (whether there is an object in the bounding box), and the category probability.

[0064] Assuming that we plan to use the YOLOv7 target detection model to perform PRPD data analysis every 5 minutes, we adopt the following strategy to accumulate the output of the YOLOv7 target detection model and use this data to train the model to more accurately judge and detect partial discharge.

[0065] Store YOLOv7 output: Store the YOLOv7 output of each analysis in a database or file system, including timestamp, detected object location, size, confidence, and category.

[0066] Annotate Data: Annotate important outputs, especially those with known partial discharge events.

[0067] Aggregation and summary statistics: Aggregate the output of the YOLOv7 object detection model over consecutive time periods, such as calculating the frequency, size change, and location change of local discharge events detected over a specific time period.

[0068] Time Series Analysis: Use time series analysis techniques to identify discharge patterns and periodic changes.

[0069] Supervised learning: Use the accumulated labeled data to train a classifier or regression model to predict future PD events or their severity.

[0070] Transfer learning: Using the standard YOLOv7 object detection model as a starting point, fine-tune the model to make it more suitable for the specific PRPD data and discharge type.

[0071] The recognition performance of the YOLOv7 object detection model is evaluated on the test set. The performance evaluation indicators include multiple types of accuracy, recall, and F1 score. The model is adjusted and optimized according to the test results, including adjusting model parameters, improving data preprocessing steps, or further optimizing the NMS algorithm.

[0072] Specific options for adjusting model parameters include:

[0073] According to the characteristics of the PRPD spectra of local discharge of different types of defects, the configuration file of the YOLOv7 target detection model, including the anchor box size and the number of categories, is adjusted to make the model more suitable for specific PRPD data and discharge types.

[0074] Specific solutions to improve the data preprocessing steps include:

[0075] Normalize pixel values: Scaling pixel values ​​to a standard range, such as 0 to 1, through min-max scaling or Z-score normalization helps the model converge faster during training and improves model performance.

[0076] Data augmentation: Artificially increasing the size of a dataset by creating modified versions of images, including flipping, rotation, scaling, and color adjustments.

[0077] Automatic Data Augmentation (AutoAugment): Uses search algorithms to build and find optimal data augmentation strategies.

[0078] Region magnification: A targeted enhancement method to improve the model’s detection performance for small objects.

[0079] CutMix and MixUp: Improve the generalization ability of the model and prevent overfitting, suitable for some tasks that do not require precise local feature positioning.

[0080] Mosaic data enhancement: synthesize multiple images to allow the model to recognize multi-scale objects.

[0081] Random noise injection: Improve the robustness of the model in noisy environments.

[0082] Random Erasing: Enhances the model’s tolerance to partial occlusion.

[0083] Perspective transformation: allows the model to adapt to the recognition of targets under different perspectives.

[0084] Specific solutions to further optimize the NMS algorithm include:

[0085] Adjust the NMS threshold:

[0086] The NMS threshold determines which bounding boxes are considered overlapping and should be suppressed. A lower threshold will result in more false positives, and a higher threshold will miss some real objects. Control the strictness of the detection by adjusting the threshold.

[0087] Using Soft-NMS:

[0088] Soft-NMS improves NMS by downgrading the confidence of overlapping bounding boxes instead of suppressing them completely, preserving more detection results while reducing the information loss caused by NMS.

[0089] Apply category specific NMS thresholds:

[0090] Different categories of objects require different NMS thresholds.

[0091] Introducing DIoU-NMS or CIoU-NMS: handling bounding boxes of different sizes and shapes more accurately.

[0092] Optimize the implementation of NMS:

[0093] By using a more efficient data structure to manage bounding boxes, the complexity of calculating IoU is reduced and the speed of NMS is accelerated.

[0094] Use multithreading or parallel processing:

[0095] Suppression based on each bounding box is independent and NMS is performed in parallel

[0096] Dynamically adjust NMS parameters:

[0097] Dynamically adjust the NMS parameters according to the complexity of the input image or the density of the object.

[0098] Combining confidence and NMS:

[0099] Before performing NMS, the bounding boxes are sorted according to their confidence scores and only those bounding boxes with confidence scores above a certain threshold are considered, reducing the number of bounding boxes that NMS needs to process.

[0100] Experiments and Evaluation:

[0101] Experiment with different NMS optimization strategies and use the validation set to evaluate their impact on model performance. Choose the strategy that provides the best balance between accuracy, recall, and F1 score.

[0102] The recognition performance of the model is regularly evaluated to ensure that it can accurately predict partial discharge events, and the model is regularly updated based on the results of the performance evaluation and newly collected data.

[0103] Embodiment 2:

[0104] This embodiment provides a multi-source partial discharge intelligent identification system based on the YOLOv7 algorithm, including:

[0105] A data processing module, used to obtain the multi-source partial discharge PRPD spectrum to be identified and perform pre-processing;

[0106] The intelligent recognition module substitutes the data into the optimized YOLOv7 target detection model to obtain the partial discharge recognition results.

[0107] Among them, the optimization steps of the YOLOv7 target detection model are as follows:

[0108] The single-source partial discharge PRPD spectrum is obtained as a training set, and the multi-source partial discharge PRPD spectrum is obtained as a test set for data preprocessing;

[0109] Use the training set to train the YOLOv7 object detection model to learn the spectral features for identifying different types of defects;

[0110] Apply the Soft-NMS algorithm and the performance evaluation index GIoU to the YOLOv7 target detection model, perform iterative training, and further optimize the model;

[0111] The recognition performance of the trained YOLOv7 target detection model is evaluated on the test set, and the model is adjusted and optimized based on the test results.

[0112] The rest is the same as in Example 1.

[0113] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A multi-source partial discharge intelligent identification method based on YOLOv7 algorithm, characterized in that: The following steps are involved: Obtain the multi-source partial discharge PRPD spectrum to be identified and perform data preprocessing; The multi-source partial discharge PRPD spectra after data preprocessing are used as the input of the YOLOv7 target detection model to obtain the partial discharge recognition results; Among them, the optimization steps of the YOLOv7 target detection model are as follows: The single-source partial discharge PRPD spectrum is obtained as a training set, and the multi-source partial discharge PRPD spectrum is obtained as a test set for data preprocessing; Using the training set to train the YOLOv7 target detection model, learning to identify spectral features of different types of defects; Apply the Soft-NMS algorithm and the performance evaluation index GIoU to the YOLOv7 target detection model, perform iterative training, and further optimize the model; The recognition performance of the trained YOLOv7 target detection model is evaluated on the test set, and the model is adjusted and optimized according to the test results.

2. The multi-source partial discharge intelligent identification method based on the YOLOv7 algorithm according to claim 1 is characterized in that: The data preprocessing includes multiple processing such as grayscale processing, normalization, dimensionality reduction processing and noise reduction processing.

3. The multi-source partial discharge intelligent identification method based on the YOLOv7 algorithm according to claim 1 is characterized in that: The main network structure of the YOLOv7 target detection model includes Input, Backbone network, Neck network and Head network.

4. The multi-source partial discharge intelligent identification method based on the YOLOv7 algorithm according to claim 3 is characterized in that: The Input includes Mosaic data enhancement and adaptive anchor box calculation.

5. The multi-source partial discharge intelligent identification method based on the YOLOv7 algorithm according to claim 1 is characterized in that: When training the YOLOv7 target detection model, the configuration file of the YOLOv7 target detection model is adjusted according to the characteristics of the partial discharge PRPD spectrum.

6. The multi-source partial discharge intelligent identification method based on the YOLOv7 algorithm according to claim 5 is characterized in that: The configuration file of the YOLOv7 object detection model includes the anchor box size and the number of categories.

7. The multi-source partial discharge intelligent identification method based on the YOLOv7 algorithm according to claim 1 is characterized in that: The Soft-NMS algorithm is used to separate and reconstruct the multi-source partial discharge PRPD spectrum.

8. The multi-source partial discharge intelligent identification method based on the YOLOv7 algorithm according to claim 1 is characterized in that: The GIoU as a performance evaluation indicator includes multiple ones of accuracy, recall and F1 score.

9. The multi-source partial discharge intelligent identification method based on the YOLOv7 algorithm according to claim 1 is characterized in that: The adjustment and optimization include adjusting the parameters of the YOLOv7 target detection model, improving the data preprocessing steps and further optimizing the NMS algorithm.

10. A multi-source partial discharge intelligent identification system based on YOLOv7 algorithm, characterized in that: include: A data processing module, used to obtain the multi-source partial discharge PRPD spectrum to be identified and perform pre-processing; The intelligent recognition module uses the multi-source partial discharge PRPD spectrum after data preprocessing as the input of the YOLOv7 target detection model to obtain the partial discharge recognition result; Among them, the optimization steps of the YOLOv7 target detection model are as follows: The single-source partial discharge PRPD spectrum is obtained as a training set, and the multi-source partial discharge PRPD spectrum is obtained as a test set for data preprocessing; Using the training set to train the YOLOv7 target detection model, learning to identify spectral features of different types of defects; Apply the Soft-NMS algorithm and the performance evaluation index GIoU to the YOLOv7 target detection model, perform iterative training, and further optimize the model; The recognition performance of the trained YOLOv7 target detection model is evaluated on the test set, and the model is adjusted and optimized according to the test results.

Images