Partial discharge type classification method and system based on multi-algorithm model

Abstract

The invention relates to a partial discharge type classification method and system based on a multi-algorithm model, and the method comprises the steps: obtaining the time series data of a partial discharge event through a YOLO algorithm, and carrying out the preprocessing; taking the time sequence data as the input of the PatchTST model to obtain a discharge prediction result; inputting the discharge prediction result into a classifier for discharge type classification to obtain a classification result; wherein the training steps of the PatchTST model are as follows: segmenting the preprocessed time sequence data into Patches with the same size, and extracting the features of the Patches by using the PatchTST model; carrying out feature mapping on each Patch by using a trainable linear parameter matrix Wp; using the Patch features subjected to feature mapping to train a Patch TST model; and a self-supervised learning method is adopted to randomly cover a part of time sequence data and train a PatchTST model to reconstruct the covered data. Compared with the prior art, in-depth analysis of the partial discharge detection system on the development trend and type classification of the discharge event is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of partial discharge detection, and in particular to a partial discharge type classification method and system based on a multi-algorithm model. Background Art

[0002] Partial discharge (PD) is an early signal of insulation failure in power equipment, and its detection is crucial to preventing equipment failure and maintaining power system safety. Traditional PD detection methods rely on manual analysis or simple automated tools, which have limitations when processing complex image data. In recent years, deep learning technology, especially the YOLO (You Only Look Once) algorithm, has shown excellent performance in image recognition and target detection. For example, patent application CN 117612081 A discloses a YOLO-based automatic detection method for excess material in aircraft fuel tanks. This method uses the YOLO target detection algorithm to automatically detect the inside of an aircraft fuel tank, and can efficiently and accurately detect excess material inside the tank, thereby improving the safety and reliability of the aircraft fuel tank. Although the YOLO algorithm performs well in target detection, its feature extraction capability is limited when processing complex PD maps. For the subtle features and patterns of PD, more advanced feature extraction and analysis techniques are required. Existing PD detection systems lack in-depth analysis of the development trend and type classification of discharge events. Summary of the invention

[0003] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a partial discharge type classification method and system based on a multi-algorithm model, which significantly improves the accuracy of partial discharge type detection.

[0004] The purpose of the present invention can be achieved by the following technical solutions:

[0005] A partial discharge type classification method based on a multi-algorithm model comprises the following steps:

[0006] The YOLO algorithm is used to obtain the time series data of partial discharge events for preprocessing;

[0007] Using the time series data of the local discharge event as input of the PatchTST model to obtain a discharge prediction result of the local discharge event;

[0008] Inputting the discharge prediction result into a classifier to classify the discharge type and obtain a classification result;

[0009] Among them, the training steps of the PatchTST model are as follows:

[0010] The preprocessed time series data is divided into patches of equal size, and the features of the patches are extracted using the PatchTST model;

[0011] Use a trainable linear parameter matrix Wp to feature map each patch and convert it into a potential representation that the Transformer encoder can process;

[0012] Use the feature-mapped Patch features to train the PatchTST model to learn local features and long-term dependencies in time series data;

[0013] A self-supervised learning method is used to randomly mask part of the time series data and train the PatchTST model to reconstruct the masked data.

[0014] Furthermore, the preprocessing includes multiple ones of denoising, normalization and processing of missing values.

[0015] Furthermore, the discharge prediction result includes a prediction of future discharge events and an analysis of existing discharge patterns.

[0016] Furthermore, the patch features are extracted by using a Transformer encoder and a multi-head attention mechanism.

[0017] Furthermore, a position encoding matrix Wpos is introduced to capture the temporal order information between the patches to predict and classify partial discharge events.

[0018] Furthermore, whether the patches overlap depends on the design and analysis requirements of the PatchTST model.

[0019] Furthermore, the classifier is trained to identify different discharge types using labeled local discharge event data, wherein the data includes various discharge pattern samples labeled with corresponding discharge types.

[0020] Furthermore, the classifier includes multiple ones of support vector machines, random forests or neural networks.

[0021] Furthermore, according to the difference between the prediction result and the actual observation data, the parameters and structure of the PatchTST model are adjusted to perform iterative optimization.

[0022] According to another aspect of the present invention, there is provided a partial discharge type classification system based on a multi-algorithm model, comprising:

[0023] A data processing module is used to collect time series data of partial discharge events and perform preprocessing;

[0024] The prediction module uses the preprocessed time series data as the input of the PatchTST model to obtain the discharge prediction results of the partial discharge event;

[0025] A classification module, using a classifier to classify the discharge mode of the partial discharge event into a discharge type to obtain a classification result;

[0026] Among them, the training steps of the PatchTST model are as follows:

[0027] The preprocessed time series data is divided into patches of equal size, and the features of the patches are extracted using the PatchTST model;

[0028] Use a trainable linear parameter matrix Wp to feature map each patch and convert it into a potential representation that the Transformer encoder can process;

[0029] Use the feature-mapped Patch features to train the PatchTST model to learn local features and long-term dependencies in time series data;

[0030] A self-supervised learning method is used to randomly mask part of the time series data and train the PatchTST model to reconstruct the masked data.

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

[0032] 1. The present invention predicts partial discharge events by combining the YOLO algorithm with the PatchTST model, effectively solving the complexity and uncertainty problems in the prediction of multivariate time series of partial discharge, significantly improving the accuracy of partial discharge spectrum recognition, and providing an effective technical means for the maintenance and fault prevention of power systems.

[0033] 2. The present invention introduces a self-supervised pre-training task, randomly masks part of the data and trains the PatchTST model to reconstruct the masked data, so that the PatchTST model achieves better results than traditional supervised learning on large data sets, and plays an important role in the analysis of time series data related to partial discharge. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 The present invention is a flowchart of a partial discharge type classification method based on a multi-algorithm model. DETAILED DESCRIPTION

[0035] 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.

[0036] Example 1

[0037] This embodiment provides a partial discharge type classification method based on a multi-algorithm model, such as Figure 1 As shown, the following steps are included:

[0038] S1. Use the YOLO algorithm to obtain the time series data of partial discharge events and perform preprocessing.

[0039] In this embodiment, the YOLOv7 (You Only Look Once version 7) target detection algorithm is used, which is the latest version of the YOLO series algorithm. When encountering overlapping of multi-source partial discharge in the PRPD spectrum, YOLOv7 can automatically detect the spectrum features from a PRPD spectrum and give their category information. The output tensor of YOLOv7 is a three-dimensional tensor containing information about the detected target, such as the category, position, confidence, etc. of the target. Specifically, if multi-scale detection is considered, the size of the output tensor may be [batch_size, grid_size, grid_size, num_anchors*(num_classes+5)], where batch_size is the batch size, grid_size is the size of the feature map, num_anchors is the number of anchor boxes, and num_classes is the number of target categories. The size of the output tensor of YOLOv7 depends on the configuration of the network structure and the size of the input image.

[0040] The time series data of partial discharge events are preprocessed, including denoising, normalization, and processing of missing values, to ensure that the data is suitable for deep learning analysis.

[0041] S2. Using the time series data of the local discharge event as the input of the PatchTST model to obtain the discharge prediction result of the local discharge event.

[0042] The training steps of the PatchTST model are as follows:

[0043] According to the requirements of the PatchTST model, the time series data is split into subsequences of equal size, namely "Patches", which can be overlapping or non-overlapping, depending on the model design and analysis requirements, and the features of each Patch are extracted. The trainable linear parameter matrix Wp is used to map the features of each Patch and convert it into a potential representation that the Transformer encoder can handle.

[0044] A trainable parameter matrix Wp is used for feature mapping, which maps the input data (e.g., patches of a time series) to a higher-dimensional latent space. The linear parameter matrix Wp is part of the model, and its parameters are updated during training through the back-propagation algorithm to learn patterns and features in the data.

[0045] The position encoding matrix Wpos is introduced to capture the temporal order information between different Patches to predict and classify local discharge events. Patches are processed using a multi-head attention mechanism and a Transformer encoder, and the extracted features are used to train the PatchTST model to learn local features and long-term dependencies in time series data; after training, the PatchTST model is used to predict local discharge events, and the output of the PatchTST model includes: prediction of future discharge events and analysis of existing discharge patterns.

[0046] The purpose of position encoding is to provide the model with position information about each element (token or patch) in the sequence. It is usually a fixed matrix that is used to encode the position information into the input of the model, ensuring that the model can capture the sequential relationship of the elements in the sequence. Position encoding can be learned or predefined (such as encoding generated by sine and cosine functions). In PatchTST, position encoding can be used to capture the relative temporal relationship between different time series patches.

[0047] A self-supervised learning method is adopted to enhance the ability of PatchTST model to capture time series patterns by randomly masking part of the data and training the PatchTST model to reconstruct the randomly masked part of the data.

[0048] The results predicted by the PatchTST model are explained, the changes of discharge events over time are analyzed, and the future discharge patterns are predicted. According to the difference between the predicted results and the actual observed data, the parameters and structure of the PatchTST model are adjusted and iteratively optimized.

[0049] S3. Input the discharge prediction result into a classifier to classify the discharge type and obtain a classification result.

[0050] The multi-head attention mechanism and Transformer encoder are used to extract features from the time series data of partial discharge events, including waveform features, frequency domain features, time domain features, etc., for training classifiers. The labeled partial discharge event data is used to train the classifier to identify different discharge types. These data contain samples of various discharge modes, and each sample is labeled with the corresponding discharge type to learn to identify different types of partial discharge events.

[0051] Classifiers include support vector machines (SVM), random forests, or neural networks, etc.

[0052] The trained classifier is used to classify the discharge type of the prediction results of the PatchTST model. The classifier identifies the type of each discharge event based on the learned pattern to obtain the classification result.

[0053] The classifier outputs the type of each partial discharge event, and these results can be used for further analysis or as a basis for maintenance decisions.

[0054] Example 2

[0055] This embodiment provides a partial discharge type classification system based on a multi-algorithm model, including:

[0056] A data processing module is used to collect time series data of partial discharge events and perform preprocessing;

[0057] The prediction module uses the preprocessed time series data as the input of the PatchTST model to obtain the discharge prediction results of the partial discharge event;

[0058] A classification module, using a classifier to classify the discharge mode of the partial discharge event into a discharge type to obtain a classification result;

[0059] Among them, the training steps of the PatchTST model are as follows:

[0060] The preprocessed time series data is divided into patches of equal size, and the features of the patches are extracted using the PatchTST model;

[0061] Use a trainable linear parameter matrix Wp to feature map each patch and convert it into a potential representation that the Transformer encoder can process;

[0062] Use the feature-mapped Patch features to train the PatchTST model to learn local features and long-term dependencies in time series data;

[0063] A self-supervised learning method is used to randomly mask part of the time series data and train the PatchTST model to reconstruct the masked data.

[0064] The rest is the same as in Example 1.

[0065] 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 partial discharge type classification method based on a multi-algorithm model, characterized in that: The following steps are involved: The YOLO algorithm is used to obtain the time series data of partial discharge events for preprocessing; Using the time series data of the local discharge event as input of the PatchTST model to obtain a discharge prediction result of the local discharge event; Inputting the discharge prediction result into a classifier to classify the discharge type and obtain a classification result; Among them, the training steps of the PatchTST model are as follows: The preprocessed time series data is divided into patches of equal size, and the features of the patches are extracted using the PatchTST model; Use a trainable linear parameter matrix Wp to feature map each patch and convert it into a potential representation that the Transformer encoder can process; Use the feature-mapped Patch features to train the PatchTST model to learn local features and long-term dependencies in time series data; A self-supervised learning method is used to randomly mask part of the time series data and train the PatchTST model to reconstruct the masked data.

2. The partial discharge type classification method based on a multi-algorithm model according to claim 1 is characterized in that: The preprocessing includes multiple ones of denoising, normalization and processing of missing values.

3. The partial discharge type classification method based on a multi-algorithm model according to claim 1 is characterized in that: The discharge prediction results include predictions of future discharge events and analysis of existing discharge patterns.

4. The partial discharge type classification method based on a multi-algorithm model according to claim 1 is characterized in that: The patch features are extracted by using a Transformer encoder and a multi-head attention mechanism.

5. The partial discharge type classification method based on a multi-algorithm model according to claim 1 is characterized in that: The position encoding matrix Wpos is introduced to capture the temporal order information between the patches to predict and classify partial discharge events.

6. The partial discharge type classification method based on a multi-algorithm model according to claim 1 is characterized in that: Whether the patches overlap depends on the design and analysis requirements of the PatchTST model.

7. The partial discharge type classification method based on a multi-algorithm model according to claim 1 is characterized in that: The classifier is trained to identify different discharge types using labeled local discharge event data, wherein the data includes various discharge pattern samples labeled with corresponding discharge types.

8. The partial discharge type classification method based on a multi-algorithm model according to claim 7 is characterized in that: The classifier includes a support vector machine, a random forest or a neural network.

9. The partial discharge type classification method based on a multi-algorithm model according to claim 1, characterized in that: According to the difference between the predicted results and the actual observed data, the parameters and structure of the PatchTST model are adjusted and iterative optimization is performed.

10. A partial discharge type classification system based on a multi-algorithm model, characterized in that: include: A data processing module is used to collect time series data of partial discharge events and perform preprocessing; The prediction module uses the preprocessed time series data as the input of the PatchTST model to obtain the discharge prediction results of the partial discharge event; A classification module, using a classifier to classify the discharge mode of the partial discharge event into a discharge type to obtain a classification result; Among them, the training steps of the PatchTST model are as follows: The preprocessed time series data is divided into patches of equal size, and the features of the patches are extracted using the PatchTST model; Use a trainable linear parameter matrix Wp to feature map each patch and convert it into a potential representation that the Transformer encoder can process; Use the feature-mapped Patch features to train the PatchTST model to learn local features and long-term dependencies in time series data; A self-supervised learning method is used to randomly mask part of the time series data and train the PatchTST model to reconstruct the masked data.

Images