A Transformer Partial Discharge Separation Algorithm Based on Self-Supervised Networks

Abstract

This invention discloses a transformer partial discharge separation algorithm based on a self-supervised network, comprising the following steps: acquiring and storing partial discharge signals of the transformer; extracting discharge waveforms using an outlier detection method; constructing a self-contrast dataset and a negative example dataset for pulse data; extracting signal spectral features using a supervised contrastive network and training the network's spectral feature extraction capability; finally, classifying the features using an iterative self-organizing clustering algorithm; setting initialization information in the iterative self-organizing clustering algorithm, clustering the feature data obtained in step five, and after classification, determining the discharge type of data in each category using expert knowledge to obtain the discharge type result; this invention, by classifying and calculating discharge signals, can effectively calculate the discharge quantity and total discharge of different discharge faults, realizing the functions of transformer operating condition monitoring and fault early warning, and effectively protecting the safety of the transformer.

Description

Technical Field

[0001] This invention relates to the field of transformer partial discharge detection technology, specifically a transformer partial discharge separation algorithm based on a self-supervised network. Background Technology

[0002] Current transformer fault detection methods mostly employ oil chromatography, which suffers from long analysis times and the need for periodic sampling and analysis of the insulating oil, making real-time monitoring of transformers difficult. Furthermore, most algorithms targeting partial discharge signals in transformers are based on simulation datasets, resulting in a lack of generalization ability.

[0003] To address the aforementioned issues, we propose a transformer partial discharge separation algorithm based on a self-supervised network to solve the problems mentioned above. Summary of the Invention

[0004] The purpose of this invention is to provide a transformer partial discharge separation algorithm based on a self-supervised network to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A transformer partial discharge separation algorithm based on self-supervised networks.

[0007] Step 1: Acquire and store the partial discharge signal of the transformer;

[0008] Step 2: Extract the discharge waveform using an outlier detection method;

[0009] Step 3: Constructing the confidence-based dataset and negative example dataset for the impulse data;

[0010] Step 4: Use a supervised comparison network to extract signal spectral features and train the network's spectral feature extraction capability;

[0011] Step 5: Finally, the iterative self-organizing clustering algorithm is used to classify the features;

[0012] Step 6: Set the initialization information in the iterative self-organizing clustering algorithm, cluster the feature data obtained in Step 5, and after the classification is completed, use expert knowledge to determine the discharge type of the data in each category to obtain the discharge type result.

[0013] As a further aspect of the present invention: Step one specifically includes: acquiring data of partial discharge pulses in a certain frequency band using signal acquisition equipment, receiver, acquisition network card, directional antenna, and other equipment, and saving the acquired partial discharge data to a laptop or local storage device.

[0014] As a further aspect of the present invention: Step two specifically includes: The partial discharge signal of the transformer has a periodicity. In each discharge cycle, there is a most similar discharge pulse data in the discharge pulse of the next cycle. By calculating the spatial distance, the most similar pulse data and the least similar pulse data are extracted for each pulse signal in other discharge cycles. The similar data constitute a dataset called the confidence pair dataset, and the dissimilar data constitute another dataset called the negative example dataset.

[0015] As a further aspect of the present invention: Step three specifically includes: selecting the discharge pulse whose discharge type needs to be determined, and calculating the spatial distance with each discharge pulse data in other discharge cycles to obtain a pulse data with the largest spatial distance and a pulse data with the smallest spatial distance. The calculation of spatial distance requires first extracting each pulse from the same dimension.

[0016] As a further aspect of the present invention: the data with the largest spatial distance is the negative example of the judgment pulse, and the data with the smallest spatial distance is the positive example of the judgment pulse. The same spatial distance calculation is performed on each pulse signal in the dataset through a cyclic operation.

[0017] As a further aspect of the present invention, step four specifically includes: inputting the pulse signal spectrum data for which the discharge type needs to be determined into the supervised comparison network model, constructing a supervised task to train the network to extract the spectral features of the signal data, wherein the supervised task involves constructing a loss function so that the features of the input data extracted by the supervised comparison network are close to the features extracted from the positive data, and far from the features extracted from the negative data; through the training of the supervised task, the network acquires the ability to extract effective features of the input signal, and through the pulse signal feature extraction process of the supervised comparison network, the feature information of the pulse data for which the discharge type needs to be determined is obtained.

[0018] As a further aspect of the present invention: Step five specifically includes: setting training parameters, including initializing the number of cluster centers, setting the minimum number of samples in each cluster, the standard deviation threshold of the data in each cluster, the shortest distance between two different cluster centers, the number of allowed iterations, etc.; clustering the obtained feature data using the set initial values ​​until the training iteration that satisfies the initial values ​​ends, and obtaining different feature categories.

[0019] As a further aspect of the present invention, the clustering specifically includes: during the clustering process, the algorithm readjusts the number of cluster centers according to the initial value, and re-clusters until the fixed requirements in the initial value are finally met to obtain the optimal clustering effect.

[0020] Compared with the prior art, the beneficial effects of the present invention are:

[0021] 1. This invention can detect the discharge status of transformers and classify the discharge signals over a period of time by combining digital signal processing technology and artificial intelligence algorithm technology. It can effectively calculate the discharge amount and total discharge of different discharge faults, realize the function of transformer operating condition monitoring and fault early warning, and effectively protect the safety of transformers.

[0022] 2. This invention solves the problem that the lack of standard waveform labels for discharge pulse data makes supervised training difficult, and also solves the problems of long detection cycles and poor real-time performance of traditional oil chromatography methods.

[0023] 3. The algorithm provided by this invention, combined with technologies such as web interface and data processing, can display the discharge status of transformers in real time through a software platform, and show the discharge status of different discharge categories. It can provide early warning for equipment that is about to experience a discharge fault, reminding staff to carry out timely maintenance and protection, and has strong practical value.

[0024] 4. In addition, this invention collects and analyzes real-time transformer discharge data, which avoids the disadvantage of poor algorithm generalization caused by using simulation datasets. Attached Figure Description

[0025] Figure 1 This is a system flowchart of an embodiment of the present invention;

[0026] Figure 2 This is a normal distribution diagram of partial discharge signals provided in an embodiment of the present invention;

[0027] Figure 3 This is a result image of outlier detection provided in an embodiment of the present invention;

[0028] Figure 4 This is a waveform diagram of a discharge pulse in an embodiment of the present invention;

[0029] Figure 5 This is a schematic diagram of the self-supplied data and negative sample data in an embodiment of the present invention;

[0030] Figure 6 This is a schematic diagram of the supervised comparison learning process in an embodiment of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] like Figure 1 As shown, one embodiment of the present invention provides a transformer partial discharge separation algorithm based on a self-supervised network, including: collecting and storing the partial discharge signal of the transformer, extracting the discharge waveform using an outlier detection method, constructing a self-contrast data set and a negative example dataset of pulse data, extracting signal spectral features using a supervised contrast network, training the network's spectral feature extraction capability, and finally classifying the features using an iterative self-organizing clustering algorithm.

[0033] The specific process includes:

[0034] Step 1: Collect the partial discharge signal of the transformer, process and analyze the data to obtain the time-domain waveform signal of the discharge;

[0035] Step 2: Observe the distribution of signal data, use an outlier detection algorithm based on normal distribution to extract the signal waveform, and remove the background noise from the original data;

[0036] Step 3: Extract the waveform signal from the discharge pulse, calculate the confidence pair data and negative example data of each pulse data for which the discharge type needs to be determined in other discharge cycles, and construct the confidence pair dataset and negative example dataset;

[0037] Step 4: Set up a supervised training task to train the data feature extraction capability of the supervised contrast network. This includes inputting the dataset obtained in Step 3 and the pulse data to be judged into the network. By setting up a supervised training task, the feature data extracted from the pulse data to be judged is similar to the feature data obtained from the positive dataset, while the feature data is not similar to the data features in the negative examples. This results in a supervised contrast network that can effectively extract data features.

[0038] Step 5: Extract feature information of the pulse data to be judged using the trained supervised contrastive network;

[0039] Step Six: Set the initialization information in the iterative self-organizing clustering algorithm and cluster the feature data obtained in Step Five. After classification is complete, use expert knowledge to determine the discharge type of the data in each category to obtain the discharge type result.

[0040] The acquisition and storage of partial discharge signals from the transformer specifically includes acquiring partial discharge pulses within a certain frequency band using signal acquisition equipment, a receiver, a network card, and a directional antenna, and saving the acquired partial discharge data to a laptop or local storage device. This partial discharge data contains a significant amount of background noise, which can affect the effectiveness of operations such as feature extraction.

[0041] Outlier detection is used to remove background noise and extract the signal waveform of partial discharge. Specifically, it includes: first, drawing a distribution map of the local data, as shown below. Figure 2 As shown, the condition that the data distribution satisfies the normal distribution is determined, and the outlier detection algorithm based on the normal distribution is used to extract the signal waveform of partial discharge.

[0042] The outlier detection process based on normal distribution includes: calculating the mean (x) and standard deviation (std(x)) of partial discharge data, where x is the input data; setting outlier detection boundaries up_board and low_board, where up_board is equal to the mean of the data plus a certain multiple of the standard deviation, typically set to 3 times; and low_board is the mean of the data minus the same multiple of the standard deviation. After setting the outlier detection boundaries, the data is judged as an outlier by whether it falls within the boundaries. The data judged as outliers are saved and a time label is added. This step can effectively remove noise from the original data and extract the waveform of the signal.

[0043] The construction of the confidence-based data and negative example datasets for the pulse data includes extracting the pulse data waveforms from the original data to obtain individual pulse signal waveforms within each cycle, selecting the signal waveforms to be judged, and calculating the spatial distance with pulse waveform data from other cycles to find the most similar and least similar data in other discharge cycles, such as... Figure 4 The image shows a pulse waveform that needs to be used to determine the type of discharge. Figure 5 sub Figure 1-3 For data with a similar pulse waveform, and Figures 4-6 The negative examples are those dissimilar to the pulse waveform data. The most similar data in each cycle is extracted to obtain the positive example dataset, and the least similar data in each cycle is extracted to obtain the negative example dataset, thus completing the dataset construction. The spatial distance can be calculated using Euclidean distance.

[0044]

[0045] Where d Space_dis The space distance between two waveforms is represented by k, which represents the data dimension of each waveform, and x represents the space distance between two waveforms. i and yi For data points in waveform data, the data of each waveform must be unified to the same dimension before calculating spatial distance.

[0046] The confidence pair dataset consists of the waveform with the smallest spatial distance from the judgment pulse data in each discharge cycle; the negative example dataset consists of the waveform data with the largest spatial distance from the judgment pulse data in each discharge cycle.

[0047] The method described above uses a supervised contrastive network to extract signal spectral features and trains the network's spectral feature extraction capabilities. This includes constructing a supervised task, using positive example data as positive samples and negative example datasets as negative samples. During the extraction of pulse features, a loss function is constructed to make the extracted features more similar to the features of positive samples in high-dimensional space and more distant from the features of negative samples. By iteratively training the network's feature extraction capabilities, feature data that is similar to the positive example data in high-dimensional space but dissimilar to the features of negative examples in high-dimensional space is obtained.

[0048] Furthermore, the iterative self-organizing clustering algorithm for feature classification specifically includes: using the feature data obtained in the previous step as input, initializing some fixed parameters, which are also the boundaries that must be followed during training iterations, and training through the constraints of these initialized data to finally cluster the features of the pulse data to be judged; wherein the initialized data includes the number of initial cluster centers, the minimum number of samples in each cluster, the standard deviation threshold of the data in each cluster, the shortest distance between two different cluster centers, the maximum number of allowed iterations, and clustering through the aggregation and splitting characteristics of the network to finally obtain the different categories of features;

[0049] Further analysis of the classification results based on expert knowledge can determine the discharge type of the discharge pulses within the data acquisition time, and calculate the total discharge amount and discharge amount of different discharge types within that time period. After obtaining different feature clustering results, the specific type of discharge category can be determined according to the generation principle of the partial discharge signal. Different discharge types match different discharge principles, resulting in differentiated information in the waveform data. This further yields the discharge type of the pulse signal in the transformer discharge data. Simultaneously, time accumulation and type judgment can determine the total discharge amount and discharge amount of different discharge types within that time period, thereby enabling real-time detection and protection of transformers. This result has significant value for discharge detection and real-time operating condition monitoring of transformers.

[0050] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Although this specification describes embodiments, not every embodiment contains only one technical solution. This method of description is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A transformer partial discharge separation algorithm based on a self-supervised network, characterized in that: Specifically, the steps include the following: Step 1: Acquire and store the partial discharge signal of the transformer; Step 2: Extract the discharge waveform using an outlier detection method; Step 3: Constructing the confidence-based dataset and negative example dataset for impulse data; Step 4: Use a supervised comparison network to extract signal spectral features and train the network's spectral feature extraction capability; Step 5: Finally, the iterative self-organizing clustering algorithm is used to classify the features; Step 6: Set the initialization information in the iterative self-organizing clustering algorithm, cluster the feature data obtained in Step 5, and after the classification is completed, use expert knowledge to determine the discharge type of the data in each category to obtain the discharge type result; The construction of the confidence pair dataset and negative example dataset of the pulse data includes extracting the pulse data waveform from the original data to obtain the individual pulse signal waveform in each cycle, selecting the signal waveform to be judged, calculating the spatial distance with the pulse waveform data in other cycles, finding the most similar and least similar data in other discharge cycles, extracting the most similar data in each discharge cycle to obtain the confidence pair dataset, and extracting the least similar data in each discharge cycle to obtain the negative example dataset, thus completing the construction of the confidence pair dataset and negative example dataset; For a single pulse signal waveform that needs to be judged, the spatial distance between it and the pulse waveform data in other discharge cycles is calculated. The positive example dataset consists of the waveform data with the smallest spatial distance to the pulse signal waveform that needs to be judged in each discharge cycle; the negative example dataset consists of the waveform data with the largest spatial distance to the pulse signal waveform that needs to be judged in each discharge cycle.

2. The transformer partial discharge separation algorithm based on a self-supervised network according to claim 1, characterized in that, Step one specifically includes: acquiring data of partial discharge pulses in a certain frequency band using signal acquisition equipment, receiver, acquisition network card and directional antenna, and saving the acquired partial discharge data to a laptop or local storage device.

3. The transformer partial discharge separation algorithm based on a self-supervised network according to claim 1, characterized in that, Step three specifically includes: The partial discharge signal of the transformer has a periodicity. In each discharge cycle, there is a most similar discharge pulse data in the discharge pulse of the next cycle. By calculating the spatial distance, the most similar pulse data and the least similar pulse data are extracted for each pulse signal in other discharge cycles. The similar data constitute a dataset called the confidence pair dataset, and the dissimilar data constitute another dataset called the negative example dataset.

4. The transformer partial discharge separation algorithm based on a self-supervised network according to claim 1, characterized in that, Step three specifically includes: selecting the discharge pulse whose discharge type needs to be determined, and calculating the spatial distance between the pulse data and each discharge pulse data in other discharge cycles to obtain the pulse data with the largest spatial distance and the pulse data with the smallest spatial distance. Before calculating the spatial distance, each pulse must be truncated to the same dimension.

5. The transformer partial discharge separation algorithm based on a self-supervised network according to claim 4, characterized in that, The data with the largest spatial distance is the negative example of the corresponding discharge pulse for determining the discharge type, and the data with the smallest spatial distance is the positive example of the corresponding discharge pulse. The same spatial distance calculation is performed on each pulse signal in the dataset through a cyclic operation.

6. The transformer partial discharge separation algorithm based on a self-supervised network according to claim 1, characterized in that, Step four specifically includes: inputting the pulse signal spectrum data for which the discharge type needs to be determined into the supervised comparison network model; constructing a supervised task to train the network's ability to extract features from the signal spectrum data; wherein the supervised task involves constructing a loss function so that the features extracted from the input data by the supervised comparison network are close to the features extracted from the positive example data and far from the features extracted from the negative example data; through the training of the supervised task, the network acquires the ability to extract effective features from the input signal; and through the pulse signal feature extraction process performed by the supervised comparison network, the feature information of the pulse data for which the discharge type needs to be determined is obtained.

7. The transformer partial discharge separation algorithm based on a self-supervised network according to claim 1, characterized in that, Step five specifically includes: setting training parameters, including initializing the number of cluster centers, setting the minimum number of samples in each cluster, the standard deviation threshold of the data in each cluster, the shortest distance between two different cluster centers, and the allowed number of iterations; clustering the obtained feature data using the set initial values ​​until the set initial values ​​are met, then ending the iterative clustering to obtain different feature categories.

8. The transformer partial discharge separation algorithm based on a self-supervised network according to claim 7, characterized in that, The clustering specifically includes: during the clustering process, the algorithm readjusts the number of cluster centers according to the initial value, and re-clusters until the fixed requirements in the initial value are met to obtain the optimal clustering effect.

Images