Partial discharge signal analysis method, system and device and readable storage medium

Abstract

The invention relates to the technical field of signal analysis, and discloses a partial discharge signal analysis method, system and device and a readable storage medium, and the method comprises the steps: a data acquisition stage: constructing a partial discharge characteristic matrix which comprises energy values of partial discharge signals at different scales and time windows, and discharge times and phase values corresponding to the time windows; a pre-training stage: pre-training a feature extractor and a classifier by using the source domain data; in the formal training stage, the source domain data and the target domain data are utilized, a pre-trained feature extractor and a domain discriminator are trained by adopting an adversarial game strategy, and an adversarial domain adaptive network is constructed; and a model updating stage: utilizing the classifier to generate pseudo labels of unlabeled samples in the target domain data, adding the pseudo label data of the target domain into the training set to re-train the feature extractor, the classifier and the domain discriminator, and realizing label distribution for each unlabeled sample in the target domain data. According to the invention, the partial discharge signals are monitored on line by using an unsupervised classification method.

Description

Technical Field

[0001] The present invention relates to the technical field of signal analysis, and particularly to a method, system, device and readable storage medium for analyzing partial discharge signals. Background Art

[0002] During the long-term operation of power equipment, its insulation system gradually ages due to various factors, thus triggering partial discharge phenomena. Partial discharge is a discharge phenomenon that occurs in a partial insulation area between two conductive electrodes with a gap. Partial discharge is an early sign of insulation failure of power equipment. If not monitored and diagnosed in time, it may lead to insulation breakdown or other serious faults, thereby affecting the stable operation of the power system. Traditional partial discharge detection methods mainly include pulse current method, ultrasonic method, optical method, etc. However, these methods are vulnerable to noise interference in complex electromagnetic environments, and their sensitivity and detection accuracy are limited.

[0003] In recent years, Ultra-High Frequency (UHF) electromagnetic wave signal analysis technology has been widely used in the field of partial discharge detection. By detecting the high-frequency electromagnetic wave signals generated by power equipment during operation, online monitoring of partial discharge signals can be achieved. However, existing UHF electromagnetic wave signal analysis methods cannot achieve efficient and accurate detection of partial discharge in power equipment due to defects such as noise interference, low detection accuracy, and lack of labeled samples. Summary of the Invention

[0004] The purpose of the present invention is to provide a method, system, device and readable storage medium for analyzing partial discharge signals based on unsupervised learning. Using the partial discharge feature matrix constructed by partial discharge signals as source domain data, and using the source domain data and its fault label information to label the target domain data under unknown working conditions, online monitoring of partial discharge signals and evaluating the health status of power equipment.

[0005] To achieve the above purpose, the present invention provides the following technical solutions:

[0006] In the first aspect, a method for analyzing partial discharge signals based on unsupervised learning is provided, including the following steps:

[0007] Data acquisition stage: Collect partial discharge signals of power equipment, perform multi-scale energy decomposition on the partial discharge signals and use a moving time window for signal division to construct a partial discharge feature matrix, where the partial discharge feature matrix includes the energy values of partial discharge signals at different scales and time windows, the number of discharges and phase values corresponding to the time windows;

[0008] Pretraining stage: The partial discharge feature matrices under different working conditions are respectively used as the source domain data with fault labels and the unlabeled target domain data. The feature extractor and classifier are pre-trained using the source domain data. The classifier is used to classify the source domain features extracted by the feature extractor;

[0009] Formal training stage: Using the source domain data and the target domain data, the feature extractor and domain discriminator are trained using the adversarial game strategy. By extracting domain-invariant features, the source domain and target domain features are aligned to construct an adversarial domain adaptation network;

[0010] Model update stage: The classifier is used to generate pseudo-labels for the unlabeled samples in the target domain data. The pseudo-labeled data in the target domain is added to the training set to retrain the feature extractor, the classifier, and the domain discriminator, and label assignment is achieved for each unlabeled sample in the target domain data.

[0011] Furthermore, in the data acquisition stage, an ultra-high frequency sensor is used to collect the ultra-high frequency electromagnetic wave signals generated by the partial discharge of the power equipment. After preprocessing the ultra-high frequency electromagnetic wave signals, the partial discharge signals are generated.

[0012] Furthermore, the frequency of the collected ultra-high frequency electromagnetic wave signals is 300 MHz - 3 GHz.

[0013] Furthermore, the preprocessing of the ultra-high frequency electromagnetic wave signals includes:

[0014] Signal sampling is performed using a high-speed data acquisition card to convert the analog ultra-high frequency electromagnetic wave signals collected by the ultra-high frequency sensor into digital ultra-high frequency electromagnetic wave signals;

[0015] The digital ultra-high frequency electromagnetic wave signals are filtered, amplified, calibrated, and denoised.

[0016] Furthermore, discrete wavelet transform is performed on the partial discharge signals, and the energy values of each layer of wavelet decomposition are calculated;

[0017] The wavelet-decomposed partial discharge signals are divided using a moving time window, and the energy values of the signals after each layer of wavelet decomposition in different time windows are obtained as the main dimension features;

[0018] The phase information and the number of discharges corresponding to the time window are obtained as additional dimension features;

[0019] The partial discharge feature matrix is constructed using the main dimension features and the additional dimension features.

[0020] Furthermore, performing discrete wavelet transform on the partial discharge signals and calculating the energy values of each layer of wavelet decomposition includes:

[0021] By means of the discrete wavelet transform method, decompose the partial discharge signal into approximation coefficients and detail coefficients of different scales:

[0022]

[0023] where x(t) is the partial discharge signal, J represents the decomposition level, A J is the wavelet approximation coefficient of the J-th layer, and D j is the wavelet detail coefficient of the j-th layer;

[0024] Calculate the energy value of each layer of wavelet decomposition of the partial discharge signal:

[0025]

[0026] where E j is the energy value of the j-th layer of wavelet decomposition, n j is the number of detail coefficients of the j-th layer of wavelet decomposition, and D j (i) is the i-th wavelet detail coefficient of the j-th layer;

[0027] Normalize the energy value of each layer of wavelet decomposition of the partial discharge signal:

[0028]

[0029] where the energy value NE j represents the proportion of the energy value of the j-th layer of wavelet decomposition in the total energy value.

[0030] Furthermore, the partial discharge feature matrix is:

[0031]

[0032] where E j,i represents the energy value of the j-th layer of wavelet decomposition of the partial discharge signal in the i-th time window, and the value range is [0, 1];

[0033] P i represents the number of discharges in the i-th time window, and the value range is [0, 1];

[0034] φ i represents the phase value in the i-th time window, and the value range is [0, 1].

[0035] Furthermore, the pre-training stage specifically includes:

[0036] Input the source domain data into the feature extractor to extract source domain features;

[0037] Input the extracted source domain features into the classifier for label classification to obtain predicted labels;

[0038] Calculate the classification loss by using the cross - entropy loss function for the predicted labels and the true labels of the source - domain samples, and back - propagate the gradient of the classification loss until the preset number of iterations is reached to obtain the pre - trained feature extractor and classifier.

[0039] Furthermore, the formal training stage specifically includes:

[0040] Input the sample data of the source domain and the target domain into the pre - trained feature extractor to extract the source - domain sample features and the target - domain sample features respectively;

[0041] Input the extracted source - domain sample features and target - domain sample features into the domain discriminator for domain classification, and use the cross - entropy loss function to calculate the domain adversarial loss;

[0042] After taking the opposite of the gradient of the domain adversarial loss through the gradient reversal layer and then performing back - propagation for adversarial training, iteratively update the network parameters until convergence. The trained feature extractor is used to extract the domain - invariant features of the source - domain and target - domain data in the high - dimensional space.

[0043] Furthermore, the model update stage specifically includes:

[0044] Input the unlabeled samples in the target - domain data into the classifier for label prediction, and generate pseudo - labels for the unlabeled samples with a confidence level higher than the threshold;

[0045] Combine the obtained target - domain pseudo - label data and the source - domain data to update the training set, and use the updated training set to retrain the feature extractor, the classifier, and the domain discriminator. Through self - training iteration, until each unlabeled sample in the target - domain data realizes label assignment.

[0046] In a second aspect, a partial discharge signal analysis system based on unsupervised learning is further provided, including:

[0047] A data acquisition module, which is used to collect the partial discharge signals of power equipment, perform multi - scale energy decomposition on the partial discharge signals and use a moving time window for signal division to construct a partial discharge feature matrix. The partial discharge feature matrix includes the energy values of the partial discharge signals at different scales and time windows, the number of discharges and phase values corresponding to the time windows;

[0048] A pre - training module, which is used to take the partial discharge feature matrices under different working conditions as the source - domain data with fault labels and the unlabeled target - domain data respectively, and use the source - domain data to pre - train the feature extractor and classifier. The classifier is used to classify the source - domain features extracted by the feature extractor;

[0049] A formal training module for training the feature extractor and the domain discriminator by using the source domain data and the target domain data and adopting an adversarial game strategy, realizing feature alignment between the source domain and the target domain by extracting domain-invariant features, and constructing an adversarial domain adaptation network;

[0050] A model update module for using the classifier to generate pseudo-labels for unlabeled samples in the target domain data, adding the pseudo-label data of the target domain to the training set to retrain the feature extractor, the classifier and the domain discriminator, and realizing label assignment for each unlabeled sample in the target domain data.

[0051] Based on the same inventive concept, the present invention also provides an electronic device, including: a memory, a processor; the processor is configured to read and execute a computer program stored in the memory to implement the foregoing method for analyzing partial discharge signals based on unsupervised learning.

[0052] Based on the same inventive concept, the present invention also provides a computer storage medium, in which computer-executable instructions are stored, and when the computer-executable instructions are executed, the foregoing method for analyzing partial discharge signals based on unsupervised learning is implemented.

[0053] The technical effects and advantages of the present invention:

[0054] (1) The present invention adopts an unsupervised classification method to analyze partial discharge signals, realizes online monitoring of partial discharge signals by assigning fault labels to each unlabeled sample in the target domain, and obtains the health status of power equipment;

[0055] (2) The present invention solves the problem of dependence on a large number of data labels in the training process of time series data, uses the labeled data in the source domain to assist the unlabeled data in the target domain for training, obtains an effective unsupervised classification model, avoids cumbersome data annotation work, and improves the training efficiency;

[0056] (3) The present invention adds adversarial training in the process of aligning the source domain and the target domain by using a feature mapping-based method, makes the adversarial training process more stable, enables the target domain to make full use of the labeled data in the source domain, and improves the unsupervised classification accuracy.

[0057] Other features and advantages of the present invention will be described in the following specification, and part of them will become obvious from the specification or be understood by implementing the present invention. The objectives and other advantages of the present invention can be achieved and obtained through the structures pointed out in the specification, the claims and the drawings. Description of the Drawings

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0059] Figure 1 It is a schematic flowchart of the partial discharge signal analysis method based on unsupervised learning in the first embodiment of the present invention;

[0060] Figure 2 It is a schematic structural diagram of the partial discharge signal analysis system based on unsupervised learning in the second embodiment of the present invention. Detailed implementation manners

[0061] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0062] The present invention discloses a partial discharge signal analysis method based on unsupervised learning, as Figure 1 shown, the method includes the following steps:

[0063] S1. Data acquisition stage: Collect the partial discharge signals of power equipment, perform multi-scale energy decomposition on the partial discharge signals and use a moving time window to divide the signals, and construct a partial discharge feature matrix, where the partial discharge feature matrix includes the energy values of the partial discharge signals at different scales and time windows, the number of discharges and phase values corresponding to the time windows;

[0064] S2. Pre-training stage: Respectively use the partial discharge feature matrices under different working conditions as the source domain data with fault labels and the unlabeled target domain data, and use the source domain data to pre-train the feature extractor and classifier. The classifier is used to classify the source domain features extracted by the feature extractor;

[0065] S3. Formal training stage: Use the source domain data and the target domain data to train the feature extractor and domain discriminator by adopting an adversarial game strategy, realize the alignment of source domain and target domain features by extracting domain-invariant features, and construct an adversarial domain adaptation network;

[0066] S4. Model Update Phase: Use the classifier to generate pseudo-labels for the unlabeled samples in the target domain data, add the pseudo-labeled data of the target domain to the training set, and retrain the feature extractor, the classifier, and the domain discriminator to assign labels to each unlabeled sample in the target domain data.

[0067] In the embodiments of the present invention, an unsupervised classification method is adopted to analyze partial discharge signals. By assigning fault labels to each unlabeled sample in the target domain, online monitoring of partial discharge signals is realized, and the health status of power equipment is obtained. In the process of aligning the source domain and the target domain based on the feature mapping method, adversarial training is added, making the process of adversarial training more stable. The target domain can make full use of the labeled data in the source domain to improve the unsupervised classification accuracy.

[0068] In the data acquisition phase of S1 above, use a ultra-high frequency sensor to collect the ultra-high frequency electromagnetic wave signals generated by partial discharge of power equipment, and generate the partial discharge signals after preprocessing the ultra-high frequency electromagnetic wave signals. According to a specific embodiment, the frequency of collecting the ultra-high frequency electromagnetic wave signals is 300 MHz - 3 GHz.

[0069] According to the embodiments of the present invention, preprocessing the ultra-high frequency electromagnetic wave signals includes: using a high-speed data acquisition card for signal sampling, converting the analog ultra-high frequency electromagnetic wave signals collected by the ultra-high frequency sensor into digital ultra-high frequency electromagnetic wave signals, and then performing filtering, amplification, calibration, and noise reduction processing on the digital ultra-high frequency electromagnetic wave signals.

[0070] The construction of the partial discharge feature matrix in step S2 above includes the following steps:

[0071] S21. Perform discrete wavelet transform on the partial discharge signals and calculate the energy values of each layer of wavelet decomposition;

[0072] S22. Use a moving time window to divide the partial discharge signals after wavelet decomposition, and obtain the energy values of the signals after each layer of wavelet decomposition in different time windows as the main dimension features;

[0073] S23. Obtain the phase information and the number of discharges corresponding to the time window as additional dimension features;

[0074] S24. Use the main dimension features and the additional dimension features to construct a partial discharge feature matrix.

[0075] Wavelet transform is a signal processing technique used to decompose a signal into wavelet basis functions of different scales (frequencies) and positions (times), obtain the local information of the signal at different times and frequencies, and at the same time retain the time and frequency characteristics of the signal. Wavelet transform is generally used to process non-steady signals. Its basic idea is to represent the signal as a linear combination of wavelet basis functions, and these wavelet basis functions can be the original wavelet mother function after translation and scaling. In practical applications, since the ultra-high frequency electromagnetic wave signal exists in the form of discrete samples in the computer, the discrete wavelet transform is used to process the ultra-high frequency electromagnetic wave signal.

[0076] According to the above step S21, perform discrete wavelet transform on the partial discharge signal and calculate the energy value of each layer of wavelet decomposition, including:

[0077] By using the discrete wavelet transform method, decompose the partial discharge signal into approximation coefficients and detail coefficients of different scales, where the approximation coefficients contain the trend information of the signal, and the detail coefficients contain the mutation information of the signal:

[0078]

[0079] Among them, x(t) is the partial discharge signal, J represents the decomposition level, A J is the wavelet approximation coefficient of the Jth layer, D j is the wavelet detail coefficient of the jth layer;

[0080] The number of layers of wavelet decomposition refers to the number of layers into which the original signal is decomposed when performing wavelet transform. When performing wavelet transform, the signal will be decomposed into wavelet coefficients of different scales and directions, and each scale and direction corresponds to one layer of wavelet decomposition. Each layer of wavelet decomposition will decompose the signal into low-frequency components and high-frequency components.

[0081] Then, calculate the energy value of each layer of wavelet decomposition of the partial discharge signal:

[0082]

[0083] Among them, E j represents the energy value of the jth layer of wavelet decomposition, n j is the number of detail coefficients of the jth layer of wavelet decomposition, D j (i) is the ith wavelet detail coefficient of the jth layer;

[0084] Then, perform normalization processing on the energy value of each layer of wavelet decomposition of the partial discharge signal:

[0085]

[0086] Among them, the energy value NE j represents the proportion of the energy value of the jth layer of wavelet decomposition in the total energy value.

[0087] According to the above step S22, a moving time window is used to divide the normalized energy values of the wavelet decomposition of each layer, and the energy values in different time windows are obtained as the main dimension features.

[0088] According to the above step S23, the time series data of the phase information and the number of discharges are obtained, and the moving time window is used for data division to obtain the phase information and the number of discharges in the corresponding time window. After normalizing the phase information and the number of discharges, they are used as additional dimension characteristics.

[0089] Calculate the normalized value of the number of discharges according to the following formula:

[0090]

[0091] where x′ is the normalized value of the original number of discharges x, and the numerical range is [0, 1]. min(x) and max(x) are the minimum and maximum values in the time series data of the number of discharges, respectively.

[0092] Calculate the normalized value of the phase according to the following formula:

[0093]

[0094] where φ′ is the normalized value of the original phase φ, and the numerical range is [0, 1].

[0095] According to the above steps, the constructed partial discharge feature matrix is:

[0096]

[0097] where E j,i represents the energy value of the jth layer wavelet decomposition of the partial discharge signal in the ith time window, which is a normalized value;

[0098] P i represents the number of discharges in the ith time window, and the value range is [0, 1];

[0099] φ i represents the phase value in the ith time window, and the value range is [0, 1].

[0100] In the embodiment of the present invention, the partial discharge feature matrix under known working conditions is used as the source domain data with fault labels, and the partial discharge feature matrix under unknown working conditions is used as the unlabeled target domain data to train the feature extractor, classifier, and domain discriminator.

[0101] Feature extractor: used to map the sample data to a specific feature space, so that the classifier can distinguish the categories of data from the source domain, while the domain discriminator cannot distinguish which domain the data comes from.

[0102] Classifier: Used to classify data from the source domain and distinguish the correct labels as much as possible.

[0103] Domain discriminator: Used to determine whether the sample features come from the source domain or the target domain. If the domain discriminator cannot determine the domain information of a mapped feature, this feature can be considered a domain-invariant representation.

[0104] In the embodiment of the present invention, the feature extractor and the classifier constitute a feedforward neural network. A domain discriminator is added after the feature extractor and connected by a gradient reversal layer GRL in the middle. During the training process, the loss function of the classifier and the loss function of the domain discriminator are minimized simultaneously, guiding the features extracted by the feature extractor to be able to identify the fault types within the source domain and, due to the characteristics of domain-invariant features, be applicable to fault diagnosis within the target domain working conditions.

[0105] The pre-training stage specifically includes: inputting the source domain data into the feature extractor to extract source domain features; inputting the extracted source domain features into the classifier for label classification to obtain predicted labels; using the cross-entropy loss function to calculate the classification loss between the predicted labels and the true labels of the source domain samples, and backpropagating the gradient of the classification loss until the preset number of iterations is reached to obtain the pre-trained feature extractor and classifier.

[0106] According to a preferred embodiment of the present invention, before using the feature extractor to extract features from the source domain data, first apply a clustering algorithm to perform clustering analysis on the source domain data to identify similar classes in the source domain (i.e., data clusters with similar feature distributions). To fully extract features from the source domain data, use a transformer to fuse the similar classes.

[0107] In the formal training stage, using the source domain data and the target domain data, adopt an adversarial game strategy to train the pre-trained feature extractor and domain discriminator, and align the source domain and target domain features by extracting domain-invariant features. The construction of the adversarial domain adaptation network specifically includes: inputting the sample data of the source domain and the target domain into the pre-trained feature extractor to extract source domain sample features and target domain sample features respectively; inputting the extracted source domain sample features and target domain sample features into the domain discriminator for domain classification, and using the cross-entropy loss function to calculate the domain adversarial loss; taking the opposite of the gradient of the domain adversarial loss through the gradient reversal layer and then performing backpropagation for adversarial training, iteratively updating the network parameters until convergence. The trained feature extractor is used to extract domain-invariant features of the source domain and target domain data in the high-dimensional space.

[0108] In an embodiment of the present invention, the feature extractor includes at least two one-dimensional convolutional layers arranged in sequence and a global average pooling layer. The one-dimensional convolutional layer is used to perform one-dimensional convolutional calculation on the source domain samples and the target domain sample data to obtain initial features, and the global average pooling layer is used to reduce the dimension of the initial features to obtain feature data. The feature data of the source domain samples is input into a classifier, and the classifier is used to predict the label category of the source domain samples according to the feature data of the source domain samples. The feature data of the source domain samples and the target domain samples is input into a domain discriminator, and the domain discriminator is used to determine whether the sample belongs to the source domain or the target domain according to the feature data of the input samples.

[0109] After adversarial training, few-shot learning is performed using the target domain data, that is, a small amount of labeled data in the target domain data is used to fine-tune the feature extractor and the classifier to further improve the feature extraction ability for the target domain data.

[0110] In the model update stage, the classifier is used to generate pseudo-labels for the unlabeled samples in the target domain data, and the pseudo-label data of the target domain is added to the training set to retrain the feature extractor, the classifier, and the domain discriminator to achieve label assignment for each unlabeled sample in the target domain data. Specifically, it includes: inputting the unlabeled samples in the target domain data into the classifier for label prediction, and generating pseudo-labels for the unlabeled samples with a confidence level higher than the threshold; combining the obtained pseudo-label data of the target domain with the source domain data to update the training set, and using the updated training set to retrain the feature extractor, the classifier, and the domain discriminator. Through self-training iteration, until label assignment is achieved for each unlabeled sample in the target domain data.

[0111] The model update stage specifically includes:

[0112] 1. Continuously collect data:

[0113] After the adaptive model is deployed, continuously monitor the performance of the classifier in the target domain, and collect the low-confidence data predicted by the classifier, that is, the unlabeled data continuously generated in the target domain. These data often contain new features or marginal samples, and analyze these data to perform adaptive updates of the model.

[0114] 2. Unsupervised feature update:

[0115] The self-training mechanism uses the current model to perform label prediction on the newly collected target domain data, and uses the high-confidence prediction samples as pseudo-labeled data to gradually expand the labeled data set of the target domain.

[0116] Adaptive update of the feature extractor, using this pseudo-labeled data to retrain or fine-tune the feature extractor so that it can adapt to the newly emerging feature changes in the target domain. By continuously learning from this new data, gradually optimize the feature extractor so that it can better capture the new feature distribution in the target domain.

[0117] 3. Continuous optimization of domain alignment:

[0118] Periodic update of adversarial training. According to the feature distribution of the new data in the target domain, periodically perform adversarial training for domain alignment. By updating the weights of the domain discriminator, ensure that the alignment of the source domain and target domain data in the feature space can adapt to the changes in the target domain. By periodically retraining the classifier, ensure its classification performance on the new features.

[0119] Fine-tune the shared feature space. Through continuous adversarial learning, further adjust the shared feature space of the source domain and target domain to ensure that the feature extractor maintains the best alignment effect between the source domain and target domain.

[0120] 4. Dynamic adjustment of few-shot learning:

[0121] As the labeled data in the target domain increases, dynamically expand the few-shot learning set, and use the few-shot extended data in the target domain to further fine-tune the classifier to improve the classifier's adaptability to new data.

[0122] The present invention solves the problem of dependence on a large number of data labels in the training process of time series data. Using the labeled data in the source domain to assist in training the unlabeled data in the target domain, an effective unsupervised classification model is obtained, avoiding cumbersome data annotation work and improving training efficiency.

[0123] The second embodiment of the present invention provides a partial discharge signal analysis system based on unsupervised learning, as Figure 2 shown, the system includes:

[0124] A data acquisition module, used to collect partial discharge signals of power equipment, perform multi-scale energy decomposition on the partial discharge signals and use a moving time window for signal division to construct a partial discharge feature matrix, where the partial discharge feature matrix includes the energy values of the partial discharge signal at different scales and time windows, the number of discharges and phase values corresponding to the time window;

[0125] A pre-training module, used to respectively use the partial discharge feature matrices under different working conditions as the source domain data with fault labels and the unlabeled target domain data, and use the source domain data to pre-train a feature extractor and a classifier. The classifier is used to perform label classification on the source domain features extracted by the feature extractor;

[0126] The formal training stage module is used to utilize the source domain data and the target domain data, and train the feature extractor and the domain discriminator by adopting an adversarial game strategy, achieve feature alignment between the source domain and the target domain by extracting domain-invariant features, and construct an adversarial domain adaptation network;

[0127] The model update stage module is used to generate pseudo-labels for the unlabeled samples in the target domain data by using the classifier, and add the pseudo-label data of the target domain to the training set to retrain the feature extractor, the classifier and the domain discriminator, so as to achieve label assignment for each unlabeled sample in the target domain data.

[0128] Regarding the system in the above embodiments, the specific manners in which each unit module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated herein.

[0129] Based on the same inventive concept, an embodiment of the present invention further provides an electronic device, including: a memory and a processor, where the processor is configured to read and execute a computer program stored in the memory to implement the foregoing method for analyzing partial discharge signals based on unsupervised learning.

[0130] Based on the same inventive concept, an embodiment of the present invention further provides a computer storage medium, in which computer-executable instructions are stored, and when the computer-executable instructions are executed, the foregoing method for analyzing partial discharge signals based on unsupervised learning is implemented.

[0131] In several embodiments provided in the present application, it should be understood that the disclosed device and method can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation. For example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling or direct coupling or communication connection to each other can be through some interfaces, and the indirect coupling or communication connection of the device or module can be in an electrical, mechanical or other form.

[0132] The module described as a separation component may or may not be physically separated. The component shown as a module may or may not be a physical module, that is, it may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. In addition, in each embodiment of the present invention, each functional module can be integrated into a processing module, or each module can exist physically alone, or two or more modules can be integrated into one module. The above integrated module can be implemented in the form of hardware or in the form of a software functional module.

[0133] If the above integrated module is implemented in the form of a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in each embodiment of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0134] It should be noted that for the foregoing method embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the present invention is not limited by the described action sequence, because according to the present invention, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0135] In the above embodiments, the descriptions of each embodiment have their own emphases. For the parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments. Finally, it should be noted that: the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A partial discharge signal analysis method based on unsupervised learning, characterized in that: The method comprises the following steps: Data acquisition stage: collecting partial discharge signals of power equipment, performing multi-scale energy decomposition on the partial discharge signals and using a moving time window to divide the signals, and constructing a partial discharge feature matrix, which includes energy values ​​of partial discharge signals at different scales and time windows, and the number of discharges and phase values ​​in the corresponding time windows; Pre-training stage: taking the partial discharge feature matrices under different working conditions as source domain data with fault labels and unlabeled target domain data, respectively, and using the source domain data to pre-train a feature extractor and a classifier, the classifier is used to classify the source domain features extracted by the feature extractor; Formal training stage: using the source domain data and the target domain data, adopting an adversarial game strategy to train the feature extractor and domain discriminator, aligning the source domain and target domain features by extracting domain-invariant features, and constructing an adversarial domain adaptive network; Model update stage: using the classifier to generate pseudo labels for unlabeled samples in the target domain data, adding the pseudo label data of the target domain to the training set to retrain the feature extractor, the classifier and the domain discriminator, and realizing label assignment for each unlabeled sample in the target domain data.

2. The method according to claim 1, characterized in that In the data acquisition stage, an ultra-high frequency sensor is used to collect ultra-high frequency electromagnetic wave signals generated by partial discharge of power equipment, and the partial discharge signal is generated after pre-processing the ultra-high frequency electromagnetic wave signals.

3. The method according to claim 2, characterized in that The frequency of collecting the ultra-high frequency electromagnetic wave signal is 300MHz-3GHz.

4. The method according to claim 2, characterized in that: Preprocessing the ultra-high frequency electromagnetic wave signal includes: A high-speed data acquisition card is used to sample signals and convert the UHF electromagnetic wave analog signals collected by the UHF sensor into UHF electromagnetic wave digital signals; The ultra-high frequency electromagnetic wave digital signal is filtered, amplified, calibrated, and noise-reduced.

5. The method according to any one of claims 1 to 4, characterized in that: The constructing of the partial discharge characteristic matrix comprises: Performing discrete wavelet transform on the local discharge signal, and calculating the energy value of each layer of wavelet decomposition; The partial discharge signal after wavelet decomposition is divided by using a moving time window, and the energy value of each layer of the wavelet decomposition signal in different time windows is obtained as the main dimension feature; Obtain the phase information and discharge times of the corresponding time window as additional dimension features; The main dimension features and the additional dimension features are used to construct a local discharge feature matrix.

6. The method according to claim 5, characterized in that Performing discrete wavelet transform on the partial discharge signal and calculating the energy value of each layer of wavelet decomposition includes: The partial discharge signal is decomposed into approximation coefficients and detail coefficients of different scales by discrete wavelet transform method: Where x(t) is the local discharge signal, J represents the number of decomposition layers, and A J is the wavelet approximation coefficient of the Jth layer, D j is the wavelet detail coefficient of the jth layer; Calculate the energy value of each layer of wavelet decomposition of partial discharge signal: Among them, E j is the energy value of the j-th layer wavelet decomposition, n j is the number of detail coefficients of the j-th layer wavelet decomposition, D j (i) is the i-th wavelet detail coefficient of the j-th layer; Normalize the energy value of each layer of wavelet decomposition of the partial discharge signal: Among them, the energy value NE j It indicates the ratio of the energy value of the j-th layer wavelet decomposition to the total energy value.

7. The method according to claim 6, characterized in that The partial discharge characteristic matrix is: Among them, E j,i It represents the energy value of the j-th wavelet decomposition of the partial discharge signal in the i-th time window, and its value range is [0,1]; P i represents the number of discharges in the i-th time window, and its value range is [0,1]; φ i Represents the phase value in the i-th time window, and its value range is [0,1].

8. The method according to claim 1, characterized in that The pre-training stage specifically includes: Input source domain data into the feature extractor to extract source domain features; Input the extracted source domain features into the classifier for label classification to obtain the predicted label; The predicted label and the true label of the source domain sample are used to calculate the classification loss using the cross entropy loss function, and the gradient of the classification loss is back-propagated until the preset number of iterations is reached to obtain the pre-trained feature extractor and classifier.

9. The method according to claim 8, characterized in that The formal training phase specifically includes: Input the sample data of the source domain and the target domain into the pre-trained feature extractor to extract the sample features of the source domain and the sample features of the target domain respectively; The extracted source domain sample features and target domain sample features are input into the domain discriminator for domain classification, and the cross entropy loss function is used to calculate the domain adversarial loss; The gradient of the domain adversarial loss is inverted through the gradient reversal layer and then back-propagated for adversarial training. The network parameters are iteratively updated until convergence. The trained feature extractor is used to extract domain-invariant features of the source domain and target domain data in high-dimensional space.

10. The method according to claim 9, characterized in that The model updating stage specifically includes: Inputting the unlabeled samples in the target domain data into the classifier for label prediction, and generating pseudo labels for the unlabeled samples above the confidence threshold; The obtained target domain pseudo-label data is combined with the source domain data, the training set is updated, and the feature extractor, the classifier and the domain discriminator are retrained using the updated training set, and self-training is iterated until each unlabeled sample in the target domain data is labeled.

11. A partial discharge signal analysis system based on unsupervised learning, characterized in that: The system comprises: A data acquisition module is used to collect partial discharge signals of power equipment, perform multi-scale energy decomposition on the partial discharge signals and use a moving time window to divide the signals, and construct a partial discharge feature matrix, wherein the partial discharge feature matrix includes energy values ​​of the partial discharge signals at different scales and time windows, and the number of discharges and phase values ​​of the corresponding time windows; A pre-training module, used to use the partial discharge feature matrices under different working conditions as source domain data with fault labels and unlabeled target domain data, and use the source domain data to pre-train a feature extractor and a classifier, wherein the classifier is used to perform label classification on the source domain features extracted by the feature extractor; A formal training module is used to use the source domain data and the target domain data to train the feature extractor and the domain discriminator using an adversarial game strategy, align the source domain and the target domain features by extracting domain-invariant features, and construct an adversarial domain adaptive network; The model updating module is used to generate pseudo labels for unlabeled samples in the target domain data using a classifier, add the pseudo label data of the target domain to a training set to retrain the feature extractor, the classifier and the domain discriminator, and implement label assignment for each unlabeled sample in the target domain data.

12. An electronic device, characterized in that: include: Memory, processor; The processor is used to read and execute the computer program stored in the memory to implement the partial discharge signal analysis method based on unsupervised learning according to any one of claims 1 to 10.

13. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed, the method for analyzing partial discharge signals based on unsupervised learning according to any one of claims 1 to 10 is implemented.

Images