Phase decomposition GIS partial discharge fault diagnosis method based on supervised contrastive learning

Abstract

A phase-decomposition GIS partial discharge fault diagnosis method based on supervised contrastive learning comprises the following steps: step (1): obtaining GIS partial discharge signals, using four simulated defects (tip discharge, suspended discharge, air gap discharge, and surface discharge) to capture the PRPS signals of the four defects; step (2): the training set of the original data is enhanced by four techniques: Gaussian noise addition, Gaussian noise scaling, random clipping, and phase shift; step (3): two data enhancement methods are randomly selected for pre-training tasks; step (4): downstream tasks use the knowledge learned from the pre-training tasks for classification and detection; step (5): using a data test set to verify its classification effect. The present invention proposes a phase-decomposition GIS partial discharge fault diagnosis method based on supervised contrastive learning, which has a small data volume requirement and a high fault diagnosis accuracy.

Description

Technical Field

[0001] The invention belongs to the technical field of GIS fault diagnosis, and in particular relates to a phase decomposition GIS partial discharge fault diagnosis method based on supervised contrast learning. Background Art

[0002] In recent years, artificial intelligence and machine learning algorithms have successfully performed pattern recognition and classification of partial discharge phenomena in GIS, such as support vector machine (SVM), random forest (RF), logistic regression (LR), etc. The pattern recognition and classification method based on PRPS has the advantages of simple measurement system and ability to distinguish signals from noise, and is robust to signal attenuation and noise interference, but the limitation of advanced feature extraction is a big problem faced in traditional machine learning. The feature extraction process is very time-consuming and relies on domain expertise.

[0003] In order to solve the above problems, deep neural networks (DNNs) have been developed, but they contain a large number of nonlinear activation functions, layers, and neurons, and their convergence speed is slow and complex. In order to solve the DNN problem, convolutional neural networks (CNNs) have been proposed. There are also some different types and variants of CNN architectures, including AlexNet, VGG, ResNet, and GoogLeNet. Although they have better learning capabilities than traditional machine learning, their fault diagnosis requires a large amount of data sets. In the partial discharge fault diagnosis of GIS, it is difficult to obtain a large amount of data based on different environments and fault categories.

[0004] In this paper, we propose a new method for GIS fault diagnosis using supervised contrastive learning (SCL) of PRPS signals. Contrastive learning aims to maximize the similarity between any two vector representations by minimizing the contrastive loss function. Its advantages include distinguishing different types of data, it can be used to change the feature distribution, SCL can eliminate task-irrelevant information by constructing positive data pairs, and uses less data sets. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a phase-decomposition GIS partial discharge fault diagnosis method based on supervised contrastive learning. The supervised contrastive learning model includes a pre-training task and a downstream task. Data enhancement is performed in the pre-training task. The enhanced data first passes through an encoder network to create a frozen encoder and then is fed back to the Projection head. After the contrast loss is minimized, the frozen weight encoder is used for the downstream task, that is, a classifier for PRPS classification is used as the downstream task. The method of the present invention has a small data volume requirement and a high fault diagnosis accuracy.

[0006] The present invention specifically provides a phase decomposition GIS partial discharge fault diagnosis method based on supervised contrastive learning. The GIS partial discharge fault identification method comprises the following steps:

[0007] Step (1): Acquisition of GIS partial discharge signals, using four simulated defects (tip discharge, suspended discharge, air gap discharge, and surface discharge), and capturing the PRPS signals of the four defects;

[0008] Step (2): The training set of the original data is enhanced by four techniques: Gaussian noise addition, Gaussian noise scaling, random cropping, and phase shifting.

[0009] Step (3): Randomly select two data enhancement methods for pre-training tasks;

[0010] Step (4): Downstream tasks use the knowledge learned from the pre-training tasks for classification and detection;

[0011] Step (5): Use the data test set to verify its classification effect.

[0012] Define the PRPS signal of the mth power cycle as x m =[x m1 ,x m2 ,...,x mN ], M is the number of power cycles, N is the number of phase angles in a power cycle, in matrix form, the PRPS signal is expressed as:

[0013]

[0014] Where X is an N×M matrix.

[0015] The four data enhancement methods in step (2) are as follows:

[0016] (21) Gaussian noise addition is expressed as:

[0017]

[0018] Among them, n1 is Gaussian distribution, x ij are the components of the PRPS matrix X in (1), For the matrix Add The Gaussian noise takes into account different noise environments.

[0019] (22) Gaussian noise scaling can be expressed as:

[0020]

[0021] n2 is Gaussian distribution, Gaussian noise scaling Considering the variation of partial discharge amplitude with fault severity, the matrix The (i,j) term is

[0022] (23) Then the clipping can be expressed as:

[0023]

[0024] Among them, RC(·) changes the irrelevant information gain components in the input matrix to 0, weakens the weight of the background (or noise) factor, and makes the model insensitive to missing values. And according to the severity of the fault, random clipping takes into account the variation of partial discharge amplitude.

[0025] (24)The phase shift can be expressed as:

[0026]

[0027] Here, PS(·) permutes the phase of the input matrix. This modification takes into account the phase synchronization error between the electrical equipment and the fault diagnosis system.

[0028] In step (3), two data enhancement methods are randomly selected as follows:

[0029] After data augmentation, four matrices are selected Two of them are used to construct For a set of T randomly sampled sample pairs / label pairs {X t ,y t} t＝1,2,...,T , the corresponding samples used for training include 2T, namely {X i ,y i} t＝1,2,...,2T ,in and Yes X t (t=1,…,T) two randomly augmented matrices, and

[0030] After data augmentation in step (3), the encoder represents different data views into an embedding space with high-level features. In the pre-training task, the encoder weights are retained and frozen to serve the downstream task. The steps are as follows:

[0031] (31) The first part is the encoder network f(.), which operates by transforming the two-dimensional matrix Transformed into a one-dimensional vector, recorded as D f is the output form of the last layer in the network. and We obtain a pair of vectors h2t-1 and h 2t ,The encoder network includes a convolutional layer for extracting high-level features and a flattening layer.

[0032] (32)Projection head g(.) is expressed as where g(.) is composed of D with a nonlinear activation ReLU function g A single linear layer consisting of units, i∈I=1,...,2T is the index of any augmented sample in the multi-viewer batch.

[0033] (33) The supervised contrast loss is expressed as:

[0034]

[0035] Among them, A(i)≡I\{i} is the set with element i eliminated from I, is the set of indicators of fault samples with the same label as the i-th label. Where τ>0 is a scalar temperature parameter, |·| is the total number of elements in the set, and the goal of the pre-training task is to minimize the contrast loss function in (6).

[0036] The downstream tasks in step (4) are as follows:

[0037] (41) The downstream task uses the pre-trained encoder f(·) and freezes the weights before classification into class Dc.

[0038] (42) The classifier network c(.) is expressed as Where c(.) consists of one hidden layer, two dropout layers and a Softmax. There are two 20% dropout layers before and after the hidden layer to avoid overfitting. The cross entropy loss function of the classifier network is expressed as:

[0039]

[0040] Where b is the index of the mini-batch. When index j is the index of the ground truth value, t j =1, in other cases t j = 0. In the present invention, the original data set is divided into a training set and a test set. The Adam optimizer is used to update the learnable parameters of the supervised contrast and cross entropy loss functions.

[0041] Compared with the prior art, the beneficial effects are:

[0042] (1) The phase decomposition GIS partial discharge fault diagnosis method based on supervised contrastive learning can distinguish different types of data. It can be used to change the feature distribution and achieve efficient feature extraction.

[0043] (2) Through supervised comparative learning, there is no need for the large amount of data required by traditional machine learning, which saves experimental costs and greatly improves the speed and feasibility of fault classification.

[0044] (3) The phase-decomposition GIS partial discharge fault diagnosis method based on supervised contrastive learning proposed in the present invention has high recognition accuracy, fast operation speed, can accurately diagnose the type of partial discharge fault in GIS, and has strong engineering practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flow chart of a phase decomposition GIS partial discharge fault diagnosis method based on supervised contrast learning of the present invention;

[0046] Figure 2 It is the PRPS signal diagram of four defects;

[0047] Figure 3 It is the structural diagram of the encoder network. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, a specific implementation method of a phase-decomposition GIS partial discharge fault diagnosis method based on supervised contrastive learning of the present invention is described in detail below with reference to the accompanying drawings.

[0049] like Figure 1 As shown, a phase decomposition GIS partial discharge fault diagnosis method based on supervised contrast learning of the present invention comprises the following steps:

[0050] Step (1): Acquisition of GIS partial discharge signals, using four simulated defects (tip discharge, suspended discharge, air gap discharge, and surface discharge), and capturing the PRPS signals of the four defects;

[0051] Step (2): The training set of the original data is enhanced by four techniques: Gaussian noise addition, Gaussian noise scaling, random cropping, and phase shifting.

[0052] Step (3): Randomly select two data enhancement methods for pre-training tasks;

[0053] Step (4): Downstream tasks use the knowledge learned from the pre-training tasks for classification and detection;

[0054] Step (5): Use the data test set to verify its classification effect.

[0055] The PRPS signals of the four defects captured in step (1) are as follows: Figure 2 shown.

[0056] Define the PRPS signal of the mth power cycle as x m =[xm1 ,x m2 ,...,x mN ], M is the number of power cycles, N is the number of phase angles in a power cycle, in matrix form, the PRPS signal is expressed as:

[0057]

[0058] Where X is an N×M matrix.

[0059] The four data enhancement methods in step (2) are as follows:

[0060] (21) Gaussian noise addition is expressed as:

[0061]

[0062] Among them, n1 is Gaussian distribution, x ij are the components of the PRPS matrix X in (1), For the matrix Add The Gaussian noise takes into account different noise environments.

[0063] (22) Gaussian noise scaling can be expressed as:

[0064]

[0065] n2 is Gaussian distribution, Gaussian noise scaling Considering the variation of partial discharge amplitude with fault severity, the matrix The (i,j) term is

[0066] (23) Random cropping can be expressed as:

[0067]

[0068] Among them, RC(·) changes the irrelevant information gain components in the input matrix to 0, weakens the weight of the background (or noise) factor, and makes the model insensitive to missing values. And according to the severity of the fault, random clipping takes into account the variation of partial discharge amplitude.

[0069] (24)The phase shift can be expressed as:

[0070]

[0071] Here, PS(·) permutes the phase of the input matrix. This modification takes into account the phase synchronization error between the electrical equipment and the fault diagnosis system.

[0072] Two data enhancement methods are randomly selected in step (3) as follows:

[0073] After data augmentation, four matrices are selected Two of them are used to construct For a set of T randomly sampled sample pairs / label pairs {X t ,y t} t＝1,2,...,T , the corresponding samples used for training include 2T, namely {X i ,y i} t＝1,2,...,2T ,in and Yes X t (t=1,…,T) two randomly augmented matrices, and

[0074] After data augmentation in step (3), the encoder represents different data views into an embedding space with high-level features. In the pre-training task, the encoder weights are retained and frozen to serve the downstream task. The steps are as follows:

[0075] (31) The first part is the encoder network f(.), which operates by transforming the two-dimensional matrix Transformed into a one-dimensional vector, recorded as D f is the output form of the last layer in the network. and We obtain a pair of vectors h 2t-1 and h 2t ,The encoder network includes a convolutional layer for extracting high-level features and a flattening layer.

[0076] (32)Projection head g(.) is expressed as where g(.) is composed of D with a nonlinear activation ReLU function g A single linear layer consisting of units, i∈I=1,...,2T is the index of any augmented sample in the multi-viewer batch.

[0077] (33) The supervised contrast loss is expressed as:

[0078]

[0079] Among them, A(i)≡I\{i} is the set with element i eliminated from I, is the indicator set of fault samples with the same label as the i-th label. Where τ>0 is a scalar temperature parameter, |·| is the total number of elements in the set, and the goal of the pre-training task is to minimize the contrast loss function in (6).

[0080] The downstream tasks in step (4) are as follows:

[0081] 6.1 Downstream Tasks Using the pre-trained encoder f(·), the weights are frozen before classification into class Dc.

[0082] 6.2 Classifier network c(.) is expressed as Where c(.) consists of one hidden layer, two dropout layers and a Softmax. There are two 20% dropout layers before and after the hidden layer to avoid overfitting. The cross entropy loss function of the classifier network is expressed as:

[0083]

[0084] Where b is the index of the mini-batch. When index j is the index of the ground truth value, t j =1, in other cases t j = 0. In the present invention, the original data set is divided into a training set and a test set. The Adam optimizer is used to update the learnable parameters of the supervised contrast and cross entropy loss functions.

[0085] The following are examples of implementation

[0086] Table 1 shows the fault types (tip, suspension, air gap, particle and noise) and the number of samples. For data augmentation, we choose mean μ1 = 0, standard deviation σ1 = 0.01 as the parameters for Gaussian noise addition, μ2 = 1, standard deviation σ2 = 0.01 for Gaussian noise scaling, and random clipping with a learning rate a = 0.05. Figure 3 The structure of the encoder network is shown, which consists of 5 convolutional blocks (each convolutional block consists of a 3×3 convolutional layer, a rectified linear unit activation layer, and a maximum pooling layer) and a D f =3584 flattening layer. After the encoder network, a single hidden layer projection head with Dg = 256 and τ = 0.06 is used for nonlinear transformation, and the projection h optimizes the supervised contrast loss in (6). For downstream tasks, the classifier network includes a hidden layer of 900 nodes and a D c The Softmax layer with =5 is used to classify the five fault types.

[0087] Table 1

[0088] Fault type Cutting Edge Suspension Air Gap Particles noise Sample size 102 84 154 42 138

[0089] In the experiments, the dataset is divided into training and test sets with 80% and 20% sample ratios, respectively. In addition, the data augmentation method doubles the training set size. During the training process, we experiment with different parameter selections to obtain the optimized hyperparameters related to mini-batch, number of layers, etc. to optimize our method. Table 2 shows the optimization of the minimum and maximum bounds of each hyperparameter, as well as their types. The proposed SCL is optimized with different parameter combinations and the best parameter combination is obtained.

[0090] Table 2

[0091] Hyperparameters Minimum Maximum mini-batch 8 128 Number of layers 2 8 Kernel size 3×3 9×9 Epochs 80 200 Learning Rate 0.0005 0.02 Classifier Network Node 500 1500 Projectionhead Node 64 512 Dropoutrate 0.1 0.6

[0092] After trying several architectures, we used 5 convolutional layers with 3×3 kernel size for the encoder network, a single hidden layer with 256 nodes for the projection head, and a hidden layer with 900 nodes for the classifier network with a dropout rate of 15%, which, combined with learning rate, mini-batch, and epochs of 0.0005, 16, and 180, achieved the highest overall classification accuracy for each fault type. In addition, the parameters of the SCL method for each combination were nonlinearly transformed using the ReLU function. In this experiment, we compared the performance of the proposed SCL with support vector machines (SVM), multilayer perceptrons (MLP), and convolutional neural networks (CNN), where the MLP used ReLU activation and Adam optimizer with up to 200 iterations. The mean of non-zero values ​​and the number of non-zero values ​​at each stage were used for feature extraction. Table 3 gives the experimental range of parameters for SVM and MLP.

[0093] Table 3

[0094]

[0095]

[0096] The regularization parameter C of SVM SVM The value range of C is 0.005~100, which avoids the overfitting problem. SVM=100, the SVM model achieved the highest accuracy. For the MLP method, different numbers of hidden layers were tried from 1 to 5, and the number of nodes in each layer was reduced to prevent overfitting. After trying several sets of different numbers of nodes and hidden layers, we found the MLP model with 4 hidden layers of (200, 150, 100, 50) nodes to achieve the highest accuracy. Here, we used the same range of hyperparameters and architectures for the SCL encoder-classifier network and the CNN method, as shown in Table 2. For the CNN model, 5 convolutional layers (3×3 kernel size, 900 nodes hidden layer and 2 20% dropout with learning rate, batch size and epoch of 0.0005, 16 and 200 respectively) achieved the highest classification accuracy.

[0097] Table 4 gives the performance analysis of SCL. In the identification of particle discharge faults, the proposed SCL outperforms SVM (the test result is 100%, while SVM is 76.20%). In terms of overall classification performance, SCL is 3.73% higher than MLP, and performs particularly well in classifying particles, air gaps, and noise. Although the accuracy of tip discharge in the proposed SCL is lower than that of CNN, the overall performance (96.73% for the proposed SCL and 91.15% for CNN) is still better than CNN. Therefore, it can be seen from Table 4 that among the four methods, the accuracy of the proposed SCL method is the best, which proves the innovation and effectiveness of the present invention.

[0098] Table 4

[0099] Fault type SVM (%) MLP (%) CNN(%) SCL(%) Tip discharge 89.21％(91 / 102) 90.20(92 / 102) 100 96.08％(98 / 102) Suspension discharge 92.86(78 / 84) 85.71(72 / 84) 85.71(72 / 84) 95.24(80 / 84) Particle discharge 76.20(32 / 42) 90.48(38 / 42) 83.33(35 / 42) 100(42 / 42) Air Gap Discharge 85.06％(131 / 154) 80.01(134 / 154) 90.91(140 / 154) 94.16(145 / 154) noise 95.65(132 / 138) 93.48(129 / 138) 90.58(125 / 138) 100 Overall recognition rate 89.23(464 / 520) 93.00(465 / 520) 91.15(474 / 520) 96.73(503 / 520)

[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention and are not intended to limit the present invention. A person skilled in the art should understand that the specific implementation of the present invention can be modified or replaced by equivalents, but these modifications or changes are within the scope of protection of the pending claims.

Claims

1. A phase decomposition GIS partial discharge fault diagnosis method based on supervised contrastive learning, characterized in that: The GIS partial discharge fault diagnosis method comprises the following steps: Step (1): Acquisition of GIS partial discharge signals, using four simulated defects: tip discharge, suspension discharge, air gap discharge, and surface discharge, and capturing PRPS signals of the four defects; Step (2): The training set of the original data is enhanced by four techniques: Gaussian noise addition, Gaussian noise scaling, random cropping, and phase shifting. Step (3): Randomly select two data enhancement methods for pre-training tasks; Step (4): Downstream tasks use the knowledge learned from the pre-training tasks for classification and detection; Step (5): Use the data test set to verify its classification effect.

2. The phase decomposition GIS partial discharge fault diagnosis method based on supervised contrastive learning according to claim 1 is characterized in that: Define the PRPS signal of the mth power cycle as x m =[x m1 ,x m2 ,...,x mN ], M is the number of power cycles, N is the number of phase angles in a power cycle, in matrix form, the PRPS signal is expressed as: Where X is an N×M matrix.

3. The phase decomposition GIS partial discharge fault diagnosis method based on supervised contrastive learning according to claim 2 is characterized in that: The four data enhancement methods in step (2) are as follows: (21) Gaussian noise addition is expressed as: Among them, n1 is Gaussian distribution, x ij are the components of the PRPS matrix X in (1), For the matrix Add The Gaussian noise takes into account different noise environments; (22) Gaussian noise scaling is expressed as: n2 is Gaussian distribution, Gaussian noise scaling Considering the variation of partial discharge amplitude with fault severity, the matrix The (i,j) term is (23) Random cropping is expressed as: Among them, RC(·) changes the irrelevant information gain components in the input matrix to 0, weakens the weight of the background or noise factor, and makes the model insensitive to missing values; and according to the severity of the fault, random clipping takes into account the change of partial discharge amplitude; (24) The phase shift is expressed as: Here, PS(·) permutes the phase of the input matrix. This modification takes into account the phase synchronization error between the electrical equipment and the fault diagnosis system.

4. The phase decomposition GIS partial discharge fault diagnosis method based on supervised contrastive learning according to claim 3 is characterized in that: Two data enhancement methods are randomly selected in step (3) as follows: After data augmentation, four matrices are selected Two of them are used to construct For a set of T randomly sampled sample pairs / label pairs {X t ,y t } t＝1,2,...,T , the corresponding samples used for training include 2T, namely {X i ,y i } t＝1,2,...,2T ,in and Yes X t (t=1,…,T) two randomly augmented matrices, and 5. The phase decomposition GIS partial discharge fault diagnosis method based on supervised contrastive learning according to claim 4 is characterized in that: After data augmentation in step (3), the encoder represents different data views into an embedding space with high-level features. In the pre-training task, the encoder weights are retained and frozen to serve the downstream task. The steps are as follows: (31) The first part is the encoder network f(.), which operates by transforming the two-dimensional matrix Transformed into a one-dimensional vector, recorded as D f is the output form of the last layer in the network. and We obtain a pair of vectors h 2t-1 and h 2t ,The encoder network includes a convolution layer for extracting high features and a flat layer; (32)Projection head g(.) is expressed as where g(.) is composed of D with a nonlinear activation ReLU function g A single linear layer consisting of units, i∈I=1,...,2T is the index of any augmented sample in the multi-view batch; (33) The supervised contrast loss is expressed as: Among them, A(i)≡I\{i} is the set with element i eliminated from I, is the indicator set of fault samples with the same label as the i-th label; where τ>0 is the scalar temperature parameter, |·| is the total number of elements in the set, and the goal of the pre-training task is to minimize the contrast loss function in (6).

6. The phase decomposition GIS partial discharge fault diagnosis method based on supervised contrastive learning according to claim 5 is characterized in that: The downstream tasks in step (4) are as follows: (41) The downstream task uses the pre-trained encoder f(·) and freezes the weights before classification into class Dc; (42) The classifier network c(.) is expressed as Where c(.) consists of a hidden layer, two dropout layers and a Softmax; there are two 20% dropout layers before and after the hidden layer to avoid overfitting; the cross entropy loss function of the classifier network is expressed as: Where b is the index of the mini-batch. When index j is the index of the ground truth value, t j =1, in other cases t j =0.

Images