CNN-GRU neural network-based cable discharge current identification method

Abstract

The invention discloses a CNN-GRU neural network-based cable discharge current identification method. The method comprises the steps of obtaining discharge current signal data of a cable; the method comprises the following steps: extracting various current characteristic quantities of a discharge current signal by adopting a convolutional neural network (CNN), constructing a current characteristic data set, and dividing the current characteristic data set into a training set and a test set; constructing a CNN-GRU neural network model, and training the CNN-GRU neural network model by using the training set; and inputting test set data into the trained CNN-GRU neural network model, and outputting an identification result of a cable discharge current signal. According to the cable discharge current identification method based on the CNN-GRU neural network, advantage complementation can be realized in feature extraction and time sequence modeling of discharge signals, and an effective solution is provided for accurate identification of the cable discharge current.

Description

Technical Field

[0001] The invention relates to the technical field of cable discharge current identification, in particular to a cable discharge current identification method based on a CNN-GRU neural network. Background Art

[0002] The stable operation of the power system is of great significance to the development of social economy. As a key component of the power transmission system, the reliability of cables is directly related to the safe operation of the power grid. Cable discharge is one of the main manifestations of cable insulation degradation. It refers to the electrical breakdown of cables during long-term use due to factors such as aging of insulation materials and moisture. During the discharge process, obvious discharge current waveforms will be generated. Accurate identification and detection of these discharge signals can help to timely troubleshoot faults and prevent accidents. However, since cables have been laid underground for a long time, their monitoring is difficult.

[0003] In recent years, deep learning technology has become an emerging method in the field of machine learning. Its core concept is to abstract and extract features from data layer by layer through multi-layer neural networks, thereby realizing automatic learning and processing of complex data. Among them, convolutional neural networks (CNN) and recurrent neural networks (RNN) are particularly outstanding in the field of signal classification. They can automatically extract complex high-dimensional features and show good adaptive capabilities. In addition, cable discharge current signals usually have time series characteristics. Gated recurrent units (GRUs), as an improved structure of recurrent neural networks (RNNs), can effectively process long time series signals and capture the temporal dependencies and changing trends in signals. Summary of the invention

[0004] The present invention provides a cable discharge current identification method based on a CNN-GRU neural network, which combines a convolutional neural network (CNN) with a gated recurrent unit (GRU), can achieve complementary advantages in feature extraction and time series modeling of discharge signals, and provides an effective solution for accurate identification of cable discharge current.

[0005] The technical solution adopted by the present invention is:

[0006] The cable discharge current identification method based on CNN-GRU neural network includes the following steps:

[0007] Step 1: Obtain the discharge current signal data of the cable;

[0008] Step 2: Use a convolutional neural network (CNN) to extract various current feature quantities of the discharge current signal, construct a current feature data set, and divide the current feature data set into a training set and a test set;

[0009] Step 3: Build a CNN-GRU neural network model and train the CNN-GRU neural network model using the training set;

[0010] Step 4: Input the test set data into the trained CNN-GRU neural network model and output the recognition result of the cable discharge current signal.

[0011] In step 1, the discharge current signal data is preprocessed as follows:

[0012] The discharge current signal data is normalized, and the amplitude of each discharge current signal is scaled to the [0,1] interval. The maximum and minimum mean normalization method is used to convert the discharge current signal data into a dimensionless per unit value to eliminate the dimension difference between the data. The formula is as follows:

[0013]

[0014] Where: X max Indicates the maximum value in the data; X min represents the minimum value in the data; X′ is the dimensionless value obtained after normalization; X i is the i-th data in the discharge current signal data to be processed.

[0015] In step 2, the current characteristic quantity includes:

[0016] ①, discharge current peak value;

[0017] ②, discharge waveform;

[0018] 3. Temperature;

[0019] ④. Number of partial discharges;

[0020] ⑤, discharge duration;

[0021] ⑥. Discharge spectrum characteristics.

[0022] In step 3, the constructed CNN-GRU neural network model includes:

[0023] (1) Convolutional neural network (CNN) includes: convolution layer, activation layer, pooling layer and fully connected layer; the convolution layer extracts local features by sliding the convolution kernel on the input data;

[0024] Z = X*W+b;

[0025] Where: Z is the output feature map, X is the input, * indicates the convolution operation, W is the convolution kernel or weight matrix, and b is the bias term.

[0026] Activation layer, after each convolution layer, a nonlinear activation function, namely ReLU activation function, is added to increase the nonlinearity of the network so that the network can learn more complex features. The expression of the activation function is as follows;

[0027] A = ReLU(Z) = max(0,Z);

[0028] The pooling layer selects the maximum value within the pooling window to reduce the dimension and computational complexity of the model;

[0029] P = max(A);

[0030] Where: P is the output value of the pooling layer; A is the output value after ReLU activation.

[0031] The fully connected layer is used to flatten the features after convolution and pooling and perform the final classification or regression task.

[0032]

[0033] Where: z j is the output value of the jth neuron; w ij is the concatenated input feature; x i is the i-th feature of the input; b j is the bias term for the output neuron j.

[0034] Usually the output of the fully connected layer will pass through a softmax activation function to obtain the classification probability. The softmax activation function is as follows:

[0035]

[0036] Where: f j is the predicted probability of the jth class; z j is the jth output value of the fully connected layer; k is the total number of categories.

[0037] In the cable discharge current signal recognition, the spatiotemporal characteristics of the discharge current waveform can be regarded as a one-dimensional signal. Through one-dimensional convolution operation, the convolutional neural network (CNN) can extract the local pattern of the signal.

[0038] (2) Gated Recurrent Unit (GRU):

[0039] The gated recurrent unit (GRU) is used to capture the temporal correlation of the cable discharge current signal, so as to accurately identify the discharge type at different moments;

[0040] The gated recurrent unit (GRU) includes: update gate, reset gate;

[0041] Update gate: It is used to control how the information of the previous time step and the information of the current candidate hidden state are mixed, and determines how much information from the past is retained in the current hidden state.

[0042] z t =λ(W z ·[h t-1 ,x t ]);

[0043] Where: z t is the update gate at time step t. λ is the sigmoid activation function, W Z is the weight matrix in the update gate, x t is the current input, h t-1 It is the hidden state at the previous moment.

[0044] Reset gate: used to control whether the current time step is affected by the previous hidden state;

[0045] r t =λ(W r ·[h t-1 ,x t ]);

[0046] Where: r t is the reset gate at time step t; W r is the weight matrix of the reset gate.

[0047] Candidate hidden state: Combine the previous hidden state with the current input and update its value;

[0048] h t =tanh(W h ·[r t *h t-1 ,x t ]);

[0049] Where: h t is the candidate hidden state at time step t; W h is the weight matrix of the candidate hidden states.

[0050] Final hidden state: weighted sum of the previous hidden state and candidate hidden state, and update it to the new hidden state;

[0051] h t =z t *h t-1 +(1-z t )*h t ;

[0052] Where: h t represents the new hidden state.

[0053] In step 3, the CNN-GRU neural network model is trained using the training set, as follows:

[0054] During the training process of the CNN-GRU neural network model, first, the network weights are initialized and forward propagation is performed to calculate the initial output of the network; then, based on the results of the forward propagation, the loss function is calculated, which is used to quantify the gap between the output of the CNN-GRU neural network model and the true value; finally, the network parameters are adjusted through the back-propagation algorithm to make the CNN-GRU neural network model gradually converge and improve the prediction accuracy.

[0055] In step 3, the CNN-GRU neural network model is trained using the training set, including the following steps:

[0056] S3.1: Initialize network weights: Randomly initialize the weights of the CNN-GRU neural network model. The present invention uses He initialization to adapt to the characteristics of ReLU through a higher weight variance, reducing the problem of gradient disappearance. The weights W are randomly sampled from a uniform distribution in the following interval:

[0057]

[0058] Where: W is the weight matrix; is uniformly distributed, with a value range of [a, b]; n in is the number of input neurons in the current layer.

[0059] S3.2: Forward propagation: Through convolution and pooling operations, feature values ​​are extracted from the discharge current signal data, and these feature values ​​are input into the gated recurrent unit (GRU) network to capture its temporal dependency; then, the output of the gated recurrent unit (GRU) is processed through a fully connected layer to finally achieve the goal of predicting the discharge current signal.

[0060] S3.3: Calculate the loss function: Since there is a certain error between the predicted value and the true value, the error should be calculated through the loss function. The formula is as follows:

[0061]

[0062] Where: L is the loss function value; N is the number of samples; y i Reflects the real discharge current signal, y i is the predicted discharge current signal.

[0063] S3.4: Back propagation: By gradually calculating the gradient of the error from the output layer to the input layer, it is determined how to adjust the various parameters in the network so that the value of the loss function gradually decreases.

[0064]

[0065] Where: is the gradient of the loss function L with respect to the weight matrix W; The partial derivative of the loss function with respect to the output y, is the partial derivative of the output layer with respect to the weight W.

[0066] S3.5: Update weights: Update and adjust the network weights with the help of the optimizer. The network weight parameters are gradually adjusted to achieve the goal of reducing the loss function value and improving the model performance. The formula is as follows.

[0067]

[0068] Where: ε is the learning rate, W new is the new weight matrix; W old is the old weight matrix; is the gradient of the loss function L with respect to the weight matrix W.

[0069] S3.6: Iteration: Continue iterating until the loss converges.

[0070] In step 4, the mean square error (M 1 ), mean absolute error (M 2 ), mean square error percentage (M 3 ), Root Mean Square Error (M 4 ), and decide 1 Determined coefficient (R 2 ) as the evaluation index to evaluate the performance of the CNN-GRU neural network model, including:

[0071] 1) Mean absolute error (M) 1 ): The model performance is evaluated by calculating the mean absolute deviation between the predicted value and the true value, using M 1 The smaller the M 1 The value indicates that the average absolute deviation between the predicted value and the true value is smaller, indicating that the model has higher prediction accuracy. The formula is as follows:

[0072]

[0073] 2) Mean absolute percentage error (M 2 ): M 2 The average deviation of the predicted value from the true value is measured in percentage form. The specific formula is shown below.

[0074]

[0075] 3) Mean square error (M 3 ): Through M 3 The value can effectively reflect the size of the error, M 3 The smaller the value, the smaller the difference between the model prediction value and the true value, which means the better the model performance. The specific formula is as follows:

[0076]

[0077] 4) Root mean square error (M 4 ): This indicator reflects the square root of the average of the squares of the differences between the predicted values ​​and the true values. The specific formula is as follows:

[0078]

[0079] 5) Coefficient of determination (R 2 ): This indicator can measure the model's fitting effect on the data, R 2 The closer the value is to 1, the better the model fits the data, indicating that the model performs better. The specific formula is as follows:

[0080]

[0081] Where: y i is the true value, y i is the predicted value, is the true mean, and n is the number of samples.

[0082] The present invention provides a cable discharge current identification method based on CNN-GRU neural network, and the technical effects are as follows:

[0083] 1) The method of the present invention firstly uses the high-frequency current method to monitor the cable discharge current signal, and selects the discharge current amplitude, discharge waveform and temperature as statistical data as characteristic values; secondly, based on the characteristic values, a cable discharge current signal recognition regression model based on the CNN-GRU neural network is established. The example analysis under different schemes shows that, considering various factors, the cable discharge current recognition technology based on the CNN-GRU neural network model has a higher accuracy. Therefore, the use of the CNN-GRU neural network model can effectively improve the final accuracy of the model.

[0084] 2) In the training process of the CNN-GRU neural network model of the present invention, the convolutional neural network (CNN) has a strong feature extraction capability and can extract key feature information from the complex waveform of the cable discharge signal. Through the superposition of multiple convolutional layers and pooling layers, CNN can automatically learn and abstract the multi-dimensional feature representation of the signal, reducing the tedious work of traditional manual feature extraction. GRU is specifically used to capture the dynamic characteristics of the cable discharge current signal over time. Cable discharge signals often have time series characteristics, and the introduction of GRU enables the model to effectively process these serialized data, and can retain more historical information at a lower computing cost, thereby capturing the changing trend of the cable under different working conditions. The cooperation between the two can more efficiently improve the capture level of current discharge phenomena, significantly improve the efficiency and accuracy of cable discharge current signal recognition, improve the power supply safety level, and further improve the quality of power supply services.

[0085] 3) The present invention adopts the high-frequency current method as one of the core technologies for monitoring cable discharge signals. This method can quickly and sensitively capture the discharge phenomenon by detecting the changes in high-frequency current signals during cable discharge. The high-frequency current method has the characteristics of high precision, fast real-time response, and strong anti-interference ability, and is particularly suitable for online monitoring in complex power environments. This monitoring method provides strong technical support for the preprocessing of cable discharge signals and lays the foundation for the stable operation of the neural network model.

[0086] 4) The cable discharge current signal recognition regression model based on the CNN-GRU neural network established by the present invention can greatly enhance the recognition ability of the complex cable discharge signal. By comparing the results of different models, it is shown that the regression model based on the CNN-GRU neural network has comprehensive advantages in processing time series signals, complex waveform feature extraction, error control, etc. The CNN-GRU model is superior to other models in recognition accuracy and efficiency, which further verifies the practicality and advancement of the technical solution of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0088] Figure 1 This is a flow chart of the cable discharge current identification method based on CNN-GRU neural network.

[0089] Figure 2 This is the experimental circuit diagram for monitoring cable discharge current signal.

[0090] Figure 3 Schematic diagram of the gated recurrent unit GRU structure.

[0091] Figure 4 Schematic diagram of the CNN-GRU neural network model structure.

[0092] Figure 5 It is a prediction effect diagram of the CNN-GRU neural network model of the present invention.

[0093] Figure 6 This is the prediction effect diagram of the random forest regression model.

[0094] Figure 7 This is the prediction effect diagram of the decision tree regression model. DETAILED DESCRIPTION

[0095] The cable discharge current identification method based on CNN-GRU neural network is applied to the cable discharge current identification of distribution network, which includes the following steps:

[0096] (I) Establishing a technical framework for cable discharge current identification:

[0097] The cable discharge current identification method established by the present invention is as follows: Figure 1 As shown. The high-frequency current method is used to monitor the discharge current signal of the cable to obtain the high-frequency current signal related to the discharge; then, various features such as current amplitude, waveform, temperature, etc. are extracted from the discharge current signal in combination with the convolutional neural network (CNN). These features will be used as the input of the model to help the neural network understand the structure and pattern of the signal; the data set is divided into a training set and a test set. The training set is used for model learning and parameter adjustment, and the test set is used for model performance evaluation to verify the accuracy and generalization ability of the model.

[0098] During the model training process, the network weights are first initialized and forward propagation is performed to calculate the initial output of the network; based on the results of the forward propagation, the loss function is calculated, which is used to quantify the gap between the model output and the true value; then, the network parameters are adjusted through the back-propagation algorithm to gradually converge the model and improve the prediction accuracy; after the convolutional network completes feature extraction, the GRU network is trained to process the timing characteristics in the cable discharge current signal. The GRU can effectively capture the pattern of signal changes over time and further improve the model's ability to recognize signals; after training, the model will output the recognition result of the cable discharge current signal, which can be used as a maintenance reference for power cables and timely detect discharge phenomena in cables.

[0099] (II) Select the characteristic value of the cable discharge current signal:

[0100] The present invention uses the high frequency current method (High Frequency Current, HFC), which has the advantages of sensitive sensing to small currents, low power loss, non-contact measurement, and strong adaptability to high frequency current signals. The HFC used in the present invention can effectively monitor the cable discharge current signal, and its schematic diagram is shown in FIG. Figure 2As shown. In the partial discharge monitoring of electrical equipment, the selection of discharge current characteristic quantity is crucial because it can reflect the changes in the insulation state, thus helping to diagnose the health of the equipment. The discharge current characteristic quantity refers to the parameter extracted from the discharge current signal that can describe the discharge behavior and characteristics. The present invention selects the following six aspects as the analysis of the cable discharge current characteristic quantity.

[0101] (1) Discharge current amplitude (peak value); the maximum value of the discharge current signal, that is, the highest peak value of the current waveform. The peak value can reflect the intensity of partial discharge and is an important indicator for judging the degree of discharge. A larger peak value indicates that the partial discharge energy is larger and there may be more serious insulation degradation.

[0102] (2) Discharge waveform: The complete time waveform of the local discharge current. By analyzing the waveform shape, the generation mechanism of the discharge can be inferred. For example, the steep rising edge and short width of a pulse discharge can indicate a transient discharge event.

[0103] (3) Temperature: discharge characteristics at different temperatures. Temperature changes may affect the intensity and nature of the discharge.

[0104] (4) Discharge frequency: the number of times partial discharge occurs within a certain period of time or cycle. The discharge frequency reflects the frequency of discharge. High-frequency discharge may indicate the gradual deterioration of the insulation material or the activity of the partial discharge source.

[0105] (5) Discharge duration: the length of time a discharge phenomenon lasts. A longer discharge duration means that the partial discharge may be more stable or its development trend is more dangerous. A longer discharge duration will also cause more damage to the insulating material.

[0106] (6) Discharge spectrum characteristics: The frequency domain characteristics of the discharge current signal are extracted through spectrum analysis methods such as Fourier transform. The discharge signal usually contains rich frequency information, and the amplitude of different frequency components can reflect the nature and location of local discharge. High-frequency components may be related to surface discharge, while low-frequency components may be related to volume discharge.

[0107] (III) Construction of CNN-GRU neural network model:

[0108] Based on the eigenvalue, a cable discharge current signal recognition regression model based on the CNN-GRU neural network is established. The convolutional neural network (CNN) will be used to extract the various features in the discharge current signal, while the GRU is responsible for capturing the changes in the working state of the current signal over time. Finally, the mean square error, mean absolute error, mean square error percentage, root mean square error and determination coefficient are selected as evaluation indicators to compare the model of the present invention with the other two models to verify the superiority of the method proposed in the present invention.

[0109] (IV) Working principle:

[0110] 1. CNN-GRU neural network model:

[0111] The cable discharge current identification technology based on CNN-GRU neural network effectively combines the advantages of convolutional neural network (CNN) and gated recurrent unit (GRU), and can effectively process data related to cable discharge current to achieve accurate discharge current identification.

[0112] (1) Convolutional Neural Network (CNN):

[0113] CNN is a deep learning algorithm widely used in image and time series signal classification. In the cable discharge current, various feature data related to the discharge current can be converted into a format suitable for CNN processing after appropriate preprocessing. CNN mainly consists of convolutional layer, activation layer, pooling layer and fully connected layer.

[0114] In the cable discharge current signal recognition, the spatiotemporal characteristics of the discharge current waveform can be regarded as a one-dimensional signal. Through one-dimensional convolution operation, the convolutional neural network (CNN) can extract the local pattern of the signal.

[0115] Convolution layer: In order to capture the local changes of the discharge current signal, convolution is used to effectively extract local features.

[0116] Z=X*W+b

[0117] Where: Z is the output feature map, X is the input, * indicates the convolution operation, W is the weight matrix, and b is the bias term.

[0118] Activation layer, after each convolution layer, a nonlinear activation function, namely ReLU activation function, is added to increase the nonlinearity of the network so that the network can learn more complex features. The expression of the activation function is as follows;

[0119] A=ReLU(Z)=max(0,Z)

[0120] Pooling layer: Generally speaking, by using the maximum pooling layer (Max Pooling), the spatial dimension of the feature map can be reduced while retaining important features.

[0121] P=max(A)

[0122] Where: P is the output value of the pooling layer; A is the output value after ReLU activation.

[0123] The fully connected layer is used to flatten the features after convolution and pooling and perform the final classification or regression task.

[0124]

[0125] Where: z j is the output value of the jth neuron; w ij is the concatenated input feature; x i is the i-th feature of the input; b j is the bias term for the output neuron j.

[0126] Usually the output of the fully connected layer will pass through a softmax activation function to obtain the classification probability. The softmax activation function is as follows:

[0127]

[0128] Where: f j is the predicted probability of the jth class; z j is the jth output value of the fully connected layer; k is the total number of categories.

[0129] The convolution layer slides over the input data through the weight matrix to extract local features; the pooling layer is used to reduce the dimension and reduce the computational complexity of the model. The fully connected layer is used to flatten the features after convolution and pooling and perform the final classification.

[0130] In the cable discharge signal recognition task, the spatiotemporal characteristics of the discharge current waveform can be regarded as a one-dimensional signal. Through one-dimensional convolution operation, CNN can extract the local pattern of the signal, thereby providing a richer feature representation for the subsequent GRU.

[0131] (2) Gated Recurrent Unit GRU:

[0132] GRU is a variant of RNN, which can effectively alleviate the gradient vanishing problem in traditional RNN and capture the changes of the core working state over time. Its network structure is shown in the figure Figure 3 As shown. It controls the flow of information through a gating mechanism (including an update gate and a reset gate), and is particularly suitable for processing long time series data. In the present invention, GRU is used to capture the time correlation of the cable discharge current signal, so as to accurately identify the discharge type at different times.

[0133] Update gate: It is used to control how the information of the previous time step and the information of the current candidate hidden state are mixed, and determines how much information from the past is retained in the current hidden state.

[0134] z t =λ(W z ·[h t-1 ,x t ])

[0135] Where: z tis the update gate at time step t. λ is the sigmoid activation function, W Z is the weight matrix in the update gate, x t is the current input, h t-1 It is the hidden state at the previous moment.

[0136] Reset gate: You can control whether the current time step is affected by the previous hidden state.

[0137] r t =λ(W r ·[h t-1 ,x t ])

[0138] Where: r t is the reset gate at time step t; W r is the weight matrix of the reset gate.

[0139] Candidate hidden state: Combine the previous hidden state with the current input and update its value;

[0140] h t =tanh(W h ·[r t *h t-1 ,x t ]);

[0141] Where: h t is the candidate hidden state at time step t; W h is the weight matrix of the candidate hidden states.

[0142] Final hidden state: weighted sum of the previous hidden state and candidate hidden state, and update it to the new hidden state;

[0143] h t =z t *h t-1 +(1-z t )*h t

[0144] Where: h t represents the new hidden state.

[0145] (3) CNN-GRU hybrid model:

[0146] The CNN-GRU model proposed in this invention first extracts the spatial features of the discharge current signal through multiple convolutional layers, and then inputs the extracted features into the GRU network to capture its temporal dependency. The overall architecture of the model is as follows: Figure 4 shown.

[0147] Initialize network weights: The model randomly initializes the weights of the CNN and GRU layers. This process provides initial diversity and flexibility for model training, preventing the model from falling into deterministic results, so that different optimization paths can be explored to better adapt to diverse data distributions and task requirements. The present invention uses He initialization to adapt to the characteristics of ReLU through a higher weight variance, reducing the problem of gradient disappearance. The weights W are randomly sampled from a uniform distribution in the following interval:

[0148]

[0149] Where: W is the weight matrix; is uniformly distributed, with a value range of [a, b]; n in is the number of input neurons in the current layer.

[0150] Forward propagation: Through convolution and pooling operations, feature values ​​are extracted from the discharge current signal data and input into the GRU network to capture its time dependency. Subsequently, the output of the GRU is processed through a fully connected layer to ultimately achieve the goal of predicting the core loss.

[0151] Calculate the loss: Since there is a certain error between the predicted value and the true value, the error must be calculated through the loss function. The formula is as follows.

[0152]

[0153] Where: L is the loss function value; N is the number of samples; y i Reflects the real discharge current signal, y i is the predicted discharge current signal.

[0154] Back propagation: By gradually calculating the gradient of the error from the output layer to the input layer, it is determined how to adjust the various parameters in the network so that the value of the loss function gradually decreases.

[0155]

[0156] Where: is the gradient of the loss function L with respect to the weight matrix W; The partial derivative of the loss function with respect to the output y, is the partial derivative of the output layer with respect to the weight W.

[0157] Update weights: Use the optimizer to update and adjust the network weights. This gradually adjusts the network weight parameters to achieve the goal of reducing the loss function value and improving model performance. The formula is as follows.

[0158]

[0159] Where: ε is the learning rate, W new is the new weight matrix; W old is the old weight matrix; is the gradient of the loss function L with respect to the weight matrix W.

[0160] Iteration: Continue iterating until the loss converges.

[0161] 2. Model data processing:

[0162] (1) Data preprocessing:

[0163] The discharge current data set used in the experiment comes from the cable discharge signal data actually collected by the power company. The data contains multiple discharge types, including partial discharge, surface discharge, and gap discharge. Each type of data is collected by sensors and labeled accordingly.

[0164] To ensure the training effect of the model, the data is first normalized and the amplitude of each discharge signal is scaled to the interval [0,1]. The present invention adopts the normalization processing method of the maximum and minimum mean method to convert the original data into dimensionless per unit value to eliminate the dimension difference between the data. The formula is as follows.

[0165]

[0166] Where: X max Indicates the maximum value in the data; X min represents the minimum value in the data; X′ is the dimensionless value obtained after normalization; X i is the i-th data in the original data to be processed.

[0167] In addition, the noise signal is filtered to ensure that the signal input into the model has a high signal-to-noise ratio.

[0168] (2) Feature extraction and evaluation indicators:

[0169] Feature extraction in traditional methods usually relies on time-frequency analysis techniques such as Fourier transform and wavelet transform. However, the present invention uses CNN to automatically extract signal features, overcoming the limitations of manual feature selection. By stacking multiple layers of convolution kernels, this method can simultaneously capture the local details and overall patterns of the discharge signal.

[0170] In order to evaluate the performance of the regression model, the mean square error (M 1 ), mean absolute error (M 2 ), mean square error percentage (M 3 ), Root Mean Square Error (M 4 ), and decide 1 Determined coefficient (R 2) as the evaluation indicator:

[0171] 1) Mean absolute error (M) 1 ): The model performance is evaluated by calculating the mean absolute deviation between the predicted value and the true value, using M 1 The smaller the M 1 The value indicates that the average absolute deviation between the predicted value and the true value is smaller, indicating that the model has higher prediction accuracy. The formula is as follows.

[0172]

[0173] 2) Mean absolute percentage error (M 2 ): M 2 The average deviation of the predicted value from the true value is measured in percentage form, focusing on the relative error between the predicted value and the true value. The specific formula is shown below.

[0174]

[0175] 3) Mean square error (M 3 ): Through M 3 The value can effectively reflect the size of the error, M 3 The smaller the value, the smaller the difference between the model prediction value and the true value, which means the better the model performance. This indicator can directly reflect the accuracy of the model prediction. The specific formula is as follows.

[0176]

[0177] 4) Root mean square error (M 4 ): This indicator reflects the square root of the average of the squares of the differences between the predicted value and the true value, comprehensively considering the errors of all sample points and providing a comprehensive assessment of the overall error of the model. The specific formula is shown below.

[0178]

[0179] 5) Coefficient of determination (R 2 ): This indicator can measure the model's fitting effect on the data, R 2 The closer the value is to 1, the better the model fits the data, indicating that the model performs better. The specific formula is as follows.

[0180]

[0181] Where: y i is the true value, y i is the predicted value, y is the true mean, and n is the number of samples.

[0182] 3. Model simulation:

[0183] In order to train the CNN-GRU model, the cross entropy loss function is used, and the Adam optimizer is used to update the model parameters. During the training process, an early termination strategy is adopted to prevent overfitting, and a 5-fold cross validation is performed on the training set to evaluate the generalization ability of the model. Its parameter settings are shown in Table 1.

[0184] Table 1 Training parameter settings

[0185]

[0186] In order to verify the effect of the technology proposed in the present invention on the identification of cable discharge current, the obtained experimental data are processed and analyzed. At the same time, in order to verify the advantages of CNN-GRU, the present invention compares several common deep learning models, including decision tree (RT) and random forest (RF).

[0187] According to the experimental simulation analysis, the following experimental results are obtained: 1 and M 2 、M 3 and M 4 These two pairs of evaluation indicators are very similar, so for these two indicators, the present invention only selects M 2 and M 3 The three models are analyzed and compared.

[0188] Table 2 Model comparison results

[0189]

[0190] It can be seen from Table 2 that the mean absolute percentage error M based on the CNN-GRU neural network model 2 The lowest, which means that the average percentage error of the predicted values ​​relative to the actual values ​​is extremely low. 3 This value is also the lowest based on the CNN-GRU neural network model, indicating that the mean square difference between the predicted value and the actual value is relatively small. 2 ) is the highest among all the models based on CNN-GRU neural network, which fully demonstrates that the model can explain the changes of dependent variables and has strong explanatory power.

[0191] Experiments show that CNN-GRU performs better in processing complex time series signals, especially in weak pattern recognition of discharge signals. Experimental results on the test set show that the accuracy of the CNN-GRU model proposed in the present invention in the discharge signal classification task reaches 99.49%, which is significantly higher than traditional machine learning methods such as. At the same time, the convergence speed of training loss also shows that the model has good stability.

[0192] In order to visualize the results, we can compare the 40 random points in the test set of the above three models with the actual values, which can clearly reflect the accuracy of the prediction results and the prediction effect. Figure 5 to Figure 7 As shown in the figure. From these three figures, we can see that CNN-GRU has obvious advantages over other models in recognition accuracy, especially in the cable current discharge signal processing, its recognition accuracy is higher. Therefore, it can be inferred that CNN-GRU can accurately identify the cable discharge current and then accurately judge the related defects. This feature has a positive impact on the maintenance and repair of cables.

[0193] (V) A cable discharge current identification method based on CNN-GRU neural network of the present invention has the following characteristics:

[0194] (1) The present invention adopts a high-frequency current method to monitor the cable discharge current signal, selects characteristic values ​​including discharge current amplitude, discharge waveform and temperature as model inputs, and constructs a neural network recognition model based on these characteristic values, thereby realizing efficient recognition of the cable discharge signal.

[0195] (2) The present invention proposes a hybrid neural network model combining a convolutional neural network (CNN) and a gated recurrent unit (GRU) for the recognition of cable discharge current signals. CNN is used to extract features from the discharge current signal, and GRU is used to capture the working state of the signal that changes over time, thereby improving the efficiency and accuracy of signal recognition. By applying this model to the cable safety management in distribution substations, the efficiency and accuracy of power cable signal recognition can be improved, providing guidance for the design, construction, operation and maintenance of power cable projects.

[0196] (3) The present invention comprehensively verifies the superiority of the proposed CNN-GRU neural network model over other models in cable discharge current signal recognition through multiple evaluation indicators such as mean square error, mean absolute error, mean square error percentage, root mean square error and determination coefficient.

[0197] (4) The present invention adopts the high-frequency current method as the cable discharge signal monitoring. The characteristics of the cable discharge signal are relatively complex, and the relevant model established based on the expert system is not accurate enough. The CNN-GRU neural network model can better analyze various characteristic values ​​and increase the accuracy of model prediction.

[0198] (5) The present invention establishes a cable discharge current signal model with time series characteristics by training the model. Tests in specific cases show that the cable discharge current recognition technology based on the CNN-GRU neural network model can effectively improve the accuracy of discharge current signal recognition.

Claims

1. A cable discharge current identification method based on CNN-GRU neural network, characterized in that: The following steps are involved: Step 1: Obtain the discharge current signal data of the cable; Step 2: Use convolutional neural network (CNN) to extract various current feature quantities of the discharge current signal, construct a current feature data set, and divide the current feature data set into a training set and a test set; Step 3: Build a CNN-GRU neural network model and train the CNN-GRU neural network model using the training set; Step 4: Input the test set data into the trained CNN-GRU neural network model and output the recognition result of the cable discharge current signal.

2. The cable discharge current identification method based on the CNN-GRU neural network according to claim 1 is characterized in that: In step 1, the discharge current signal data is preprocessed as follows: The discharge current signal data is normalized, and the amplitude of each discharge current signal is scaled to the [0,1] interval. The maximum and minimum mean normalization method is used to convert the discharge current signal data into a dimensionless per unit value to eliminate the dimension difference between the data. The formula is as follows: Where: X max Indicates the maximum value in the data; X min represents the minimum value in the data; X′ is the dimensionless value obtained after normalization; X i is the i-th data in the discharge current signal data to be processed.

3. The cable discharge current identification method based on the CNN-GRU neural network according to claim 1 is characterized in that: In step 2, the current characteristic quantity includes: ①, discharge current peak value; ②, discharge waveform; 3. Temperature; ④. Number of partial discharges; ⑤, discharge duration; ⑥. Discharge spectrum characteristics.

4. The cable discharge current identification method based on the CNN-GRU neural network according to claim 1 is characterized in that: In step 3, the constructed CNN-GRU neural network model includes: (1) Convolutional neural network (CNN) includes: convolutional layer, activation layer, pooling layer and fully connected layer; Convolutional layer, which extracts local features by sliding the convolution kernel over the input data; Z = X*W+b; Where: Z is the output feature map, X is the input, * indicates the convolution operation, W is the convolution kernel or weight matrix, and b is the bias term; Activation layer, after each convolution layer, a nonlinear activation function, namely ReLU activation function, is added to increase the nonlinearity of the network so that the network can learn more complex features. The expression of the activation function is as follows; A = ReLU(Z) = max(0,Z); The pooling layer selects the maximum value within the pooling window to reduce the dimension and computational complexity of the model; P = max(A); Where: P is the output value of the pooling layer; A is the output value after ReLU activation; The fully connected layer is used to flatten the features after convolution and pooling and perform the final classification or regression task; Where: z j is the output value of the jth neuron; w ij is the concatenated input feature; x i is the i-th feature of the input; b j is the bias term of output neuron j; Usually the output of the fully connected layer will pass through a softmax activation function to obtain the classification probability. The softmax activation function is as follows: Where: f j is the predicted probability of the jth class; z j is the jth output value of the fully connected layer; k is the total number of categories; In the cable discharge current signal recognition, the spatiotemporal characteristics of the discharge current waveform can be regarded as a one-dimensional signal. Through the one-dimensional convolution operation, the convolutional neural network (CNN) can extract the local pattern of the signal. (2) Gated Recurrent Unit GRU: The gated recurrent unit (GRU) is used to capture the temporal correlation of the cable discharge current signal, so as to accurately identify the discharge type at different times. The gated recurrent unit GRU includes: update gate and reset gate; Update gate: used to control how the information of the previous time step is mixed with the information of the current candidate hidden state, and determines how much information from the past is retained in the current hidden state; z t =λ(W z ·[h t-1 ,x t ]); Where: z t is the update gate at time step t, λ is the sigmoid activation function, W Z is the weight matrix in the update gate, x t is the current input, h t-1 is the hidden state of the previous moment; Reset gate: used to control whether the current time step is affected by the previous hidden state; r t =λ(W r ·[h t-1 ,x t ]); Where: r t is the reset gate at time step t; W r is the weight matrix of the reset gate; Candidate hidden state: Combine the previous hidden state with the current input and update its value; h t =tanh(W h ·[r t *h t-1 ,x t ]); Where: h t is the candidate hidden state at time step t; W h is the weight matrix of the candidate hidden states; Final hidden state: weighted sum of the previous hidden state and candidate hidden state, and update it to the new hidden state; h t =z t *h t-1 +(1-z t )*h t ; Where: h t represents the new hidden state.

5. The cable discharge current identification method based on CNN-GRU neural network according to claim 4 is characterized in that: In step 3, the CNN-GRU neural network model is trained using the training set, as follows: During the training process of the CNN-GRU neural network model, first, the network weights are initialized and forward propagation is performed to calculate the initial output of the network; then, based on the results of the forward propagation, the loss function is calculated, which is used to quantify the gap between the output of the CNN-GRU neural network model and the true value; finally, the network parameters are adjusted through the back-propagation algorithm to make the CNN-GRU neural network model gradually converge and improve the prediction accuracy.

6. The cable discharge current identification method based on CNN-GRU neural network according to claim 5 is characterized in that: In step 3, the CNN-GRU neural network model is trained using the training set, including the following steps: S3.1: Initialize network weights: Randomly initialize the weights of the CNN-GRU neural network model. The present invention uses He initialization to adapt to the characteristics of ReLU through a higher weight variance, reducing the problem of gradient disappearance. The weights W are randomly sampled from a uniform distribution in the following interval: Where: W is the weight matrix; is uniformly distributed, with a value range of [a, b]; n in is the number of input neurons in the current layer; S3.2: Forward propagation: Through convolution and pooling operations, feature values ​​are extracted from the discharge current signal data, and these feature values ​​are input into the gated recurrent unit GRU network to capture its temporal dependency; then, the output of the gated recurrent unit GRU is processed through the fully connected layer to finally achieve the goal of predicting the discharge current signal; S3.3: Calculate the loss function: Since there is a certain error between the predicted value and the true value, the error should be calculated through the loss function. The formula is as follows: Where: L is the loss function value; N is the number of samples; y i Reflects the real discharge current signal, y i is the predicted discharge current signal; S3.4: Back propagation: By gradually calculating the gradient of the error from the output layer to the input layer, it is determined how to adjust the various parameters in the network so that the value of the loss function gradually decreases; Where: w is the gradient of the loss function L with respect to the weight matrix W; The partial derivative of the loss function with respect to the output y, is the partial derivative of the output layer with respect to the weight W; S3.5: Update weights: Update and adjust the network weights with the help of the optimizer; thereby gradually adjusting the network weight parameters to achieve the goal of reducing the loss function value and improving model performance. The formula is as follows: W new =W old -ε·▽ w L; Where: ε is the learning rate, W new is the new weight matrix; W old is the old weight matrix; ▽ w is the gradient of the loss function L with respect to the weight matrix W; S3.6: Iteration: Continue iterating until the loss converges.

7. The cable discharge current identification method based on CNN-GRU neural network according to claim 1 is characterized in that: In step 4, the mean square error M1, mean absolute error M2, mean square error percentage M3, root mean square error M4, and determination coefficient R are selected. 2 As evaluation indicators, the performance of the CNN-GRU neural network model is evaluated, including: 1) Mean absolute error M1: The model performance is evaluated by calculating the mean absolute deviation between the predicted value and the true value, and the M1 value is used for measurement; a smaller M1 value indicates a smaller mean absolute deviation between the predicted value and the true value, indicating that the model has a higher prediction accuracy; the formula is as follows: 2) Mean absolute percentage error M2: M2 measures the average deviation of the predicted value from the true value in percentage form. The specific formula is as follows: 3) Mean square error M3: The M3 value can effectively reflect the size of the error. The smaller the M3 value, the smaller the difference between the model prediction value and the true value, which means that the model performance is better. The specific formula is as follows: 4) Root mean square error M4: This indicator reflects the square root of the average of the squares of the differences between the predicted value and the true value. The specific formula is as follows: 5) Determination coefficient R 2 : This indicator can measure the model's fitting effect on the data, R 2 The closer the value is to 1, the better the model fits the data, which means the better the performance of the model. The specific formula is as follows: Where: y i is the true value, y i is the predicted value, is the true mean, and n is the number of samples.

Images