Method for detecting deformation of transformer winding based on multi-scale multi-channel CNN-LSTM model

Abstract

The invention relates to the technical field of electrical equipment fault diagnosis, and discloses a method for detecting transformer winding deformation based on a multi-size multi-channel CNN-LSTM model. Firstly, excitation signals, response signals and fault simulation signals of a transformer needing to be monitored under different frequencies and different working conditions are obtained. Afterwards, the signals are subjected to normalization processing, amplitude frequency parameters under different frequencies are obtained according to a formula, an amplitude frequency response curve is drawn, and a frequency response curve database under different working conditions is established; and dividing the database, and training a multi-size multi-channel CNN-LSTM model. And finally, collecting frequency response data of the transformer in actual operation, and obtaining the operation state of the current winding through the trained model. The frequency response generated when the transformer winding deforms is analyzed by using the multi-scale multi-channel CNN-LSTM model, so that the calculation amount of the model is not increased, the characteristics of each level of the image can be expressed, the signal characteristics are comprehensively mined, and the accuracy of winding deformation diagnosis is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical equipment fault diagnosis, and in particular to a method for detecting transformer winding deformation based on a multi-scale multi-channel CNN-LSTM model. Background Art

[0002] Transformers play a vital role in the power grid. Their main functions include voltage conversion, current transmission, power transmission, isolation and insulation, voltage stability adjustment, safety protection, power loss reduction, and phase conversion. Transformers can increase or decrease voltage as needed, effectively transmit electrical energy to power consumption areas, and meet the voltage level requirements of different power consumption equipment.

[0003] Transformers play an indispensable role in power grids, and the prediction and health management of their faults are of great significance for ensuring the stable operation of power grids, reducing operation and maintenance costs, and improving the reliability and efficiency of the entire power system. First of all, as the core equipment for power conversion, transformers may cause huge economic losses and social impacts once they fail. Through fault prediction, timely measures can be taken before the failure occurs to avoid accidents, thereby improving the reliable operation of transformers. Secondly, transformer failures are more complex, with high maintenance costs, and may affect the safety and stability of the power system. Therefore, health management and preventive maintenance of transformers can reduce the probability of failures, reduce maintenance costs, and ensure the safe and stable operation of the power grid.

[0004] Among transformer winding mechanical failures, the incidence of winding deformation is higher than other failures and will have very serious consequences, including changes in axial and radial dimensions, body displacement, winding distortion, bulging and inter-turn short circuits, etc. There are many methods to detect winding deformation, such as low-voltage pulse method, frequency response analysis method, short-circuit reactance method, and vibration detection method. Among them, the frequency response analysis method has formed a standard and is widely used in actual detection, but it needs to be judged in combination with manual experience. Therefore, it is necessary to combine multi-size and multi-channel models to fully mine data information and further improve diagnostic efficiency and accuracy. Summary of the invention

[0005] To solve the above problems, the present invention proposes a method for detecting transformer winding deformation based on a multi-size and multi-channel CNN-LSTM model.

[0006] The technical solution adopted by the present invention is:

[0007] A method for detecting transformer winding deformation based on a multi-scale multi-channel CNN-LSTM model, characterized in that it comprises the following steps:

[0008] Step 1, obtaining the excitation signal, response signal and fault simulation signal of the required monitoring transformer under different frequencies and different working conditions, and realizing frequency response signal acquisition and transmission;

[0009] Step 2, normalize these signals, obtain the amplitude frequency parameters at different frequencies according to the formula and draw the amplitude frequency response curve, establish a frequency response curve database for different working conditions, and realize normalization and frequency response curve drawing;

[0010] Step 3, divide the database;

[0011] Step 4, setting the parameters of the multi-scale multi-channel CNN-LSTM model, and using the database to train the multi-scale multi-channel CNN-LSTM model;

[0012] Step 5: Collect the frequency response data of the transformer in actual operation, and obtain the operating status of the current winding through the trained model.

[0013] Further, step 1 comprises the following steps:

[0014] Signal acquisition: Inject an excitation signal of 100kHz to 1MHz into the lower transformer to obtain its electrical signal response;

[0015] Signal transmission: The signal is transmitted to the signal amplifier through a low-noise cable, and the response signal and input signal are displayed and recorded on an oscilloscope;

[0016] Data management: Upload the waveform data detected by the oscilloscope to the cloud data management platform for storage and subsequent analysis;

[0017] Simulation: Use multi-physics field models to simulate the working state of the transformer under different fault conditions and generate corresponding fault response signals.

[0018] Further, step 2 comprises the following steps:

[0019] Data normalization: Z-score normalization is applied to eliminate the influence of environmental factors, and the original data is processed through the formula to make its mean 0 and standard deviation 1;

[0020] Amplitude frequency response parameter calculation: Calculate the amplitude response under normal operation and fault conditions according to the amplitude frequency response parameter formula to evaluate transformer performance;

[0021] Curve drawing: Draw amplitude frequency response curves from 100kHz to 1MHz, including response curves of normal state and different fault states, for performance evaluation and fault diagnosis.

[0022] Furthermore, the calculation formula for Z-score normalization is as follows:

[0023]

[0024] In the formula, A is the original data; μ is the mean of the data; σ is the standard deviation of the data; A norm Represents the data after Z-score normalization.

[0025] Furthermore, the formula for the amplitude frequency response parameter is as follows:

[0026]

[0027] Where P is the amplitude frequency response parameter, U1 is the amplitude of the transformer response signal, and U2 is the amplitude of the input excitation signal.

[0028] Furthermore, in step 3, partitioning the database includes the following steps:

[0029] Data construction: construct various amplitude frequency response curves into a fault diagnosis data set that combines virtual and actual data;

[0030] Sample sampling: The sliding window method is used to perform overlapping segmentation sampling on the normalized data, and several samples are extracted from each working state;

[0031] Dataset division: 70% of the samples are used as training sets, 15% of the samples are used as validation sets, and 15% of the samples are used as test sets.

[0032] Furthermore, the multi-scale multi-channel CNN-LSTM model in step 4 has a model structure including multiple convolutional neural networks CNN and long short-term memory networks LSTM; the CNN network includes three branches, branch one contains three convolutional modules, branch two contains convolutional module 2 and convolutional module 3, and branch three contains only convolutional module 2;

[0033] The convolution modules are set as follows: convolution module 1 contains 64 convolution kernels of size 3×3, convolution module 2 contains 128 convolution kernels of size 3×3, and convolution module 3 contains two layers of 3×3×256 convolution kernels.

[0034] Furthermore, in step 4, the settings of the multi-scale multi-channel CNN-LSTM model parameters include convolution kernel size, activation function, pooling parameters, optimizer, loss function and learning rate;

[0035] Activation function: Select ReLU function;

[0036] Pooling parameters: The size of the pooling window should be smaller than the convolution kernel, and the pooling step size should be smaller than the convolution kernel step size;

[0037] Optimizer: Use Adamax optimizer;

[0038] Loss function: optimized cross entropy loss function;

[0039] Learning rate: The initial learning rate is set to 0.0001.

[0040] Furthermore, the loss function L CrossEntropy The optimized cross entropy loss function is selected, and the expression is as follows:

[0041] f(x) ReLU =max(0,x) (3)

[0042] Where x is the input value; f(x) is the function defined on the input (x), which specifically implements the logic of the activation function; ReLU is the Rectified Linear Unit, which represents the activation function; max() means returning the maximum of its input values; and 0 is a constant value.

[0043]

[0044] Where, L CrossEntropy Cross entropy loss function; m is the number of samples; is a feature vector, representing the jth feature of the i-th sample; is the true label corresponding to each sample, indicating the true category label corresponding to the i-th sample; n is the artificially set fault label of each type, such as: 5% axial displacement is recorded as 1, 15% axial displacement is recorded as 2, 5% radial displacement is recorded as 3, 15% radial displacement is recorded as 4, ..., and so on; R(θ) is the regularization term; λ is the regularization coefficient; θ represents the model parameter.

[0045]

[0046] Where R(θ) is the regularization term; θ represents the model parameter; k is the kth individual parameter that constitutes the model parameter; Represents θ k Performs a square operation.

[0047] Further, step 5 comprises the following steps:

[0048] Model training: Use the training set and validation set to adjust and train the model parameters;

[0049] Model testing: Verify the model's accuracy in identifying winding faults through the test set;

[0050] Real-time analysis: The trained model is used to analyze the frequency response signal acquired in real time to obtain the operating status of the transformer winding through the classifier.

[0051] The beneficial effects of the present invention are:

[0052] The present invention proposes a detection method for transformer winding deformation based on a multi-scale multi-channel CNN-LSTM model. First, the excitation signal, response signal and fault simulation signal of the required monitoring transformer under different frequencies and different working conditions are obtained. After that, these signals are normalized, and the amplitude frequency parameters under different frequencies are obtained according to the formula and the amplitude frequency response curve is drawn to establish a frequency response curve database of different working conditions. The database is divided and the multi-scale multi-channel CNN-LSTM model is trained. Finally, the frequency response data of the transformer in actual operation is collected, and the operating status of the current winding is obtained through the trained model. The frequency response generated when the transformer winding is deformed is analyzed using a multi-scale multi-channel CNN-LSTM model, which does not increase the calculation amount of the model, but also can express the characteristics of each level of the image, comprehensively mine the signal characteristics, and improve the accuracy of the diagnosis of winding deformation. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flow chart of a method for detecting transformer winding deformation based on a multi-scale multi-channel CNN-LSTM model of the present invention;

[0054] Figure 2 It is a schematic diagram of a signal acquisition and transmission platform of the method of the present invention;

[0055] Figure 3 Schematic diagram of the multi-size multi-channel CNN-LSTM fault diagnosis model of the method of the present invention. DETAILED DESCRIPTION

[0056] The present invention is described below based on embodiment, but it should be noted that the present invention is not limited to this embodiment. In the detailed description of the present invention below, some specific details are described in detail. However, for the part not described in detail, those skilled in the art can also fully understand the present invention.

[0057] This embodiment provides a method for detecting transformer winding deformation based on a multi-scale multi-channel CNN-LSTM model. Figure 1 As shown, the method for detecting transformer winding deformation based on the multi-scale multi-channel CNN-LSTM model includes the following steps:

[0058] Step 1: obtain the excitation signal, response signal and fault simulation signal of the required monitoring transformer under different frequencies and different working conditions, and realize frequency response signal acquisition and transmission.

[0059] Specifically, Figure 2As shown, signal acquisition: inject different frequency excitation signals from 100kHz to 1MHz into the transformer to obtain its electrical signal response. Signal transmission: the response electrical signal is transmitted to the signal amplifier through a low-noise cable and displayed on the oscilloscope as the measured response signal; at the same time, the oscilloscope is used to record the measured input signal of the response. Data management: after the oscilloscope is connected to the network, the detected waveform data is transmitted to the cloud data management platform for storage.

[0060] Simulation: The transformer multi-physics field model is used to simulate the working state of the transformer when the winding is deformed, including axial displacement fault, radial dimensional deformation, body displacement, winding distortion, bulging and inter-turn short circuit. The sampling rate is kept the same as the measured signal acquisition process to obtain the response signal of the transformer under different working conditions as the fault response signal of the transformer.

[0061] Step 2: normalize these signals, obtain the amplitude-frequency parameters at different frequencies according to the formula and draw the amplitude-frequency response curve, establish a frequency response curve database for different working conditions, and realize normalization and frequency response curve drawing.

[0062] Data normalization: To eliminate abnormal signal peaks caused by environmental factors, Z-score normalization was used to obtain normally distributed data with a mean of 0 and a standard deviation of 1.

[0063] The calculation formula for Z-score normalization is as follows:

[0064]

[0065] In the formula, A is the original data, is the mean of the data; σ is the standard deviation of the data; A norm Represents the data after Z-score normalization.

[0066] The amplitude frequency response parameter calculation is based on formula (1) to normalize the measured response signal, measured input signal, and fault response signal obtained in the first step, and then obtain the amplitude frequency response parameters under different working conditions at different frequencies to evaluate the transformer performance.

[0067] The formula for the amplitude frequency response parameter is as follows:

[0068]

[0069] Where P is the amplitude frequency response parameter; U1 is the amplitude of the transformer response signal, and U2 is the amplitude of the input excitation signal.

[0070] Curve drawing: Draw the amplitude-frequency response curve under normal operation with a frequency of 100kHz to 1MHz, and the amplitude-frequency response curve of the winding under different fault conditions, so as to carry out performance evaluation and fault diagnosis.

[0071] Step 3: partition the database.

[0072] Data construction: The various amplitude-frequency response curves obtained are constructed into a virtual-real fault diagnosis data set.

[0073] Sample sampling: In order to increase the sample data of the deep learning network, the normalized data is sampled using the sliding window method. To ensure that each sample can fully reflect the fluctuation characteristics, overlapping segmentation sampling is performed, and 500 samples of the response data under each working state are obtained.

[0074] Dataset division: 70% of the samples are used as training sets, 15% of the samples are used as validation sets, and 15% of the samples are used as test sets.

[0075] Step 4: Set the parameters of the multi-scale multi-channel CNN-LSTM model and use the database to train the multi-scale multi-channel CNN-LSTM model.

[0076] The multi-scale multi-channel CNN-LSTM model is composed of multiple convolutional neural networks (CNNs) fused with long short-term memory networks (LSTMs). The parameters to be determined include convolution kernel size, activation function, pooling parameters, optimizer, loss function, and learning rate.

[0077] Model structure:

[0078] like Figure 3 As shown in the figure, the multi-scale multi-channel CNN network has three branches. Convolution module 1 contains 64 convolution kernels of size 3×3, convolution module 2 contains 128 convolution kernels of size 3×3, and convolution module 3 contains two layers of 3×3×256 convolution kernels. Among them, branch 1 contains three convolution modules, branch 2 contains convolution module 2 and convolution module 3, and the branch module only contains convolution module 2.

[0079] The size of the pooling window should be smaller than the size of the convolution kernel, and the pooling step size should be smaller than the step size of the convolution kernel. The Adamax optimizer is selected as the optimizer. The Adamax optimizer does not need to adjust the learning rate during training, so just set the initial learning rate to 0.0001 to avoid poor model results. The ReLU function is selected as the activation function.

[0080] Loss function L CrossEntropy The optimized cross entropy loss function is selected, and the expression is as follows:

[0081] f(x) ReLU =max(0,x) (3)

[0082] In the formula, x is the input value; f(x) is the function defined on the input x, which specifically implements the logic of the activation function; ReLU is the rectified linear unit, which represents the activation function; max() means returning the maximum of its input values; 0 is a constant value.

[0083]

[0084] Where, L CrossEntropy Cross entropy loss function; m is the number of samples; is a feature vector, representing the jth feature of the i-th sample; is the true label corresponding to each sample, indicating the true category label corresponding to the i-th sample; n is the artificially set fault label of each type, such as: 5% axial displacement is recorded as 1, 15% axial displacement is recorded as 2, 5% radial displacement is recorded as 3, 15% radial displacement is recorded as 4, ..., and so on; R(θ) is the regularization term; λ is the regularization coefficient; θ represents the model parameter.

[0085]

[0086] Where R(θ) is the regularization term; θ represents the model parameter; k is the kth individual parameter that constitutes the model parameter; Represents the k Performs a square operation.

[0087] Step 5: Collect the frequency response data of the transformer in actual operation, and obtain the operating status of the current winding through the trained model.

[0088] Model training: Use the divided data set to train and test the model, adjust the model parameters and train according to the training set and validation set; Model testing: Use the test set to verify the model's accuracy in identifying winding faults; Real-time analysis: Use the trained model to analyze and process the frequency response signal obtained in real time, and obtain the operating status of the transformer winding through the classifier.

[0089] Compared with the currently popular deep learning methods, the multi-scale multi-channel CNN-LSTM model has significantly reduced the number of convergence steps in training the model, and has improved automated feature extraction, multi-source data processing, local and global feature recognition, and adaptability. Compared with traditional fault diagnosis methods, such as the low-voltage pulse method and the short-circuit reactance method, this model has significant advantages in diagnostic efficiency and accuracy, reduced maintenance costs, real-time monitoring and prediction, and adaptive learning capabilities.

[0090] At the same time, it should be pointed out that the above description is only a typical embodiment of the present invention and does not limit the present invention in any form. Any changes made to the above embodiments based on the ideas and technical essence of the present invention are within the protection scope of the present invention.

Claims

1. A method for detecting transformer winding deformation based on a multi-scale multi-channel CNN-LSTM model, characterized in that: The following steps are involved: Step 1, obtaining the excitation signal, response signal and fault simulation signal of the required monitoring transformer under different frequencies and different working conditions, and realizing frequency response signal acquisition and transmission; Step 2, normalize these signals, obtain the amplitude frequency parameters at different frequencies according to the formula and draw the amplitude frequency response curve, establish a frequency response curve database for different working conditions, and realize normalization and frequency response curve drawing; Step 3, divide the database; Step 4, setting the parameters of the multi-scale multi-channel CNN-LSTM model, and using the database to train the multi-scale multi-channel CNN-LSTM model; Step 5: Collect the frequency response data of the transformer in actual operation, and obtain the operating status of the current winding through the trained model.

2. The method for detecting transformer winding deformation based on a multi-scale multi-channel CNN-LSTM model according to claim 1 is characterized in that: Step 1 includes the following steps: Signal acquisition: Inject an excitation signal of 100kHz to 1MHz into the lower transformer to obtain its electrical signal response; Signal transmission: The signal is transmitted to the signal amplifier through a low-noise cable, and the response signal and input signal are displayed and recorded on an oscilloscope; Data management: Upload the waveform data detected by the oscilloscope to the cloud data management platform for storage and subsequent analysis; Simulation: Use multi-physics field models to simulate the working state of the transformer under different fault conditions and generate corresponding fault response signals.

3. The method for detecting transformer winding deformation based on a multi-scale multi-channel CNN-LSTM model according to claim 1, characterized in that: Step 2 includes the following steps: Data normalization: Z-score normalization is applied to eliminate the influence of environmental factors, and the original data is processed through the formula to make its mean 0 and standard deviation 1; Amplitude frequency response parameter calculation: Calculate the amplitude response under normal operation and fault conditions according to the amplitude frequency response parameter formula to evaluate transformer performance; Curve drawing: Draw amplitude frequency response curves from 100kHz to 1MHz, including response curves of normal state and different fault states, for performance evaluation and fault diagnosis.

4. The method for detecting transformer winding deformation based on a multi-scale multi-channel CNN-LSTM model according to claim 3 is characterized in that: The calculation formula for Z-score normalization is as follows: In the formula, A is the original data; μ is the mean of the data; σ is the standard deviation of the data; A norm Represents the data after Z-score normalization.

5. The method for detecting transformer winding deformation based on a multi-scale multi-channel CNN-LSTM model according to claim 3 is characterized in that: The formula for the amplitude frequency response parameter is as follows: Where P is the amplitude frequency response parameter, U1 is the amplitude of the transformer response signal, and U2 is the amplitude of the input excitation signal.

6. The method for detecting transformer winding deformation based on a multi-scale multi-channel CNN-LSTM model according to claim 1, characterized in that: In step 3, partitioning the database includes the following steps: Data construction: construct various amplitude frequency response curves into a fault diagnosis data set that combines virtual and actual data; Sample sampling: The sliding window method is used to perform overlapping segmentation sampling on the normalized data, and several samples are extracted from each working state; Dataset division: 70% of the samples are used as training sets, 15% of the samples are used as validation sets, and 15% of the samples are used as test sets.

7. The method for detecting transformer winding deformation based on a multi-scale multi-channel CNN-LSTM model according to claim 1, characterized in that: The multi-scale multi-channel CNN-LSTM model in step 4 has a model structure including multiple convolutional neural networks CNN and long short-term memory networks LSTM; the CNN network includes three branches, branch one contains three convolution modules, branch two contains convolution module 2 and convolution module 3, and branch three contains only convolution module 2; the convolution modules are set as follows: convolution module 1 contains 64 convolution kernels of size 3×3, convolution module 2 contains 128 convolution kernels of size 3×3, and convolution module 3 contains two layers of 3×3×256 convolution kernels.

8. The method for detecting transformer winding deformation based on a multi-scale multi-channel CNN-LSTM model according to claim 1, characterized in that: In step 4, the multi-scale multi-channel CNN-LSTM model parameter settings include convolution kernel size, activation function, pooling parameters, optimizer, loss function and learning rate; Activation function: Select ReLU function; Pooling parameters: The size of the pooling window should be smaller than the convolution kernel, and the pooling step size should be smaller than the convolution kernel step size; Optimizer: Use Adamax optimizer; Loss function: optimized cross entropy loss function; Learning rate: The initial learning rate is set to 0.0001.

9. The method for detecting transformer winding deformation based on a multi-scale multi-channel CNN-LSTM model according to claim 8, characterized in that: Loss function L CrossEntropy The optimized cross entropy loss function is selected, and the expression is as follows: f(x) ReLU =max(0,x) In the formula, x is the input value; f(x) is the function defined on the input (x), which specifically implements the logic of the activation function; ReLU is the Rectified Linear Unit, which represents the activation function; max() means returning the maximum of its input values; 0 is a constant value; Where, L CrossEntropy Cross entropy loss function; m is the number of samples; is a feature vector, representing the jth feature of the i-th sample; is the true label corresponding to each sample, represents the true category label corresponding to the i-th sample; n is the various types of fault labels set manually; R(θ) is the regularization term; λ is the regularization coefficient; θ represents the model parameter; Where R(θ) is the regularization term; θ represents the model parameter; k is the kth individual parameter that makes up the model parameters; Represents θ k Performs a square operation.

10. The method for detecting transformer winding deformation based on a multi-scale multi-channel CNN-LSTM model according to claim 6, characterized in that: Step 5 includes the following steps: Model training: Use the training set and validation set to adjust and train the model parameters; Model testing: Verify the model's accuracy in identifying winding faults through the test set; Real-time analysis: The trained model is used to analyze the frequency response signal acquired in real time to obtain the operating status of the transformer winding through the classifier.

Images