Partial discharge defect category intelligent identification method fusing optical and electrical characteristics

Abstract

A GIS partial discharge defect category intelligent identification method fusing optical and electrical characteristics comprises the following steps: acquiring an optical signal of partial discharge through an optical detection system, and extracting light intensity and spectral distribution characteristics; meanwhile, an ultrahigh frequency (UHF) sensor is used for collecting electromagnetic wave signals, and time domain and frequency domain characteristics of the electromagnetic wave signals are extracted; de-noising the small optical signal by adopting spatial filtering, and processing an electromagnetic wave signal by adopting a wavelet de-noising method; fusing the denoised signals by using a multi-modal deep neural network; splicing the two types of features and carrying out dimensionality reduction through a principal component analysis method to obtain a fusion feature vector; and finally, constructing a dual-channel CNN model for training and evaluation, optimizing the model by using a cross entropy loss function and an Adam optimizer, adjusting hyper-parameters and a network structure through a verification set, improving the generalization ability of the model, and finally obtaining a GIS partial discharge defect category classification model. According to the invention, precise and intelligent identification of GIS partial discharge defects is realized.

Description

Technical Field

[0001] The invention belongs to the field of partial discharge defect category identification of gas insulated switchgear (GIS), and specifically proposes an intelligent partial discharge defect category identification method integrating optical and electrical characteristics. Background Art

[0002] The reliability and safety of power equipment such as GIS is the basis for the stable operation of the power system. As a key component of electrical equipment, the insulation system is crucial to maintain a good condition to extend the life of the equipment and ensure safe operation. However, the long-term operation of the equipment and the cumulative impact of environmental factors will cause the insulation material to gradually age. This aging process can easily induce partial discharge (PD) in GIS. Partial discharge, as a non-penetrating discharge activity in a local area inside or on the surface of electrical equipment, does not directly penetrate the insulation layer, but it will significantly accelerate the aging process of the insulation material. In severe cases, it may even cause insulation failure, thereby triggering equipment failure or even shutdown events. Therefore, early detection and accurate identification of partial discharge are of immeasurable value for the preventive maintenance of electrical equipment and ensuring the stable operation of the system.

[0003] Traditional partial discharge detection methods mainly rely on electrical signal analysis, such as current pulse monitoring, electromagnetic wave monitoring, ultrasonic detection, ultra-high frequency (UHF) detection and other means. Although partial discharge can be detected to a certain extent, a single electrical signal has limitations. The propagation path of partial discharge signals is easily restricted by complex equipment structures and changing environmental conditions. In addition, the widespread existence of noise and interference makes electrical signals face challenges in reflecting the physical characteristics of the discharge process. Although electrical signals can provide certain key information in the discharge process, it is often difficult to achieve high-precision identification by electrical signals alone in comprehensively describing the morphological characteristics of partial discharge, especially in achieving high-precision classification of discharge types.

[0004] In view of this, it is necessary to study a detection method that can comprehensively and deeply capture GIS partial discharge signals, so as to improve the accuracy of partial discharge detection. Summary of the invention

[0005] In order to solve the shortcomings of the prior art, the present invention proposes an intelligent identification method for partial discharge defect categories that integrates optical and electrical features. By extracting the optical signal features and UHF band features during partial discharge, and processing them through feature fusion algorithms, more comprehensive partial discharge feature data can be obtained. The fused feature data is trained using a deep learning algorithm to construct a defect category classification model, thereby improving the accuracy and reliability of GIS partial discharge detection. The present invention combines the signal features in the optical spectrum and electrical UHF, and can more comprehensively capture the signal features of GIS partial discharge, thereby improving the accuracy and reliability of GIS partial discharge detection.

[0006] A GIS partial discharge defect category intelligent identification method integrating optical and electrical characteristics is characterized in that it includes the following steps:

[0007] Step 1. Signal acquisition step: using an optical detection system to collect optical signals generated during partial discharge, and extracting light intensity and spectral distribution characteristics; at the same time, using a UHF sensor to collect electromagnetic wave signals generated during partial discharge, and extracting time domain and frequency domain characteristics;

[0008] Step 2. Signal denoising step: the collected optical signal is denoised by using a spatial filtering method; the collected electromagnetic wave signal is denoised by using a wavelet denoising method;

[0009] Step 3. Feature fusion step: using a fusion method based on a multimodal deep neural network, a dual-channel structure is designed to process the denoised optical signal and the electromagnetic wave signal respectively; wherein the optical signal channel uses a convolutional neural network (CNN) to extract spatial features; the electromagnetic wave signal is converted into a time-frequency image by one-dimensional convolution or short-time Fourier transform (STFT), and then the time-frequency features are extracted by CNN; the extracted optical features and UHF features are spliced, and the dimension is reduced by principal component analysis to obtain a fused feature vector;

[0010] Step 4. Model training and evaluation steps: prepare labeled optical signal and electromagnetic wave signal data sets, and divide them into training set, validation set and test set; build a dual-channel convolutional neural network model, use the fused feature vector as input, select the cross entropy loss function and Adam optimizer for training; calculate the loss through forward propagation, update the model weights through back propagation, and use the validation set to evaluate the model performance; adjust the hyperparameters and network structure to improve the generalization ability of the model, and finally use the test set to evaluate the model performance and obtain the defect category classification model.

[0011] Furthermore, the optical detection system includes a fluorescent optical fiber, a transmission optical fiber, a photomultiplier tube and a power supply unit.

[0012] Furthermore, the spatial filtering method adopts mean filtering to smooth the image by calculating the average value of the pixel neighborhood.

[0013] Furthermore, the wavelet denoising method includes performing multi-level wavelet transform on the UHF signal to decompose it into different frequency components, removing noise components below a certain threshold by thresholding, and then reconstructing the signal by inverse wavelet transform.

[0014] Furthermore, in the feature fusion step, the feature extraction of the optical signal and the UHF signal is composed of several convolutional layers and pooling layers, respectively, for extracting spatial features and time-frequency features; the fused feature vector is nonlinearly transformed through a fully connected layer to learn higher-level features.

[0015] Furthermore, the construction of the defect classification model specifically includes:

[0016] Step 4.1 Data preparation: Collect, clean and annotate optical and electromagnetic wave signal data, perform data enhancement, and divide them into training set, validation set and test set.

[0017] Step 4.2 Model design: Select and build a multimodal fusion model suitable for the task.

[0018] Step 4.3 Loss function and optimizer selection: Select appropriate loss function and optimizer for model training. The loss function is used to measure the gap between the model prediction result and the actual label, and the optimizer is used to adjust the model weight to minimize the loss function.

[0019] Step 4.4 Training and evaluation: Input the training data into the model, calculate the prediction results and loss through forward propagation, and update the model weights through the backpropagation algorithm based on the loss and the selected optimizer. During the training process, regularly use the validation set to evaluate the performance of the model to monitor the training progress of the model and avoid overfitting;

[0020] Step 4.5: Model tuning: adjust hyperparameters, network structure or perform regularization to improve the generalization ability of the model; wherein the hyperparameters include learning rate, batch size, and number of network layers;

[0021] Step 4.6 Testing and deployment: Evaluate the performance of the model on the test set to verify its generalization ability. Deploy the trained model to the actual application scenario and use new data for classification.

[0022] Furthermore, in the step 4.1 data preparation, the data enhancement is performed by rotating, scaling, and flipping to increase the diversity and quantity of the data set.

[0023] Furthermore, in the model design of step 4.2, a dual-channel convolutional neural network model is constructed, and the fused feature vector is used as input.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] 1) By integrating the multi-dimensional characteristics of optical signals and electromagnetic wave signals (ultra-high frequency (UHF) signals), the present invention can more comprehensively capture the complex phenomena in the GIS partial discharge process. The optical signal provides the spatial distribution and spectral characteristics of the discharge, while the electromagnetic wave signal reveals the time domain and frequency domain characteristics of the discharge. The comprehensive application of this multimodal information can significantly improve the accuracy and reliability of defect identification.

[0026] 2) A signal denoising step is introduced, and spatial filtering and wavelet denoising methods are used for optical signals and electromagnetic wave signals respectively, which effectively reduces noise interference, improves signal quality, and contributes to the accuracy of subsequent feature extraction, thereby further improving the performance of the recognition model.

[0027] 3) A fusion method based on multimodal deep neural network is adopted, and a dual-channel structure is designed to process the denoised optical signal and electromagnetic wave signal respectively. Through advanced algorithms such as convolutional neural network (CNN), the spatial characteristics of the optical signal and the time-frequency characteristics of the electromagnetic wave signal are effectively extracted. The dimension is reduced by principal component analysis to obtain a more representative fusion feature vector, which not only improves the efficiency of feature extraction, but also enhances the expressiveness and robustness of the features.

[0028] 4) The cross entropy loss function and Adam optimizer were used for training to effectively avoid overfitting. At the same time, the model performance was further optimized and the generalization ability of the model was improved by adjusting the hyperparameters and network structure.

[0029] 5) The present invention not only improves the intelligence level of GIS partial discharge defect identification by integrating optical and electrical features, adopting advanced signal processing and feature fusion strategies, as well as optimized model training and evaluation methods, but also provides a strong guarantee for the safe operation of the power system. Through real-time monitoring and intelligent identification of GIS partial discharge defects, potential safety hazards can be discovered in a timely manner, equipment failures and power outages can be reduced, and the stability and reliability of the power system can be improved, which has important practical application value and promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a flow chart of a method for intelligently identifying partial discharge defect types by integrating optical and electrical features according to an embodiment of the present invention;

[0031] Figure 2 This is a training flow chart for constructing a defect classification model by using a deep learning algorithm to train the fused feature vector. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] An intelligent identification method for partial discharge defect categories that integrates optical and electrical features, such as Figure 1 As shown, the following steps are included:

[0034] Step 1. Use an optical detection system to collect optical signals generated during partial discharge and extract light intensity and spectral distribution characteristics; at the same time, use an ultra-high frequency (UHF) sensor to collect electromagnetic wave signals generated during partial discharge and extract time domain and frequency domain characteristics.

[0035] This embodiment uses fluorescent optical fiber, transmission optical fiber, photomultiplier tube, and power supply unit to form an optical detection system to collect spectrum and then extract the light intensity and spectral distribution characteristics in the spectrum.

[0036] Light intensity is an important feature of spectral signals, usually referring to the luminous flux received per unit area, which can be expressed by the following formula:

[0037]

[0038] Spectral distribution refers to the distribution of light intensity at different wavelengths, as shown in formula (2):

[0039] S=∫I(λ)dλ (2)

[0040] By performing Fourier transform on the spectral signal for frequency domain analysis, its spectrum characteristics can be obtained. Assuming that the spectral signal is I(t), its Fourier transform is:

[0041]

[0042] The signal I(f) obtained by Fourier transform can be decomposed into amplitude and phase:

[0043] I(f)=|I(f)|e jφ(f) (4)

[0044] For the extraction of ultra-high frequency features, a built-in UHF sensor with low-frequency suppression function is used to detect the electromagnetic wave signal generated by partial discharge. By sampling the partial discharge UHF signals of different types of defect models for 50 consecutive power frequency cycles, the partial discharge pulse time series is extracted, the pulse amplitude and phase information is obtained, and the number of discharges in each amplitude interval and phase interval is counted to form the discharge characteristics. Matrix. According to By using the matrix, we can obtain the three-dimensional PRPD and PRPS spectra of various defect model discharges, and map the three-dimensional PRPD spectra onto the two-dimensional plane to obtain the two-dimensional PRPD spectra. And through the collected spectra information, we can extract the time domain characteristics and frequency domain characteristics of the UHF signal.

[0045] Assume that the UHF signal generated by partial discharge is v(t), where t is time. The acquisition device records the signal in a discrete sampling manner with a sampling frequency of fs. The discretized signal v[n] can be expressed as:

[0046] v[n]=v(t n )=v(n·T s )(n=0,1,2,…,N-1) (5)

[0047] The frequency distribution of UHF signal can be converted from time domain signal to frequency domain through Fourier transform. Let the time domain signal be v(t), and its Fourier transform be V(f):

[0048]

[0049] The signal characteristics of GIS partial discharge are fully extracted through optical detection system and UHF detection system.

[0050] Step 2. De-noise the collected optical and electrical signals to eliminate environmental noise and equipment interference. The specific method is as follows:

[0051] Preprocess the collected signal to improve the accuracy of the signal. For optical signals, spatial filtering is used to remove noise. The basic idea of ​​spatial filtering is to process the image through a convolution kernel (or filter). The convolution kernel is a small matrix that defines how to weight and process each pixel of the image. The specific process is as follows:

[0052] 1) Select the convolution kernel: Select a suitable convolution kernel, which determines the type and effect of filtering.

[0053] 2) Convolution operation: Slide the convolution kernel on the image and calculate the convolution result at each position.

[0054] 3) Update pixel values: Replace the value of the central pixel with the result of convolution to obtain the processed image.

[0055] The mean filtering method is used to implement the filtering operation. The image is smoothed by calculating the average value of the pixel neighborhood. The convolution kernel is usually:

[0056]

[0057] Calculate the average value of all pixels in the convolution window and assign the average value to the current pixel I filtered (x,y)

[0058]

[0059] Update the calculated average value to the new image I filtered (x,y).

[0060] For UHF signals, wavelet denoising is used to reduce the impact of noise. The basic idea of ​​wavelet denoising is to use wavelet transform to decompose the signal into different frequency components, then remove the noise components below a certain threshold through thresholding, and finally reconstruct the signal. Unlike Fourier transform, wavelet transform can provide localized information of time and frequency, so it is particularly effective for denoising non-stationary signals.

[0061] Wavelet transform decomposes the signal into components of multiple scales. By performing multi-level wavelet transform on the signal, information of different frequencies can be obtained. Its mathematical expression is:

[0062]

[0063] Where W(j,k) is the wavelet transform coefficient, ψ j,k (t) is the wavelet function, j and k represent the scale and position respectively.

[0064] The wavelet coefficients are processed by setting a threshold λ. For the wavelet coefficient W(j,k), if its absolute value is lower than the threshold, it is set to zero; otherwise, the coefficient is retained.

[0065] This is achieved by setting a soft threshold:

[0066]

[0067] After threshold processing, the denoised coefficients are reconstructed into time domain signals using inverse wavelet transform. Its mathematical expression is:

[0068]

[0069] The denoised signal is then standardized and normalized to reduce the interference of noise on the signal.

[0070] Step 3 uses a fusion method based on a multimodal deep neural network to achieve the fusion of optical features and UHF features. The specific method is:

[0071] Optical signals and UHF signals are complementary in terms of temporal and spatial resolution, frequency response, etc. Optical signals can provide precise information on discharge location and intensity, while UHF signals can capture changes in electromagnetic fields, thus providing richer information in different dimensions.

[0072] In order to solve the problem of the fusion of optical features and ultra-high frequency (UHF) features, a fusion method based on multimodal deep neural network is proposed. The model is designed as a dual-channel structure, which inputs optical signals and UHF signals respectively.

[0073] Optical signal input channel: The input optical signal is usually an image (such as an optical image of partial discharge), and a convolutional neural network (CNN) is used to extract the spatial features of the optical signal.

[0074] UHF signal input channel: UHF signal is an electromagnetic signal in the time domain or frequency domain. It is converted into a time-frequency image through 1D convolution or short-time Fourier transform (STFT), and then its time-frequency features are extracted through CNN.

[0075] Each channel consists of several convolutional layers and pooling layers, which are used to extract the features of optical signals and UHF signals. For optical feature extraction: the convolutional layer extracts spatial features (such as the shape, position, intensity change of the discharge spot, etc.). The pooling layer performs feature downsampling to reduce the dimension of the features while retaining important features.

[0076] For UHF feature extraction: UHF signal is subjected to one-dimensional convolution to extract its spectrum and time series features. Alternatively, the time domain signal is converted into a time-frequency image through short-time Fourier transform (STFT), and then the time-frequency domain features are extracted through two-dimensional convolution.

[0077] After extracting the respective characteristics of optical and UHF signals, they are fused in the network.

[0078] Combine the optical signal characteristics with the UHF signal characteristics:

[0079] F fused =[F optical ,F UHF ] (12)

[0080] Among them, F optical is the eigenvector of the optical signal, F UHF is the eigenvector of the UHF signal, F fused is the concatenated fused feature vector. The fused vector is reduced in dimension using principal component analysis to reduce the feature dimension and retain the most important information. The fused feature vector can be used in machine learning models for classification and recognition.

[0081] The fused feature vector is transformed nonlinearly through the fully connected layer to learn higher-level features. To prevent overfitting, the Dropout strategy can be used after the fully connected layer to randomly shut down some neurons.

[0082] Step 4. Use the deep learning algorithm to train the data after feature fusion and build a defect classification model. The specific method is as follows:

[0083] 1) Data preparation: Collect, clean and annotate optical and UHF signal data, perform data enhancement, and divide them into training sets, validation sets and test sets.

[0084] 2) Model design: Select and build a multimodal fusion model suitable for the task.

[0085] 3) Loss function and optimizer selection: Select appropriate loss function and optimizer for model training.

[0086] 4) Training and evaluation: Calculate the loss through forward propagation, update the model weights through backpropagation, and use the validation set to evaluate the model performance.

[0087] 5) Model tuning: Adjust hyperparameters, network structure, or perform regularization to improve the generalization ability of the model.

[0088] 6) Test and deploy: Finally, test and deploy the model and use new data for classification.

[0089] The process is as follows:

[0090] 70% of the collected data is used as a training set to train the model; 15% is used as a validation set to adjust hyperparameters and monitor model performance; the remaining 15% is used as a test set to evaluate the final performance of the model. The model selected is a dual-channel convolutional neural network.

[0091] The loss function is chosen as the cross entropy loss function, which is suitable for multi-classification tasks:

[0092]

[0093] Among them, y i is the true label of the i-th category, p i is the probability that the model predicts that class.

[0094] Select the Adam optimizer, which combines the advantages of momentum gradient descent and adaptive learning rate adjustment to accelerate convergence and avoid falling into local optimality. Adam's update rule is:

[0095]

[0096] Among them, m t and v tare the momentum and variance estimates of the gradient, and η is the learning rate.

[0097] During the training phase, the optical signal and the UHF signal are input, and forward propagated through the convolution layer, the pooling layer, and the fully connected layer to calculate the output prediction value. The predicted value is compared with the true label and the cross entropy loss is calculated. Through the back-propagation algorithm, the gradient of the loss relative to the model parameters is calculated, and the optimizer is used to update the parameters. Through multiple, small batches of data training, the training process is accelerated and memory consumption is reduced. At the end of each epoch, the performance of the model is evaluated using the validation set to monitor the convergence of the model and prevent overfitting.

[0098] During the training process, hyperparameters such as learning rate, batch size, network depth, and convolution kernel size are constantly adjusted to obtain the best model performance. After the training is completed, the model is finally evaluated using the test set to calculate performance indicators such as the model's accuracy, recall, and F1 score.

[0099] The patent of this invention obtains more comprehensive partial discharge feature data by fusing optical and ultra-high frequency features. It also uses a dual-channel convolutional neural network to fuse the features, trains the fused feature data, and constructs a defect classification model, which can improve the accuracy and reliability of GIS partial discharge detection.

[0100] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A GIS partial discharge defect category intelligent identification method integrating optical and electrical characteristics, characterized in that: The following steps are involved: Step 1. Signal acquisition: Use an optical detection system to collect optical signals generated during partial discharge, and extract light intensity and spectral distribution characteristics; at the same time, use a UHF sensor to collect electromagnetic wave signals generated during partial discharge, and extract time domain and frequency domain characteristics; Step 2. Signal denoising: The collected optical signal is denoised by using a spatial filtering method; the collected electromagnetic wave signal is denoised by using a wavelet denoising method; Step 3. Feature fusion: Using a fusion method based on a multimodal deep neural network, a dual-channel structure is designed to process the denoised optical signal and electromagnetic wave signal respectively; wherein the optical signal channel uses CNN to extract spatial features; the electromagnetic wave signal is converted into a time-frequency image through one-dimensional convolution or STFT, and then CNN is used to extract time-frequency features; The extracted spatial features and time-frequency features are concatenated and dimension reduction is performed through principal component analysis to obtain the fused feature vector; Step 4. Use the deep learning algorithm to train the fused feature vector to build a defect category classification model.

2. The intelligent identification method of GIS partial discharge defect type integrating optical and electrical characteristics according to claim 1 is characterized in that: The optical detection system includes a fluorescent optical fiber, a transmission optical fiber, a photomultiplier tube and a power supply unit.

3. The intelligent identification method of GIS partial discharge defect type integrating optical and electrical characteristics according to claim 1 is characterized in that: The spatial filtering method adopts mean filtering to smooth the image by calculating the average value of the pixel neighborhood.

4. The intelligent identification method of GIS partial discharge defect type integrating optical and electrical characteristics according to claim 1 is characterized in that: The wavelet denoising method includes performing multi-level wavelet transform on the electromagnetic wave signal to decompose it into different frequency components, removing the noise components below a certain threshold through thresholding, and then reconstructing the signal using inverse wavelet transform.

5. The intelligent identification method of GIS partial discharge defect type integrating optical and electrical characteristics according to claim 1 is characterized in that: In the feature fusion step, the feature extraction of the optical signal and the electromagnetic wave signal is respectively composed of a number of convolutional layers and pooling layers, which are used to extract spatial features and time-frequency features; The fused feature vector is transformed nonlinearly through a fully connected layer to learn higher-level features.

6. The intelligent identification method of GIS partial discharge defect type integrating optical and electrical characteristics according to claim 1 is characterized in that: The construction of the defect classification model specifically includes: Step 4.1 Data preparation: Collect, clean and annotate optical and electromagnetic wave signal data, perform data enhancement, and divide them into training set, validation set and test set. Step 4.2 Model design: Select and build a multimodal fusion model suitable for the task. Step 4.3 Loss function and optimizer selection: Select appropriate loss function and optimizer for model training. The loss function is used to measure the gap between the model prediction result and the actual label, and the optimizer is used to adjust the model weight to minimize the loss function. Step 4.4 Training and evaluation: Input the training data into the model, calculate the prediction results and loss through forward propagation, and update the model weights through the backpropagation algorithm based on the loss and the selected optimizer. During the training process, regularly use the validation set to evaluate the performance of the model to monitor the training progress of the model and avoid overfitting; Step 4.5: Model tuning: adjust hyperparameters, network structure or perform regularization to improve the generalization ability of the model; wherein the hyperparameters include learning rate, batch size, and number of network layers; Step 4.6 Testing and deployment: Evaluate the performance of the model on the test set to verify its generalization ability. Deploy the trained model to the actual application scenario and use new data for classification.

7. The intelligent identification method of GIS partial discharge defect type integrating optical and electrical characteristics according to claim 6 is characterized in that: In the step 4.1 data preparation, the data enhancement is to increase the diversity and quantity of the data set by rotating, scaling, and flipping.

8. The intelligent identification method of GIS partial discharge defect type integrating optical and electrical characteristics according to claim 6 is characterized in that: In the model design of step 4.2, a dual-channel convolutional neural network model is constructed, and the fused feature vector is used as input.

Images