A partial discharge ultrasonic detection and positioning method based on convolutional neural network

Abstract

The application provides a partial discharge ultrasonic detection and positioning method based on a convolutional neural network, first, two data sets are constructed, one is a large amount of simulation data generated by a simulation method, and the other is real data actually collected in a real power scene; then, the data in the simulation data set and the real data set are preprocessed and feature extraction is performed; then, a model of the convolutional neural network is built, the network is pre-trained by using the simulation data set, and then the real data set is used for parameter adjustment; finally, the trained network model is used for testing and actual application. The application can realize remote continuous online detection, can be applied to a portable detection instrument, realizes mobile detection, can more easily realize simultaneous detection of multiple partial discharges, and has higher precision and good robustness in partial discharge detection and positioning.

Description

Technical Field

[0001] This invention relates to the field of electrical equipment fault detection, and in particular to a method for ultrasonic detection and localization of partial discharge. Background Technology

[0002] With the rapid development of electrical systems, maintaining their stable operation is of paramount importance. Due to insulation degradation, electrical equipment may experience localized discharges under the influence of an electric field during operation. The presence of partial discharge indicates potential insulation deterioration. If not monitored and maintained promptly, prolonged partial discharge can cause related components to overheat, accelerating the aging process and leading to more severe damage. Timely and effective partial discharge detection in the early stages of a fault can reduce maintenance costs, extend equipment lifespan, and prevent major accidents.

[0003] Currently, partial discharge detection mainly relies on electrical measurement methods such as ultra-high frequency detection and pulsed current methods. These methods acquire electrical signals, extract features such as the phase-resolved partial discharge (PRPD) distribution, the phase-resolved pulse sequence (PRPS) distribution, and wavelet transform, and then utilize models like neural networks and support vector machines for identification and detection. However, these electrical measurement methods are contact-based, posing certain safety concerns and preventing mobile detection. With increasing safety requirements in modern power systems, non-contact detection methods are safer and enable remote, continuous online monitoring, aligning with the development trend of unattended power monitoring systems. Therefore, non-contact partial discharge detection has become a current research focus.

[0004] Non-contact partial discharge detection includes ultrasonic detection and infrared imaging detection. Ultrasonic detection detects and locates partial discharges by collecting and analyzing the ultrasonic signals generated during discharge. Currently, this field is still in its early stages of development both domestically and internationally. It typically uses multi-channel microphone arrays to pick up ultrasonic signals and then employs beamforming for localization. This process requires manually setting thresholds or the number of beam peaks to determine the number of abnormal locations. However, in real-world scenarios, the number of abnormal locations cannot be predicted in advance. Furthermore, the surrounding environment of power systems contains noise in the ultrasonic band, which can interfere with traditional beamforming methods.

[0005] In recent years, with the rise of machine learning, especially deep neural network algorithms, researchers in many industrial fields have begun to learn and utilize machine learning algorithms, significantly improving the performance of related applications. For example, in the field of speech signal processing, traditional speaker localization uses beamforming methods, while recent research has mainly focused on deep learning-based methods. Compared to traditional beamforming methods, deep neural network methods can more easily achieve multi-target estimation without requiring manual setting of detection thresholds or pre-setting of the number of beam peaks. Furthermore, due to the stronger modeling capabilities and certain noise reduction abilities of deep neural networks, deep learning models exhibit higher localization accuracy and better robustness. However, deep learning methods often rely on large-scale datasets, and for certain industrial scenarios, especially for the partial discharge problem in power systems, obtaining large amounts of real-world data presents a challenge.

[0006] Existing technologies already include methods for partial discharge detection based on convolutional neural networks (CNNs). The patent "Partial Discharge Fault Monitoring and Assessment Device and Method" utilizes a CNN for partial discharge detection and then employs a traditional beamforming algorithm for localization; however, this method cannot simultaneously utilize a CNN for detection and beamforming. The paper "Deep Learning Pattern Recognition Method for Partial Discharge Ultrasonic Signals Based on Similarity Matrix Blind Source Separation and Convolutional Neural Networks" first extracts features from the acquired partial discharge data, then obtains the foreground video spectrum through blind source separation, and finally performs recognition and classification using a CNN, but does not perform localization. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, this invention provides a method for ultrasonic detection of partial discharge based on convolutional neural networks. First, two datasets are constructed: one is a large amount of simulated data generated through simulation, and the other is real data collected in a real power scenario. Next, the data in both the simulated and real datasets are preprocessed and feature extracted. Then, a convolutional neural network model is built, pre-trained using the simulated dataset, and then its parameters are adjusted using the real dataset. Finally, the trained network model is used for testing and practical applications.

[0008] The technical solution adopted by this invention to solve its technical problem includes the following steps:

[0009] Step 1: Generate multi-channel ultrasonic signals x based on the characteristics of the microphone array. N =[x1,x2,...,x N], where N is the number of array elements; assuming the horizontal angle of the partial discharge position relative to the microphone array is α and the pitch angle is β, the horizontal angle α and the pitch angle β are transformed into a two-dimensional one-hot matrix y, which is used as the expected output of the network; the simulation dataset is obtained as [x N [,y]; Divide the simulation dataset into a training set and a validation set;

[0010] Step 2: Using the same microphone array as in Step 1, acquire multi-channel ultrasonic data x′ in a real-world scenario. N =[x1′,x2′,...,x′ N Simultaneously, the horizontal angle α′ and pitch angle β′ are annotated, and the horizontal angle α′ and pitch angle β′ are transformed into a two-dimensional one-hot matrix y′; the real dataset is obtained as [x′ N [,y′]; Divide the real dataset into training set, validation set and test set;

[0011] Step 3: Filter and extract features from the data in the simulation dataset and the real dataset;

[0012] Step 4: Build a convolutional neural network. The input of the convolutional neural network is the data features extracted in step 3. The network first uses two convolutional layers to perform feature transformation, then stacks multiple convolutional modules to perform further high-level feature transformation, and finally uses two more convolutional layers to output an estimate of the two-dimensional one-hot matrix y.

[0013] Step 5: Pre-train the convolutional neural network using a simulation dataset, and then fine-tune the parameters using a real dataset.

[0014] Step 6, transfer the real dataset [x′] N The test set data of y′ is fed into the trained network obtained in step 5 to obtain the final localization accuracy.

[0015] In step 1, the ratio of training set to validation set data is 8:2.

[0016] In step 2, the ratio of training set, validation set, and test set data is 8:1:1.

[0017] Step 3 involves high-pass filtering of the data in the simulation dataset and the real dataset to filter out audible frequency bands below 20kHz.

[0018] Step 3 involves feature extraction from the filtered data and performing a short-time Fourier transform on each channel of the multi-channel signal.

[0019] In step 4, the input to the convolutional neural network is the real and imaginary parts of the short-time Fourier transform of the multi-channel signal.

[0020] The convolutional neural network in step 4 uses the basic modules of the Squeeze-Excitation (SE) network, which are multiple convolutional modules stacked in the middle. That is, an SE block is cascaded after the residual convolutional block.

[0021] Step 5 utilizes the simulation dataset [x] N The neural network model is pre-trained on the training set of [y, y] with the negative log-likelihood loss function, until the performance of the model on the validation set of the simulation dataset no longer improves.

[0022] Step 5 utilizes the real dataset [x′] N The training set data of y′ is used to adjust the parameters of the neural network model. The loss function is the negative log-likelihood loss function. Training stops when the model no longer improves its performance on the validation set of the real dataset.

[0023] The beneficial effects of this invention are: it utilizes convolutional neural networks to process ultrasonic signals to achieve partial discharge detection. Compared to widely used electrical measurement methods, the ultrasonic-based method enables remote, continuous online detection and can also be applied to portable detection instruments for mobile detection. Furthermore, the ultrasonic-based method is safer. Compared to traditional beamforming methods, the convolutional neural network-based method can more easily achieve simultaneous detection of multiple partial discharges without manually setting thresholds or pre-setting the number of beam peaks. In addition, thanks to the powerful modeling capabilities of deep neural networks, the accuracy of partial discharge detection and localization is higher, and the inherent noise reduction capabilities of deep neural networks enhance the robustness of this method.

[0024] Compared to the patent "Partial Discharge Fault Monitoring and Assessment Device and Method", this invention can simultaneously perform detection and beamforming using a convolutional neural network, possessing the advantages of beamforming from deep neural networks. Compared to the paper "Deep Learning Pattern Recognition Method for Partial Discharge Ultrasonic Signals Based on Similar Matrix Blind Source Separation and Convolutional Neural Network", this invention can simultaneously perform partial discharge signal detection and localization using a convolutional neural network. Attached Figure Description

[0025] Figure 1 This is a system block diagram of the present invention;

[0026] Figure 2 This is a diagram of a convolutional neural network structure used in the partial discharge ultrasonic method. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments. The present invention includes, but is not limited to, the following embodiments.

[0028] This invention aims to propose a method for partial discharge detection and localization based on convolutional neural networks. This method uses the short-time Fourier transform of multi-channel ultrasonic signals as input to obtain the horizontal and vertical angle information of a partial discharge. The method includes the following steps:

[0029] Step 1: Generate a large amount of simulation data through simulation methods.

[0030] Based on the characteristics of the microphone array, a multi-channel ultrasonic signal x is generated. N =[x1,x2,...,x N The microphone array configuration is selected as needed, such as a planar array or a spiral array, where N is the number of array elements. Assuming the horizontal angle α and the pitch angle β relative to the microphone array at the partial discharge location are at the given position, these angles are transformed into a two-dimensional one-hot matrix y, which serves as the network's desired output. Therefore, the final simulation dataset is [x...]. N [y]. Further, the dataset is divided into training and validation sets, with a suggested ratio of 8:2.

[0031] Step 2: Collect actual partial discharge data.

[0032] Using the same microphone array as in step 1, multi-channel ultrasonic data x′ is acquired in a real-world scenario. N =[x1′,x2′,...,x′ N Simultaneously, manual annotation is performed to obtain the horizontal angle α′ and pitch angle β′ at this point, and the horizontal angle α and pitch angle β are transformed into a two-dimensional one-hot matrix y′. The final real dataset is [x′...]. N ,y′]. Further, the dataset is divided into training, validation, and test sets, with a suggested ratio of 8:1:1.

[0033] Step 3, Preprocessing and Feature Extraction.

[0034] Step 3-1: Perform high-pass filtering on the data in the simulation dataset and the real dataset to filter out audible frequency bands below 20kHz.

[0035] Step 3-2: Feature extraction is performed on the filtered data. Specifically, a short-time Fourier transform is performed on each channel of the input multi-channel signal.

[0036] Step 4: Build a convolutional neural network.

[0037] The input to a convolutional neural network is the real and imaginary parts of the short-time Fourier transform of a multi-channel signal. The network first uses two convolutional layers for feature transformation, then stacks multiple convolutional modules for further high-level feature transformation, and finally uses two more convolutional layers to output an estimate of the two-dimensional one-hot matrix y.

[0038] Step 5: Train the neural network.

[0039] Step 5-1, using the simulation dataset [x N The neural network model is pre-trained on the training set of [y, y] with the negative log-likelihood loss function, until the performance of the model on the validation set of the simulation dataset no longer improves.

[0040] Step 5-2, after pre-training, use the real dataset [x′] N The neural network model is trained using the training data of y′, and the loss function remains the negative log-likelihood loss function. Training is stopped when the model no longer improves its performance on the validation set of the real dataset.

[0041] Step 6, Model Testing.

[0042] The real dataset [x′ N The test set data of y′ is fed into the trained network obtained in step 5 to obtain the final localization accuracy.

[0043] Preferably, in step 4, the convolutional neural network uses the basic modules of the Squeeze-Excitation (SE) network, which are stacked in the middle, i.e., SE blocks are cascaded after the residual convolutional blocks.

[0044] This invention provides a method for partial discharge detection and localization based on a convolutional neural network, comprising the following steps:

[0045] Step 1: Generate a large amount of simulation data using simulation methods. Based on the characteristics of the microphone array, generate multi-channel ultrasonic signals x. N =[x1,x2,...,x N The microphone array uses a spiral array, where N is the number of array elements, specifically N equals 64. Assuming the horizontal angle of the partial discharge position relative to the microphone array is α∈[0,180] and the pitch angle is β∈[0,40], we transform the horizontal angle α and pitch angle β into a two-dimensional one-hot matrix y, with a size of 180x40, as the expected output of the network. Therefore, the final simulation dataset is [x...]. N Furthermore, the dataset is divided into a training set and a validation set in an 8:2 ratio.

[0046] Step 2: Acquire real partial discharge data. Using the same microphone array as in Step 1, acquire multi-channel ultrasonic data x′ in a real-world scenario. N =[x1′,x2′,...,x′ N Simultaneously, manual annotation is performed to obtain the horizontal angle α′∈[0,180] and the pitch angle β′∈[0,40]. The horizontal angle α and pitch angle β are then transformed into a two-dimensional one-hot matrix y′, also with a size of 180x40. Finally, the actual dataset is [x′...]. N Furthermore, the dataset is divided into training, validation, and test sets in an 8:1:1 ratio.

[0047] Step 3: Preprocessing and feature extraction;

[0048] Step 3-1: Perform high-pass filtering on the data in the simulation dataset and the real dataset to filter out the audible frequency band below 20kHz.

[0049] Step 3-2: Feature extraction from the filtered data. Specifically, a short-time Fourier transform is performed on each channel of the input multi-channel signal. The window length is 20ms, and the window overlap ratio is 50%.

[0050] Step 4: Construct the Convolutional Neural Network. The input to the convolutional neural network is the real and imaginary parts of the short-time Fourier transform of the multi-channel signal. Since the input is a 64-channel ultrasonic signal, the input size is 128 after juxtaposing the real and imaginary parts. Next, the network first performs feature transformation using two convolutional layers, then stacks five convolutional modules for further high-level feature transformation, and finally uses two more convolutional layers to output an estimate of the two-dimensional one-hot matrix y. The five stacked convolutional modules adopt the basic modules of the Squeeze-Excitation (SE) network, i.e., an SE block is cascaded after the residual convolutional block.

[0051] Step 5: Train the neural network;

[0052] Step 5-1: First, use the simulation dataset [x] N The neural network model is pre-trained on the training set of [y, y], with the training loss function being the negative log-likelihood loss function and the optimizer being AdamW, until the performance of the model on the validation set of the simulation dataset no longer improves.

[0053] Step 5-2: After pre-training, use the real dataset [x′] N The neural network model is trained using the training set data of y′, with AdamW as the optimizer and the loss function remaining the negative log-likelihood loss function. Training is stopped when the model no longer improves its performance on the validation set of the real dataset.

[0054] Step 6: Model Testing. Use the real dataset [x′] N The test set data of y′ is fed into the trained network obtained in step 5 to obtain the final localization accuracy.

Claims

1. A method for ultrasonic detection and localization of partial discharge based on convolutional neural networks, characterized in that, Includes the following steps: Step 1: Generate multi-channel ultrasonic signals based on the characteristics of the microphone array. ,in, Let be the number of array elements; assuming the horizontal angle between the partial discharge location and the microphone array is . The pitch angle is , horizontal angle and pitch angle The output is transformed into a two-dimensional one-hot matrix y, which serves as the expected output of the network; the simulation dataset is obtained as follows. The simulation dataset is divided into a training set and a validation set. Step 2: Using the same microphone array as in Step 1, acquire multi-channel ultrasonic data in a real-world scenario. At the same time, the horizontal angle at this time is marked. and pitch angle , horizontal angle and pitch angle Transform it into a two-dimensional one-hot matrix. ; Obtain the real dataset as The real dataset is divided into training, validation, and test sets. Step 3: Filter and extract features from the data in the simulation dataset and the real dataset; Step 4: Build a convolutional neural network. The input of the convolutional neural network is the data features extracted in step 3. The network first uses two convolutional layers to perform feature transformation, then stacks multiple convolutional modules to perform further high-level feature transformation, and finally uses two more convolutional layers to output an estimate of the two-dimensional one-hot matrix y. The convolutional neural network uses SE network modules, which are stacked in the middle. That is, SE blocks are cascaded after residual convolutional blocks. Step 5: Pre-train the convolutional neural network using a simulation dataset, and then fine-tune the parameters using a real dataset. Step 6, use the real dataset The test set data is fed into the trained network obtained in step 5 to obtain the final localization accuracy.

2. The method for partial discharge ultrasonic detection and localization based on convolutional neural networks according to claim 1, characterized in that, In step 1, the ratio of training set to validation set data is 8:

2.

3. The method for partial discharge ultrasonic detection and localization based on convolutional neural networks according to claim 1, characterized in that, In step 2, the ratio of training set, validation set, and test set data is 8:1:

1.

4. The method for partial discharge ultrasonic detection and localization based on convolutional neural networks according to claim 1, characterized in that, Step 3 involves high-pass filtering of the data in the simulation dataset and the real dataset to filter out audible frequency bands below 20 kHz.

5. The method for partial discharge ultrasonic detection and localization based on convolutional neural networks according to claim 1, characterized in that, Step 3 involves feature extraction of the filtered data and performing a short-time Fourier transform on each channel of the multi-channel signal.

6. The method for partial discharge ultrasonic detection and localization based on convolutional neural networks according to claim 1, characterized in that, In step 4, the input to the convolutional neural network is the real and imaginary parts of the short-time Fourier transform of the multi-channel signal.

7. The method for partial discharge ultrasonic detection and localization based on convolutional neural networks according to claim 1, characterized in that, Step 5 utilizes a simulation dataset. The neural network model is pre-trained on the training set with a negative log-likelihood loss function until the model's performance on the validation set of the simulation dataset no longer improves.

8. The method for partial discharge ultrasonic detection and localization based on convolutional neural networks according to claim 1, characterized in that, Step 5 utilizes real datasets. The neural network model is used to adjust parameters using the training set data, with the loss function being the negative log-likelihood loss function. Training stops when the model's performance no longer improves on the validation set of the real dataset.

Images