Learning dictionary-based partial discharge signal extraction method
Through a learning dictionary-based method, K-SVD dictionary and sparse autoencoder extract local discharge signals in the power system, solving the problem of signal detection difficulties caused by noise interference, and realizing more accurate signal extraction in complex environments to ensure that The safe operation of power equipment.
Patent Information
- Application Number
- CN202410990817.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-23
- Publication Date
- 2024-11-15
AI Technical Summary
In power systems, local discharge signal detection is difficult due to noise interference, and the prior art is difficult to accurately extract signals in complex environments, especially when the noise and useful signals have similarity in the time domain, filtering technology is easy to accidentally filter out useful signals. , Time-frequency domain analysis and mathematical analysis methods are not effective in some cases, and machine learning algorithms are difficult to obtain good training results due to small samples and variability in data.
Using a learning dictionary method, the dictionary is updated column by constructing a K-SVD dictionary and sparse autoencoder, and the dictionary is updated column by column. The sparse autoencoder is used to sparsely decompose and reconstruct the updated dictionary and noise-containing signals, and extract the low-noise locality. Discharge signals, sparse autoencoder maps and restores data through unsupervised neural network structures, ensuring the sparseness and accuracy of encodings.
Steadily extract more accurate discharge signals in complex environments, improving the safety and reliability of power system equipment, avoiding serious consequences of insulation failure, and ensuring the accuracy and effectiveness of detection results.
Smart Images

Figure CN118964910A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of local amplification signal extraction, and in particular relates to a method for extracting local discharge signals based on a learning dictionary. Background Art
[0002] Partial discharge (PD) can reflect the insulation condition of equipment and its possible failure rate. PD detection has become an important part of online monitoring of power equipment. In power systems, PD is caused by various defects in the insulation material, such as voids and electrical trees. If these phenomena are not detected and processed in time, these PD activities may lead to insulation failure of the power system, or even cause more serious problems, such as fire or equipment explosion.
[0003] However, it is not easy to detect and extract PD signals: in the early insulation failure stage of the equipment, the PD signal is usually very weak, and there may be a lot of electromagnetic noise at the detection site. As a result, the detected PD signal is usually interfered by a lot of noise, which seriously affects the effect of partial discharge monitoring and the determination of insulation conditions. At the same time, with the increasing complexity of the current power system, the interference of PD signals is also increasing, so it is necessary to develop more efficient methods to distinguish between true partial discharge signals and noise generated by other sources.
[0004] The detection of PD signals faces interference from many complex factors, among which noise pollution is particularly serious, mainly reflected in white Gaussian noise and narrowband noise. Therefore, researchers are particularly concerned about how to eliminate noise, determine edges, and accurately extract the waveform of PD signals. In recent years, a series of progress has been made in the processing technology of PD signals. At present, various types of methods have been proposed to reduce white noise and narrowband noise in PD signals, including time domain processing methods, time-frequency domain analysis methods, mathematical analysis methods, and machine learning algorithms.
[0005] As a basic technology in the time domain analysis of PD signals, filtering technology can extract valuable information and eliminate unnecessary noise to a certain extent. However, the useful signal and noise in the noisy PD signal have certain similarities in the time domain and are difficult to distinguish. Therefore, when using filters to process signals, it is very likely that part of the useful signal will be mistakenly filtered out. Once the useful signal is lost, it will have an adverse effect on the fault analysis and judgment after the useful signal is extracted.
[0006] Wavelet method and EMD method are both very representative time-frequency domain analysis methods. If the mother wavelet is not selected properly, it may affect the extraction accuracy of PD signal. EMD method cannot completely eliminate the influence of modal aliasing.
[0007] Mathematical analysis based on the SVD concept has been used many times in the processing of PD signals. The advantage of the SVD algorithm is that it can effectively extract useful features from complex data and reduce the dimension of the data without losing too much information. However, when the characteristics of the PD signal are not obvious and difficult to distinguish from noise, SVD may not be able to distinguish which is noise and which is a valuable signal. At the same time, the SVD algorithm principle is based on the mathematical principle of linear algebra, but the generation and propagation of PD signals may involve nonlinear processes, thereby generating nonlinear data. For nonlinear data, the processing effect of SVD is not good.
[0008] Machine learning algorithms have been continuously proposed and applied, providing a new idea for the field of signal processing. However, in PD signal processing, there are few training samples available and the data is variable. Under such circumstances, it is difficult for machine learning algorithms to obtain good training results. Summary of the invention
[0009] In order to solve the above problems, the present invention proposes a method for extracting local discharge signals based on a learning dictionary, which can stably extract more accurate discharge signals in a more complex environment.
[0010] To achieve the above object, the technical solution adopted by the present invention is: a method for extracting partial discharge signals based on a learning dictionary, comprising the steps of:
[0011] S10, construct a basic dictionary;
[0012] S20, using a sparse autoencoder to update the basic dictionary column by column until the dictionary update is completed;
[0013] S30, sparsely decompose and reconstruct the updated dictionary and the noisy PD signal to extract the low-noise partial discharge signal.
[0014] Furthermore, in step S10, the constructed basic dictionary adopts a K-SVD dictionary.
[0015] Furthermore, constructing the K-SVD dictionary includes the steps of:
[0016] S11, dictionary initialization: initialized by generating a random matrix and then performing column normalization to form the initial dictionary;
[0017] S12, sparse coding: sparse coding is performed on the initial dictionary through the MP greedy algorithm;
[0018] S13, dictionary update: optimize the representation capability of dictionary atoms through iteration to adapt to the characteristics of PD signals.
[0019] Furthermore, the sparse autoencoder adopts an unsupervised neural network structure to map the input data to a lower-dimensional feature space through the encoder and restore it to the original space through the decoder, thereby achieving data compression and reconstruction.
[0020] Furthermore, the sparse autoencoder processing process comprises the steps of:
[0021] The base dictionary is input as raw data into the encoder, which maps the input data into a lower-dimensional feature space;
[0022] Decode it using the first decoder, restore it to the original space, and then input it into the second decoder;
[0023] The second decoder outputs data under the action of the activation function and performs compression and reconstruction.
[0024] The beneficial effects of adopting this technical solution are:
[0025] The present invention uses a more advanced sparse autoencoder learning model in combination with a learning dictionary to more accurately capture the characteristics of the PD signal, thereby successfully detecting and extracting low-noise PD signals from complex environments without the need to use a specific dictionary. The sparse autoencoder can learn the deep-level characteristics of the input data and ensure the sparsity of the encoding, thereby providing better performance when representing PD signals.
[0026] The present invention can stably extract more accurate discharge signals in more complex environments, provide better protection for power system equipment, ensure its safe and reliable operation, and thus avoid serious consequences caused by insulation failure. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 A schematic flow chart of a method for extracting a partial discharge signal based on a learning dictionary according to the present invention;
[0028] Figure 2 Schematic diagram of the process of the sparse autoencoder in an embodiment of the present invention. DETAILED DESCRIPTION
[0029] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described below with reference to the accompanying drawings.
[0030] In this embodiment, see Figure 1 As shown, the present invention proposes a method for extracting partial discharge signals based on a learning dictionary, comprising the steps of:
[0031] S10, construct a basic dictionary;
[0032] S20, using a sparse autoencoder to update the basic dictionary column by column until the dictionary update is completed;
[0033] S30, sparsely decompose and reconstruct the updated dictionary and the noisy PD signal to extract the low-noise partial discharge signal.
[0034] As an optimization solution of the above embodiment, in step S10, the constructed basic dictionary adopts a K-SVD dictionary.
[0035] K-SVD is based on the idea of K-means. Each column in the dictionary and its corresponding sparse data are optimized separately. The optimization method uses the SVD algorithm.
[0036] Specifically, constructing the K-SVD dictionary includes the following steps:
[0037] S11, dictionary initialization: initialized by generating a random matrix and then performing column normalization to form the initial dictionary;
[0038] S12, sparse coding: sparse coding is performed on the initial dictionary through the MP greedy algorithm;
[0039] S13, dictionary update: optimize the representation capability of dictionary atoms through iteration to adapt to the characteristics of PD signals.
[0040] K-SVD is an iterative algorithm that gradually improves the representation ability of the dictionary by alternately performing sparse coding and dictionary updates. Its adaptive nature allows the basic dictionary to capture more details from the signal, better grasp the core features of the PD signal, and continuously update itself. In addition, K-SVD also supports regularization, which helps control the overfitting of the dictionary and ensure stable and accurate signal extraction.
[0041] The present invention generates a matrix filled with random numbers as an initial dictionary. Because the matrix is drawn from a standard normal distribution, the values in the matrix will be concentrated around 0, but there will also be some larger or smaller values. The random matrix columns are then normalized to ensure that the length (or "energy") of each column is 1. Such an initialization method is the most commonly used initial dictionary generation method in K-SVD, and it does not utilize any prior knowledge about the original signal. At the same time, this random initialization method is unsupervised, which means that the method does not rely on any specific attributes of the training data.
[0042] As an optimization solution of the above embodiment, the sparse autoencoder adopts an unsupervised neural network structure, maps the input data to a lower-dimensional feature space through the encoder, and restores it to the original space through the decoder, so as to achieve data compression and reconstruction;
[0043] This structure can effectively capture the intrinsic structure and main features of the data; sparsity means that when representing data, only a few neurons are active or have a high response.
[0044] The K-SVD training model derived from K-means is not very suitable for PD signals because the characteristics of the signal are submerged in the noise, and the K-SVD algorithm is difficult to identify and extract the original signal features. Therefore, the sparse autoencoder is used after the K-SVD dictionary is established. The sparse autoencoder can learn the deep features of the input data and ensure the sparsity of the encoding. By combining the K-SVD dictionary and the sparse autoencoder, better performance is provided when representing PD signals.
[0045] Specifically, Figure 2 As shown, the sparse autoencoder processing process includes the steps of:
[0046] The base dictionary is input as raw data into the encoder, which maps the input data into a lower-dimensional feature space;
[0047] Decode it using the first decoder, restore it to the original space, and then input it into the second decoder;
[0048] The second decoder outputs data under the action of the activation function for compression and reconstruction. An activation function, such as ReLU, is added to the output layer of the decoder to increase the nonlinear range of the reconstructed signal and thus calibrate the amplitude.
[0049] In sparse autoencoders, the sparsity of the encoding is ensured by adding a sparsity constraint or penalty.
[0050] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.
Claims
1. A method for extracting partial discharge signals based on a learning dictionary, characterized in that: Includes steps: S10, construct a basic dictionary; S20, using a sparse autoencoder to update the basic dictionary column by column until the dictionary update is completed; S30, sparsely decompose and reconstruct the updated dictionary and the noisy PD signal to extract the low-noise partial discharge signal.
2. The method for extracting partial discharge signals based on a learning dictionary according to claim 1, characterized in that: In step S10, the constructed basic dictionary adopts a K-SVD dictionary.
3. The method for extracting partial discharge signals based on a learning dictionary according to claim 2, characterized in that: Constructing the K-SVD dictionary includes the following steps: S11, dictionary initialization: initialized by generating a random matrix and then performing column normalization to form the initial dictionary; S12, sparse coding: sparse coding is performed on the initial dictionary through the MP greedy algorithm; S13, dictionary update: optimize the representation capability of dictionary atoms through iteration to adapt to the characteristics of PD signals.
4. The method for extracting partial discharge signals based on a learning dictionary according to claim 3, characterized in that: The sparse autoencoder uses an unsupervised neural network structure to map the input data to a lower-dimensional feature space through the encoder and restore it to the original space through the decoder, thereby achieving data compression and reconstruction.
5. The method for extracting partial discharge signals based on a learning dictionary according to claim 4, characterized in that: The sparse autoencoder processing process includes the steps of: The base dictionary is input as raw data into the encoder, which maps the input data into a lower-dimensional feature space; Decode it using the first decoder, restore it to the original space, and then input it into the second decoder; The second decoder outputs data under the action of the activation function and performs compression and reconstruction.
