A method for pattern recognition of UHF partial discharge signals based on time-frequency fusion features

Abstract

The present invention discloses a method for pattern recognition of ultra-high frequency partial discharge signals based on time-frequency fusion features, which belongs to the technical field of partial discharge pattern recognition. The method integrates the time domain information after similarity alignment by dynamic time warping algorithm and the time-shift invariant and multi-scale frequency domain features extracted by wavelet scattering, enriching the feature representation of partial discharge pulse signals. Specifically, it includes step one: collecting partial discharge signals caused by transformer insulation defects; step two: preprocessing data, periodic interception, data normalization, denoising, pulse signal extraction, and data set construction; step three: using dynamic time warping algorithm to perform similarity alignment enhancement on the training set; step four: constructing a wavelet scattering feature extraction network, extracting the frequency domain features of the data and performing dimensional transformation; step five: fusing the time-frequency domain feature information, using the fused data to train the classifier, and finally using the parameter-optimized model to perform pattern recognition on the partial discharge type.

Description

Technical Field

[0001] The invention relates to the technical field of partial discharge pattern recognition, and in particular to a method for ultra-high frequency partial discharge signal pattern recognition based on time-frequency fusion characteristics. Background Art

[0002] Rapid and accurate identification of partial discharge patterns is essential to prevent interruptions associated with insulation faults in industrial equipment. Ultra-high frequency (UHF) methods have been widely used to diagnose various industrial signals due to their good versatility and non-contact capabilities. Compared with the traditional method of manually constructing data features, the method of automatically learning data features using deep learning technology has more efficient pattern recognition efficiency and less workload; compared with the traditional method of insulating oil chemical analysis, it has stronger real-time data analysis capabilities, and this technology has obvious advantages.

[0003] Before using deep learning technology to automatically extract data features, the data needs to be effectively processed in order to fully utilize the feature extraction capabilities of the deep learning model and achieve better recognition results. At present, the most widely used feature transformation methods for deep learning model pattern recognition are: one-dimensional signal waveform, phase amplitude diagram (PRPD), time-frequency diagram, Fourier transform, EMD decomposition, symmetric point pattern diagram and other technologies. These feature transformation methods all have certain problems. A large amount of fault data is required to construct a PRPD diagram and reflect the pattern characteristics of different types of discharge signals; other feature transformation methods only use the feature information in a single information domain of the fault signal in the time domain or frequency domain. Experiments have shown that the time and frequency domains contain effective feature information of partial discharge signals, and rich feature expressions have a positive effect on the pattern recognition of deep learning models.

[0004] The above problems limit the accuracy of UHF partial discharge signal pattern recognition, increase the possibility of misjudgment of partial discharge signal type, and thus affect the correctness of the analysis of different types of partial discharge discharge and insulation defect damage degree of monitoring data. To address the above problems, this paper proposes a UHF partial discharge signal pattern recognition method based on time-frequency fusion features. Summary of the invention

[0005] The purpose of the present invention is to solve the above-mentioned problem and to provide a method for identifying UHF partial discharge signal patterns based on time-frequency fusion features.

[0006] To achieve the above object, the technical solution adopted by the present invention is:

[0007] The dynamic time warping algorithm is used to align the pulse training data for similarity, reduce the impact of waveform distortion on the signal, enrich the feature expression of the partial discharge signal through time-frequency fusion features, and finally train the classifier to learn the time-frequency fusion feature data to achieve accurate recognition of the partial discharge signal pattern, thereby reducing the probability of misjudgment of the signal type and achieving accurate analysis of the working conditions of electrical equipment such as transformers. The following steps are included:

[0008] S1. Build different types of defect models and apply high voltages of different levels to the defect models, find a suitable pressure range, make the collection of partial discharge signals meet the threshold requirements, and collect different types of partial discharge ultra-high frequency signals through high-frequency current transformers;

[0009] S2. Data preprocessing: complete cycle capture of partial discharge signal, data normalization, data denoising, partial discharge pulse signal data extraction, and data set construction.

[0010] S3. Use the dynamic time warping algorithm to perform similarity alignment on the partial discharge data in the training set to increase the intra-class similarity of the same type of partial discharge data and reduce the impact of signal expansion and deformation, so as to extract effective time domain features of partial discharge signals.

[0011] S4. For the partial discharge data set enhanced by the dynamic time warping algorithm, a wavelet scattering feature extraction network is used to extract features of the partial discharge pulse signal to obtain frequency domain feature information of the partial discharge;

[0012] S5. The time domain information and the frequency domain information are fused to obtain the time-frequency fusion feature vector of the partial discharge data. The fused feature data is passed into the classification model for training. After parameter optimization, a classifier model that can be used for partial discharge pattern recognition is obtained. The pattern recognition task can be completed by inputting the partial discharge signal.

[0013] Furthermore, the step S3 includes the following sub-steps:

[0014] S31. Visually traverse different types of partial discharge data in the training set. In each category file, select the data with relatively standard discharge waveform as the standard pulse waveform standard_paulse. Use the standard pulse waveform to cyclically perform dynamic time warping algorithm on the remaining waveform data in the file to obtain the cumulative distance matrix By accumulating the distance matrix, we can get the minimum similarity of each pulse waveform data after similarity alignment with the standard waveform standard_paulse, and we can get the corresponding extension path under the minimum similarity by back-tracing the cumulative distance matrix.

[0015] The calculation formula for each element in the cumulative distance matrix is:

[0016] D [ i ] [ j ] = d ( x i , y j ) + min ( D [ i − 1 ] [ j ] , D [ i ] [ j − 1 ] , D [ i − 1 ] [ j − 1 ] )

[0017] Respectively represent the index values ​​of the elements in the two comparison signal sequences, D [ i ] [ j ] Represents an index The minimum distance value of the corresponding element, express The actual distance of the corresponding element.

[0018] S32. For each pulse waveform data, data expansion is performed according to the calculated minimum similarity expansion path. The pulse waveform data in the training set after expansion has a closer similarity. This processing method can eliminate the influence caused by partial waveform distortion and enhance the inter-class similarity of pulse data.

[0019] Furthermore, the step S4 includes the following sub-steps:

[0020] S41. Build a wavelet scattering feature extraction network, preferably with three layers, and obtain the high-frequency scattering features of the enhanced partial discharge waveform data through wavelet convolution, modular nonlinearity, and pooling operations. The high-frequency scattering features contain the frequency domain feature information of the partial discharge waveform.

[0021] S42, performing dimension transformation on the obtained high-frequency scattering features to obtain a one-dimensional frequency domain feature vector, where different intervals of the one-dimensional data represent different high-frequency information features.

[0022] The specific time-frequency domain feature fusion process of S5 is as follows:

[0023] After transforming the dimension of the high-frequency scattering feature, a frequency domain feature vector containing different frequency components is obtained. This vector is spliced ​​with the time domain waveform data enhanced by the dynamic time warping algorithm to obtain feature data that integrates the time and frequency domain feature information, and the feature data is used to construct a data set.

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

[0025] 1. The present invention uses a dynamic time warping algorithm to perform similarity alignment on data of the same category in the constructed partial discharge pulse training set, which can effectively reduce the impact caused by the expansion and contraction distortion of the waveform in the data set;

[0026] 2. By fusing enhanced time domain data and frequency domain features extracted by wavelet scattering, the expression of partial discharge pulse features is enriched, providing a data basis for feature learning of the classifier and giving full play to the role of the network;

[0027] 3. This method avoids the problem of manually selecting appropriate wavelet bases in the time-frequency transformation process to obtain better frequency domain data, thus reducing the workload. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of the system flow of the present invention;

[0029] Figure 2 is a similarity alignment graph of the dynamic time warping algorithm of the present invention;

[0030] Figure 3 It is a flow chart of feature extraction of wavelet scattering feature extraction network in the present invention;

[0031] Figure 4 4 is a classification accuracy curve of the classifier in the embodiment of the present invention. DETAILED DESCRIPTION

[0032] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods.

[0033] according to Figure 1 As shown, the present invention provides a method for identifying a UHF partial discharge signal pattern based on time-frequency fusion features, comprising the following steps:

[0034] S1. Build different types of defect models and apply high voltages of different levels to the defect models, find a suitable pressure range, make the collection of partial discharge signals meet the threshold requirements, and collect different types of partial discharge ultra-high frequency signals through high-frequency current transformers;

[0035] S2. Data preprocessing: complete cycle capture of partial discharge signal, data normalization, data denoising, partial discharge pulse signal data extraction, and data set construction.

[0036] S3. Dynamic time warping algorithm is used to enhance the partial discharge data in the training set, increase the intra-class similarity of partial discharge data of the same type, and reduce the inter-class similarity between different partial discharge types, so as to extract effective partial discharge signal time domain features, such as Figure 2 shown.

[0037] Since the state of the partial discharge pulse waveform is different each time, the complete pulse waveform collected may be of the same discharge type, but the change in the collection environment will cause waveform expansion and a certain degree of distortion, affecting the training of the classifier model. Therefore, the present invention uses a dynamic time warping algorithm to align all pulses in the same category for similarity, and expand the same type of discharge pulses according to the selected standard waveform with the minimum similarity, so as to improve the inter-class similarity between the same types. The specific steps are:

[0038] Visually traverse different types of partial discharge data in the training set. In each category file, select the data with relatively standard discharge waveform as the standard pulse waveform standard_paulse. Use the standard pulse waveform to cyclically perform dynamic time warping algorithm on the remaining waveform data in the file to obtain the cumulative distance matrix By accumulating the distance matrix, the minimum similarity of each pulse waveform data after similarity alignment with the standard waveform standard_paulse can be obtained, and the expansion path corresponding to the minimum similarity can be obtained by backtracking the cumulative distance matrix.

[0039] The calculation formula for each element in the cumulative distance matrix is:

[0040] D [ i ] [ j ] = d ( x i , y j ) + min ( D [ i − 1 ] [ j ] , D [ i ] [ j − 1 ] , D [ i − 1 ] [ j − 1 ] )

[0041] Represent the index values ​​of the elements in the two sequences respectively. D [ i ] [ j ] Represents an index The minimum distance value of the corresponding element, express The actual distance of the corresponding element.

[0042] For each pulse waveform data, data expansion is performed according to the calculated minimum similarity expansion path. The pulse waveform data in the expanded training set has closer similarity. This processing method can eliminate the influence caused by partial waveform distortion and enhance the inter-class similarity of pulse data.

[0043] S4. For the partial discharge data set enhanced by the dynamic time warping algorithm, the wavelet scattering feature extraction network is used to extract the features of the partial discharge pulse signal to obtain the frequency domain feature information of the partial discharge. The calculation process is as follows: Figure 3 As shown;

[0044] The specific steps of S4 are:

[0045] Build a wavelet scattering feature extraction network, preferably with three layers, and obtain the high-frequency scattering features of the enhanced partial discharge waveform data through wavelet convolution, modulus nonlinearity, and pooling operations. The high-frequency scattering features contain the frequency domain feature information of the partial discharge waveform. Perform dimensional transformation on the obtained high-frequency scattering features to obtain a one-dimensional frequency domain feature vector.

[0046] S5. The time domain information and the frequency domain information are fused to obtain the time-frequency fusion feature vector of the partial discharge data. The fused feature data is passed into the classification model for training. After parameter optimization, a classifier model that can be used for partial discharge pattern recognition is obtained. The pattern recognition task can be completed by inputting the partial discharge signal.

[0047] The specific steps are:

[0048] After the high-frequency scattering feature is transformed by dimension, a frequency domain feature vector containing different frequency components is obtained. This vector is spliced ​​with the time domain waveform data enhanced by the dynamic time warping algorithm to obtain feature data that integrates the time and frequency domain feature information, and the feature data is used to construct a data set. The data set is used to train a one-dimensional classification neural network. The structure diagram of the one-dimensional classification neural network used in the embodiment of the present invention is shown in Table 1. The table records in detail the structure and output vector shape of each layer of the classifier used in the present invention, as well as the parameter quantity of each layer structure. The structure is used as a classifier to perform classification test on the data set. However, it should be noted that the optional classifier is not limited to this structure. Here, only a case reference is provided to describe the technical method in more detail. Finally, the trained classifier model can be used to complete the pattern recognition task.

[0049] Table 1

[0050]

[0051] Example:

[0052] The present invention constructs four typical insulation defect structures, namely metal discharge, air gap discharge, surface discharge and needle tip discharge. When high voltage is applied to the insulation defect structure, a UHF method is used to collect local discharge data of different insulation types. The sampling rate is set to 250MSa / s, and 600 pieces of discharge partial discharge pulse data generated by each typical insulation defect are obtained. The number of data points of each partial discharge pulse is set to 500. A data set is constructed according to a ratio of 4:1. The total data volume of the verification set is 800, and the total data volume of the training set is 1600, which is recorded as data set 1; wavelet scattering feature extraction is performed on the original signal waveform, and the parameter QualityFactors is set to [8,1] to obtain a 65x8 scattering feature vector. The scattering feature of each pulse is expanded and used as partial discharge feature data to construct a scattering feature data set, which is recorded as data set 2; the method of the present invention adopts a dynamic time warping algorithm to perform similarity alignment enhancement on the signal, and then fuses it with the obtained scattering frequency feature, and the fused feature is used as data to construct a data set, which is recorded as data set 3.

[0053] In order to compare the effectiveness of three different partial discharge features for pattern recognition, this embodiment constructs a classifier based on a neural network. The pattern recognition accuracy of three different data sets is shown in the attached figure. Figure 4As shown, it can be observed that both the original signal waveform and the scattering characteristics can enable the classifier to achieve a certain classification accuracy, indicating that both data represent certain partial discharge characteristics; but the time-frequency features after fusion processed by the method of this paper have better classification effect in the classifier, proving that the fused time-frequency features can contain richer feature information of partial discharge data, providing a data basis for pattern recognition of deep learning models, and at the same time proving the feasibility of the present invention.

[0054] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

[0055] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A method for ultra-high frequency partial discharge signal pattern recognition based on time-frequency fusion features; comprising the following steps: S1. Build different types of defect models and apply high voltages of different levels to the defect models, find a suitable pressure range, make the collection of partial discharge signals meet the threshold requirements, and collect different types of partial discharge ultra-high frequency signals through high-frequency current transformers; S2, data preprocessing, complete cycle interception of partial discharge signal, data normalization, data denoising, partial discharge pulse signal data extraction, and data set construction; S3. Use the dynamic time warping algorithm to perform similarity alignment on the partial discharge data in the training set, increase the intra-class similarity of the same type of partial discharge data, and reduce the impact caused by signal expansion and deformation, so as to extract effective partial discharge signal time domain features; S4. For the partial discharge data set enhanced by the dynamic time warping algorithm, a wavelet scattering feature extraction network is used to extract features of the partial discharge pulse signal to obtain frequency domain feature information of the partial discharge; S5, fusing the time domain information and the frequency domain information to obtain the time-frequency fusion feature vector of the partial discharge data, passing the fused feature data into the classification model for training, and obtaining a classifier model that can be used for partial discharge pattern recognition after parameter optimization, and completing the pattern recognition task by inputting the partial discharge signal; Furthermore, the step S3 includes the following sub-steps: S31. Visually traverse different types of partial discharge data in the training set. In each category file, select the data with relatively standard discharge waveform as the standard pulse waveform standard_paulse. Use the standard pulse waveform to cyclically perform dynamic time warping algorithm on the remaining waveform data in the file to obtain the cumulative distance matrix , the minimum similarity after similarity alignment between each pulse waveform data and the standard waveform standard_paulse can be obtained by accumulating the distance matrix, and the corresponding extension path under the minimum similarity can be obtained by back-tracing the accumulated distance matrix; The calculation formula for each element in the cumulative distance matrix is: Respectively represent the index values ​​of the elements in the two comparison signal sequences, Represents an index The minimum distance value of the corresponding element, express The actual distance of the corresponding element; S32, for each pulse waveform data, data expansion is performed according to the calculated minimum similarity expansion path, and the pulse waveform data in the training set after expansion has a closer similarity. This processing method can eliminate the influence caused by partial waveform distortion and enhance the inter-class similarity of pulse data; Furthermore, the step S4 includes the following sub-steps: S41, building a wavelet scattering feature extraction network, preferably with three layers, and obtaining high-frequency scattering features of the enhanced partial discharge waveform data through wavelet convolution, modulus nonlinearity, and pooling operations, wherein the high-frequency scattering features include frequency domain feature information of the partial discharge waveform; S42, performing dimensional transformation on the obtained high-frequency scattering features to obtain a one-dimensional frequency domain feature vector, where different intervals of the one-dimensional data represent different high-frequency information features; The specific time-frequency domain feature fusion process of S5 is as follows: After transforming the dimension of the high-frequency scattering feature, a frequency domain feature vector containing different frequency components is obtained. This vector is spliced ​​with the time domain waveform data enhanced by the dynamic time warping algorithm to obtain feature data that integrates the time and frequency domain feature information, and the feature data is used to construct a data set.

Images