Transformer wide-area voiceprint diagnosis monitoring method based on atlas domain

Abstract

The invention discloses a transformer wide-area voiceprint diagnosis monitoring method based on an atlas domain, and belongs to the technical field of transformer diagnosis, and the method comprises the following steps: S1, equipment voiceprint feature extraction; s2, carrying out equipment voiceprint recognition; s3, establishing a voiceprint feature database; and S4, monitoring the database. The voiceprint of the power transformer is collected, the noise signal of the transformer is collected, a related preprocessing and feature extraction algorithm is developed, a transfer learning artificial intelligence recognition network model is designed based on the characteristics of the noise signal of the transformer, and intelligent diagnosis and recognition of working conditions such as direct current magnetic bias, winding deformation and overload of the transformer are achieved. And the actual substation operation and maintenance requirements can be met. According to the transformer voiceprint feature database establishment method, along with system field deployment application and popularization, transformer operation abnormal working condition voiceprint samples are collected, an algorithm model is continuously optimized and improved, database content is enriched, and a comprehensive and reliable voiceprint feature database is established step by step.

Description

Technical Field

[0001] The present invention belongs to the technical field of transformer diagnosis, and in particular relates to a transformer wide-area voiceprint diagnosis and monitoring method based on a spectrum domain. Background Art

[0002] With the development of economy and society and the advancement of urbanization, the power industry has been booming. As the main equipment of power grid, power transformers have also been widely used in transmission and distribution networks. The stable operation of power grid and the efficient supply of electric energy are related to the development of national economy and the quality of life of people. Power transformer is one of the most important equipment in all links of power grid. Once a failure occurs, it will seriously affect the operation of power grid system and cause significant economic losses. Power transformer is large in size, complex in structure and expensive. Most of its major failures and serious accidents are caused by the accumulation of latent failures such as internal partial discharge, local overheating, winding deformation, loose mechanical parts, and aging of equipment insulation over time. Therefore, it is an important task for power operation and maintenance personnel to monitor the operation status of power transformer online, detect and find latent faults, analyze the development trend of latent faults based on monitoring data, accurately predict the occurrence of accidents, and timely inspect and maintain power transformers. The main means of online operation live detection of power transformers include voltage and current and other electrical quantity monitoring, spectral diagnosis, oil chromatography diagnosis, ultrasonic detection, infrared thermal imaging diagnosis, etc. Some monitoring sensor equipment has also been deployed in the substation, which has improved the ability of transformer online monitoring and fault analysis to a certain extent. However, the detection of equipment operation sound still needs to rely on the experience of patrol personnel to make abnormal judgments. There is little research on the intelligent perception of transformer sound state attributes, and the mature application of voiceprint online monitoring is still blank. The sound emitted by the transformer during operation contains a lot of state information. There are large differences in the sound signals emitted in normal state and abnormal fault. Different fault types can produce voiceprint information with inherent acoustic characteristics. Through voiceprint feature extraction and analysis, internal faults of the transformer can be effectively detected, such as DC bias magnetism, overload, mechanical looseness, transformer winding deformation, etc., which can effectively supplement the effectiveness of conventional monitoring technology. Experienced operation and maintenance personnel can qualitatively judge whether the transformer has a fault by the sound emitted by the transformer, but the human ear monitoring method is easily affected by subjectivity, has large errors, and cannot be monitored for a long time, and cannot achieve quantitative scientific analysis. Therefore, the use of voiceprint sensor monitoring and intelligent diagnosis to collect, process and analyze transformer sound signals can quantitatively monitor transformer operation voiceprint data in real time, solve the shortcomings of manual monitoring, and effectively diagnose and evaluate the health status of the transformer and analyze and predict abnormal faults. Summary of the invention

[0003] The purpose of the present invention is to overcome the deficiencies of the prior art and to provide a transformer wide-area voiceprint diagnosis and monitoring method based on the spectrum domain, which solves the problems in the above-mentioned background technology.

[0004] The object of the present invention is achieved by: a transformer wide-area voiceprint diagnosis and monitoring method based on the spectrum domain, which comprises the following steps:

[0005] S1, device voiceprint feature extraction;

[0006] S2, device voiceprint recognition;

[0007] S3, establish a voiceprint feature database;

[0008] S4. Monitor the database.

[0009] Furthermore, the device is a transformer; the feature extraction includes sound signal preprocessing, feature extraction and feature connection, and the sound signal preprocessing includes framing and pre-emphasis, the purpose of which is to remove the interference of redundant noise and prepare for the next step of short-time feature extraction; the feature extraction part includes short-time time domain features and short-time frequency domain features, and short-time energy, short-time average zero-crossing rate, short-time energy entropy, and Mel cepstral coefficients are used to describe the framing features; the extracted features are connected into feature vectors as multi-feature fusion data.

[0010] Furthermore, the feature connection improves the processing algorithm through the optimization of the MFCC feature vector. First, a weighted optimization method based on F-score is introduced into the MFCC feature vector group, and the recognition rate of the transformer noise signal is improved by tuning the F-score of the feature vector; secondly, the feature vector formed by the connection of short-time time domain and frequency domain features is reduced in dimension through PCA, and the dimension of the feature vector is reduced as much as possible under the premise of ensuring that information is not lost, thereby reducing the computational complexity; the F-score is an indicator in the field of metrology and statistics, which is used to examine the effectiveness of each dimensional component in multidimensional data; a component with a high F-score indicates that the component has a small variance in the same type of data and a large variance in different types of data, and is more efficient when used for data classification; the calculation formula for the F-score value F(k) of the component of the kth dimension of the feature vector is:

[0011]

[0012] Where N is the number of objects to be identified, u k (i) is the k-th dimension component of the feature vector of the i-th object, u k is the mean; x k (i) is the k-th dimension component of all samples of the i-th identified object, n i is the number of samples of the i-th identified object.

[0013] Furthermore, the voiceprint recognition includes modeling the target working condition after extracting the transformer voiceprint feature, comparing and judging the noise signal to be tested, and identifying the type of fault condition; it includes identifying transformer DC bias magnetization, overload, overexcitation, partial discharge, mechanical looseness, transformer winding deformation, on-load tap changer fault and circuit breaker operating mechanism abnormality, and using the deep neural network as the baseline to improve the DNN model for the transformer noise feature, and after the DNN model training is completed, remove the last classification layer, and use the output result of the last hidden layer as the voiceprint template for comparison; for different working conditions, randomly take out 5 to 15 data from the training data set, obtain the target template through the network model, and cluster the obtained target template using the k-means clustering algorithm; cluster the target template output corresponding to the above 5 to 15 data with k being 2, and after clustering, select the cluster where the majority of samples are located as the target template p representing this type of working condition, then extract the features of the noise signal to be detected, input the trained network model, and obtain the output O of the last hidden layer, and finally, use the cosine similarity metric to compare the similarity between the output O and the target template p.

[0014] Furthermore, the voiceprint feature database established includes a transformer working state acoustic feature database, which contains the acoustic features of the transformer in normal state and different fault states, and uses it as sample data for artificial intelligence algorithm model and voiceprint intelligent diagnosis; the neural network model is trained using labeled acoustic data, and the transformer voltage, current, load, oil temperature, winding temperature and other state parameters are used as inputs of the neural network model, and the neural network model gives an output prediction on whether the transformer is abnormal; judgments are made through the experience and knowledge of power experts, and the experience and knowledge of the experts are imparted to the neural network, and the neural network model is updated. After repeated iterative training, the neural network can also give corresponding predictions for unknown situations; the artificial intelligence algorithm is then used to predict the working state of the transformer and the state parameters of the transformer, and for each abnormal state, the database includes the distinguishing feature type of its abnormal state, the used power experience knowledge is stored in the database, and a database interface for multiple attribute predictions is reserved.

[0015] Furthermore, the monitoring of the database includes voiceprint monitoring, integrated noise signal storage, voiceprint feature extraction and voiceprint recognition algorithm, to achieve online monitoring and intelligent diagnosis of transformer sound, including voiceprint display, equipment and data management, real-time monitoring, active alarm, manual diagnosis and voiceprint sample library; using the voiceprint recognition algorithm, a comprehensive integrated diagnosis is performed on the online monitored voiceprint data, including spectrum analysis, event detection, load identification and time-frequency domain analysis, to form a comprehensive diagnosis report on the equipment operation status; at the same time, it has the transformer equipment management function, which can manage and view the overall operating condition information of the transformer, and generate a maintenance work order based on the comprehensive voiceprint intelligent diagnosis results, manufacturer, voltage level, operation life, winding type and cooling method.

[0016] Furthermore, the establishment of the voiceprint feature database includes transformer discharge fault diagnosis, using synchronous compression wavelet transform to enhance the energy aggregation in time-frequency and obtain clear time-frequency distribution characteristics; assuming that the signal f(t)∈L 2 (R), then the continuous wavelet transform WT(a,b) of the signal f(t) is:

[0017]

[0018] Where: a, b represent the scale and displacement factor respectively; represents the wavelet basis function; is the conjugate of ψ(·);

[0019] The frequency domain expression of WT(a,b) is:

[0020]

[0021] Where: They represent the Fourier transform of the time domain signals f(ω) and ψ(ω), respectively; W represents the frequency; represents an imaginary number;

[0022] Based on the above two equations, we can get the instantaneous frequency ω of the signal: f for:

[0023]

[0024] The wavelet coefficients are compressed and rearranged to obtain the synchronous compressed wavelet transform SWT f The formula is;

[0025] SWT f (b,ω)=∫WT f (a,b)a -3 / 2 δ[ω(a,b)-ω]da

[0026] Where δ(·) represents the Dirac function.

[0027] Furthermore, the voiceprint feature extraction includes transformer voiceprint feature extraction, reversible and multi-scale time-frequency analysis is performed through short-time Fourier transform and Wigner distribution, the overall signal is obtained on the time-frequency phase plane, and a frequency-related adjustment factor is introduced into the window function to adjust the time-frequency resolution, and high-frequency features and time-frequency information are extracted; the sound signal is subjected to generalized S transform to obtain a time-frequency graph, and the texture features of the signal time-frequency graph are extracted to characterize the state information of the transformer. The S transform of the stationary signal x(t) is:

[0028]

[0029] Where: t is time, f is frequency, t is the Gaussian window position parameter, ω(f,τ-1) is the Gaussian window function; since the Gaussian window function of different signals at the same frequency is fixed, in order to improve the energy concentration of time-frequency, parameters α and β are introduced. The basic formula is:

[0030]

[0031] In the formula, α and β are adjustment factors and are positive numbers.

[0032] Furthermore, it includes a voiceprint time-frequency spectrum diagram, which is a two-dimensional energy spectrum diagram formed by framing the sound signal, adding windows and performing discrete Fourier transformation, and stacking it according to the time dimension; the voiceprint time-frequency spectrum has the time domain and frequency domain characteristics of the sound signal, as well as intuitive image characteristics, which is convenient for subsequent deep learning processing. The core voiceprint signal is cut into samples with a duration of 1 s, and the overlapping segmentation method is used for framing. The frame length is set to 4096 frames, and the frame shift is set to 512 frames; in order to effectively reduce signal distortion, the window function selects the Hamming window, and its formula is:

[0033]

[0034] The framed and windowed samples are then subjected to short-time discrete Fourier transform to obtain a voiceprint time-frequency spectrum. The horizontal axis of the voiceprint time-frequency spectrum is time, and the vertical axis is frequency. The color of the graph represents the content of the current frequency component at the current moment, which is represented by the power spectral density.

[0035] Furthermore, including data set comparison test, through classification recognition algorithms SVM, NN and CNN, the recognition accuracy comparison test is carried out on the ESC-50 data set using traditional classification recognition algorithm and transfer learning network; first, the 2000 audios of the ESC-50 data set are divided into 5 stacks, each with 400, covering all 50 classes, and ensuring that the audios of the same scene at different time periods only appear in one stack to avoid data contamination; secondly, the training and test cross-validation methods are used in the 5 stacks of data respectively, and the training data volume and the test data volume are cross-folded 50%, 80% of each stack is selected as the training set, and the remaining 20% ​​is used as the test set, and then, the results of the 5 stacks are averaged to obtain the mean recognition accuracy. Combined with the characteristics of transformer voiceprint data, the recognition accuracy of the transfer learning deep neural network algorithm is improved from 42.63% of the general CNN algorithm to 91.80%; the classifier recognition accuracy is obtained through the ESC-50 data set.

[0036] The beneficial effects of the present invention are as follows: by collecting the voiceprint of the power transformer, the transformer noise signal is collected, and the relevant preprocessing and feature extraction algorithms are developed. Based on the characteristics of the transformer noise signal, a transfer learning artificial intelligence recognition network model is designed to realize the intelligent diagnosis and recognition of the working conditions such as DC bias magnetism, winding deformation, and overload of the transformer. Through data set testing and substation deployment and application, the effectiveness and reliability of the voiceprint collection device and related algorithms are verified, which can meet the actual substation operation and maintenance needs. The method for establishing a transformer voiceprint feature database, with the on-site deployment, application and promotion of the system, collects voiceprint samples of abnormal transformer operation conditions, continuously optimizes and improves the algorithm model, enriches the database content, and gradually establishes a comprehensive and reliable voiceprint feature database. The accuracy rate of transformer abnormal working condition recognition can reach about 70%. Using the voiceprint recognition method, on the one hand, it can independently identify specific categories of defects, effectively solve the potential defects that cannot be identified by means such as images or online monitoring, and on the other hand, it can be used as an effective supplement to other monitoring methods such as partial discharge to improve the comprehensive monitoring capability. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a transformer voiceprint feature extraction flow chart of the present invention;

[0038] Figure 2 is a noise signal preprocessing flow chart of the present invention;

[0039] Figure 3 It is the MFCC feature extraction flow chart of the present invention;

[0040] Figure 4 It is a diagram of the voiceprint database construction process of the present invention;

[0041] Figure 5 It is a flow chart of the gray wolf optimization algorithm of the present invention;

[0042] Figure 6 It is a diagram of the process of obtaining the Mel time-frequency spectrum of the voiceprint signal of the present invention;

[0043] Figure 7 This is a diagram of the voiceprint time-frequency spectrum drawing process of the present invention. DETAILED DESCRIPTION

[0044] The present invention is further described in detail below in conjunction with the accompanying drawings. It should be pointed out that it is only for the purpose of more clearly describing and explaining the present invention.

[0045] Example 1

[0046] like Figure 1-7 As shown, this embodiment discloses a transformer wide-area voiceprint diagnosis and monitoring method based on the spectrum domain, which includes the following steps:

[0047] S1. Device voiceprint feature extraction; the device is a transformer; the feature extraction includes sound signal preprocessing, feature extraction and feature concatenation.

[0048] Transformer noise signal preprocessing, transformer noise signals are usually collected for a long time, and it is difficult to directly analyze them in the time domain and frequency domain. Therefore, it is necessary to preprocess them first, and divide the transformer sound signals collected in a period of time into several short-time signals before processing. Signal preprocessing includes two steps: framing and windowing. In order to ensure the continuity of two adjacent frames, there is generally overlap between the two frames. During the framing process, if the frame length is too short, the feature vector scale is small and the representativeness is poor; if the frame length is too long, the sound signal changes too much, affecting the accuracy of the feature vector. Compared with the speech signal, the transformer noise signal is relatively stable. Under the premise of not seriously affecting the recognition efficiency, the frame length can be appropriately increased to obtain higher accuracy. According to experimental test analysis, the frame length of about 500 ms is most suitable for transformer noise. In addition, considering that the noise signal of the transformer under the same working conditions is relatively stable, the continuity of the adjacent two frames after framing is good, the overlap rate of adjacent frames can be appropriately reduced to reduce the amount of calculation. According to experimental test analysis, the overlap rate of adjacent frames can be taken as 40%. In order to extract the frame features, it is necessary to perform discrete Fourier transform on it. If the discrete Fourier transform is directly performed on the frame signal, it will produce large distortion. Therefore, each frame signal is first windowed and then Fourier transformed to increase the continuity of the two ends of the signal, thereby reducing the distortion caused by Fourier transform. Commonly used windowing functions include rectangular window, Hamming window and Hanning window. The Hamming window is used to process the frame signal. The Hamming window has good low-pass capability and can better reflect the frequency characteristics of short-time signals.

[0049] Transformer noise feature extraction, after the transformer noise signal is framed and preprocessed, each frame signal is regarded as a short-time stationary signal, and the short-time domain features and short-time domain features are extracted respectively. The short-time domain features mainly include short-time energy, short-time average zero-crossing rate, and short-time energy entropy. The short-time frequency features use Mel-Frequency Cepstral Co-efficien ts (MFCC). In addition, according to the operating conditions of the transformer, the energy similarity, vibration correlation, vibration stability and frequency complexity are designed. Short-time energy, calculate the energy of each frame noise signal, and the vector composed of the short-time energy of all frames is the short-time energy feature; short-time average zero-crossing rate, calculate the number of times the signal waveform crosses the horizontal axis in each frame, and the vector composed of the short-time zero-crossing rate of all frames is the short-time average zero-crossing rate feature; short-time energy entropy, calculate the uniformity of energy distribution in each frame, and the vector composed of the short-time energy entropy of all frames is the short-time energy entropy feature. Mel frequency cepstrum coefficient: The process of obtaining MFCC mainly includes five processes: fast Fourier transform, modulus value, Mel filtering, logarithmic transform, and discrete cosine transform. First, a frame signal is transformed by FFT, and its modulus value is taken. Then, it is transformed to the Mel frequency domain through n Mel filter groups, and then logarithmic transform is performed. Finally, a group of n coefficients c(i) obtained by DCT operation constitute the MFCC feature vector of the frame signal. An MFCC feature vector is obtained for each preprocessed frame signal, and together they form an MFCC feature vector group. The horizontal axis of the voiceprint feature graph extracted from the transformer noise signal represents time, and the vertical axis represents the noise Mel frequency. The brighter the color, the greater the sound energy.

[0050] Texture, as a natural attribute of images, is an important characteristic component of the visual system. The generalized S transform is developed from the short-time Fourier transform (STFT) and the Wigner distribution (WVD). As a lossless reversible transform and multi-scale time-frequency analysis method, it is easy to obtain the time-frequency composition of the overall signal on the time-frequency phase plane. It makes up for the defects of the single resolution of STFT and the phase of WT and there is no cross-term interference problem. The frequency-related adjustment factor is introduced in the window function to adjust the time-frequency resolution. It can accurately express non-stationary sound signals whose frequency components change with time, and extract high-frequency features and rich time-frequency information. Generalized S transform, generalized S transform is performed on the sound signal to obtain a time-frequency diagram, and the texture features of the signal time-frequency diagram are extracted to represent the state information of the transformer. The S transform of the stationary signal x(t) is:

[0051]

[0052] Where: t is time, f is frequency, t is the Gaussian window position parameter, ω(f,τ-1) is the Gaussian window function; since the Gaussian window function of different signals at the same frequency is fixed, in order to improve the energy concentration of time-frequency, parameters α and β are introduced. The basic formula is:

[0053]

[0054] In the formula, α and β are adjustment factors and are positive numbers.

[0055] Transformer acoustic signal denoising: Since transformer acoustic signals come from component vibration and environmental noise, it is necessary to purify the acoustic signals in order to improve identification accuracy. Wavelet denoising or CEEMDAN can be selected to filter out transformer noise signals and compare the denoising results. The CEEMDAN algorithm is based on EEMD and weakens the modal aliasing problem by adaptively adding white noise, thereby improving decomposition efficiency. It is suitable for processing nonlinear and non-stationary signals through time-frequency modal decomposition.

[0056] Wavelet denoising, the principle of wavelet (WT) denoising is to use the multi-resolution, decorrelation, basis selection flexibility and low entropy of WT to make the time-frequency window variable, and separate the noise and signal according to their different characteristics. The key to WT denoising is to retain the wavelet coefficients of the control signal, remove the wavelet coefficients of the noise, and inversely transform the remaining wavelet coefficients to obtain the noise-reduced signal. Common denoising methods are divided into three categories: WT modulus maximum denoising, WT correlation denoising, and WT threshold denoising. WT modulus maximum mainly removes the noise modulus maximum points and reconstructs the remaining maximum points based on the difference in the distribution law of the signal and noise at each scale to obtain the denoised signal. Correlation-based denoising mainly distinguishes the wavelet coefficient types according to the size of the correlation coefficient based on the strength of the correlation between the noise and the useful signal at each scale, removes the noise wavelet coefficients, and reconstructs the remaining coefficients to obtain the purified signal. Threshold denoising filters the wavelet coefficients by setting a threshold according to the distribution characteristics of the useful signal and noise energy amplitude after WT decomposition, and reconstructs the signal.

[0057] Transformer voiceprint feature connection, through existing research and experimental test verification, it is known that the short-time domain and frequency domain features have different meanings, different dimensions, and different contribution rates to voiceprint recognition. If the above feature vectors are directly used for subsequent analysis, the amount of calculation is large, and sometimes even the recognition rate is reduced. Therefore, the present invention designs an optimized and improved processing algorithm for MFCC feature vectors based on the characteristics of transformer noise signals. First, a weighted optimization method based on F-score is introduced into the MFCC feature vector group to improve the recognition rate of transformer noise signals by tuning the F-score of the feature vector. Secondly, the feature vector connected by short-time domain and frequency domain features is reduced in dimension through PCA (Principal Component Analysis). Under the premise of ensuring that information is not lost, the dimension of the feature vector is reduced as much as possible, thereby reducing the computational complexity.

[0058] Sound signal preprocessing includes framing and pre-emphasis, the purpose of which is to remove the interference of redundant noise and prepare for the next step of short-time feature extraction; the feature extraction part includes short-time time domain features and short-time frequency domain features, and short-time energy, short-time average zero-crossing rate, short-time energy entropy, and Mel cepstral coefficients are used to describe the framing features; the extracted features are connected into feature vectors as data for multi-feature fusion. The feature connection improves the processing algorithm through the optimization of MFCC feature vectors. First, a weighted optimization method based on F-score is introduced into the MFCC feature vector group, and the recognition rate of transformer noise signal is improved by tuning the F-score of the feature vector; secondly, the feature vector formed by the connection of short-time time domain and frequency domain features is reduced in dimension through PCA, and the dimension of the feature vector is reduced as much as possible under the premise of ensuring that information is not lost, thereby reducing the computational complexity; the F-score is an indicator in the field of metrology and statistics, which is used to examine the effectiveness of each dimensional component in multidimensional data; a component with a high F-score indicates that the component has a small variance in the same type of data and a large variance in different types of data, and is more efficient when used for data classification; the calculation formula for the F-score value F(k) of the component of the kth dimension of the feature vector is:

[0059]

[0060] Where N is the number of objects to be identified, u k (i) is the k-th dimension component of the feature vector of the i-th object, u k is the mean; x k (i) is the k-th dimension component of all samples of the i-th identified object, n i is the number of samples of the i-th identified object.

[0061] S2, equipment voiceprint recognition; the voiceprint recognition includes the extraction of transformer voiceprint features, modeling the target working condition, comparing and judging the noise signal to be tested, and identifying the type of fault working condition; it includes transformer DC bias magnetization, overload, overexcitation, partial discharge, mechanical looseness, transformer winding deformation, on-load tap changer fault and circuit breaker operating mechanism abnormality, and using deep neural network (Deep Neural Networks-DNN) as the baseline, improving the DNN model for transformer noise features, and after the DNN model training is completed, removing the last classification layer, and using the output result of the last hidden layer as the voiceprint template for comparison. Network structure, the network structure of transformer voiceprint feature analysis based on DNN, the network consists of four fully connected hidden layers, the input is the result of the connection of the frame feature vector extracted from the transformer noise signal, and the output is the One-hot vector corresponding to different working conditions, using the ReLu (RectifiedlLinear Units) function as the activation function, the classification uses the SoftMax function, and the objective function uses the cross entropy function.

[0062] Based on the model training of transfer learning, in the actual operation of the transformer, abnormal fault phenomena rarely occur, and it is difficult to collect a large amount of sample data under different abnormal working conditions. Therefore, in the actual training of the network model, due to the small number of negative samples, direct training is prone to overfitting. In addition, if the voiceprint model of different working conditions of a certain type of transformer that has been trained is used to monitor other types of transformers, it is necessary to retrain the new monitoring transformer. Based on the assumption that acoustic features have a large degree of commonality, the present invention uses the transfer learning method according to the idea of ​​parameter migration to train and learn the noise features of transformers of different models. First, the ESC-50 dataset is used for model pre-training, the FC1 layer and the FC2 layer are fixed, and the output is used as a feature extractor, and then the transformer voiceprint dataset is used to fine-tune the model. Specifically, after pre-training the network model with ESC-50 general sound data, the basic voiceprint network framework parameters are obtained, and the parameters of the FC1 and FC2 layers of the pre-trained model are retained. On this basis, the parameters of the FC3 and FC4 layers are further trained using the transformer-specific voiceprint data. After the second training, all parameters are fixed, and the last output of the hidden layer is used as the template of the monitoring target for comparison. For different working conditions, 5 to 15 data in the training data set are randomly selected, and the target template is obtained through the network model. The obtained target template is clustered using the k-means clustering algorithm; the target template output corresponding to the above 5 to 15 data is clustered with k being 2. After clustering, the cluster where most samples are located is selected as the target template p representing this type of working condition. Then, the noise signal to be detected is feature extracted and input into the trained network model to obtain the output O of the last hidden layer. Finally, the cosine similarity metric is used to compare the similarity between the output O and the target template p.

[0063] S3. Establish a voiceprint feature database; the voiceprint feature database includes the following contents: audio samples include audio files of the transformer in normal and abnormal working states. For each abnormal audio file, the database indicates its abnormal type and its audio features. The transformer working state includes the normal working state and abnormal working state of the transformer, as well as the corresponding abnormal state type. The transformer state parameters mainly include the transformer voltage, current, load, oil temperature, winding temperature and other parameters. On the one hand, they come from manually labeled data, and on the other hand, the prediction of state parameters is obtained by the output of the neural network. Potential attributes include audio features that are easy to distinguish when the transformer is in different working states, or potential feature vectors output by artificial intelligence algorithm models. These feature vectors are closely related to the working state of the transformer and can characterize certain properties of the transformer. Electric power experience knowledge and artificial intelligence algorithm models store these expert experience knowledge.

[0064] The voiceprint feature database established includes a transformer working state acoustic feature database, which contains the acoustic features of the transformer in normal state and different fault states, and uses it as sample data for artificial intelligence algorithm model and voiceprint intelligent diagnosis; the neural network model is trained using labeled acoustic data, and the transformer voltage, current, load, oil temperature, winding temperature and other state parameters are used as inputs of the neural network model, and the neural network model gives an output prediction on whether the transformer is abnormal; judgments are made through the experience and knowledge of power experts, and the experience and knowledge of experts are imparted to the neural network, and the neural network model is updated. After repeated iterative training, the neural network can also give corresponding predictions for unknown situations; the artificial intelligence algorithm is then used to predict the working state of the transformer and the state parameters of the transformer, and for each abnormal state, the database includes the distinguishing feature type of its abnormal state, the used power experience knowledge is stored in the database, and a database interface for multiple attribute predictions is reserved. The voiceprint feature database established includes transformer discharge fault diagnosis, using synchronous compression wavelet transform to enhance the energy aggregation on time-frequency and thus obtain clear time-frequency distribution characteristics; assuming that the signal f(t)∈L 2 (R), then the continuous wavelet transform WT(a,b) of the signal f(t) is:

[0065]

[0066] Where: a, b represent the scale and displacement factor respectively; represents the wavelet basis function; is the conjugate of ψ(·);

[0067] The frequency domain expression of WT(a,b) is:

[0068]

[0069] Where: They represent the Fourier transform of the time domain signals f(ω) and ψ(ω), respectively; W represents the frequency; represents an imaginary number;

[0070] Based on the above two equations, we can get the instantaneous frequency ω of the signal: f for:

[0071]

[0072] The wavelet coefficients are compressed and rearranged to obtain the synchronous compressed wavelet transform SWT f The formula is;

[0073] SWT f (b,ω)=∫WT f (a,b)a -3 / 2 δ[ω(a,b)-ω]da

[0074] Where δ(·) represents the Dirac function.

[0075] S4. Monitor the database; the monitoring of the database includes voiceprint monitoring, integrated noise signal storage, voiceprint feature extraction and voiceprint recognition algorithm, to achieve online monitoring and intelligent diagnosis of transformer sound, including voiceprint display, equipment and data management, real-time monitoring, active alarm, manual diagnosis and voiceprint sample library; use the voiceprint recognition algorithm to conduct a comprehensive and integrated diagnosis of the online monitored voiceprint data, including spectrum analysis, event detection, load identification and time-frequency domain analysis, to form a comprehensive diagnosis report on the equipment operation status; at the same time, it has the transformer equipment management function, which can manage and view the overall operating condition information of the transformer, and generate maintenance work orders based on the comprehensive voiceprint intelligent diagnosis results, manufacturer, voltage level, years of operation, winding type and cooling method. The voiceprint feature extraction includes transformer voiceprint feature extraction, reversible and multi-scale time-frequency analysis through short-time Fourier transform and Wigner distribution, and the overall signal is obtained on the time-frequency phase plane. The frequency-related adjustment factor is introduced in the window function to adjust the time-frequency resolution and extract high-frequency features and time-frequency information;

[0076] The voiceprint time-frequency spectrum is a two-dimensional energy spectrum formed by framing the sound signal, adding windows and performing discrete Fourier transformation, and stacking it according to the time dimension. The voiceprint time-frequency spectrum has the time-domain and frequency-domain characteristics of the sound signal, as well as intuitive image characteristics, which is convenient for subsequent deep learning processing. The core voiceprint signal is cut into samples with a duration of 1 s, and the overlapping segmentation method is used for framing. The frame length is set to 4096 frames, and the frame shift is set to 512 frames. In order to effectively reduce signal distortion, the window function selects the Hamming window, and its formula is:

[0077]

[0078] The framed and windowed samples are then subjected to short-time discrete Fourier transform to obtain a voiceprint time-frequency spectrum. The horizontal axis of the voiceprint time-frequency spectrum is time, and the vertical axis is frequency. The color of the graph represents the content of the current frequency component at the current moment, which is represented by the power spectral density.

[0079] Dataset comparison test, through the classification recognition algorithm SVM, NN and CNN, the recognition accuracy comparison test was carried out on the ESC-50 dataset using the traditional classification recognition algorithm and the transfer learning network; first, the 2000 audios of the ESC-50 dataset were divided into 5 stacks, each with 400 audios, covering all 50 classes, and ensuring that the audios of the same scene at different times only appear in one stack to avoid data pollution; secondly, the training and test cross-validation methods were used in the 5 stacks of data, and the training data volume and the test data volume were cross-folded 50%, 80% of each stack was selected as the training set, and the remaining 20% ​​was used as the test set, and then the average results of the 5 stacks were averaged to obtain the mean recognition accuracy. Combined with the characteristics of the transformer voiceprint data, the recognition accuracy of the transfer learning deep neural network algorithm was improved from 42.63% of the general CNN algorithm to 91.80%; the classifier recognition accuracy was obtained using the ESC-50 dataset.

[0080] A transformer wide-area voiceprint diagnostic monitoring device based on the spectrum domain, the diagnostic detection device is based on the online monitoring and analysis diagnosis system of the power transformer voiceprint feature, mainly including a high-sensitivity voiceprint acquisition device, a data switch, a management server, an algorithm server, and an intelligent diagnosis platform. Based on voiceprint acquisition, voiceprint feature database, voiceprint feature extraction and recognition, fault recognition model, etc., transformer online monitoring, status analysis, fault recognition, active warning and other applications are realized.

[0081] Including voiceprint collection device: using MEMS acoustic sensors and embedded high-speed processors, the voiceprint collection device is developed to realize the functions of sound signal collection, preprocessing, and transmission. Data switch: access to multiple voiceprint collection devices to realize high-speed transmission of voiceprint data. Management server: store the voiceprint data collected by the front end, run the voiceprint intelligent diagnosis platform, and be responsible for the call and interaction of the voiceprint algorithm model. Algorithm server: manage the voiceprint feature database, run the transformer fault voiceprint recognition model, extract the noise signal characteristics through transformer voiceprint recognition, establish voiceprint models of different fault noises, and realize the fault voiceprint recognition function of transformer abnormal analysis. Intelligent diagnosis platform: provide human-computer interaction visualization function, realize transformer attribute parameter management, audio information management, transformer analysis results, fault active alarm, diagnosis report and maintenance work order automatic generation and management functions. Acoustic sensor array: The sound sensor uses MEMS microphones, and is designed with replaceable microphone modules. It has an internal integrated temperature sensor. With the synchronous acquisition of sound signals, it can realize data correction based on ambient temperature in the back-end software, and perform semi-enclosed silicone rubber positioning, and cooperate with low-pressure foam mold materials to achieve double-layer directional noise reduction. Signal conditioning module: It consists of impedance matching, bandpass filter, gain amplification and other parts, which condition the output signal of the sensor and complete the matching, filtering and amplification of small signals. Signal processing module: It uses an embedded ARM platform and realizes digital sampling of sound signals through the ADC conversion module. The built-in main control program forms the sampled data into a natural beam, and sends the processing results through the network port to the voiceprint intelligent diagnosis platform and algorithm server for identification and diagnosis. At the same time, the signal processing module runs an embedded operating system to realize a multi-task processing mechanism, and provides external interfaces such as RS232 serial port and JTAG debugging. Power module: 12V DC power supply is adopted. Through the positive and negative pole reverse connection prevention design and electromagnetic compatibility (EMC) protection design, the safety and reliability of the collection device are improved. It has EMC4 level electromagnetic protection. At the same time, the power conversion unit can convert 12V into 5V, 3.3V and other DC voltages required by each module. The shell of the voiceprint collection device is made of lightweight aluminum alloy, integrated with waterproof and sound-permeable membrane, anti-aging rubber gasket and other waterproof and dustproof environmental adaptability measures, which can achieve IP67 level protection.

[0082] The present invention collects transformer noise signals by collecting voiceprints of power transformers, and develops relevant preprocessing and feature extraction algorithms. Based on the characteristics of transformer noise signals, a transfer learning artificial intelligence recognition network model is designed to realize intelligent diagnosis and recognition of transformer DC bias, winding deformation, overload and other working conditions. Through data set testing and substation deployment and application, the effectiveness and reliability of the voiceprint collection device and related algorithms are verified, which can meet the actual substation operation and maintenance needs. The method for establishing a transformer voiceprint feature database, with the on-site deployment, application and promotion of the system, collects voiceprint samples of abnormal transformer operation conditions, continuously optimizes and improves the algorithm model, enriches the database content, and gradually establishes a comprehensive and reliable voiceprint feature database. The accuracy rate of transformer abnormal working condition recognition can reach about 70%. Using the voiceprint recognition method, on the one hand, it can independently identify specific categories of defects, effectively solve the potential defects that cannot be identified by means such as images or online monitoring, and on the other hand, it can be used as an effective supplement to other monitoring methods such as partial discharge to improve the comprehensive monitoring capability.

[0083] The above are only preferred specific implementation modes of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solutions and concepts of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A transformer wide-area voiceprint diagnosis and monitoring method based on the spectrum domain, characterized in that: The following steps are involved: S1, device voiceprint feature extraction; S2, device voiceprint recognition; S3, establish a voiceprint feature database; S4. Monitor the database.

2. The transformer wide-area voiceprint diagnosis and monitoring method based on the spectrum domain according to claim 1 is characterized by: The device is a transformer; the feature extraction includes sound signal preprocessing, feature extraction and feature connection. The sound signal preprocessing includes framing and pre-emphasis, the purpose of which is to remove the interference of redundant noise and prepare for the next step of short-time feature extraction; the feature extraction part includes short-time time domain features and short-time frequency domain features, and short-time energy, short-time average zero-crossing rate, short-time energy entropy and Mel cepstral coefficients are used to describe the framing features; the extracted features are connected into feature vectors as multi-feature fusion data.

3. The transformer wide-area voiceprint diagnosis and monitoring method based on the spectrum domain according to claim 2 is characterized by: The feature connection is improved by optimizing the processing algorithm of the MFCC feature vector. First, a weighted optimization method based on F-score is introduced into the MFCC feature vector group, and the recognition rate of the transformer noise signal is improved by tuning the F-score of the feature vector; secondly, the feature vector formed by the connection of short-time time domain and frequency domain features is reduced in dimension by PCA, and the dimension of the feature vector is reduced while ensuring that information is not lost, thereby reducing the computational complexity; the F-score is an indicator in the field of metrology and statistics, which is used to examine the effectiveness of each dimensional component in multidimensional data; a component with a high F-score indicates that the component has a small variance in the same type of data and a large variance in different types of data, and is more efficient when used for data classification; the calculation formula for the F-score value F(k) of the component of the kth dimension of the feature vector is: Where N is the number of objects to be identified, u k (i) is the k-th dimension component of the feature vector of the i-th object, u k is the mean; x k (i) is the k-th dimension component of all samples of the i-th identified object, n i is the number of samples of the i-th identified object.

4. The transformer wide-area voiceprint diagnosis and monitoring method based on the spectrum domain according to claim 1 is characterized by: The voiceprint recognition includes extracting transformer voiceprint features, modeling the target working condition, comparing and judging the noise signal to be measured, and identifying the type of fault working condition; It includes the identification of transformer DC bias, overload, overexcitation, partial discharge, mechanical looseness, transformer winding deformation, on-load tap changer fault and circuit breaker operating mechanism abnormality. It uses deep neural network as a baseline to improve the DNN model for transformer noise characteristics. After the DNN model training is completed, the last classification layer is removed and the output result of the last hidden layer is used as the voiceprint template for comparison. For different working conditions, 5 to 15 data are randomly taken out from the training data set, and the target template is obtained through the network model. The k-means clustering algorithm is used to cluster the obtained target template; the target template output corresponding to the above 5 to 15 data is clustered with k=2. After clustering, the cluster where most samples are located is selected as the target template p representing this type of working condition. Then, the feature of the noise signal to be detected is extracted and input into the trained network model to obtain the output O of the last hidden layer. Finally, the cosine similarity measurement is used to compare the similarity between the output O and the target template p.

5. The transformer wide-area voiceprint diagnosis and monitoring method based on the spectrum domain according to claim 1 is characterized by: The voiceprint feature database established includes a transformer working state acoustic feature database, which contains the acoustic features of the transformer in normal state and different fault states, and uses it as sample data for artificial intelligence algorithm model and voiceprint intelligent diagnosis; the neural network model is trained using labeled acoustic data, and the transformer state parameters such as voltage, current, load, oil temperature, winding temperature, etc. are used as inputs of the neural network model, and the neural network model gives an output prediction on whether the transformer is abnormal; judgments are made through the experience and knowledge of power experts, and the experience and knowledge of the experts are imparted to the neural network, and the neural network model is updated. After repeated iterative training, the neural network can also give corresponding predictions for unknown situations; the artificial intelligence algorithm is then used to predict the working state of the transformer and the state parameters of the transformer, and for each abnormal state, the database includes the distinguishing feature type of its abnormal state, the used power experience knowledge is stored in the database, and a database interface for multiple attribute predictions is reserved.

6. The transformer wide-area voiceprint diagnosis and monitoring method based on the spectrum domain according to claim 1 is characterized by: The monitoring of the database includes voiceprint monitoring, integrated noise signal storage, voiceprint feature extraction and voiceprint recognition algorithm, to achieve online monitoring and intelligent diagnosis of transformer sound, including voiceprint display, equipment and data management, real-time monitoring, active alarm, manual diagnosis and voiceprint sample library; using the voiceprint recognition algorithm, a comprehensive integrated diagnosis is performed on the online monitored voiceprint data, including spectrum analysis, event detection, load identification and time-frequency domain analysis, to form a comprehensive diagnosis report on the equipment operation status; at the same time, it has the transformer equipment management function, which can manage and view the overall operating condition information of the transformer, and generate a maintenance work order based on the voiceprint intelligent diagnosis results, manufacturer, voltage level, operation life, winding type and cooling method.

7. The transformer wide-area voiceprint diagnosis and monitoring method based on the spectrum domain according to claim 5 is characterized by: The establishment of the voiceprint feature database includes transformer discharge fault diagnosis, using synchronous compression wavelet transform to enhance the energy aggregation in time-frequency and obtain clear time-frequency distribution characteristics; assuming that the signal f(t)∈L 2 (R), then the continuous wavelet transform WT(a,b) of the signal f(t) is: Where: a, b represent the scale and displacement factor respectively; represents the wavelet basis function; is the conjugate of ψ(·); The frequency domain expression of WT(a,b) is: Where: They represent the Fourier transform of the time domain signals f(ω) and ψ(ω), respectively; W represents the frequency; represents an imaginary number; Based on the above two equations, we can get the instantaneous frequency ω of the signal: f for: The wavelet coefficients are compressed and rearranged to obtain the synchronous compressed wavelet transform SWT f The formula is; SWT f (b,ω)=∫WT f (a,b)a -3 / 2 δ[ω(a,b)-ω]da Where δ(·) represents the Dirac function.

8. The transformer wide-area voiceprint diagnosis and monitoring method based on the spectrum domain according to claim 6 is characterized by: The voiceprint feature extraction includes transformer voiceprint feature extraction, reversible and multi-scale time-frequency analysis is performed through short-time Fourier transform and Wigner distribution, the overall signal is obtained on the time-frequency phase plane, and a frequency-related adjustment factor is introduced into the window function to adjust the time-frequency resolution, and extract high-frequency features and time-frequency information; The generalized S transform of the sound signal is used to obtain the time-frequency diagram. The texture features of the signal time-frequency diagram are extracted to represent the state information of the transformer. The S transform of the stationary signal x(t) is: Where: t is time, f is frequency, t is the Gaussian window position parameter, ω(f,τ-1) is the Gaussian window function; since the Gaussian window function of different signals at the same frequency is fixed, in order to improve the energy concentration of time-frequency, parameters α and β are introduced. The basic formula is: In the formula, α and β are adjustment factors and are positive numbers.

9. The transformer wide-area voiceprint diagnosis and monitoring method based on the spectrum domain according to claim 3 is characterized by: It includes a voiceprint time-frequency spectrum diagram, which is a two-dimensional energy spectrum diagram formed by framing the sound signal, adding windows and performing discrete Fourier transformation, and stacking it according to the time dimension; the voiceprint time-frequency spectrum has the time domain and frequency domain characteristics of the sound signal, as well as intuitive image characteristics, which is convenient for subsequent deep learning processing. The core voiceprint signal is cut into samples with a duration of 1s, and the overlapping segmentation method is used for framing. The frame length is set to 4096 frames and the frame shift is set to 512 frames; in order to effectively reduce signal distortion, the window function selects the Hamming window, and its formula is: The framed and windowed samples are then subjected to short-time discrete Fourier transform to obtain a voiceprint time-frequency spectrum. The horizontal axis of the voiceprint time-frequency spectrum is time, and the vertical axis is frequency. The color of the graph represents the content of the current frequency component at the current moment, which is represented by the power spectral density.

10. The transformer wide-area voiceprint diagnosis and monitoring method based on the spectrum domain according to claim 5 is characterized by: Including data set comparison test, through the classification recognition algorithm SVM, NN and CNN, the recognition accuracy comparison test is carried out on the ESC-50 data set using the traditional classification recognition algorithm and the transfer learning network; first, the 2000 audios of the ESC-50 data set are divided into 5 stacks, each with 400, covering all 50 classes, and ensuring that the audio of the same scene at different times only appears in one stack to avoid data pollution; secondly, the training and test cross-validation methods are used in the 5 stacks of data respectively, and the training data volume and the test data volume are cross-folded 50%, 80% of each stack is selected as the training set, and the remaining 20% ​​is used as the test set, and then, the average results of the 5 stacks are averaged to obtain the mean recognition accuracy. Combined with the characteristics of the transformer voiceprint data, the recognition accuracy of the transfer learning deep neural network algorithm is improved from 42.63% of the general CNN algorithm to 91.80%; the classifier recognition accuracy is obtained through the ESC-50 data set.

Images