CN113707176A - Transformer fault detection method based on acoustic signal and deep learning technology - Google Patents

Transformer fault detection method based on acoustic signal and deep learning technology Download PDF

Info

Publication number
CN113707176A
CN113707176A CN202111026413.3A CN202111026413A CN113707176A CN 113707176 A CN113707176 A CN 113707176A CN 202111026413 A CN202111026413 A CN 202111026413A CN 113707176 A CN113707176 A CN 113707176A
Authority
CN
China
Prior art keywords
transformer
data
fault detection
signal
sound
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202111026413.3A
Other languages
Chinese (zh)
Other versions
CN113707176B (en
Inventor
程汪刘
倪修峰
童旸
鲍文霞
曹成功
高志国
李飞
卢俊结
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Anhui Nanrui Jiyuan Power Grid Technology Co ltd
Anhui University
Tongling Power Supply Co of State Grid Anhui Electric Power Co Ltd
Original Assignee
Anhui Nanrui Jiyuan Power Grid Technology Co ltd
Anhui University
Tongling Power Supply Co of State Grid Anhui Electric Power Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Anhui Nanrui Jiyuan Power Grid Technology Co ltd, Anhui University, Tongling Power Supply Co of State Grid Anhui Electric Power Co Ltd filed Critical Anhui Nanrui Jiyuan Power Grid Technology Co ltd
Priority to CN202111026413.3A priority Critical patent/CN113707176B/en
Publication of CN113707176A publication Critical patent/CN113707176A/en
Application granted granted Critical
Publication of CN113707176B publication Critical patent/CN113707176B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/03—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the type of extracted parameters
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L15/00—Speech recognition
    • G10L15/06—Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice
    • G10L15/063—Training
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/27—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique
    • G10L25/30—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique using neural networks
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/48—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use
    • G10L25/51—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use for comparison or discrimination
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L15/00—Speech recognition
    • G10L15/06—Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice
    • G10L15/063—Training
    • G10L2015/0631—Creating reference templates; Clustering
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
    • Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
    • Y04S10/00—Systems supporting electrical power generation, transmission or distribution
    • Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
    • Y04S10/52—Outage or fault management, e.g. fault detection or location

Landscapes

  • Engineering & Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Human Computer Interaction (AREA)
  • Physics & Mathematics (AREA)
  • Acoustics & Sound (AREA)
  • Multimedia (AREA)
  • Signal Processing (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)

Abstract

The invention relates to a transformer fault detection method based on acoustic signals and a deep learning technology, and compared with the prior art, the transformer fault detection method overcomes the defect that accurate detection of transformer faults is difficult to perform by using voiceprint signals. The invention comprises the following steps: acquiring sound data of the power transformer; preprocessing acoustic signals in a training sample set; extracting sound characteristics of the sound signal data; constructing a transformer fault detection model; training a transformer fault detection model; acquiring and preprocessing acoustic signal data of a transformer to be detected; and obtaining a fault detection result of the transformer to be detected. The invention can detect the fault of the transformer based on the voiceprint signal.

Description

Transformer fault detection method based on acoustic signal and deep learning technology
Technical Field
The invention relates to the technical field of transformer fault detection, in particular to a transformer fault detection method based on an acoustic signal and deep learning technology.
Background
The transformer in the power grid has the characteristics of large usage amount, various types and specifications, and long service time, which causes the frequency of the transformer in the power grid system to be over high. According to statistics, about 5 transformers fail every 200 transformers with a run time of more than 4 years. Therefore, troubleshooting and repairing the transformer becomes an important procedure for the stable operation of the power grid.
For the fault elimination of the transformer, the traditional processing method needs to go to a place to be eliminated manually, and whether the transformer has a fault is diagnosed according to the sound emitted by the transformer by means of manual experience. This method not only needs a lot of time and effort, but also is interfered by human factors, and may cause fault diagnosis errors.
In recent years, deep learning is increasingly advanced, and the signal processing method has the characteristic of high efficiency. If the sound of the transformer can be processed in a deep learning mode, the characteristics of the sound emitted by the transformer can be automatically analyzed, and the characteristics are classified, so that the transformer with the fault can be quickly diagnosed, the labor cost can be greatly reduced, and the stability of a power grid can be maintained more quickly.
In the prior art, although there are some voiceprint detection methods, most of them use a gaussian mixture model and a hidden markov model as effective acoustic models to perform voiceprint detection, after the development of deep learning, the methods include convolutional Neural networks, self-encoders, Recurrent Neural Networks (RNNs), and more deep learning networks are applied in the field of voice detection. From the indexes of accuracy rate, accuracy and the like of sound detection, the voiceprint detection method based on deep learning is far better than the voiceprint detection method of the traditional acoustic model, but vibration data are mostly adopted in the voiceprint detection method based on deep learning in the prior art, and audio data are not utilized for fault detection.
Therefore, how to better apply the deep learning technology to the transformer voiceprint fault detection has become an urgent technical problem to be solved.
Disclosure of Invention
The invention aims to solve the defect that the transformer fault is difficult to accurately detect by using a voiceprint signal in the prior art, and provides a transformer fault detection method based on an acoustic signal and deep learning technology to solve the problem.
In order to achieve the purpose, the technical scheme of the invention is as follows:
a transformer fault detection method based on an acoustic signal and deep learning technology comprises the following steps:
11) acquiring sound data of the power transformer: acquiring sound data of the transformer by a voiceprint acquisition sensor in a field, marking the sound data into a normal class and a fault class, and defining the sound data as a training sample set;
12) preprocessing acoustic signals in a training sample set: carrying out denoising pretreatment on the collected power transformer sound data by using a pretreatment method of segmentation, framing, sound windowing and a self-adaptive filtering method; carrying out data enhancement on the acoustic signal through cutting, noise adding and tuning;
13) sound feature extraction of sound signal data: extracting sound features of the preprocessed power transformer sound data by adopting a Mel cepstrum coefficient, and extracting an MFCC coefficient;
14) constructing a transformer fault detection model: constructing a transformer fault detection model based on the characteristics of the double-gating convolution network model and the transformer acoustic signal;
15) training a transformer fault detection model: inputting the extracted MFCC coefficients into a transformer fault detection model for training;
16) acquiring and preprocessing acoustic signal data of the transformer to be detected: acquiring transformer acoustic signal data to be detected, performing denoising pretreatment, and extracting MFCC coefficients from the transformer acoustic signal data to be detected after the denoising pretreatment;
17) obtaining a fault detection result of the transformer to be detected: and inputting the MFCC coefficient into the trained transformer fault detection model to obtain a transformer fault detection result.
The pre-processing of acoustic signals within the training sample set comprises the steps of:
21) carrying out segmentation operation on the collected sound data s (t) of the power transformer:
the obtained power transformer sound data is segmented,
is divided into s (t) ═ s1(t),s2(t),...,sq(t),...sr(t) }, calculating the total length L of the voiceprint data, and the calculation formula is as follows:
L＝time×fSSample＝r×rL，
wherein fs issampleFor the sampling frequency of the sound, time is the sampling time, r is the number of segments, rLIs the segment length;
22) for segmented transformer sound data sq(t) performing framing:
setting the voiceprint frame length of the transformer to be 500ms, and performing framing processing to obtain
sq(t)＝{sq1(t),sq2(t),...,sqp(t),...sqLength(t)}，
Wherein, the Length of each frame is set to 500ms, and each segment is divided into a Length frame;
23) windowing the sound of the transformer after framing:
windowing the frame data by end point smoothing, and windowing the frame by using a Hamming window, wherein a function w (t) of the Hamming window is represented as follows:
Figure BDA0003243486160000031
wherein M is the frame length and t is the time;
obtaining a time domain signal f for each frameqp(t), the expression of which is as follows:
fqp(t)＝sqp(t)*w(n)，
wherein f isqp(t) is the time domain signal of the p frame of the q-th segment signal, w (n) is the windowing function, sqp(t) is a signal value of a p frame of the q-th segment signal;
24) denoising the windowed sound by using an adaptive filter:
to fqp(t) sampling to obtain Xi(n) digital signal sequence, and setting initial filter values as follows:
Figure BDA0003243486160000032
wherein, W (k) is the optimal weight coefficient, mu is the convergence factor, and lambda is the maximum eigenvalue of the correlation matrix;
calculating an actual output estimate of the filter output:
y(k)＝WT(k)Xi(n)，
where y (k) is the actual output estimate, WT(k) For transposition of optimal weight coefficients, Xi(n) is an input signal sequence;
e (k) is an error signal, for calculating the error:
e(k)＝d(k)-y(k)，
where d (k) is the desired output value, the filter coefficients at time k +1 are updated:
W(k+1)＝W(k)+μe(k)X(k)，
wherein, W (k +1) is the optimal weight coefficient at the time of k +1, W (k) is the optimal weight coefficient at the time of k, e (k) is the error at the time of k, and x (k) is the input sequence at the time of k;
continuously iterating and solving by using a steepest descent method to minimize an error signal, and obtaining output y (k) after adaptive filtering and denoising;
25) and (k) performing data enhancement on the acoustic signal through cutting, noise adding and tuning processing on the output y (k) of the adaptive filtering and denoising.
The sound feature extraction of the sound signal data comprises the following steps:
31) and (c) performing inverse transformation on the y (k) subjected to data enhancement to generate s (t), and performing pre-emphasis operation on s (t), wherein the calculation formula is as follows:
y(z)＝s(z)H(z)，
where y (Z) is the output of the pre-emphasis, s (Z) is the Z-domain representation of the signal s (t), h (Z) is the transfer function of the high-pass filter:
Figure BDA0003243486160000041
the high-pass filter is in the z-domain, where μ has a value between 0.9 and 1.0;
32) for pre-emphasized transformer sound data sq(t) performing framing:
setting the length of a transformer voiceprint frame as 30ms, performing framing processing on the transformer voice data, and setting the processed structure as each frame for 30 ms;
33) windowing the framed sound data: windowing is carried out on the frames by using a hamming window, and if the output obtained by the previous two steps of preprocessing is S (n) and the window function is W (n), the calculation formula is as follows:
S'(n)＝S(n)×W(n)；
wherein the window function W (n) is:
Figure BDA0003243486160000051
in the formula, N represents N time, the value range of N is more than or equal to 0 and less than or equal to N-1, N is the number of sampling points, different Hamming windows can be obtained according to different a values, N represents a sequence independent variable, and a is a set constant;
34) and performing fast Fourier transform on the windowed data:
performing FFT on each frame signal to obtain frequency domain signal x (k):
Figure BDA0003243486160000052
where x (k) represents the frequency domain output, x (N) represents the time domain input, and N is the number of sampling points;
35) carrying out Mel filtering on the data after the fast Fourier transform, wherein the conversion formula of the Mel filtering is as follows:
Figure BDA0003243486160000053
wherein f represents a physical frequency, mel (f) represents a mel frequency; filtering the obtained Mel frequency through M Mel-scale triangular filter banks, wherein the formula of the filter bank is as follows:
Figure BDA0003243486160000054
in this formula:
Figure BDA0003243486160000055
and then calculating the logarithmic energy output by each filter bank:
Figure BDA0003243486160000056
wherein E (m) is logarithmic energy, Hm(k) Is a filter bank;
36) performing cepstrum analysis on the Mel filtered data to extract MFCC coefficients C (n), wherein the MFCC coefficients C (n) are extracted by performing Discrete Cosine Transform (DCT):
Figure BDA0003243486160000061
where M represents the number of filters.
The method for constructing the transformer fault detection model comprises the following steps:
41) setting a transformer fault detection model based on a double-gate convolution neural network, wherein the transformer fault detection model comprises two convolution gate control layers, two pooling layers, a full-connection layer and an output layer;
42) setting a convolution gating layer: the convolution gating layer extracts features through convolution operation of input data and convolution kernels to obtain a feature map, and the depth of the feature map depends on the number of set convolution kernels;
assume that the input is X ∈ RA×BWhere a and B represent the length and width of the input data, respectively, the convolution operation is defined as:
Figure BDA0003243486160000062
wherein:
Figure BDA0003243486160000063
represents the jth characteristic plan of the first convolutional layer,
Figure BDA0003243486160000064
represents the ith plane characteristic diagram of the (l-1) th convolution layer, M represents the number of the plane characteristic diagrams,
Figure BDA0003243486160000065
which represents the convolution kernel or kernels, is,
Figure BDA0003243486160000066
which represents the offset of the bias voltage,
Figure BDA0003243486160000067
and
Figure BDA0003243486160000068
the specific parameters are determined by training optimization, f represents a excitation layer function, common excitation functions include ReLu and sigmoid functions, and the expression is as follows:
Figure BDA0003243486160000069
the output of the gating is represented as:
Figure BDA00032434861600000610
wherein: h (X) denotes the output of Gated CNN, W and V denote different convolution kernels, b and c denote different biases,
Figure BDA00032434861600000611
expressing the corresponding multiplication of elements of the matrix, expressing sigma to represent sigmoid gating function, and expressing X to represent the output of the previous layer;
43) designing a pooling layer: setting a pooling layer by adopting a maximum pooling or average pooling method;
44) and building a full connection layer and adopting a softmax classifier as an output layer for final classification output.
The training of the transformer fault detection model comprises the following steps:
51) inputting the MFCC coefficients extracted from the training sample set into a transformer fault detection model;
52) and continuously iterating forward propagation and backward propagation of the MFCC coefficients in the transformer fault detection model, and performing parameter optimization to obtain the trained transformer fault detection model.
Advantageous effects
Compared with the prior art, the transformer fault detection method based on the acoustic signal and the deep learning technology can detect the fault of the transformer based on the voiceprint signal (non-vibration signal); the method comprises the steps of constructing a transformer working condition sound data set by collecting transformer sounds and classifying and marking the transformer sounds as 'normal' and 'fault'; for the interference of environmental noise, preprocessing a transformer working condition sound data set for denoising; in consideration of the practicability of the training model, in order to enhance the robustness of a subsequently built neural network, a plurality of lines of data enhancement operations are carried out on data; then, extracting the sound data set of the transformer through MFCC characteristics and sending the sound data set into a network for training; when the basic CNN network is used for fault detection with complex conditions, the phenomena of low training efficiency and overfitting exist due to the fact that the audio data volume is small or the audio duration is long, and the detection accuracy is low.
Aiming at the transformer fault detection effect, the invention designs a double-gate convolution neural network transformer sound detection model, and provides a favorable method basis and detection model realization for realizing transformer fault on-line monitoring.
Drawings
FIG. 1 is a sequence diagram of the method of the present invention;
FIG. 2 is a logic flow diagram of a method implementation of the present invention;
fig. 3 is a structural diagram of a double-gated convolutional neural network according to the present invention.
Detailed Description
So that the manner in which the above recited features of the present invention can be understood and readily understood, a more particular description of the invention, briefly summarized above, may be had by reference to embodiments, some of which are illustrated in the appended drawings, wherein:
as shown in fig. 1 and fig. 2, the method for detecting a transformer fault based on an acoustic signal and a deep learning technique according to the present invention includes the following steps:
firstly, acquiring sound data of the power transformer. The sound data of the transformer is acquired through the field acquisition of the voiceprint acquisition sensor, and is divided into a normal class and a fault class through marking, and the normal class and the fault class are defined as a training sample set. In a laboratory link, the acquired sound data of the transformer can be divided into two types, namely a normal working state and a fault working state, an experiment database is established, and a training set and a test set of the experiment database are divided according to a certain proportion.
And secondly, pre-processing the acoustic signals in the training sample set. Carrying out denoising pretreatment on the collected power transformer sound data by using a pretreatment method of segmentation, framing, sound windowing and a self-adaptive filtering method; and then carrying out data enhancement on the acoustic signal through cutting, noise adding and tuning processing. The method comprises the following specific steps:
(1) carrying out segmentation operation on the collected sound data s (t) of the power transformer:
the obtained power transformer sound data is segmented,
is divided into s (t) ═ s1(t),s2(t),...,sq(t),...sr(t) }, calculating the total length L of the voiceprint data, and the calculation formula is as follows:
L＝time×fSSample＝r×rL，
wherein fs issampleFor the sampling frequency of the sound, time is the sampling time, r is the number of segments, rLIs the segment length.
(2) For segmented transformer sound data sq(t) performing framing:
setting the voiceprint frame length of the transformer to be 500ms, and performing framing processing to obtain
sq(t)＝{sq1(t),sq2(t),...,sqp(t),...sqLength(t)}，
Wherein, the Length of each frame is set as 500ms, and each segment is divided into a Length frame.
(3) Windowing the sound of the transformer after framing:
windowing the frame data by end point smoothing, and windowing the frame by using a Hamming window, wherein a function w (t) of the Hamming window is represented as follows:
Figure BDA0003243486160000081
wherein M is the frame length and t is the time;
obtaining a time domain for each frameSignal fqp(t), the expression of which is as follows:
fqp(t)＝sqp(t)*w(n)，
wherein f isqp(t) is the time domain signal of the p frame of the q-th segment signal, w (n) is the windowing function, sqp(t) is a signal value of a p frame of the q-th segment signal.
(4) Denoising the windowed sound by using an adaptive filter:
to fqp(t) sampling to obtain Xi(n) digital signal sequence, and setting initial filter values as follows:
Figure BDA0003243486160000091
wherein, W (k) is the optimal weight coefficient, mu is the convergence factor, and lambda is the maximum eigenvalue of the correlation matrix;
calculating an actual output estimate of the filter output:
y(k)＝WT(k)Xi(n)，
where y (k) is the actual output estimate, WT(k) For transposition of optimal weight coefficients, Xi(n) is an input signal sequence;
e (k) is an error signal, for calculating the error:
e(k)＝d(k)-y(k)，
where d (k) is the desired output value, the filter coefficients at time k +1 are updated:
W(k+1)＝W(k)+μe(k)X(k)，
wherein, W (k +1) is the optimal weight coefficient at the time of k +1, W (k) is the optimal weight coefficient at the time of k, e (k) is the error at the time of k, and x (k) is the input sequence at the time of k;
and continuously iterating and solving by using a steepest descent method to minimize the error signal and obtain the output y (k) after the adaptive filtering and denoising.
(5) And (k) performing data enhancement on the acoustic signal through cutting, noise adding and tuning processing on the output y (k) of the adaptive filtering and denoising.
Thirdly, extracting the sound characteristics of the sound signal data: as with speech recognition, there are two methods of application for audio data during fault detection, the first is to directly input a one-dimensional signal for processing, and the second is to convert the audio data into an image to extract audio features for subsequent processing, where one of the most commonly used features is the Mel-scale Frequency Coefficients, MFCC for short. The sound feature extraction is carried out on the preprocessed power transformer sound data by adopting a Maillard cepstrum coefficient, an MFCC coefficient is extracted, and the extracted MFCC coefficient is sent to a network at the later stage to generate a feature map. The method comprises the following specific steps:
(1) and (c) performing inverse transformation on the y (k) subjected to data enhancement to generate s (t), and performing pre-emphasis operation on s (t), wherein the calculation formula is as follows:
y(z)＝s(z)H(z)，
where y (Z) is the output of the pre-emphasis, s (Z) is the Z-domain representation of the signal s (t), h (Z) is the transfer function of the high-pass filter:
Figure BDA0003243486160000101
the high pass filter is in the z-domain, where μ has a value between 0.9 and 1.0. The sound data passes through a high-pass filter, so that the energy of a high-frequency part is improved, the energy of the high-frequency part is kept consistent with that of a low-frequency part as much as possible, and the frequency spectrum can be obtained by the same signal-to-noise ratio in the whole frequency band.
(2) For pre-emphasized transformer sound data sq(t) performing framing:
the length of the transformer voiceprint frame is set to be 30ms, the framing processing of the transformer voice data is carried out, and the structure after the processing is 30ms of each frame.
(3) Windowing the framed sound data. The purpose of the windowing function is to increase the continuity of the left and right ends of each frame of sound by multiplying each frame of sound by the windowing function.
Windowing is carried out on the frames by using a hamming window, and if the output obtained by the previous two steps of preprocessing is S (n) and the window function is W (n), the calculation formula is as follows:
S'(n)＝S(n)×W(n)；
wherein the window function W (n) is:
Figure BDA0003243486160000111
in the formula, N represents N time, the value range of N is more than or equal to 0 and less than or equal to N-1, N is the number of sampling points, different Hamming windows can be obtained according to different a values, N represents a sequence independent variable, and a is a set constant.
(4) And carrying out fast Fourier transform on the windowed data. Fast Fourier Transform (FFT) is a common way to transform a time-domain signal into the frequency domain.
Performing FFT on each frame signal to obtain frequency domain signal x (k):
Figure BDA0003243486160000112
where x (k) represents the frequency domain output, x (N) represents the time domain input, and N is the number of sample points.
(5) Carrying out Mel filtering on the data after the fast Fourier transform, wherein the conversion formula of the Mel filtering is as follows:
Figure BDA0003243486160000113
wherein f represents a physical frequency, mel (f) represents a mel frequency; filtering the obtained Mel frequency through M Mel-scale triangular filter banks, wherein the formula of the filter bank is as follows:
Figure BDA0003243486160000114
in this formula:
Figure BDA0003243486160000115
and then calculating the logarithmic energy output by each filter bank:
Figure BDA0003243486160000116
wherein E (m) is logarithmic energy, Hm(k) Is a filter bank.
(6) Performing cepstrum analysis on the data after the Mel filtering to extract an MFCC coefficient C (n), wherein the MFCC coefficient is extracted by performing Discrete Cosine Transform (DCT):
Figure BDA0003243486160000121
where M represents the number of filters. The larger M, the more feature values are extracted per frame, the more information amount, and thus the more accurately the signal can be described.
Fourthly, constructing a transformer fault detection model: and constructing a transformer fault detection model based on the characteristics of the double-gating convolution network model and the transformer acoustic signal.
The transformer acoustic signal can reflect the working condition of the transformer, and the traditional manual troubleshooting of the transformer needs to be carried out on-site acoustic detection manually with experience. The mode for detecting the transformer fault has certain hysteresis, namely the transformer at a certain position needs to be checked one by one through experienced manual intervention after the transformer is in fault under normal conditions. Manpower with abundant transformer troubleshooting experience needs years of actual operation and cultivation, and a large amount of time and cost are consumed. The transformer on-line monitoring system can quickly and conveniently monitor faults by using a deep learning method, the deep learning method is similar to the traditional fault detection method at present, most of the deep learning method still uses vibration data, and the research on fault detection by only using audio data is less. Compared with vibration data, the audio data is easier to obtain and low in cost, and is more suitable for large-scale practical application. Therefore, the invention provides a method for detecting the fault by using the audio data of the transformer based on the double-gate convolution neural network, and designs a model for detecting and identifying the fault of the transformer based on the acoustic signal.
Because the transformer is usually placed outdoors, the acoustic signal of the transformer often has great noise, and when the energy of the noise is too large, the real acoustic signal sent by the local machine of the transformer is submerged, so that the trained detection model has low effect. The invention carries out denoising pretreatment on the collected acoustic signals in the previous steps, reduces the noise energy by adopting a self-adaptive filtering method, and reserves the local acoustic signals of the transformer. On the premise of removing noise, the sound data of the transformer adopts a plurality of data enhancement modes to expand the data set samples, so that a sample set with more complex conditions is input when the network is trained, and the robustness of a transformer detection model is enhanced.
For the input processing of the transformer audio data, if the audio is directly input into the network training in a one-dimensional mode, the memory required by the training is too large, and the physical realization is difficult. If the voice data in the one-dimensional mode is used as a time domain signal and is directly input into the network, the network can only extract a small part of feature information, which is not beneficial to the training of the network model. Therefore, before data is input into the network, an MFCC coefficient feature extraction mode is adopted to convert a time domain signal into a time frequency signal, and the time frequency signal is input into the network in a two-dimensional signal mode. The MFCC coefficients are time-frequency signals and contain more characteristic information, so that the network can extract more effective characteristics, and the trained model is better.
The general convolutional neural network is affected by the audio format, so that the training efficiency is low and the overfitting phenomenon often occurs, and the fault detection accuracy is low. Different from general convolutional neural network, the invention adopts the network structure idea of gate-controlled convolution and adds a gate-controlled switch to the output of the convolutional network. The gate control switch is used for realizing a buffer control effect when information is transmitted in training, so that the overfitting phenomenon is solved and the training efficiency is improved. Aiming at the characteristics of the sound of the transformer, the invention adds a convolution gating layer on the basis of a single gating convolution network and designs a double gating convolution neural network. The double-gated convolutional neural network adopts two convolutional gating layers, so that deep level characteristics of the sound of the transformer local machine can be effectively extracted, the training effect is improved, and a training model with a good effect is obtained.
The method comprises the following specific steps:
(1) as shown in fig. 3, a transformer fault detection model is set based on a double-gated convolutional neural network, which includes two convolutional gating layers, two pooling layers, a full-link layer, and an output layer.
(2) Setting a convolution gating layer: the convolution gating layer extracts features through convolution operation of input data and convolution kernels to obtain a feature map, and the depth of the feature map depends on the number of set convolution kernels;
assume that the input is X ∈ RA×BWhere a and B represent the length and width of the input data, respectively, the convolution operation is defined as:
Figure BDA0003243486160000131
wherein:
Figure BDA0003243486160000132
represents the jth characteristic plan of the first convolutional layer,
Figure BDA0003243486160000133
represents the ith plane characteristic diagram of the (l-1) th convolution layer, M represents the number of the plane characteristic diagrams,
Figure BDA0003243486160000134
which represents the convolution kernel or kernels, is,
Figure BDA0003243486160000135
which represents the offset of the bias voltage,
Figure BDA0003243486160000136
and
Figure BDA0003243486160000137
is determined by training optimization, f represents the excitation layer function, the excitation in common useThe functions include ReLu and sigmoid functions, and the expression is as follows:
Figure BDA0003243486160000138
the output of the gating is represented as:
Figure BDA0003243486160000139
wherein: h (X) denotes the output of Gated CNN, W and V denote different convolution kernels, b and c denote different biases,
Figure BDA0003243486160000141
and elements representing the matrix are multiplied correspondingly, sigma represents a sigmoid gating function, and X represents the output of the previous layer. Through the gate control structure, the gradient dispersion of the model is reduced, the training of the model is promoted, the structure of the model is simplified, and meanwhile, the nonlinear representation capability of the model is reserved;
(3) designing a pooling layer: when the convolution kernel is small, the feature map after convolution layer is still large, at this time, the dimensionality reduction needs to be performed through pooling operation, the number of features is reduced, the robustness of the network to the input features is enhanced, and the pooling layer is set by adopting a maximum pooling or average pooling method.
(4) And building a full connection layer and adopting a softmax classifier as an output layer for final classification output.
And fifthly, training a transformer fault detection model: and inputting the extracted MFCC coefficients into a transformer fault detection model for training. The method comprises the following specific steps:
(1) inputting the MFCC coefficients extracted from the training sample set into a transformer fault detection model;
(2) and continuously iterating forward propagation and backward propagation of the MFCC coefficients in the transformer fault detection model, and performing parameter optimization to obtain the trained transformer fault detection model.
Sixthly, acquiring and preprocessing the acoustic signal data of the transformer to be detected: acquiring transformer acoustic signal data to be detected, performing denoising pretreatment, and extracting MFCC coefficients from the transformer acoustic signal data to be detected after the denoising pretreatment.
Seventhly, obtaining a fault detection result of the transformer to be detected: and inputting the MFCC coefficient into the trained transformer fault detection model to obtain a transformer fault detection result.
The foregoing shows and describes the general principles, essential features, and advantages of the invention. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, which are merely illustrative of the principles of the invention, but that various changes and modifications may be made without departing from the spirit and scope of the invention, which fall within the scope of the invention as claimed. The scope of the invention is defined by the appended claims and equivalents thereof.

Claims (5)

1. A transformer fault detection method based on an acoustic signal and deep learning technology is characterized by comprising the following steps:
11) acquiring sound data of the power transformer: acquiring sound data of the transformer by a voiceprint acquisition sensor in a field, marking the sound data into a normal class and a fault class, and defining the sound data as a training sample set;
12) preprocessing acoustic signals in a training sample set: carrying out denoising pretreatment on the collected power transformer sound data by using a pretreatment method of segmentation, framing, sound windowing and a self-adaptive filtering method; carrying out data enhancement on the acoustic signal through cutting, noise adding and tuning;
13) sound feature extraction of sound signal data: extracting sound features of the preprocessed power transformer sound data by adopting a Mel cepstrum coefficient, and extracting an MFCC coefficient;
14) constructing a transformer fault detection model: constructing a transformer fault detection model based on the characteristics of the double-gating convolution network model and the transformer acoustic signal;
15) training a transformer fault detection model: inputting the extracted MFCC coefficients into a transformer fault detection model for training;
16) acquiring and preprocessing acoustic signal data of the transformer to be detected: acquiring transformer acoustic signal data to be detected, performing denoising pretreatment, and extracting MFCC coefficients from the transformer acoustic signal data to be detected after the denoising pretreatment;
17) obtaining a fault detection result of the transformer to be detected: and inputting the MFCC coefficient into the trained transformer fault detection model to obtain a transformer fault detection result.
2. The method for detecting the fault of the transformer based on the acoustic signal and the deep learning technology as claimed in claim 1, wherein the pre-processing of the acoustic signal in the training sample set comprises the following steps:
21) carrying out segmentation operation on the collected sound data s (t) of the power transformer:
the obtained power transformer sound data is segmented,
is divided into s (t) ═ s1(t),s2(t),...,sq(t),...sr(t) }, calculating the total length L of the voiceprint data, and the calculation formula is as follows:
L＝time×fSSample＝r×rL，
wherein fs issampleFor the sampling frequency of the sound, time is the sampling time, r is the number of segments, rLIs the segment length;
22) for segmented transformer sound data sq(t) performing framing:
setting the voiceprint frame length of the transformer to be 500ms, and performing framing processing to obtain
sq(t)＝{sq1(t),sq2(t),...,sqp(t),...sqLength(t)}，
Wherein, the Length of each frame is set to 500ms, and each segment is divided into a Length frame;
23) windowing the sound of the transformer after framing:
windowing the frame data by end point smoothing, and windowing the frame by using a Hamming window, wherein a function w (t) of the Hamming window is represented as follows:
Figure FDA0003243486150000021
wherein M is the frame length and t is the time;
obtaining a time domain signal f for each frameqp(t), the expression of which is as follows:
fqp(t)＝sqp(t)*w(n)，
wherein f isqp(t) is the time domain signal of the p frame of the q-th segment signal, w (n) is the windowing function, sqp(t) is a signal value of a p frame of the q-th segment signal;
24) denoising the windowed sound by using an adaptive filter:
to fqp(t) sampling to obtain Xi(n) digital signal sequence, and setting initial filter values as follows:
Figure FDA0003243486150000022
wherein, W (k) is the optimal weight coefficient, mu is the convergence factor, and lambda is the maximum eigenvalue of the correlation matrix;
calculating an actual output estimate of the filter output:
y(k)＝WT(k)Xi(n)，
where y (k) is the actual output estimate, WT(k) For transposition of optimal weight coefficients, Xi(n) is an input signal sequence;
e (k) is an error signal, for calculating the error:
e(k)＝d(k)-y(k)，
where d (k) is the desired output value, the filter coefficients at time k +1 are updated:
W(k+1)＝W(k)+μe(k)X(k)，
wherein, W (k +1) is the optimal weight coefficient at the time of k +1, W (k) is the optimal weight coefficient at the time of k, e (k) is the error at the time of k, and x (k) is the input sequence at the time of k;
continuously iterating and solving by using a steepest descent method to minimize an error signal, and obtaining output y (k) after adaptive filtering and denoising;
25) and (k) performing data enhancement on the acoustic signal through cutting, noise adding and tuning processing on the output y (k) of the adaptive filtering and denoising.
3. The transformer fault detection method based on the acoustic signal and deep learning technology as claimed in claim 1, wherein the acoustic feature extraction of the acoustic signal data comprises the following steps:
31) and (c) performing inverse transformation on the y (k) subjected to data enhancement to generate s (t), and performing pre-emphasis operation on s (t), wherein the calculation formula is as follows:
y(z)＝s(z)H(z)，
where y (Z) is the output of the pre-emphasis, s (Z) is the Z-domain representation of the signal s (t), h (Z) is the transfer function of the high-pass filter:
H(Z)＝1-|μz-1，
the high-pass filter is in the z-domain, where μ has a value between 0.9 and 1.0;
32) for pre-emphasized transformer sound data sq(t) performing framing:
setting the length of a transformer voiceprint frame as 30ms, performing framing processing on the transformer voice data, and setting the processed structure as each frame for 30 ms;
33) windowing the framed sound data: windowing is carried out on the frames by using a hamming window, and if the output obtained by the previous two steps of preprocessing is S (n) and the window function is W (n), the calculation formula is as follows:
S'(n)＝S(n)×W(n)；
wherein the window function W (n) is:
Figure FDA0003243486150000041
in the formula, N represents N time, the value range of N is more than or equal to 0 and less than or equal to N-1, N is the number of sampling points, different Hamming windows can be obtained according to different a values, N represents a sequence independent variable, and a is a set constant;
34) and performing fast Fourier transform on the windowed data:
performing FFT on each frame signal to obtain a frequency domain signal w (k):
Figure FDA0003243486150000042
where x (k) represents the frequency domain output, x (N) represents the time domain input, and N is the number of sampling points;
35) carrying out Mel filtering on the data after the fast Fourier transform, wherein the conversion formula of the Mel filtering is as follows:
Figure FDA0003243486150000043
wherein f represents a physical frequency, mel (f) represents a mel frequency; filtering the obtained Mel frequency through M Mel-scale triangular filter banks, wherein the formula of the filter bank is as follows:
Figure FDA0003243486150000044
in this formula:
Figure FDA0003243486150000045
and then calculating the logarithmic energy output by each filter bank:
Figure FDA0003243486150000046
wherein E (m) is logarithmic energy, Hm(k) Is a filter bank;
36) performing cepstrum analysis on the Mel filtered data to extract MFCC coefficients C (n), wherein the MFCC coefficients C (n) are extracted by performing Discrete Cosine Transform (DCT):
Figure FDA0003243486150000047
where M represents the number of filters.
4. The method for detecting the transformer fault based on the acoustic signal and the deep learning technology as claimed in claim 1, wherein the step of constructing the transformer fault detection model comprises the following steps:
41) setting a transformer fault detection model based on a double-gate convolution neural network, wherein the transformer fault detection model comprises two convolution gate control layers, two pooling layers, a full-connection layer and an output layer;
42) setting a convolution gating layer: the convolution gating layer extracts features through convolution operation of input data and convolution kernels to obtain a feature map, and the depth of the feature map depends on the number of set convolution kernels;
assume that the input is X ∈ RA×BWhere a and B represent the length and width of the input data, respectively, the convolution operation is defined as:
Figure FDA0003243486150000051
wherein:
Figure FDA0003243486150000052
represents the jth characteristic plan of the first convolutional layer,
Figure FDA0003243486150000053
represents the ith plane characteristic diagram of the (l-1) th convolution layer, M represents the number of the plane characteristic diagrams,
Figure FDA0003243486150000054
which represents the convolution kernel or kernels, is,
Figure FDA0003243486150000055
which represents the offset of the bias voltage,
Figure FDA0003243486150000056
and
Figure FDA0003243486150000057
the specific parameters are determined by training optimization, f represents a excitation layer function, the excitation function is a sigmoid function, and the expression of the excitation function is as follows:
Figure FDA0003243486150000058
the output of the gating is represented as:
Figure FDA0003243486150000059
wherein: h (X) denotes the output of Gated CNN, W and V denote different convolution kernels, b and c denote different biases,
Figure FDA00032434861500000510
expressing the corresponding multiplication of elements of the matrix, expressing sigma to represent sigmoid gating function, and expressing X to represent the output of the previous layer;
43) designing a pooling layer: setting a pooling layer by adopting a maximum pooling or average pooling method;
44) and building a full connection layer and adopting a softmax classifier as an output layer for final classification output.
5. The method for detecting the transformer fault based on the acoustic signal and the deep learning technology as claimed in claim 1, wherein the training of the transformer fault detection model comprises the following steps:
51) inputting the MFCC coefficients extracted from the training sample set into a transformer fault detection model;
52) and continuously iterating forward propagation and backward propagation of the MFCC coefficients in the transformer fault detection model, and performing parameter optimization to obtain the trained transformer fault detection model.
CN202111026413.3A 2021-09-02 2021-09-02 Transformer fault detection method based on acoustic signal and deep learning technology Active CN113707176B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111026413.3A CN113707176B (en) 2021-09-02 2021-09-02 Transformer fault detection method based on acoustic signal and deep learning technology

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111026413.3A CN113707176B (en) 2021-09-02 2021-09-02 Transformer fault detection method based on acoustic signal and deep learning technology

Publications (2)

Publication Number Publication Date
CN113707176A true CN113707176A (en) 2021-11-26
CN113707176B CN113707176B (en) 2022-09-09

Family

ID=78657439

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111026413.3A Active CN113707176B (en) 2021-09-02 2021-09-02 Transformer fault detection method based on acoustic signal and deep learning technology

Country Status (1)

Country Link
CN (1) CN113707176B (en)

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113466616A (en) * 2021-06-22 2021-10-01 海南电网有限责任公司乐东供电局 Method and device for quickly positioning cable fault point
CN114155878A (en) * 2021-12-03 2022-03-08 北京中科智易科技有限公司 Artificial intelligence detection system, method and computer program
CN114543983A (en) * 2022-03-29 2022-05-27 阿里云计算有限公司 Vibration signal identification method and device
CN114638256A (en) * 2022-02-22 2022-06-17 合肥华威自动化有限公司 Transformer fault detection method and system based on sound wave signals and attention network
CN115358110A (en) * 2022-07-25 2022-11-18 国网江苏省电力有限公司淮安供电分公司 Transformer fault detection system based on acoustic sensor array
CN115392293A (en) * 2022-08-01 2022-11-25 中国南方电网有限责任公司超高压输电公司昆明局 Transformer fault monitoring method and device, computer equipment and storage medium
CN115424635A (en) * 2022-11-03 2022-12-02 南京凯盛国际工程有限公司 Cement plant equipment fault diagnosis method based on sound characteristics
CN116189711A (en) * 2023-04-26 2023-05-30 四川省机场集团有限公司 Transformer fault identification method and device based on acoustic wave signal monitoring
CN116645978A (en) * 2023-06-20 2023-08-25 方心科技股份有限公司 Electric power fault sound class increment learning system and method based on super-computing parallel environment
CN117894317A (en) * 2024-03-14 2024-04-16 沈阳智帮电气设备有限公司 Box-type transformer on-line monitoring method and system based on voiceprint analysis
CN117894317B (en) * 2024-03-14 2024-05-24 沈阳智帮电气设备有限公司 Box-type transformer on-line monitoring method and system based on voiceprint analysis

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109961034A (en) * 2019-03-18 2019-07-02 西安电子科技大学 Video object detection method based on convolution gating cycle neural unit
CN110335617A (en) * 2019-05-24 2019-10-15 国网新疆电力有限公司乌鲁木齐供电公司 A kind of noise analysis method in substation
CN110534118A (en) * 2019-07-29 2019-12-03 安徽继远软件有限公司 Transformer/reactor method for diagnosing faults based on Application on Voiceprint Recognition and neural network
US20210048487A1 (en) * 2019-08-12 2021-02-18 Wuhan University Power transformer winding fault positioning method based on deep convolutional neural network integrated with visual identification
CN112910812A (en) * 2021-02-25 2021-06-04 电子科技大学 Modulation mode identification method for deep learning based on space-time feature extraction
CN113192532A (en) * 2021-03-29 2021-07-30 安徽理工大学 Mine hoist fault acoustic analysis method based on MFCC-CNN

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109961034A (en) * 2019-03-18 2019-07-02 西安电子科技大学 Video object detection method based on convolution gating cycle neural unit
CN110335617A (en) * 2019-05-24 2019-10-15 国网新疆电力有限公司乌鲁木齐供电公司 A kind of noise analysis method in substation
CN110534118A (en) * 2019-07-29 2019-12-03 安徽继远软件有限公司 Transformer/reactor method for diagnosing faults based on Application on Voiceprint Recognition and neural network
US20210048487A1 (en) * 2019-08-12 2021-02-18 Wuhan University Power transformer winding fault positioning method based on deep convolutional neural network integrated with visual identification
CN112910812A (en) * 2021-02-25 2021-06-04 电子科技大学 Modulation mode identification method for deep learning based on space-time feature extraction
CN113192532A (en) * 2021-03-29 2021-07-30 安徽理工大学 Mine hoist fault acoustic analysis method based on MFCC-CNN

Cited By (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113466616A (en) * 2021-06-22 2021-10-01 海南电网有限责任公司乐东供电局 Method and device for quickly positioning cable fault point
CN114155878A (en) * 2021-12-03 2022-03-08 北京中科智易科技有限公司 Artificial intelligence detection system, method and computer program
CN114155878B (en) * 2021-12-03 2022-06-10 北京中科智易科技有限公司 Artificial intelligence detection system, method and computer program
CN114638256A (en) * 2022-02-22 2022-06-17 合肥华威自动化有限公司 Transformer fault detection method and system based on sound wave signals and attention network
CN114638256B (en) * 2022-02-22 2024-05-31 合肥华威自动化有限公司 Transformer fault detection method and system based on acoustic wave signals and attention network
CN114543983A (en) * 2022-03-29 2022-05-27 阿里云计算有限公司 Vibration signal identification method and device
CN115358110A (en) * 2022-07-25 2022-11-18 国网江苏省电力有限公司淮安供电分公司 Transformer fault detection system based on acoustic sensor array
CN115392293A (en) * 2022-08-01 2022-11-25 中国南方电网有限责任公司超高压输电公司昆明局 Transformer fault monitoring method and device, computer equipment and storage medium
CN115424635B (en) * 2022-11-03 2023-02-10 南京凯盛国际工程有限公司 Cement plant equipment fault diagnosis method based on sound characteristics
CN115424635A (en) * 2022-11-03 2022-12-02 南京凯盛国际工程有限公司 Cement plant equipment fault diagnosis method based on sound characteristics
CN116189711A (en) * 2023-04-26 2023-05-30 四川省机场集团有限公司 Transformer fault identification method and device based on acoustic wave signal monitoring
CN116189711B (en) * 2023-04-26 2023-06-30 四川省机场集团有限公司 Transformer fault identification method and device based on acoustic wave signal monitoring
CN116645978A (en) * 2023-06-20 2023-08-25 方心科技股份有限公司 Electric power fault sound class increment learning system and method based on super-computing parallel environment
CN116645978B (en) * 2023-06-20 2024-02-02 方心科技股份有限公司 Electric power fault sound class increment learning system and method based on super-computing parallel environment
CN117894317A (en) * 2024-03-14 2024-04-16 沈阳智帮电气设备有限公司 Box-type transformer on-line monitoring method and system based on voiceprint analysis
CN117894317B (en) * 2024-03-14 2024-05-24 沈阳智帮电气设备有限公司 Box-type transformer on-line monitoring method and system based on voiceprint analysis

Also Published As

Publication number Publication date
CN113707176B (en) 2022-09-09

Similar Documents

Publication Publication Date Title
CN113707176B (en) Transformer fault detection method based on acoustic signal and deep learning technology
CN110245608B (en) Underwater target identification method based on half tensor product neural network
CN109357749B (en) DNN algorithm-based power equipment audio signal analysis method
CN108827605B (en) Mechanical fault feature automatic extraction method based on improved sparse filtering
CN108630209B (en) Marine organism identification method based on feature fusion and deep confidence network
CN107545890A (en) A kind of sound event recognition method
CN108922513A (en) Speech differentiation method, apparatus, computer equipment and storage medium
CN105206270A (en) Isolated digit speech recognition classification system and method combining principal component analysis (PCA) with restricted Boltzmann machine (RBM)
KR102276964B1 (en) Apparatus and Method for Classifying Animal Species Noise Robust
CN109256118B (en) End-to-end Chinese dialect identification system and method based on generative auditory model
CN112326210A (en) Large motor fault diagnosis method combining sound vibration signals with 1D-CNN
CN113192532A (en) Mine hoist fault acoustic analysis method based on MFCC-CNN
CN115602152B (en) Voice enhancement method based on multi-stage attention network
CN111899757A (en) Single-channel voice separation method and system for target speaker extraction
CN112017682A (en) Single-channel voice simultaneous noise reduction and reverberation removal system
CN113488060B (en) Voiceprint recognition method and system based on variation information bottleneck
WO2019232833A1 (en) Speech differentiating method and device, computer device and storage medium
CN112086100B (en) Quantization error entropy based urban noise identification method of multilayer random neural network
CN112183225B (en) Underwater target signal feature extraction method based on probability latent semantic analysis
CN115758082A (en) Fault diagnosis method for rail transit transformer
CN116741148A (en) Voice recognition system based on digital twinning
CN115376526A (en) Power equipment fault detection method and system based on voiceprint recognition
CN108806725A (en) Speech differentiation method, apparatus, computer equipment and storage medium
CN115457980A (en) Automatic voice quality evaluation method and system without reference voice
CN111540368A (en) Stable bird sound extraction method and device and computer readable storage medium

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant