CN111044814B - Method and system for identifying transformer direct-current magnetic bias abnormality - Google Patents
Method and system for identifying transformer direct-current magnetic bias abnormality Download PDFInfo
- Publication number
- CN111044814B CN111044814B CN201911191490.7A CN201911191490A CN111044814B CN 111044814 B CN111044814 B CN 111044814B CN 201911191490 A CN201911191490 A CN 201911191490A CN 111044814 B CN111044814 B CN 111044814B
- Authority
- CN
- China
- Prior art keywords
- signal
- transformer
- interference noise
- sound
- noise
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)
Abstract
The invention discloses a method and a system for identifying transformer direct current magnetic bias abnormity, wherein the method comprises the steps of separating interference noise signals of collected noise-containing signals of a transformer to obtain separated de-noising signals; extracting the characteristics of the de-noising signal to obtain a de-noising signal after dimension reduction, and converting the de-noising signal after dimension reduction into an acoustic signal time-frequency spectrogram; and converting the sound signal time-frequency spectrogram into a cepstrum coefficient, and identifying sound spectrum abnormality according to a time sequence and the cepstrum coefficient.
Description
Technical Field
The invention relates to the technical field of transformer direct-current magnetic bias abnormity identification, in particular to a method and a system for identifying transformer direct-current magnetic bias abnormity.
Background
With the gradual expansion of the scale of the ultrahigh voltage transmission project in China, the number and the voltage grade of direct current transmission lines are continuously increased, the direct current magnetic biasing problem generated by alternating current and direct current hybrid transmission is increasingly severe, and the direct current magnetic biasing problem of a three-phase group type single-phase transformer commonly adopted in the ultrahigh voltage transmission project is particularly prominent. Because the magnetic hysteresis loop of the iron core has nonlinear characteristics, direct current magnetic bias can cause serious distortion of exciting current, which can not only cause the iron core and metal accessories of the power transformer to generate local overheating and further cause thermal aging acceleration of transformer insulating materials, but also cause magnetostriction to be aggravated, thereby causing vibration enhancement of the transformer iron core, and if the treatment is not carried out in time, the problems of mechanical structure stability such as loosening of an iron core clamping piece and the like can be caused. Therefore, aiming at key power transmission equipment such as a transformer in an alternating current-direct current hybrid power transmission system, the on-line monitoring work of direct current magnetic biasing is emphasized, and data support and reference are provided for the condition evaluation and maintenance plan formulation of the equipment.
At present, the research on the dc magnetic bias state of a transformer can be divided into two types, electrical quantity and non-electrical quantity. The research of the direct current magnetic biasing electrical quantity of the transformer depends on calculating the exciting current of the single-phase transformer under the direct current magnetic biasing working condition and carrying out finite element simulation or constructing a circuit-magnetic circuit model of the transformer. The method has the advantages of strong interpretability, but the existing method has the problems of high requirement on the precision of model parameters, difficulty in obtaining the parameters, complex model and the like, so that engineering application is difficult to realize; the research on the non-electrical quantity of the transformer is mostly concentrated in the vibration and sound category, the structures of an iron core, a winding and the like of equipment such as the transformer can vibrate and generate mechanical waves in the operation process, and the generated vibration and sound signals contain a large amount of equipment state information. The method has the advantages that the sensor does not need to generate electromagnetic coupling with the transformer, and has strong advantages in the work of online monitoring, uninterrupted power detection and the like of the transformer, and the defects that the coupling between the state of the transformer and the vibration and noise is very complex, and the traditional signal analysis means is difficult to obtain the key characteristic quantity which is widely applicable to various transformers, so that an effective mode identification method cannot be formed.
Vibration volume and sound volume all contain a large amount of equipment state information, because the vibration signal of transformer and sound signal compare has stronger interference killing feature, consequently the study of vibration signal is mostly concentrated on in the state monitoring research of transformer at present, but vibration signal acquisition is more strict to the requirement of stationing the position, and less stationing skew will lead to the measuring result to produce very big change, and this is unfavorable for different model transformer vibration data unification, and the problem that the space sensitivity is too high can then be solved well to the sound signal. The existing acoustic signal analysis method mostly adopts traditional machine learning methods such as sound level measurement, wavelet transformation, empirical mode decomposition and the like, so that loss of voiceprint characteristics is easily caused, and continuous online monitoring cannot be realized.
Therefore, a technique is needed to identify the dc magnetic bias abnormality of the transformer.
Disclosure of Invention
The technical scheme of the invention provides a method and a system for identifying the direct current magnetic bias abnormality of a transformer, which are used for solving the problem of how to identify the direct current magnetic bias abnormality of the transformer.
In order to solve the above problem, the present invention provides a method for identifying a dc magnetic bias abnormality of a transformer, wherein the method comprises:
carrying out interference noise signal separation on the acquired noise-containing signals of the transformer to obtain separated de-noise signals;
extracting the characteristics of the de-noising signal to obtain a de-noising signal after dimension reduction, and converting the de-noising signal after dimension reduction into an acoustic signal time-frequency spectrogram;
and converting the sound signal time-frequency spectrogram into a cepstrum coefficient, and identifying the abnormality of the sound spectrum according to the time sequence and the cepstrum coefficient.
Preferably, the performing interference noise signal separation on the acquired noise-containing signal of the transformer to obtain a separated de-noised signal includes:
separating instantaneous interference noise signals by a blind source separation method based on a similarity matrix;
and separating continuous interference noise signals by a two-channel difference method.
Preferably, the interference noise signal includes:
an instantaneous interference noise signal and/or a continuous interference noise signal.
Preferably, the method further comprises the following steps: and respectively arranging acoustic signal acquisition points of the transformer at the central positions of the two long end surfaces of the transformer.
Preferably, the separating the persistent interference noise signal by a two-channel difference method includes:
fourier transformation is respectively carried out on a first transformer noise-containing signal collected by a first collection point and a second transformer noise-containing signal collected by a second collection point, so that a first frequency domain signal and a second frequency domain signal are obtained;
acquiring the continuous interference noise of any acquisition point and the frequency domain signal of the sound of the transformer body based on the continuous interference noise and the sound intensity change rate of the sound of the transformer body on the basis of the intensity attenuation of sound propagation of each frequency band in the first frequency domain signal and the second frequency domain signal and no frequency shift;
and separating a continuous interference noise signal according to the continuous interference noise and the frequency domain signal of the sound of the transformer body.
According to another aspect of the present invention, there is provided a system for identifying a dc magnetic bias abnormality of a transformer, the system comprising:
the separation unit is used for carrying out interference noise signal separation on the acquired noise-containing signals of the transformer to obtain separated de-noised signals;
the dimension reduction unit is used for extracting the characteristics of the de-noising signal to obtain the de-noising signal after dimension reduction and converting the de-noising signal after dimension reduction into an acoustic signal time-frequency spectrogram;
and the identification unit is used for identifying the abnormality of the sound frequency spectrum according to the time sequence and the cepstrum coefficient by converting the sound signal time-frequency spectrogram into the cepstrum coefficient.
Preferably, the separation unit is configured to perform interference noise signal separation on the acquired noise-containing signal of the transformer, obtain a separated de-noising signal, and further configured to:
separating instantaneous interference noise signals by a blind source separation method based on a similar matrix;
and separating continuous interference noise signals by a two-channel difference method.
Preferably, the interference noise signal includes:
an instantaneous interference noise signal and/or a continuous interference noise signal.
Preferably, the system further comprises an acquisition unit for: and respectively arranging acoustic signal acquisition points of the transformer at the central positions of the two long end surfaces of the transformer.
Preferably, the separation unit is configured to separate the persistent interference noise signal by a two-channel difference method, and is further configured to:
carrying out Fourier transform on a first transformer noise-containing signal acquired by a first acquisition point and a second transformer noise-containing signal acquired by a second acquisition point respectively to obtain a first frequency domain signal and a second frequency domain signal;
acquiring continuous interference noise of any acquisition point and a frequency domain signal of the sound of the transformer body based on continuous interference noise and the sound intensity change rate of the sound of the transformer body, wherein the continuous interference noise of any acquisition point and the frequency domain signal of the sound of the transformer body are acquired based on intensity attenuation and no frequency shift of sound transmission of each frequency band in the first frequency domain signal and the second frequency domain signal;
and separating a continuous interference noise signal according to the continuous interference noise and the frequency domain signal of the sound of the transformer body.
The technical scheme of the invention provides a method and a system for identifying the DC magnetic bias abnormality of a transformer, wherein the method comprises the steps of separating interference noise signals of collected noise-containing signals of the transformer to obtain separated de-noising signals; the method comprises the steps of obtaining a dimension-reduced de-noising signal by extracting features of the de-noising signal, and converting the dimension-reduced de-noising signal into an acoustic signal time-frequency spectrogram; and converting the acoustic signal time-frequency spectrogram into a cepstrum coefficient, and identifying the abnormality of the cepstrum coefficient according to the time sequence and the cepstrum coefficient.
Drawings
A more complete understanding of exemplary embodiments of the present invention may be had by reference to the following drawings in which:
FIG. 1 is a flowchart of a method for identifying DC magnetic bias anomalies in a transformer according to a preferred embodiment of the present invention;
FIG. 2 is a schematic diagram of a step-by-step acoustic interference signal blind source separation process according to a preferred embodiment of the present invention;
FIG. 3 is a flow chart of a two-channel difference method according to a preferred embodiment of the present invention;
FIG. 4 is a schematic diagram illustrating the separation effect of the mixed signal of the sound of the transformer body and the sound of the cooling fan according to the preferred embodiment of the present invention;
FIG. 5 is a schematic diagram of a time-frequency domain conversion of a signal according to a preferred embodiment of the present invention;
FIG. 6 is a schematic diagram of a 50FMCCs-GRU based anomaly identification system according to the preferred embodiment of the present invention; and
fig. 7 is a system configuration diagram for identifying a dc magnetic bias abnormality of a transformer according to a preferred embodiment of the present invention.
Detailed Description
Example embodiments of the present invention will now be described with reference to the accompanying drawings, however, the invention may be embodied in many different forms and should not be construed as limited to the embodiments set forth herein, which are provided for a complete and complete disclosure of the invention and to fully convey the scope of the invention to those skilled in the art. The terminology used in the exemplary embodiments illustrated in the accompanying drawings is not intended to be limiting of the invention. In the drawings, the same unit/element is denoted by the same reference numeral.
Unless otherwise defined, terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. Further, it will be understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and will not be interpreted in an idealized or overly formal sense.
Fig. 1 is a flowchart of a method for identifying dc magnetic bias anomalies of a transformer according to a preferred embodiment of the invention. According to the method and the device, the direct current magnetic bias identification method and the related system design work of the power transformer based on acoustic fingerprint deep learning are developed based on the operation and inspection requirements of power grid equipment and by combining the operation and inspection professional characteristics of companies and the development trend of constructing an intelligent operation and inspection system. The intelligent identification of the DC magnetic biasing working condition of the large-scale power transformer is realized through the fusion of an artificial intelligence technology and the traditional operation and inspection service. The intelligent support system can provide support for equipment operation and maintenance, overhaul and production management intelligence in the ubiquitous power Internet of things construction, realize equipment intellectualization, operation and maintenance intelligence, overhaul intelligence and production management intelligence, improve equipment state management and control force and operation and inspection management penetrating power, realize innovation development and efficiency improvement of data-driven operation and inspection business, and comprehensively promote innovation of operation and inspection working modes and production management modes. As shown in fig. 1, a method for identifying dc magnetic bias abnormality of a transformer includes:
preferably, in step 101: and carrying out interference noise signal separation on the acquired noise-containing signals of the transformer to obtain separated de-noise signals. Preferably, the interference noise signal separation is performed on the acquired noise-containing signal of the transformer, and the separated noise-removed signal is obtained, including: separating instantaneous interference noise signals by a blind source separation method based on a similarity matrix; and separating continuous interference noise signals by a two-channel difference method. Preferably, the interference noise signal comprises: an instantaneous interference noise signal and/or a continuous interference noise signal.
Preferably, the method further comprises the following steps: and respectively arranging an acoustic signal acquisition point of the transformer at the central position of the two long end surfaces of the transformer. Preferably, the separating the continuous interference noise signal by the two-channel difference method includes: carrying out Fourier transform on a first transformer noise-containing signal acquired by a first acquisition point and a second transformer noise-containing signal acquired by a second acquisition point respectively to obtain a first frequency domain signal and a second frequency domain signal; based on the intensity attenuation of sound propagation of each frequency band in the first frequency domain signal and the second frequency domain signal and no frequency shift, acquiring the continuous interference noise of any acquisition point and the frequency domain signal of the sound of the transformer body based on the continuous interference noise and the sound intensity change rate of the sound of the transformer body; and separating the continuous interference noise signal according to the continuous interference noise and the frequency domain signal of the sound of the transformer body.
The method firstly analyzes the interference noise source, and the voiceprint diagnosis of the transformer is not widely applied all the time, and mainly comprises the following steps: in the process of collecting sound data on site, various interference noises often exist, and troubles are brought to the processing and fault diagnosis of the sound of the transformer body. Therefore, the noise removal processing should be performed on the sound signal first when deep neural network training or practical application is performed. Through the field sound signal collection of the transformer substation, the interference signals possibly existing in the transformer operation environment are classified as follows according to the characteristics of the interference signals:
TABLE 1 Classification and dominance of interfering signals
The type of interference signal varies because the transformer acoustic signal is subject to more complex interference. First, the interference bands of corona discharges in the sustained weak interference class and of bird sounds in the transient interference class are not intersected by the transformer body band (0-4000 Hz) and can therefore be disregarded. For other interferences, the following two algorithms are used to process the sound signal of the substation: 1) For instantaneous interference signals, a blind source separation method based on a similar matrix is adopted; 2) For continuous strong interference signals, a two-channel difference method is adopted.
The application relates to a blind source separation method based on a similarity matrix to separate transient signals. A method for separating and extracting various source signals from aliased signals in the case of blind source separation into a small number of a priori source signals. This method is commonly used to separate unstable signals from continuously stable signals.
In the sound collection process of the transformer, transient interference such as bird song and the like occupies most of interference types of the transformer substation. The common characteristics of such interference are short duration and concentrated energy distribution of the interfering signal, while the transformer body sound is a continuous and stable signal. Therefore, the method and the device strip instantaneous interference signals from original sounds by using a blind source separation method based on the similarity matrix, thereby eliminating irrelevant interference and improving the accuracy and efficiency of a subsequent identification algorithm.
Firstly, distance norms are taken from the frequency spectrum W according to columns and normalized, the similarity between signals of each frame is calculated to obtain the cosine similarity between the feature vectors of the ith frame and the kth frame in a similarity matrix S, each point (i, k) of the similarity matrix S corresponds to the feature vector of the ith frame and the kth frame in the frequency spectrum W, and the calculation formula of the similarity matrix is as follows:
wherein, the variable m is the frame number, m is more than or equal to 1 and less than or equal to n, and n is the frame length of the time frequency spectrum.
The similarity matrix is insensitive to periodic signals by calculating the difference between frames, and transient interference sound signals are sparseness and volatility relative to the sound signals of the transformer body. And defining a repeated spectrum model V as a median filtering value of each frequency band of the matrix S, wherein the median filtering value represents the vector correlation degree of each frame of the frequency spectrum and the similar frame of the frequency spectrum. The repeated spectrum V used for filtering should be less than or equal to the original spectrum W, i.e.:
F＝min(V，W) (2)
the repeated spectrum model obtained through median filtering is not accurate, in order to separate interference signals more accurately, a wiener filtering method (minimum mean square error filtering) is adopted to improve the repeated spectrum model, and the standard equation of the wiener filtering is as follows:
R xs (m)＝Σ i h(i)R xx (m-i),m≥0 (3)
wherein R is xs As a function of the cross-correlation of the desired signal with the actual signal, R xx Is an autocorrelation function, and h is the optimal filter coefficient under the minimum mean square error.
And finally, extracting the repetition characteristics of the original signal by using time-frequency masking, namely separating the frequency spectrum of the background time from the frequency spectrum of the foreground time. In order to verify the effectiveness of the algorithm, the method intercepts fragments containing various interferences in the field measured data, and compares the instantaneous interference signals obtained by blind source separation, the transformer body signals and the time-frequency spectrogram of the original field measured signals. As shown in fig. 2. Fig. 2 is a first diagram of transformer acoustic signals mixed with footstep acoustic signals, which can be separated from stationary background signals by a blind source separation method based on a similarity matrix.
This application separates continuous interference signal through the binary channels difference method, need use two same model sound sensor to carry out synchronous collection at the signal acquisition stage. The sensor arrangement position needs to guarantee that the positions of two sensors have a certain distance, and aims to guarantee that the ratio of the sound of the fan to the sound of the transformer body is different in two groups of received signals. Because the two sensors are positioned differently and are positioned in the same medium, the sound propagation of each frequency band only has intensity attenuation but does not have frequency shift. If the sound intensity change rate of the fan and the transformer body is obtained, the frequency domain signal of the fan sound and the transformer body at any channel can be obtained, and therefore the purpose of separating the fan sound and the transformer body sound is achieved.
Firstly, fourier transform is carried out on two-channel signals to obtain frequency domain signals, and the frequency domain signals pass through a 50Hz frequency multiplication comb filter. The signals collected by each channel are composed of signals sent by two sound sources of the fan and the transformer body;
next, the sound source composition of each frequency component in the two-channel signal needs to be divided. The sound source composition can be divided into three independent subsets according to each frequency band: p 1 Is composed of a transformerSet of frequency point sequence numbers, P, of volume generation 2 Set of sequence numbers, P, for frequency points generated purely by the fan 3 The method comprises the following steps of (1) collecting the serial numbers of frequency points formed by a transformer and a fan;
because the sound intensity of two frequency points belonging to the same sound source can be attenuated according to the same change rate in the course of propagation, the ratio of sound intensity of two frequency points can not be changed after the position is changed, and the ratio of sound intensity of two frequency points belonging to different sound sources can be changed after the position is changed, according to said condition, it can be determined to respectively define set P 1 And P 2 (ii) a The method can calculate the sound intensity change rate; and finally, calculating to obtain frequency domain vectors of the fan sound and the transformer body sound after separation in the two channels of signals according to the sound intensity change rate.
The specific steps are as shown in fig. 3, and two sound sensors of the same type are required to be used for synchronous acquisition in the signal acquisition stage. The sensor arrangement position needs to guarantee that the positions of two sensors have a certain distance, and aims to guarantee that the ratio of the sound of the fan to the sound of the transformer body is different in two groups of received signals. Because the two sensors are positioned differently and are in the same medium, the sound propagation in each frequency band only has intensity attenuation but does not have frequency shift. If the sound intensity change rate of the fan and the transformer body is obtained, the frequency domain signal of the fan sound and the transformer body at any channel can be obtained, and therefore the purpose of separating the fan sound and the transformer body sound is achieved. The analysis effect is shown in fig. 4, the separation effect of the mixed signal of the sound of the transformer body and the sound of the cooling fan is that the separation result obtained by the comb filter in combination with the difference method is basically consistent with the original sound of the transformer body.
Preferably, at step 102: and extracting the characteristics of the de-noised signals to obtain the de-noised signals after dimension reduction, and converting the de-noised signals after dimension reduction into an acoustic signal time-frequency spectrogram.
The original signal is subjected to dimensionality reduction processing, and the original signal contains much information due to the fact that the transformer sound contains much redundant information. In order to fully utilize computer computing resources and improve the performance of the identification model, firstly, the original sound signal of the transformer needs to be subjected to dimensionality reduction. Because the vibration of the transformer is a source of sound generation, in order to obtain a better-performance dimension reduction processing mode, the vibration law of the running state of the transformer needs to be analyzed first.
In conventional speech recognition algorithms, the speech signal is typically feature extracted using Mel-Frequency cepstral coefficients (MFCCs). The auditory system of human ears has different frequency perception sensitivities for each frequency band, and the MFCCs has the main advantage that Mel nonlinear frequency spectrum obtained based on human auditory perception experiments is used for carrying out feature extraction on sound signals. In the application, the MFCCs thought is referred to, and a feature extraction method for transformer voice mode recognition is provided according to the frequency characteristics of the voice signals of the transformer, so that the dimension reduction processing of the voice signals is realized.
(1) Firstly, the original signal is processed by framing, windowing and short-time discrete Fourier transform
Framing: the analysis of the vibration rule of the transformer operating state by the previous subsection shows that the frame length when short-time fast Fourier transform is carried out is not less than 0.02s in order to ensure that information is not missed because the acoustic signal of the transformer is mainly 50Hz frequency multiplication. Based on the conclusion, the method takes the transformer sound signal with the time length of 1s as a sample, frames the sample, and divides each frame into 97 frames according to the time domain, wherein the frame length of each frame is 0.04s, and the frame shift is 0.01 s.
Windowing: windowing is carried out on each frame of time domain data, and the purpose is to weaken distortion influence caused by Fourier transform to be carried out next time. The application selects a Hamming window (hamming) with better time and frequency aggregation characteristics:
wherein N is the length of the Hamming window.
Short-time discrete Fourier transform (STFT): then, performing STFT on the windowed sound signal, wherein the calculation formula is as follows:
the process is shown in figure 3.
(2) Periodic energy spectrum estimation for each frame
P i (k) And (4) an energy spectrum estimation array.
(3) Design 50Hz frequency multiplication triangular filter bank
According to the analysis of the sound frequency characteristics of the transformer, a 50Hz frequency multiplication triangular filter bank is designed, firstly, the 50Hz frequency multiplication frequency value is converted to a 50Hz interval frequency list and is expressed by f (m):
m＝(1，2，3，...，F max /100) (7)
wherein r is s Is the sampling rate of the audio file; m is the serial number of the filter, and the total number of the filters is limited by the upper limit F of the frequency range needing to extract the features max And (6) determining. Based on the measured acoustic data of the transformer core, most of the energy is concentrated in the range of 0 to 4 kHz. It should be noted that: for the converter transformer, the energy of the part of the running sound signal above 4kHz cannot be ignored, and if the application object is the converter transformer, the number of filters can be increased, and the information of a higher frequency band is included. The invention is followed by a common power transformer only, not a converter transformer, and therefore the number of filter banks here is 4000/100= 40.
The filter bank is constructed in the following way: the first filter starts at the first point of f (m), the second gets the maximum and the third returns to zero. The second filter starts at the second point of f (m), reaches a maximum at the third point, returns to zero at the fourth point, and so on. The filter bank expression can thus be derived as follows:
energy spectrum P using designed filter bank i (k) The resulting X (k) has been filtered to emphasize energy near the 50Hz octave. Then, performing cepstrum analysis on X (k): taking logarithm first, then performing Discrete Cosine Transform (DCT), so as to obtain the Cepstrum coefficient composed by each frame, which is called 50Hz Frequency Multiplication Cepstrum coefficient for convenience of expression, and is abbreviated as 50FMCCs (50 Hz-Frequency Multiplication Cepstrum Coefficients). The 50FMCCs of all frames are combined into a matrix R (k, t), where k denotes the cepstral coefficient number associated with the frequency domain and t denotes the frame number associated with the time domain. The specific steps are shown in fig. 5.
Preferably, in step 103: and converting the sound signal time-frequency spectrogram into a cepstrum coefficient, and identifying the abnormality of the sound spectrum according to the time sequence and the cepstrum coefficient.
Because the sound signal differentiation condition of the transformer is very serious, even if the transformer is a three-phase split type transformer produced by the same manufacturer, the sound signal time-frequency characteristics of each independent single-phase transformer have certain difference, and the deep learning has the biggest advantage for the traditional machine learning that the generalization of the model is improved along with the expansion of the sample. Therefore, the present application will use a gate cycle Unit (GRU) in the deep neural network to perform fault mode identification on the transformer sound sequence. However, the amount of the frequency domain and the amount of the time domain are large, and the amount of calculation is increased sharply and the operation time is increased greatly when the amount of the frequency domain and the amount of the time domain are directly input into the deep neural network as the input amount. In order to ensure that the characteristics of the transformer samples are preserved while the data input is compressed, the application uses 50FMCCs as the GRU input, and the combination of the two forms a 50FMCCs-GRU identification system, as shown in FIG. 6.
(1)50FMCCs
Firstly, denoising an original sample with the duration of 1s and the frequency band cut to 4 kHz. The feature vectors are then compressed to obtain 50FMCCs, where 50FMCCs consists of 97 sets of feature vectors, each set of feature vectors having a length of 40.
(2)GRU
And sequentially inputting the characteristic vectors into each node X _ n of the GRU input layer in a time sequence, wherein n represents the serial number of the node. After data is input from the input layer, the data enters the GRU subunit, then enters each node H _ n of the hidden layer, and finally is output to the output layer node o. And comparing the label value y with the output value o to obtain an error L, continuously iterating according to an error reverse transfer mode, and finally determining each weight parameter and bias parameter in the network structure.
In order to verify the effectiveness of the 50FMCCs-GRU identification model, the voice data of the 500kV alternating current transformer in the direct current large current ground entering test period are collected by the 500kV toasting substation to form a data set.
In the aspect of noise signal acquisition, according to the IEC60651 standard, the noise measurement frequency range should cover 25 Hz-16 kHz, and the acquisition equipment adopts a combined system of an electret capacitor type gun-shaped pointing microphone and a recorder to ensure the reliability of recorded data: the gun-shaped directional microphone has stronger directivity, and the frequency response range meets the sampling requirement; the recorder with 24bit 96kHz recording capability can record as many noise details as possible in high-frequency response and wide dynamic range, and is connected with the gun microphone through the Kangnong interface, so that the stability and the anti-interference capability of data transmission are ensured. The microphone arrangement positions are symmetrically arranged and are located in the middle of two long end face sides of each transformer. The pointing direction of the microphone is perpendicular to the long end face of the transformer, and the microphone is 50cm away from the wall of the oil tank of the transformer and 150cm away from the ground.
In the aspect of data set division, the characteristic that a 500kV Jingting becomes a three-phase split type transformer is utilized, the data of the A phase and the B phase are used as a training set, and the data of the C phase is used as a test set. Because the influence factors of the sound signals of the transformer are more, although three single-phase transformers are produced by the same manufacturer, the voiceprints still have larger differences. The method also ensures that the training set and the test set have good independence, and the experimental result can reflect the generalization of the model to a certain degree. The final data set sample numbers are shown in table 2.
Table 2 data set sample distribution
(3) Comparison of different recognition model effects
In order to compare the performances of different recognition models, hyper-parameter optimization is respectively carried out on a 50FMCCs-GRU model and other recognition models, and the results of the iteration 100 steps of the deep learning recognition model are compared with the results of traditional Machine learning algorithms such as a Support Vector Machine (SVM), a k-Nearest Neighbor (kNN), a k-Nearest Neighbor and the like. The results are shown in Table 3.
TABLE 3 comparison of results and effects of hyper-parametric tuning for several recognition models based on this example
As can be seen from table 3, the algorithm based on deep learning is generally superior to the conventional machine learning algorithm in terms of both operation time and accuracy. The 50FMCCs feature extraction can greatly improve the operation accuracy, and the situation of non-convergence exists when the time domain data is simply used as the input quantity of deep learning models such as GRU, RNN and LSTM, and the accuracy is extremely low. Therefore, the 50FMCCs provided by the method have certain application value. The comparison among 50FMCCs-GRU, 50FMCCs-LSTM and 50FMCCs-RNN shows that GRU has strong advantage in calculation accuracy compared with other two circulation deep neural networks, and the calculation time of each batch in each iteration of 50FMCCs-GRU, 50FMCCs-LSTM and 50FMCCs-RNN is respectively 0.42s, 0.23s and 0.56s in view of calculation rate. Since the structure of the GRU is more compact relative to the RNN, the calculation rate is faster than that of the LSTM while ensuring accuracy.
Fig. 7 is a diagram illustrating a system for identifying dc magnetic bias abnormality of a transformer according to a preferred embodiment of the present invention. As shown in fig. 7, the present application provides a system for identifying dc magnetic bias anomalies of a transformer, the system comprising:
the separating unit 701 is configured to perform interference noise signal separation on the acquired noise-containing signal of the transformer, and obtain a separated de-noise signal. Preferably, the separation unit 701 is configured to perform interference noise signal separation on the acquired noise-containing signal of the transformer, obtain a separated denoised signal, and further configured to: separating instantaneous interference noise signals by a blind source separation method based on a similarity matrix; and separating continuous interference noise signals by a two-channel difference method. Preferably, the interference noise signal comprises: an instantaneous interference noise signal and/or a continuous interference noise signal. Preferably, the system further comprises an acquisition unit for: and respectively arranging an acoustic signal acquisition point of the transformer at the central position of the two long end surfaces of the transformer.
And a dimension reduction unit 702, configured to obtain the dimension-reduced denoising signal by performing feature extraction on the denoising signal, and convert the dimension-reduced denoising signal into an acoustic signal time-frequency spectrogram.
The identifying unit 703 is configured to convert the time-frequency spectrogram of the acoustic signal into a cepstrum coefficient, and identify an abnormality of the sound spectrum according to the time sequence and the cepstrum coefficient.
Preferably, the separation unit is configured to separate the persistent interference noise signal by a two-channel difference method, and is further configured to: carrying out Fourier transform on a first transformer noise-containing signal acquired by a first acquisition point and a second transformer noise-containing signal acquired by a second acquisition point respectively to obtain a first frequency domain signal and a second frequency domain signal; based on the intensity attenuation of sound propagation of each frequency band in the first frequency domain signal and the second frequency domain signal and no frequency shift, acquiring the continuous interference noise of any acquisition point and the frequency domain signal of the sound of the transformer body based on the continuous interference noise and the sound intensity change rate of the sound of the transformer body; and separating the continuous interference noise signal according to the continuous interference noise and the frequency domain signal of the sound of the transformer body.
The system 700 for identifying dc magnetic bias anomalies of a transformer according to the preferred embodiment of the present invention corresponds to the method 100 for identifying dc magnetic bias anomalies of a transformer according to the preferred embodiment of the present invention, and will not be described herein again.
The invention has been described with reference to a few embodiments. However, other embodiments of the invention than the ones disclosed above are equally possible within the scope of these appended patent claims, as these are known to those skilled in the art.
Generally, all terms used in the claims are to be interpreted according to their ordinary meaning in the technical field, unless explicitly defined otherwise herein. All references to "a/an/the [ device, component, etc ]" are to be interpreted openly as referring to at least one instance of said device, component, etc., unless explicitly stated otherwise. The steps of any method disclosed herein do not have to be performed in the exact order disclosed, unless explicitly stated.
Claims (10)
1. A method of identifying transformer dc magnetic bias anomalies, the method comprising:
carrying out interference noise signal separation on the acquired noise-containing signals of the transformer to obtain separated de-noise signals;
the method comprises the steps of extracting the characteristics of the de-noising signal, obtaining the de-noising signal after dimension reduction, and converting the de-noising signal after dimension reduction into an acoustic signal time-frequency spectrogram, and comprises the following steps:
framing, windowing and short-time discrete Fourier transform are carried out on the de-noised signal;
carrying out periodic energy spectrum estimation on each frame of the converted de-noised signals;
establishing a frequency multiplication triangular filter bank based on the periodic energy spectrum estimation, and extracting features based on the frequency multiplication triangular filter bank;
and converting the acoustic signal time-frequency spectrogram into a cepstrum coefficient, and identifying the abnormality of the acoustic spectrum through a 50FMCCs-GRU identification model according to the time sequence and the cepstrum coefficient.
2. The method according to claim 1, wherein the step of performing interference noise signal separation on the collected noise-containing signals of the transformer to obtain separated noise-removed signals includes:
separating instantaneous interference noise signals by a blind source separation method based on a similar matrix;
and separating continuous interference noise signals by a two-channel difference method.
3. The method of claim 1, the interference noise signal comprising:
an instantaneous interference noise signal and/or a continuous interference noise signal.
4. The method of claim 2, further comprising: and respectively arranging acoustic signal acquisition points of the transformer at the central positions of the two long end surfaces of the transformer.
5. The method of claim 4, the separating persistent interference noise signals by a two-channel differential method, comprising:
fourier transformation is respectively carried out on a first transformer noise-containing signal collected by a first collection point and a second transformer noise-containing signal collected by a second collection point, so that a first frequency domain signal and a second frequency domain signal are obtained;
acquiring continuous interference noise of any acquisition point and a frequency domain signal of the sound of the transformer body based on continuous interference noise and the sound intensity change rate of the sound of the transformer body, wherein the continuous interference noise of any acquisition point and the frequency domain signal of the sound of the transformer body are acquired based on intensity attenuation and no frequency shift of sound transmission of each frequency band in the first frequency domain signal and the second frequency domain signal;
and separating a continuous interference noise signal according to the continuous interference noise and the frequency domain signal of the sound of the transformer body.
6. A system for identifying transformer dc magnetic bias anomalies, the system comprising:
the separation unit is used for carrying out interference noise signal separation on the collected noise-containing signals of the transformer to obtain separated de-noise signals;
the dimension reduction unit is used for extracting the characteristics of the de-noise signal to obtain the de-noise signal after dimension reduction, and converting the de-noise signal after dimension reduction into an acoustic signal time-frequency spectrogram, and comprises:
performing framing, windowing and short-time discrete Fourier transform on the de-noised signal;
carrying out periodic energy spectrum estimation on each frame of the converted de-noised signals;
establishing a frequency multiplication triangular filter bank based on the periodic energy spectrum estimation, and extracting features based on the frequency multiplication triangular filter bank;
and the identification unit is used for converting the sound signal time-frequency spectrogram into a cepstrum coefficient and identifying the abnormality of the sound frequency spectrum through a 50FMCCs-GRU identification model according to the time sequence and the cepstrum coefficient.
7. The system of claim 6, wherein the separation unit is configured to perform interference noise signal separation on the acquired noise-containing signal of the transformer, obtain a separated de-noised signal, and further configured to:
separating instantaneous interference noise signals by a blind source separation method based on a similar matrix;
and separating continuous interference noise signals by a two-channel difference method.
8. The system of claim 6, the interference noise signal comprising:
an instantaneous interference noise signal and/or a continuous interference noise signal.
9. The system of claim 7, further comprising an acquisition unit to: and respectively arranging acoustic signal acquisition points of the transformer at the central positions of the two long end surfaces of the transformer.
10. The system of claim 7, the separation unit to separate the continuous interference noise signal by a two-channel differential method, further to:
carrying out Fourier transform on a first transformer noise-containing signal acquired by a first acquisition point and a second transformer noise-containing signal acquired by a second acquisition point respectively to obtain a first frequency domain signal and a second frequency domain signal;
acquiring the continuous interference noise of any acquisition point and the frequency domain signal of the sound of the transformer body based on the continuous interference noise and the sound intensity change rate of the sound of the transformer body on the basis of the intensity attenuation of sound propagation of each frequency band in the first frequency domain signal and the second frequency domain signal and no frequency shift;
and separating a continuous interference noise signal according to the continuous interference noise and the frequency domain signal of the sound of the transformer body.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911191490.7A CN111044814B (en) | 2019-11-28 | 2019-11-28 | Method and system for identifying transformer direct-current magnetic bias abnormality |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911191490.7A CN111044814B (en) | 2019-11-28 | 2019-11-28 | Method and system for identifying transformer direct-current magnetic bias abnormality |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111044814A CN111044814A (en) | 2020-04-21 |
| CN111044814B true CN111044814B (en) | 2023-03-14 |
Family
ID=70233946
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911191490.7A Active CN111044814B (en) | 2019-11-28 | 2019-11-28 | Method and system for identifying transformer direct-current magnetic bias abnormality |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111044814B (en) |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111579056A (en) * | 2020-05-19 | 2020-08-25 | 北京快鱼电子股份公司 | Transformer direct-current magnetic bias prediction method and system |
| CN112633147A (en) * | 2020-12-22 | 2021-04-09 | 电子科技大学 | Magnetic anomaly detection method based on multi-feature fusion and intelligent iterative optimization algorithm |
| CN113567096B (en) * | 2021-06-15 | 2022-04-19 | 南方电网科学研究院有限责任公司 | Method and device for evaluating running state of converter transformer under no-load pressurization working condition with line |
| CN113807592B (en) * | 2021-09-22 | 2023-07-07 | 长春工程学院 | Method, system and equipment for predicting direct-current magnetic bias current of neutral point grounding transformer |
| CN114113896B (en) * | 2021-11-29 | 2024-05-28 | 郑定坤 | Modularized multi-level converter type high-voltage direct-current transmission fault diagnosis method |
| CN114492543A (en) * | 2022-04-12 | 2022-05-13 | 国网山东省电力公司东营供电公司 | Partial discharge classification and identification method based on time-frequency domain characteristics |
| CN115186721B (en) * | 2022-09-13 | 2022-12-09 | 国网江西省电力有限公司电力科学研究院 | IMF-based dynamic calculation method for accumulated effect of direct current magnetic bias of transformer |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20020022257A (en) * | 2000-09-19 | 2002-03-27 | 오길록 | The Harmonic-Noise Speech Coding Algorhthm Using Cepstrum Analysis Method |
| CN101246043A (en) * | 2008-03-28 | 2008-08-20 | 清华大学 | On-line monitoring method for vibration and noise of AC power transformer influenced by DC magnetic biasing |
| CN104330660A (en) * | 2014-10-23 | 2015-02-04 | 中国南方电网有限责任公司电网技术研究中心 | Power transformer DC magnetic bias state detection method based on noise method |
| CN105785189A (en) * | 2016-01-07 | 2016-07-20 | 国网宁夏电力公司电力科学研究院 | Portable transformer direct current magnetic bias monitor |
| JP2018036113A (en) * | 2016-08-30 | 2018-03-08 | 学校法人 芝浦工業大学 | Signal processing device, and abnormality diagnostic device and abnormality diagnostic method of power apparatus |
| CN108241098A (en) * | 2016-12-27 | 2018-07-03 | 大唐观音岩水电开发有限公司 | Power transformer D.C. magnetic biasing monitors system and monitoring method |
-
2019
- 2019-11-28 CN CN201911191490.7A patent/CN111044814B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20020022257A (en) * | 2000-09-19 | 2002-03-27 | 오길록 | The Harmonic-Noise Speech Coding Algorhthm Using Cepstrum Analysis Method |
| CN101246043A (en) * | 2008-03-28 | 2008-08-20 | 清华大学 | On-line monitoring method for vibration and noise of AC power transformer influenced by DC magnetic biasing |
| CN104330660A (en) * | 2014-10-23 | 2015-02-04 | 中国南方电网有限责任公司电网技术研究中心 | Power transformer DC magnetic bias state detection method based on noise method |
| CN105785189A (en) * | 2016-01-07 | 2016-07-20 | 国网宁夏电力公司电力科学研究院 | Portable transformer direct current magnetic bias monitor |
| JP2018036113A (en) * | 2016-08-30 | 2018-03-08 | 学校法人 芝浦工業大学 | Signal processing device, and abnormality diagnostic device and abnormality diagnostic method of power apparatus |
| CN108241098A (en) * | 2016-12-27 | 2018-07-03 | 大唐观音岩水电开发有限公司 | Power transformer D.C. magnetic biasing monitors system and monitoring method |
Non-Patent Citations (4)
| Title |
|---|
| Multitaper Estimation of Frequency-Warped Cepstra With Application to Speaker Verification;Johan Sandberg et al.;《IEEE SIGNAL PROCESSING LETTERS》;20100430;第17卷(第4期);343-346 * |
| 基于变压器故障噪声监测诊断技术的音频特征提取方法研究;曾琦器 等;《机电信息》;20190715(第2019年20期);25-27 * |
| 基于声纹及振动的变压器故障诊断技术研究;余长厅 等;《高压电器》;20191116;第55卷(第11期);248-254 * |
| 基于改进MFCC和VQ的变压器声纹识别模型;王丰华 等;《中国电机工程学报》;20170305;第37卷(第5期);1535-1542 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111044814A (en) | 2020-04-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111044814B (en) | Method and system for identifying transformer direct-current magnetic bias abnormality | |
| CN112257521B (en) | CNN underwater acoustic signal target identification method based on data enhancement and time-frequency separation | |
| Yegnanarayana et al. | Processing of reverberant speech for time-delay estimation | |
| Han et al. | A recursive sparse representation strategy for bearing fault diagnosis | |
| CN101426169B (en) | Time-domain tracking filter quickly detecting acoustic response parameter of sounding body and system | |
| CN112326210A (en) | Large motor fault diagnosis method combining sound vibration signals with 1D-CNN | |
| CN104964738A (en) | Noise separation-based power transformer noise assessment method | |
| CN112329914B (en) | Fault diagnosis method and device for buried transformer substation and electronic equipment | |
| Wanli et al. | The research of feature extraction based on MFCC for speaker recognition | |
| CN111912519B (en) | Transformer fault diagnosis method and device based on voiceprint frequency spectrum separation | |
| CN116576956A (en) | Multisource vibration signal separation method based on distributed optical fiber acoustic wave sensing | |
| Rahman et al. | Dynamic time warping assisted svm classifier for bangla speech recognition | |
| CN112052712A (en) | Power equipment state monitoring and fault identification method and system | |
| Ruan et al. | Use AF-CNN for end-to-end fiber vibration signal recognition | |
| Chu et al. | A noise-robust FFT-based auditory spectrum with application in audio classification | |
| Dang et al. | Cochlear filter cepstral coefficients of acoustic signals for mechanical faults identification of power transformer | |
| CN103778914A (en) | Anti-noise voice identification method and device based on signal-to-noise ratio weighing template characteristic matching | |
| CN115064182A (en) | Fan fault feature identification method of self-adaptive Mel filter in strong noise environment | |
| CN112863517A (en) | Speech recognition method based on perceptual spectrum convergence rate | |
| Kuan et al. | A robust BFCC feature extraction for ASR system. | |
| Vani et al. | Hilbert Huang transform based speech recognition | |
| CN111968671A (en) | Low-altitude sound target comprehensive identification method and device based on multi-dimensional feature space | |
| Lei et al. | A fault diagnosis approach of reciprocating compressor gas valve based on local mean decomposition and autoregressive-generalized autoregressive conditional heteroscedasticity model | |
| Luo et al. | Insulation fault identification of vacuum circuit breakers based on improved MFCC and SVM | |
| Patil et al. | Audio environment identification |
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 |