Non-intrusive fault diagnosis method and system using transformer sound signals
Through the improved wavelet packet energy extraction and hidden Markov model, combined with transformer acoustic signals for fault diagnosis, the problem of identifying the operating status of the transformer in the prior art is solved, and rapid and accurate fault identification and management efficiency improvement is achieved.
Patent Information
- Application Number
- CN202411181646.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-27
- Publication Date
- 2024-12-06
AI Technical Summary
The prior art is difficult to quickly and accurately identify the operating status of the transformer through non-invasive online detection, especially when the power system is expanded and the amount of acoustic signal data is huge, resulting in a bottleneck in the rational use of data.
The improved wavelet packet energy extraction method and the Hidden Markov model (HMM) are combined to obtain the transformer acoustic signals for preprocessing, extracting features and constructing an HMM classifier to achieve non-invasive diagnosis of faults.
Effectively extract the acoustic signal characteristics, reduce the calculation amount, improve the accuracy and efficiency of transformer fault diagnosis, can quickly identify different fault conditions, and improve equipment management efficiency.
Smart Images

Figure CN119086058A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of electrical engineering technology, and more particularly to a non-intrusive fault diagnosis method and system using transformer acoustic signals. Background Art
[0002] Power transformers are the core equipment of the power system. Whether they can operate safely and reliably plays a vital role in the operation of the entire power supply and distribution system. Among them, the core and windings are important components of the transformer. It is inevitable that failures will occur during long-term operation. Relevant statistics show that core and winding failures account for more than half of the total transformer failures. In view of this, it is necessary to strengthen the monitoring of core and winding failures.
[0003] At present, the more mature transformer online detection methods include: dissolved gas analysis, infrared temperature measurement, partial discharge online detection, frequency response analysis, vibration analysis, etc. These traditional methods help to ensure the safe operation of the power system, but the non-intrusive intelligent online detection technology that has received widespread attention in recent years has the advantages of simple equipment, fast measurement speed, non-contact measurement, and no impact on the normal operation of the equipment.
[0004] Transformer soundprint fault diagnosis usually takes the sound signal emitted during normal operation as the benchmark. The collected real-time operating status sound signal is compared with the benchmark signal to qualitatively identify the transformer operating condition. The transformer's sound signal mainly comes from the winding and core, so it can directly reflect changes in its internal structure, such as internal discharge and loose device screws. This is something that the above-mentioned traditional detection methods cannot accomplish.
[0005] However, due to the expansion of the scale of power systems and the increasing diversification of data sources, the amount of information contained in acoustic signals is huge, and the amount of data required for acoustic signal analysis is extremely large, which makes the rational use of data encounter bottlenecks. Therefore, more and more researchers have begun to study algorithm models that can be applied to transformer voiceprint diagnosis, hoping to combine online detection technology with machine learning theory, and use advanced models or algorithms to quickly and accurately extract effective feature information from massive and multivariate data to evaluate and describe the transformer operating conditions.
[0006] Therefore, how to perform non-invasive online intelligent detection of transformers based on sound signal analysis to quickly and accurately identify the operating status of power transformers in order to effectively improve equipment management efficiency is an urgent problem that technical personnel in this field need to solve. Summary of the invention
[0007] In view of this, the present invention provides a non-intrusive fault diagnosis method and system using transformer acoustic signals to solve the technical problems mentioned in the background technology.
[0008] In order to achieve the above object, the present invention adopts the following technical solution:
[0009] A non-intrusive fault diagnosis method using transformer acoustic signals comprises the following steps:
[0010] S1. Collect and pre-process the sound signals emitted by the transformer under different operating conditions;
[0011] S2. Based on the different energy distribution of each frequency band of the sound signal emitted by the transformer under different working conditions, the sound signal characteristics are extracted by improving the wavelet packet energy extraction method;
[0012] S3. The extracted feature quantity is input into the HMM classifier which has been trained and constructed by the hidden Markov model to obtain the target fault classification result.
[0013] Preferably, the preprocessing process of the transformer sound signal in step S1 includes frame processing and windowing processing;
[0014] The frame processing is to divide the speech signal into short frames, regard each frame signal as a steady-state signal, and set the adjacent frames to overlap partially to make the parameters of each frame transition smoothly;
[0015] The windowing process is to add a Hamming window to the sound signal.
[0016] Preferably, the formula of the Hamming window is:
[0017]
[0018] Where a is a constant. Changing the value of a can produce different Hamming windows.
[0019] Preferably, the specific contents of step S2 include:
[0020] S21. Perform a wavelet packet decomposition and reconstruction on the preprocessed sound signal;
[0021] S22. Calculate the energy value of each frequency band after decomposition, perform a second wavelet packet decomposition and reconstruction according to the energy proportion of the first frequency band, and calculate the energy value of each frequency band after decomposition;
[0022] S23. Perform normalized percentage processing on the feature vector according to the energy value to obtain the sound signal feature vector.
[0023] Preferably, the specific content of step S21 is:
[0024] Decompose the signal f(x) into i layers, is the i-th wavelet packet coefficient on the j-th layer, then:
[0025]
[0026]
[0027]
[0028] Where t=1,2,…,2 J , i=1,2,…,2 J , J = log2N, h k-2J is a low-pass filter, g k-2J is a high-pass filter;
[0029] The reconstruction formula of the i-th layer wavelet packet is:
[0030]
[0031] Where, j = J, J-1, ..., 1, 0;
[0032] The energy calculation formula for the i-th interval in the N frequency bands decomposed in step S22 is:
[0033]
[0034] Among them, f(x) is the original signal, E i is the energy of the ith frequency band, P i is the discrete point amplitude, P i (k) represents the kth wavelet packet coefficient corresponding to the node after wavelet packet decomposition;
[0035] The feature vector in step S23 is:
[0036]
[0037] Preferably, the specific contents of step S3, completing model training and constructing an HMM classifier through a hidden Markov model, include:
[0038] S31. Select initial parameters according to the connection mode of left-right Markov chain. (l) Under the condition, O (l)= O1 (l) , O2 (l) ···,O T (l) , l = 1, ···, L, use the forward-backward algorithm to calculate the probability of the observation sequence occurring;
[0039] S32. Under the condition of only obtaining the observation sequence, the Baum-Welch algorithm is used to initialize the model parameters and modify the reestimation formula to obtain the reestimation formula for multiple observations;
[0040] S33. Update the parameters through the re-estimation formula, and after multiple iterations, find the optimal solution for the parameters that maximize the probability of the observed value. The model parameters at this time are the final model parameters, and the model training is realized;
[0041] S34. Use the trained model to calculate the probability of the observation sequence, and determine the corresponding operating state by comparing the solved probability to generate fault diagnosis, thereby obtaining the constructed HMM classifier.
[0042] Preferably, calculating the probability of occurrence of the observation sequence using the forward-backward algorithm includes forward probability calculation and backward probability calculation; wherein:
[0043] The forward probability is:
[0044]
[0045] The backward probability is:
[0046]
[0047] The probability of L independent observation sequences occurring is:
[0048]
[0049] Among them, λ=(π,A,B), is the model parameter, π is the initial probability vector, and the matrix A=(a ij ) N×N and B=(b ik ) N×M It describes the probability transfer relationship between hidden states and between hidden states and observed values, π i is the probability of the initial i-th state appearing, b i (o1) is the probability under the observation state o1.
[0050] Preferably, the modified reestimation formula to obtain the reestimation formula for multiple observations is specifically:
[0051]
[0052] in, is the probability of the initial i-th state appearing after normalization, is the initial value of the forward probability of the lth sequence at the i-th state after normalization, is the initial value of the backward probability of the lth sequence at the i-th state after normalization;
[0053]
[0054] in, is the normalized forward probability of the lth sequence in the i-th state at time t, aij is the transition matrix from the jth state at time t to the ith state at time t, is the normalized value of the lth sequence in state j at time t+1 t+1 The probability of the observed situation occurring, is the normalized backward probability of the lth sequence at the jth state at time t+1;
[0055]
[0056] in, is the normalized forward probability of the lth sequence in the i-th state at time t, is the normalized backward probability of the lth sequence at the jth state at time t.
[0057] A non-invasive fault diagnosis system using transformer acoustic signals, based on the non-invasive fault diagnosis method using transformer acoustic signals, comprises: an acoustic signal acquisition module, a preprocessing module, an acoustic signal feature extraction module, an HMM classifier and a model training module;
[0058] A sound signal acquisition module is used to collect sound signals emitted by the transformer under different operating conditions;
[0059] A preprocessing module, used for preprocessing the collected sound signals;
[0060] The sound signal feature extraction module is used to extract the sound signal features by improving the wavelet packet energy extraction method based on the different energy distributions of each frequency band of the sound signal emitted by the transformer under different working conditions;
[0061] Model training module, used to complete model training through hidden Markov model and build HMM classifier;
[0062] The HMM classifier is used to input the extracted feature quantity into the constructed HMM classifier to obtain the target fault classification result.
[0063] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements a non-intrusive fault diagnosis method using transformer acoustic signals.
[0064] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a non-invasive fault diagnosis method and system using transformer acoustic signals. Based on the simplicity and non-invasive characteristics of acoustic signal acquisition, and combined with the rapidity and efficiency of hidden Markov model training, the method has better performance in using acoustic signals for transformer fault diagnosis. Specifically, the IWPE-HMM framework uses the hidden Markov model as the backbone framework, and uses the improved wavelet packet energy extraction method to effectively extract feature information from the acoustic signal and reduce the amount of calculation, thereby being able to well distinguish different transformer fault conditions. BRIEF DESCRIPTION OF THE DRAWINGS
[0065] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0066] Figure 1 A schematic flow chart of a non-invasive fault diagnosis method using transformer acoustic signals provided by the present invention;
[0067] Figure 2 A schematic diagram of acoustic signal framing provided by the present invention;
[0068] Figure 3 Schematic diagram of the improved wavelet packet energy extraction provided by the present invention.
[0069] Figure 4 This is a transformer fault diagnosis flow chart provided by the present invention.
[0070] Figure 5 This is a schematic diagram of the left-right HMM structure provided by the present invention. DETAILED DESCRIPTION
[0071] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0072] The embodiment of the present invention discloses a non-intrusive fault diagnosis method using transformer acoustic signals, such as Figure 1 , including the following steps:
[0073] S1. Collect and pre-process the sound signals emitted by the transformer under different operating conditions;
[0074] S2. Based on the different energy distribution of each frequency band of the sound signal emitted by the transformer under different working conditions, the sound signal characteristics are extracted by improving the wavelet packet energy extraction method;
[0075] S3. The extracted feature quantity is input into the HMM classifier which has been trained and constructed by the hidden Markov model to obtain the target fault classification result.
[0076] In order to further implement the above technical solution, Figure 2 , the preprocessing process of the transformer sound signal in step S1 includes frame processing and windowing processing;
[0077] In the range of 10 to 30 ms, the characteristic parameters of speech spectrum remain basically unchanged. The frame processing is to divide the speech signal into short frames, and regard each frame signal as a steady-state signal. In order to avoid missing information between frames, the adjacent frames are set to overlap partially to make the parameters of each frame transition smoothly.
[0078] In order to reduce signal distortion and improve the continuity between frames, the windowing process is to add a Hamming window to the sound signal.
[0079] In order to further implement the above technical solution, the formula of the Hamming window is:
[0080]
[0081] Where a is a constant. Changing the value of a can produce different Hamming windows.
[0082] In practical applications, the energy distribution of each frequency band of the sound signal emitted by the transformer under different working conditions is different, so the wavelet packet frequency band energy parameters can be used to construct the characteristic quantity of the sound signal. However, the frequency of the transformer body sound and mechanical fault sound is mainly distributed in the lower frequency part below 2000Hz, and the sampling rate of the acquisition equipment is mostly tens of kilohertz. If the decomposition is performed directly according to the wavelet packet frequency band energy method, when the number of decomposition layers is small, the frequency band range contained in the front frequency band is wide, and the transformer sound signal is at low frequency, so the energy proportion will be large, the feature is not obvious, and it is not conducive to subsequent identification; if the number of decomposition layers is increased, the number of frequency bands will increase exponentially, the amount of calculation will increase dramatically, and the frequency distribution characteristics of the transformer sound signal indicate that there is no need to pay attention to the high-frequency part. Too many layers will increase the difficulty of identification. The limitations of the wavelet algorithm will cause the sound information of the high-frequency part to be completely lost. To solve this problem, an improved wavelet packet algorithm is proposed.
[0083] In order to further implement the above technical solution, the specific contents of step S2 include:
[0084] S21. Perform a wavelet packet decomposition and reconstruction on the preprocessed sound signal;
[0085] S22. Calculate the energy value of each frequency band after decomposition, perform a second wavelet packet decomposition and reconstruction according to the energy proportion of the first frequency band, and calculate the energy value of each frequency band after decomposition;
[0086] S23. Perform normalized percentage processing on the feature vector according to the energy value to obtain the sound signal feature vector.
[0087] In order to further implement the above technical solution, Figure 3 , the specific content of step S21 is:
[0088] Decompose the signal f(x) into i layers, is the i-th wavelet packet coefficient on the j-th layer, then:
[0089]
[0090]
[0091]
[0092] Where t=1,2,…,2 J , i=1,2,…,2 J , J = log2N, h k-2J is a low-pass filter, g k-2J is a high-pass filter;
[0093] The reconstruction formula of the i-th layer wavelet packet is:
[0094]
[0095] Where, j = J, J-1, ..., 1, 0;
[0096] The energy calculation formula for the i-th interval in the N frequency bands decomposed in step S22 is:
[0097]
[0098] Among them, f(x) is the original signal, E i is the energy of the ith frequency band, P i is the discrete point amplitude, P i (k) represents the kth wavelet packet coefficient corresponding to the node after wavelet packet decomposition;
[0099] The feature vector in step S23 is:
[0100]
[0101] In order to further implement the above technical solution, Figure 4, step S3, the specific contents of completing model training and constructing HMM classifier through hidden Markov model include:
[0102] S31. Select initial parameters according to the connection mode of left-right Markov chain. (l) Under the condition, O (l)= O1 (l) , O2 (l) ···,O T (l) , l = 1, ···, L, use the forward-backward algorithm to calculate the probability of the observation sequence occurring;
[0103] S32. Under the condition of only obtaining the observation sequence, the Baum-Welch algorithm is used to initialize the model parameters and modify the reestimation formula to obtain the reestimation formula for multiple observations;
[0104] S33. The parameters are updated by re-estimating the formula, and the parameters with the highest probability of the observed value are optimized after multiple iterations. The model parameters at this time are the final model parameters, and the model training is implemented to obtain the constructed HMM classifier;
[0105] S34. Use the trained model to calculate the probability of the observation sequence, and determine the corresponding operating state by comparing the solved probability to generate fault diagnosis, thereby obtaining the constructed HMM classifier.
[0106] In this embodiment, the forward-backward algorithm calculation formula is as follows:
[0107] Initialize the forward probability:
[0108] Initial value, to calculate the probability of being in each hidden state at the first time point:
[0109] α1(i)=π i b i (o1), 1≤i≤N
[0110] Among them, π i represents the probability of the initial i-th state appearing, b i (o1) is the probability under the observation state o1;
[0111] Recursion:
[0112]
[0113] Among them, α t is the observation sequence to time t, and the probability of being in state j at time t, a ij is the transition matrix from the jth state at time t to the ith state at time t+1; b j (ot+1 ) is the probability of the corresponding observation situation occurring in state j;
[0114] Termination, sum the probabilities of all states:
[0115]
[0116] Similarly, initialize the backward probability:
[0117] β T (i)=1
[0118] Recursion:
[0119]
[0120] Among them, b j (ot +1 ) is the observation sequence at time t+1 (ot +1 ), βt +1 (j) is the backward probability of the observation sequence after state j;
[0121] Backward Probability:
[0122] P(O|λ)=π i b i (o1)β1(i);
[0123] When λ = (π, A, B) and the observation sequence are known, the probability that the HMM is in state i at time t and in state j at time t+1 is defined as:
[0124] ξ t (i,j)=P(q t =i,q t+1 =j|O,λ)
[0125] Then we can deduce:
[0126] ξ t (i, j) = |α t (i)a ij b j (o t+1 )β t+1 (j)]| / P(O|λ)
[0127] Therefore, the probability of being in state i at time t is:
[0128]
[0129] Update the parameters using the revaluation formula:
[0130]
[0131]
[0132]
[0133] in, represents the expected number of transitions from state i, represents the expected number of transitions from state i to state j;
[0134] According to the observation sequence O and the set initial parameter λ0, the new parameter is obtained by re-estimating the formula Under the new parameter model, This indicates that the new model parameters after reestimation represent the observed sequence O better than the original parameters, and then the reestimation process is repeated until Until convergence, the model parameters at this time are the final model parameters.
[0135] In combination with the actual operation of the transformer, a left-right Markov chain is adopted. Each state can only be transferred from the left to the right. In order to better describe the dynamic change process of the transformer operation, an intermediate state is introduced between the previous and next states as a transition. Since the Markov chain is represented by the state transfer matrix A and the initial probability vector π, its initial parameters should be selected according to the connection method of the left-right Markov chain.
[0136] When using the above-mentioned left-right Markov chain, more than one observation sequence is used when training the model in the actual description of the state evolution process, such as Figure 5 , at this time, it is also necessary to make corrections to the revaluation formula mentioned above.
[0137] Assume that L independent observation sequences O are used (l) , O (l)= O1 (l) , O2 (l) ···,O T (l) , l = 1, ···, L, at this time, the probability of the observed sequence occurring is:
[0138]
[0139] Among them, λ=(π,A,B), is the model parameter, π is the initial probability vector, and the matrix A=(a ij ) N×N and B=(b ik ) N×M It describes the probability transfer relationship between hidden states and between hidden states and observed values respectively;
[0140] A complete HMM is usually described by the following five parameters, where N and M are the number of hidden states and observable values, respectively; π is the initial probability vector; and the matrix A = (a ij ) N×N and B=(b ik ) N×M It describes the probability transfer relationship between hidden states and between hidden states and observation values respectively; since N and M remain unchanged during the model training process, it is generally abbreviated as λ = (π, A, B).
[0141] At the same time, in order to prevent the problem of too small values in the calculation process, the forward and backward factors γ and μ involved in each step of the iteration are normalized and calibrated as α* and β*, thereby obtaining the multi-observation formula shown below.
[0142] In order to further implement the above technical solution, the revaluation formula is modified to obtain the revaluation formula of multiple observations:
[0143]
[0144] in, is the probability of the initial i-th state appearing after normalization, is the initial value of the forward probability of the lth sequence at the i-th state after normalization, is the initial value of the backward probability of the lth sequence at the i-th state after normalization;
[0145]
[0146] in, is the normalized forward probability of the lth sequence in the i-th state at time t, a ij is the transition matrix from the jth state at time t to the ith state at time t, is the normalized value of the lth sequence in state j at time t+1 t+1 The probability of the observed situation occurring, is the normalized backward probability of the lth sequence at the jth state at time t+1;
[0147]
[0148] in, is the normalized forward probability of the lth sequence in the i-th state at time t, is the normalized backward probability of the lth sequence at the jth state at time t;
[0149] The model training is completed according to the revised re-estimation formula and the HMM classifier is constructed for fault diagnosis.
[0150] In another embodiment, a capacitive microphone is used as an acoustic signal acquisition device to collect data with a sampling rate of 24kHz. The acquisition object is a dry-type transformer used for power distribution with a voltage level of 10kV. In addition, in addition to collecting acoustic signals of normal operation of the transformer, acoustic signal data of loose transformer core and winding faults and multi-point grounding faults of the transformer core are also collected.
[0151] In the acoustic signal preprocessing, the frame length is set to 500ms and the frame shift is set to half of the frame length to ensure the continuity between acoustic signal frames. In the windowing process, the value of a is selected as a 0.46 Hamming window.
[0152] In improving the wavelet packet feature extraction, thanks to the advantages of wavelet packet decomposition, it is not necessary to set the decomposition according to a fixed number of decomposition layers, but to obtain very accurate feature information according to the difference characteristics of different sound signals. Based on the premise of the spectral distribution characteristics of the signal collected for the object, the number of layers for the first decomposition is taken as 4, and the number of layers for the second decomposition is taken as 3. When used for other objects, specific adjustments can be made according to your needs.
[0153] After the feature information is extracted, it is randomly divided into two data sets, one for training and the other for testing. During the test, a set of known fault data is first used to train its corresponding classification model as part of the HMM classifier. Then the training of other parts is completed to build a complete classifier. Finally, the data of the test set is input into the constructed classifier to obtain the final fault diagnosis result.
[0154] A non-intrusive fault diagnosis system using transformer acoustic signals, based on a non-intrusive fault diagnosis method using transformer acoustic signals, comprising: an acoustic signal acquisition module, a preprocessing module, an acoustic signal feature extraction module, an HMM classifier and a model training module;
[0155] A sound signal acquisition module is used to collect sound signals emitted by the transformer under different operating conditions;
[0156] A preprocessing module, used for preprocessing the collected sound signals;
[0157] The sound signal feature extraction module is used to extract the sound signal features by improving the wavelet packet energy extraction method based on the different energy distributions of each frequency band of the sound signal emitted by the transformer under different working conditions;
[0158] Model training module, used to complete model training through hidden Markov model and build HMM classifier;
[0159] The HMM classifier is used to input the extracted feature quantity into the constructed HMM classifier to obtain the target fault classification result.
[0160] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements a non-intrusive fault diagnosis method using transformer acoustic signals.
[0161] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0162] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A non-intrusive fault diagnosis method using transformer acoustic signals, It is characterized in that The following steps are involved: S1. Collect and pre-process the sound signals emitted by the transformer under different operating conditions; S2. Based on the different energy distribution of each frequency band of the sound signal emitted by the transformer under different working conditions, the sound signal characteristics are extracted by improving the wavelet packet energy extraction method; S3. The extracted feature quantity is input into the HMM classifier which has been trained and constructed by the hidden Markov model to obtain the target fault classification result.
2. A non-intrusive fault diagnosis method using transformer acoustic signals according to claim 1, It is characterized in that The preprocessing process of the transformer sound signal in step S1 includes frame processing and windowing processing; The frame processing is to divide the speech signal into short frames, regard each frame signal as a steady-state signal, and set the adjacent frames to overlap partially to make the parameters of each frame transition smoothly; The windowing process is to add a Hamming window to the sound signal.
3. A non-intrusive fault diagnosis method using transformer acoustic signals according to claim 2, It is characterized in that The formula for the Hamming window is: Where a is a constant. Changing the value of a can produce different Hamming windows.
4. The non-intrusive fault diagnosis method using transformer acoustic signals according to claim 1, It is characterized in that The specific contents of step S2 include: S21. Perform a wavelet packet decomposition and reconstruction on the preprocessed sound signal; S22. Calculate the energy value of each frequency band after decomposition, perform a second wavelet packet decomposition and reconstruction according to the energy proportion of the first frequency band, and calculate the energy value of each frequency band after decomposition; S23. Perform normalized percentage processing on the feature vector according to the energy value to obtain the sound signal feature vector.
5. A non-intrusive fault diagnosis method using transformer acoustic signals according to claim 4, It is characterized in that The specific content of step S21 is: Decompose the signal f(x) into i layers, is the i-th wavelet packet coefficient on the j-th layer, then: Where t=1,2,…,2 J , i=1,2,…,2 J , J = log 2 N,h k-2J is a low-pass filter, g k-2J is a high-pass filter; The reconstruction formula of the i-th layer wavelet packet is: Where, j = J, J-1, ..., 1, 0; The energy calculation formula for the i-th interval in the N frequency bands decomposed in step S22 is: Among them, f(x) is the original signal, E i is the energy of the ith frequency band, P i is the discrete point amplitude, P i (k) represents the kth wavelet packet coefficient corresponding to the node after wavelet packet decomposition; The feature vector in step S23 is:
6. The non-intrusive fault diagnosis method using transformer acoustic signals according to claim 1, It is characterized in that Step S3, completing model training through hidden Markov model and constructing HMM classifier includes: S31. Select initial parameters according to the connection mode of left-right Markov chain. (l) Under the condition, O (l) =O 1 (l) , O 2 (l) ···,O T (l) , l = 1, ···, L, use the forward-backward algorithm to calculate the probability of the observation sequence occurring; S32. Under the condition of only obtaining the observation sequence, the Baum-Welch algorithm is used to initialize the model parameters and modify the reestimation formula to obtain the reestimation formula for multiple observations; S33. Update the parameters through the re-estimation formula, and after multiple iterations, find the optimal solution for the parameters that maximize the probability of the observed value. The model parameters at this time are the final model parameters, and the model training is realized; S34. Use the trained model to calculate the probability of the observation sequence, and determine the corresponding operating state by comparing the solved probability to generate fault diagnosis, thereby obtaining the constructed HMM classifier.
7. A non-intrusive fault diagnosis method using transformer acoustic signals according to claim 6, It is characterized in that The forward-backward algorithm is used to calculate the probability of the occurrence of the observation sequence, including forward probability calculation and backward probability calculation; where: The forward probability is: The backward probability is: The probability of L independent observation sequences occurring is: Among them, λ=(π,A,B), is the model parameter, π is the initial probability vector, and the matrix A=(a ij ) N×N and B=(b ik ) N×M It describes the probability transfer relationship between hidden states and between hidden states and observed values, π i is the probability of the initial i-th state appearing, b i (o 1 ) is o 1 The probability of the observed state.
8. The non-intrusive fault diagnosis method using transformer acoustic signals according to claim 6, It is characterized in that The modified revaluation formula to obtain the revaluation formula for multiple observations is as follows: in, is the probability of the initial i-th state appearing after normalization, is the initial value of the forward probability of the lth sequence at the i-th state after normalization, is the initial value of the backward probability of the lth sequence at the i-th state after normalization; in, is the normalized forward probability of the lth sequence in the i-th state at time t, a ij is the transition matrix from the jth state at time t to the ith state at time t, is the normalized value of the lth sequence in state j at time t+1 t+1 The probability of the observed situation occurring, is the normalized backward probability of the lth sequence at the jth state at time t+1; in, is the normalized forward probability of the lth sequence in the i-th state at time t, is the normalized backward probability of the lth sequence at the jth state at time t.
9. A non-intrusive fault diagnosis system using transformer acoustic signals. It is characterized in that A non-intrusive fault diagnosis method using transformer acoustic signals based on any one of claims 1 to 8, comprising: an acoustic signal acquisition module, a preprocessing module, an acoustic signal feature extraction module, an HMM classifier and a model training module; A sound signal acquisition module is used to collect sound signals emitted by the transformer under different operating conditions; A preprocessing module, used for preprocessing the collected sound signals; The sound signal feature extraction module is used to extract the sound signal features by improving the wavelet packet energy extraction method based on the different energy distributions of each frequency band of the sound signal emitted by the transformer under different working conditions; Model training module, used to complete model training through hidden Markov model and build HMM classifier; The HMM classifier is used to input the extracted feature quantity into the constructed HMM classifier to obtain the target fault classification result.
10. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the non-intrusive fault diagnosis method using transformer acoustic signals as described in any one of claims 1 to 8 is implemented.
Citation Information
Cited By
Photovoltaic box transformer substation abnormal operation state diagnosis and fault location method based on sound signals
CN121415805A
