Transformer fault voiceprint recognition method based on voiceprint denoising and DBO-SVM
By combining wavelet threshold denoising and variational mode decomposition, combined with the dung beetle algorithm to optimize the support vector machine parameters, a transformer fault diagnosis model is built, which solves the problem of lack of actual noise consideration in the voiceprint feature value in transformer fault recognition, and improves the accuracy of identification Rate and robustness enable online monitoring and cost reduction.
Patent Information
- Application Number
- CN202411024841.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-29
- Publication Date
- 2024-12-06
AI Technical Summary
In the prior art, in the identification of transformer faults, the voiceprint characteristic values lack consideration for actual noise, resulting in large identification errors, and strict position requirements for vibration sensors, high cost, and difficult to widely use in practical applications.
The wavelet threshold denoising and variational mode decomposition are combined, and the support vector machine parameters are optimized through the dung beetle algorithm to build a transformer fault diagnosis model to improve the accuracy of fault identification.
It improves the accuracy and robustness of the identification of the core looseness of the transformer, reduces the impact of other noises on the sound marks, reduces the cost, and realizes online monitoring.
Smart Images

Figure CN119091906A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of transformer fault detection, and in particular relates to a transformer fault voiceprint recognition method based on voiceprint denoising and DBO-SVM. Background Art
[0002] Transformers carry the important tasks of voltage and current conversion and voltage stabilization in the power grid. They play an irreplaceable role in the power system. Ensuring the normal operation of transformers is the premise for the safety and stability of the power system. As the power transformer operates in the power grid for longer, its internal mechanical structure will also change. Common transformer mechanical failures include loose windings, loose cores, and main insulation failures. These failures can lead to a series of hidden dangers such as accelerated insulation aging and poor anti-interference ability. According to statistics, most of the transformer damage failures are caused by transformer winding failures and loose cores. Therefore, in order to ensure the reliability of power supply to the power grid, it is of great significance to further monitor and study the transformer core windings.
[0003] There are currently many methods for identifying transformer faults, and the mainstream methods are oil testing, gas detection, and vibration signal monitoring. These methods are effective and reliable for transformer mechanical faults, but the collection of vibration signals has very strict requirements on the location of vibration sensors, and a small offset will result in a large error. However, the collection of vibration signals requires very strict requirements on the location of vibration sensors, and a small offset will result in a large error. In addition, a certain number of vibration sensors are required to measure vibration signals for large transformers, which greatly increases the cost and is difficult to be widely used in real life. The oil testing method cannot achieve real-time monitoring, and may have certain limitations for some faults that require timely response. Acoustic signals can achieve non-contact large-scale monitoring with relatively low cost, and can also achieve online monitoring.
[0004] In the prior art, the patent "A transformer core winding loose fault diagnosis method based on transformer vibration spectrum amplitude" application number / patent number: 202311688568.2, publicly designed a transformer core winding loose fault diagnosis method based on transformer vibration spectrum amplitude, mainly based on the transformer vibration amplitude platform, and gives the judgment result based on the vibration signal data under the transformer open circuit and short circuit conditions. The patent "Monitoring method and system under loose state of transformer core based on finite element theory" application number / patent number: 202311246790.7, publicly designed a transformer core loose state monitoring method and system based on finite element theory, and reflects the degree of core looseness through the change of elastic modulus of core preload to achieve real-time monitoring. The patent "A transformer soundprint abnormality detection method" application number / patent number: 202110872885.4, proposes a denoising model, extracts Mel spectrum features from the denoised sound signal, and uses the detection model G-MADE to score the features, and finally judges whether the transformer is correct based on the score.
[0005] The above patent considers that the abnormal soundprint has a certain reference significance for the loose core fault of the transformer, but the soundprint characteristic value adopted lacks consideration of other noises in the actual substation, which affects the actual collected soundprint results. At present, there is less raw data of transformer sound signals, and it is difficult to accumulate public data sets. In addition, the neural network algorithm has local minima and generalization ability. There are very strict requirements for the location of vibration sensors, and a small offset result will produce a large error.
[0006] Therefore, it is necessary to design a transformer fault voiceprint recognition method based on voiceprint denoising and DBO-SVM to solve the above problems. Summary of the invention
[0007] The technical problem to be solved by the present invention is to provide a transformer fault voiceprint recognition method based on voiceprint denoising and DBO-SVM. The method uses a combination of wavelet threshold denoising and variational mode decomposition to denoise the original voiceprint, and uses the dung beetle algorithm to optimize the fault diagnosis model parameters of the SVM that can solve the classification problem under small samples, thereby improving the accuracy of fault recognition.
[0008] In order to achieve the above technical effects, the technical solution adopted by the present invention is:
[0009] The transformer fault voiceprint recognition method based on voiceprint denoising and DBO-SVM includes the following steps:
[0010] S1, collects the soundprint signals of transformer core under different looseness conditions through no-load experiments;
[0011] S2, processing the collected voiceprint signal based on variational mode decomposition VMD and wavelet threshold denoising;
[0012] The signal is decomposed into 12 intrinsic mode function IMF components through VMD;
[0013] The IMF component with the largest kurtosis value deviation among the IMF components is selected for wavelet threshold denoising, and the denoised IMF component is reconstructed with the remaining unprocessed IMF components to obtain the denoised signal;
[0014] Extract the MFCC feature value of the voiceprint signal based on the Mel-frequency cepstral coefficient;
[0015] S3, based on the MFCC eigenvalues of the voiceprint signal, the Dung Beetle Algorithm (DBO) is used to optimize the support vector machine (SVM) parameters to build a transformer core loosening diagnosis model, and the optimal kernel function and penalty factor are obtained through the optimization algorithm;
[0016] S4, dividing the voiceprint signal collected in step S1 into a training set and a test set according to a fixed ratio, training the transformer core loosening diagnosis model through the training set, and testing the transformer core loosening diagnosis model through the test set, finally obtaining a trained transformer core loosening diagnosis model, and performing transformer fault voiceprint recognition through the transformer core loosening diagnosis model.
[0017] Preferably, in step S2, wavelet threshold denoising includes:
[0018] Select appropriate wavelet function and decomposition level to perform wavelet decomposition on the IMF component to obtain wavelet coefficients, set the soft threshold function as the threshold function, retain the required wavelet coefficients, and reconstruct the signal by inverse transform using the processed wavelet coefficients.
[0019] Preferably, in step S2, extracting MFCC feature values of the voiceprint signal based on Mel-frequency cepstral coefficients includes:
[0020] The collected transformer voiceprint signal is preprocessed to obtain a time domain frame signal. The preprocessing includes pre-emphasis, framing and windowing. The window function used in windowing is Hamming window.
[0021] The preprocessed voiceprint signal is subjected to fast Fourier transform to obtain a linear spectrum, and then the time-domain framed signal after fast Fourier transform is passed through a Mel filter bank and the logarithmic energy is calculated, and finally the required MFCC eigenvalue parameters are obtained through discrete cosine transform.
[0022] Preferably, in step S3, the transformer core loosening diagnosis model is constructed by optimizing the support vector machine parameters using the dung beetle algorithm, and the optimal kernel function and penalty factor are obtained by the optimization algorithm, including:
[0023] Population parameter initialization:
[0024] Set the population size, maximum number of iterations, upper and lower bounds, and set the ratio of the four populations of rolling dung beetles, breeding dung beetles, small dung beetles, and thieving dung beetles. The preferred ratio is 6:6:7:11.
[0025] Calculate the fitness of the dung beetle in each iteration, and select the error rate between the SVM prediction result and the actual result as the fitness function;
[0026] Update the position of the dung beetle and make boundary judgments for each dung beetle; calculate the fitness and optimal solution of the dung beetle in each iteration, determine whether the accuracy requirements and the number of iterations are met, and update the position and speed of each particle individual;
[0027] Based on the new optimal solution, the position and speed of the dung beetle are updated until the conditions meet the accuracy requirements and the number of iterations, and the optimal parameters are obtained;
[0028] Based on the MFCC eigenvalues and the optimal parameter support vector machine, and trained with the training set, the transformer core loose fault diagnosis model is obtained.
[0029] Preferably, the transformer fault voiceprint recognition system based on voiceprint denoising and DBO-SVM includes a voiceprint processing module, which is used to perform noise reduction processing on the collected on-site voiceprint to obtain a clean single voiceprint;
[0030] Information extraction module, used to extract MFCC feature values of voiceprint signals;
[0031] A model training module, used to build the transformer core loosening diagnosis model, and train the transformer core loosening diagnosis model according to the training set divided by the voiceprint signal data;
[0032] The identification and judgment module is used to perform fault diagnosis based on the transformer core loosening fault diagnosis model.
[0033] Preferably, the voiceprint processing module extracts the transformer voiceprint signal feature value based on the Mel-frequency cepstral coefficient, specifically:
[0034] The signal is decomposed into 12 intrinsic mode function IMF components through VMD;
[0035] The IMF component with the largest kurtosis value deviation among the IMF components is selected for wavelet threshold denoising, and the appropriate wavelet function and decomposition layer number are selected to perform wavelet decomposition on the signal to obtain the wavelet coefficients. The soft threshold function is set as the threshold function, the required wavelet coefficients are retained, and the processed wavelet coefficients are used to reconstruct the signal through inverse transformation; finally, the denoised IMF component and the remaining unprocessed IMF components are reconstructed.
[0036] The beneficial effects of the present invention are:
[0037] 1. The present invention uses variational mode decomposition and wavelet threshold denoising to process the original soundprint, and then optimizes the support vector machine parameters through the dung beetle algorithm to construct a fault diagnosis model, which improves the accuracy and robustness of transformer core loosening faults and has a certain effect on actual production.
[0038] 2. The present invention measures the voiceprint signals of the transformer core under different loose conditions through no-load experiments, extracts the characteristic values of the transformer voiceprint signals based on the Mel-frequency cepstral coefficients, and finally optimizes the support vector machine parameters through the dung beetle optimization algorithm. It has high fault monitoring and recognition accuracy and can greatly reduce the influence of other noises on the original voiceprint. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 It is a schematic diagram of a flow chart of the present invention;
[0040] Figure 2 This is a time domain diagram of a transformer's normal operating soundprint in an embodiment of the present invention;
[0041] Figure 3 is a kurtosis diagram obtained based on variational mode decomposition after adding white noise in an embodiment of the present invention;
[0042] Figure 4 is a time domain diagram of transformer voiceprint after wavelet threshold denoising in an embodiment of the present invention;
[0043] Figure 5 is a flow chart of MFCC extraction in an embodiment of the present invention;
[0044] Figure 6 1 is a schematic diagram of the principle of optimizing support vector machine parameters using a dung beetle algorithm in an embodiment of the present invention;
[0045] Figure 7a to Figure 7e : is a three-dimensional graph of the transformer core soundprint characteristic values in an embodiment of the present invention; wherein a represents the MFCC feature in a normal state, b represents the MFCC feature when the core is 25% loose, c represents the MFCC feature when the core is 50% loose, d represents the MFCC feature when the core is 75% loose, and e represents the MFCC feature when the core is 100% loose;
[0046] Figure 8 is a diagram of model recognition results in an embodiment of the present invention;
[0047] Fig. 9 is a schematic diagram of a system framework provided by the system in an embodiment of the present invention;
[0048] Fig.10 It is a conceptual diagram of the boundary strategy in an embodiment of the present invention. DETAILED DESCRIPTION
[0049] Embodiment 1:
[0050] like Figure 1 As shown, the transformer fault voiceprint recognition method based on voiceprint denoising and DBO-SVM includes the following steps:
[0051] S1, collects the soundprint signals of transformer core under different looseness conditions through no-load experiments;
[0052] S2, processing the collected voiceprint signal based on variational mode decomposition VMD and wavelet threshold denoising;
[0053] The signal is decomposed into 12 intrinsic mode function IMF components through VMD;
[0054] The IMF component with the largest kurtosis value deviation among the IMF components is selected for wavelet threshold denoising, and the denoised IMF component is reconstructed with the remaining unprocessed IMF components to obtain the denoised signal;
[0055] Extract the MFCC feature value of the voiceprint signal based on the Mel-frequency cepstral coefficient;
[0056] S3, based on the MFCC eigenvalues of the voiceprint signal, the Dung Beetle Algorithm (DBO) is used to optimize the support vector machine (SVM) parameters to build a transformer core loosening diagnosis model, and the optimal kernel function and penalty factor are obtained through the optimization algorithm;
[0057] S4, dividing the voiceprint signal collected in step S1 into a training set and a test set according to a fixed ratio, training the transformer core loosening diagnosis model through the training set, and testing the transformer core loosening diagnosis model through the test set, finally obtaining a trained transformer core loosening diagnosis model, and performing transformer fault voiceprint recognition through the transformer core loosening diagnosis model.
[0058] Preferably, in step S2, wavelet threshold denoising includes:
[0059] Select appropriate wavelet function and decomposition level to perform wavelet decomposition on the IMF component to obtain wavelet coefficients, set the soft threshold function as the threshold function, retain the required wavelet coefficients, and reconstruct the signal by inverse transform using the processed wavelet coefficients.
[0060] Preferably, in step S2, extracting MFCC feature values of the voiceprint signal based on Mel-frequency cepstral coefficients includes:
[0061] The collected transformer voiceprint signal is preprocessed to obtain a time domain frame signal. The preprocessing includes pre-emphasis, framing and windowing. The window function used in windowing is Hamming window.
[0062] The preprocessed voiceprint signal is subjected to fast Fourier transform to obtain a linear spectrum, and then the time-domain framed signal after fast Fourier transform is passed through a Mel filter bank and the logarithmic energy is calculated, and finally the required MFCC eigenvalue parameters are obtained through discrete cosine transform.
[0063] Preferably, in step S3, the transformer core loosening diagnosis model is constructed by optimizing the support vector machine parameters using the dung beetle algorithm, and the optimal kernel function and penalty factor are obtained by the optimization algorithm, including:
[0064] Population parameter initialization:
[0065] Set the population size, maximum number of iterations, upper and lower bounds, and set the ratio of the four populations of rolling dung beetles, breeding dung beetles, small dung beetles, and thieving dung beetles. The preferred ratio is 6:6:7:11.
[0066] Calculate the fitness of the dung beetle in each iteration, and select the error rate between the SVM prediction result and the actual result as the fitness function;
[0067] Update the position of the dung beetle and make boundary judgments for each dung beetle; calculate the fitness and optimal solution of the dung beetle in each iteration, determine whether the accuracy requirements and the number of iterations are met, and update the position and speed of each particle individual;
[0068] Based on the new optimal solution, the position and speed of the dung beetle are updated until the conditions meet the accuracy requirements and the number of iterations, and the optimal parameters are obtained;
[0069] Based on the MFCC eigenvalues and the optimal parameter support vector machine, and trained with the training set, the transformer core loose fault diagnosis model is obtained.
[0070] Preferably, the transformer fault voiceprint recognition system based on voiceprint denoising and DBO-SVM includes a voiceprint processing module, which is used to perform noise reduction processing on the collected on-site voiceprint to obtain a clean single voiceprint;
[0071] Information extraction module, used to extract MFCC feature values of voiceprint signals;
[0072] A model training module, used to build the transformer core loosening diagnosis model, and train the transformer core loosening diagnosis model according to the training set divided by the voiceprint signal data;
[0073] The identification and judgment module is used to perform fault diagnosis based on the transformer core loosening fault diagnosis model.
[0074] Preferably, the voiceprint processing module extracts the transformer voiceprint signal feature value based on the Mel-frequency cepstral coefficient, specifically:
[0075] The signal is decomposed into 12 intrinsic mode function IMF components through VMD;
[0076] The IMF component with the largest kurtosis value deviation among the IMF components is selected for wavelet threshold denoising, and the appropriate wavelet function and decomposition layer number are selected to perform wavelet decomposition on the signal to obtain the wavelet coefficients. The soft threshold function is set as the threshold function, the required wavelet coefficients are retained, and the processed wavelet coefficients are used to reconstruct the signal through inverse transformation; finally, the denoised IMF component and the remaining unprocessed IMF components are reconstructed.
[0077] Embodiment 2:
[0078] The transformer sound is mainly generated by the vibration of the transformer body and the vibration of the components installed on the transformer body. The transformer body vibration noise generally comes from the magnetostriction of the iron core and the leakage magnetic force between the iron core laminations. With the improvement of modern iron core lamination technology, the vibration caused by the leakage magnetic force can be ignored. As the transformer runs for a longer time, the tightness of the iron core will also change; studies have shown that the MFCC Mel-Cepstral coefficient is based on the human cochlear auditory model, so it can be used to obtain more soundprint information to diagnose the looseness of the transformer core fault.
[0079] Figure 2 This is a time domain diagram of the transformer operation soundprint collected and the spectrum diagram analyzed.
[0080] Figure 3 The figure shows a kurtosis diagram obtained based on variational mode decomposition after adding white noise, specifically:
[0081] The signal is decomposed into 12 intrinsic mode functions (IMFs) through VMD, and the IMF12 component with the largest kurtosis value deviation from 3 in the mixed signal is selected for wavelet threshold denoising.
[0082] Select appropriate wavelet function and decomposition layer number to perform wavelet decomposition on the signal, obtain wavelet coefficients, set soft threshold function as threshold function, retain required wavelet coefficients, use processed wavelet coefficients to reconstruct the signal through inverse transformation. Finally, reconstruct the denoised IMF components and the remaining unprocessed IMF components.
[0083] Figure 5 The flowchart for extracting MFCC feature parameters is shown as follows:
[0084] Step 1: Perform fast Fourier transform on the pre-processed acoustic signal;
[0085] Step 2: Pass through Mel filter;
[0086] Step 3: Perform logarithmic compression to obtain signal energy;
[0087] Step 4: Finally obtain the MFCC feature parameters by performing discrete cosine transformation on the signal.
[0088] Figure 6 The optimization of the support vector machine parameters based on the dung beetle optimization algorithm is specifically as follows:
[0089] Initialize population parameters, set population size, maximum number of iterations, upper and lower bounds, and set the population ratios of rolling dung beetles, breeding dung beetles, small dung beetles, and thieving dung beetles to 6:6:7:11; calculate the fitness of dung beetles in each iteration, and select the error rate between the SVM prediction result and the actual result as the fitness function;
[0090] Update the position of the dung beetle and make boundary judgments for each dung beetle; calculate the fitness and optimal solution of the dung beetle in each iteration, determine whether the accuracy requirements and the number of iterations are met, and update the position and speed of each particle individual;
[0091] Based on the new optimal solution, the position and speed of the dung beetle are updated until the conditions meet the accuracy requirements and the number of iterations, and the optimal parameters are obtained;
[0092] In this example, the transformer soundprint is processed for noise reduction to minimize the interference of external environmental noise. The soundprint feature value is extracted based on the Mel cepstral coefficient, and the required information is accurately obtained from the extracted soundprint. Finally, a three-dimensional image of the core looseness is obtained. Figure 7a to Figure 7e shown.
[0093] Finally, this example uses the dung beetle optimization algorithm to optimize the two parameters of the support vector machine. Selecting the appropriate penalty factor C and kernel function g can effectively improve the SVM classification accuracy.
[0094] To further illustrate the application scenario of the present invention, a specific application of this example is given, which includes:
[0095] 1. Process the collected voiceprint signal through variational mode decomposition (VMD) and wavelet threshold denoising:
[0096] The audio collected in the actual substation is affected by the surrounding uncontrollable environmental noise and is difficult to use directly. It is necessary to further denoise the audio. This article takes a certain actual voiceprint as an example. The specific process is as follows:
[0097] 1) Decompose the original voiceprint signal to obtain the intrinsic mode function u k (t):
[0098] u k (t) = A k (t)cosφ k (t); (1)
[0099] In the formula, φk (t) is a non-decreasing phase function; A k (t) is the envelope function, k = 1, 2, ..., n;
[0100] Gaussian smoothing is used to estimate the width, and constraints are introduced to minimize the sum of the bandwidths of each component, and the sum of each component is the original signal. The constraint model is:
[0101]
[0102] Among them, ω is the center frequency, k is the number of decomposition layers; u k ω k Respectively represent each eigenmode component and the center frequency corresponding to each mode component; represents the unit impulse function, t represents the time, The correction makes the spectrum of each modal component modulated to the corresponding baseband;
[0103] After constructing the variational problem, we solve it and introduce the augmented Lagrangian function to reconstruct equation (2) into an unconstrained variational problem:
[0104]
[0105] Among them, α is the quadratic penalty factor, which determines the accuracy of signal reconstruction; λ is the Lagrange multiplication operator; f(t) represents the input signal, and λ(t) represents the Lagrange multiplication operator corresponding to time;
[0106] Finally, the multiplication operator alternating direction method is used to solve the optimal solution of the model by updating the parameters:
[0107]
[0108] in, The current remaining amount Wiener filtering; n is the number of iterations; represents the Fourier transform of λ(t)
[0109] Further, Figure 4 The IMF12 component that deviates too much from 3 is denoised by wavelet threshold, and finally the denoised IMF12 and the remaining unprocessed IMF components are reconstructed to obtain the required voiceprint Figure 4 .
[0110] 2. Extract Mel cepstral coefficients:
[0111] Studies have shown that MFCC is based on the human cochlear auditory model, so it can be used to obtain more soundprint information to diagnose transformer core loosening faults; Figure 5 This is the MFCC extraction flow chart. The specific method is
[0112] as follows:
[0113] 1) Perform fast Fourier transform on the preprocessed voiceprint signal, transform the preprocessed time domain frame signal by fast Fourier transform (FFT) to obtain the linear spectrum, then pass it through the m-dimensional Mel filter, and calculate the logarithmic energy s(m) by formula (6):
[0114] s(m)=ln(X n (k) 2 H m (k)); (6)
[0115] Where m represents the dimension of the Mel filter, and this paper takes m=13; Xn(k) represents the linear spectrum after FFT transformation, and Hm(k) is the filter parameter:
[0116]
[0117] Where f(m) is the center frequency of the filter; finally, s(m) is transformed through discrete cosine equation (8) to obtain the required MFCC
[0118] Characteristic parameter C(n):
[0119]
[0120] Where n represents the order of the MFCC coefficient, usually 12-16; N represents the frame length, and M represents the number of filters.
[0121] Furthermore, the preprocessing of the voiceprint signal includes pre-emphasis, framing, and windowing, specifically:
[0122] Pre-emphasis: Pre-emphasis is used to increase the high-frequency components of the voiceprint signal to prevent the loss of high-frequency information due to attenuation during the propagation of the voiceprint. The expression is:
[0123] H(z)=1-μz -1 ; (9)
[0124] Where μ is the pre-emphasis coefficient, which is usually between 0.9375 and 0.97, and is taken as 0.95 in this paper;
[0125] Indicates the input signal.
[0126] Framing: Voiceprint signals have short-term stability. In order to ensure the continuity and autocorrelation between frames, framing is required. At the same time, if the frame length is too long, the signal change will obviously affect the accuracy of the feature vector, and if the frame length is too short, some information will be lost. In this paper, N= per frame is set to 25ms, and the overlap rate is set to 50%;
[0127] Windowing: Adding a window function to each frame avoids excessive distortion. This paper selects the Hamming window with less frequency leakage as the added window function; its formula is:
[0128]
[0129] Where N is the length of the Hamming window and n represents the number of points.
[0130] 3. Dung beetle optimization algorithm to optimize support vector machine parameters:
[0131] The parameter optimization process of the present invention is as follows Figure 6 As shown in the figure, the dung beetle optimization algorithm DBO is a heuristic optimization algorithm based on insect behavior, which takes into account both global exploration and local development, and has the characteristics of fast convergence speed and high solution accuracy.
[0132] By simulating the foraging, stealing, reproduction, ball rolling and dancing behaviors of dung beetles in the process of searching for food, the best solution is found by optimizing the solution of the problem, specifically:
[0133] The global space can be explored by rolling dung beetles. Their behaviors are divided into two modes: rolling and dancing:
[0134] Rolling ball means that in the absence of obstacles, the moving position of the dung beetle is affected by the light intensity, and the position update formula is:
[0135]
[0136] Where X i t represents the position of dung beetle i after the tth iteration, α is a natural coefficient assigned a value of -1 or 1, k∈(0,0.2] is a deviation coefficient, and b∈(0,1) is a constant value; X worse represents the global worst position.
[0137] When there is an obstacle in front, the new direction is obtained by constructing a tangent function by simulating the dancing behavior of the dung beetle.
[0138] The update formula is:
[0139]
[0140] where θ∈[0,π].
[0141] The egg-laying dung beetle will transport the dung ball to a suitable location to lay eggs, and limit its movement range to the safe area class to achieve local area development. The boundary selection strategy is used to simulate the safe egg-laying area:
[0142] Lb * =max(X gbest ×(1-R),Lb);(13)
[0143] Ub * =min(X gbest ×(1+R),Ub);(14)
[0144]
[0145] Where X gbest is the current local optimal position, Lb * is the lower bound of the spawning area, Ub * is the upper limit of the spawning area, Lb
[0146] is the lower bound of the optimization problem, Ub is the upper bound of the optimization problem, T max is the maximum number of iterations.
[0147] like Fig.10 The following is a conceptual diagram of the boundary strategy, where X gbest The black circles around the red circle are dung beetle eggs, and the remaining red circles are the upper and lower limits of the boundary; when the egg-laying area is determined, the dung beetle will lay eggs in this area. At this time, the position update formula is:
[0148]
[0149] where b 1 、b 2 is a 1×D independent random array, where D is the dimension of the optimization problem; B i t For the dung beetle i laying eggs in the t
[0150] The position after iterations.
[0151] When foraging, the dung beetle chooses the best area to forage. The position update formula is:
[0152]
[0153] Among them C 1 is a random number that follows a normal distribution, C 2 is a random vector belonging to (0,1); Indicates the position information of the dung beetle at the tth iteration.
[0154] The upper and lower limits of the optimal region are defined as:
[0155] Lb l =max(X best ×(1-R),Lb);(18)
[0156] Ub l =min(X best ×(1+R),Ub);(19)
[0157] Where X best is the global optimal solution, Lb l is the lower bound of the optimal foraging area, Ub l is the upper bound of the optimal foraging area.
[0158] According to the law of competition in nature, some dung beetles will steal the food of other dung beetles. When the thieving dung beetles are in the best position, best When nearby, it competes for food with other dung beetles. Its position X thief The update formula is:
[0159]
[0160] Where g is a 1×D independent random vector obeying a random distribution, and S is a constant.
[0161] 4. Transformer core loosening fault diagnosis model based on dung beetle algorithm optimized support vector machine:
[0162] The process of this example is as follows: the collected voiceprint signal is decomposed into 12 intrinsic mode functions (IMFs) through VMD, and the IMF12 components whose IMF component kurtosis value deviates too much from 3 in the mixed signal are selected for wavelet threshold denoising; the appropriate wavelet function and decomposition layer are selected to perform wavelet decomposition on the signal to obtain wavelet coefficients, and the soft threshold function is set as the threshold function, the required wavelet coefficients are retained, and the processed wavelet coefficients are used to reconstruct the signal through inverse transformation. Finally, the denoised IMF components are reconstructed with the remaining unprocessed IMF components.
[0163] The MFCC feature values are extracted from the denoised voiceprint signal. The data set is divided into a training set and a validation set using the cross-validation method. The model performance is then trained and evaluated under different parameter combinations. The trained model is tested with the validation set, and the classification accuracy is obtained as the performance evaluation indicator of the classifier.
[0164] In this specific example, the recognition accuracy of the fault model is as follows: Figure 8 As shown, the average accuracy rate reaches 96.25%. Under normal circumstances, the model recognition accuracy rate reaches 96.4%; when the core is 25% loose, the accuracy rate is 100%; when the core is 50% loose, the accuracy rate is 95.3%; when the core is 75% loose, the accuracy rate is 98.4%; when the core is 100% loose, the accuracy rate is 92.2%.
[0165] Fig. 9 This is a schematic diagram of the system framework provided by this system. The system includes: voiceprint processing module, information extraction module, model training module, and recognition and judgment module; among which:
[0166] The voiceprint processing module is used to perform noise reduction processing on the collected on-site voiceprint to obtain a clean single voiceprint;
[0167] The information extraction module is used to extract the voiceprint MFCC feature value;
[0168] The model training module is used to build the DBO-SVM model and train the DBO-SVM model according to the voiceprint sample data;
[0169] The identification and judgment module is used to perform fault diagnosis based on the transformer core loose fault diagnosis model.
[0170] This voiceprint recognition system has been improved to accurately monitor and identify transformer faults. The voiceprint signals of the transformer core under different loose conditions are measured through no-load experiments, and then the original voiceprints are processed through variational mode decomposition and wavelet threshold denoising. The characteristic values of the transformer voiceprint signals are extracted based on the Mel-frequency cepstral coefficients, and finally the fault diagnosis model is constructed by optimizing the support vector machine parameters through the dung beetle algorithm.
Claims
1. A transformer fault voiceprint recognition method based on voiceprint denoising and DBO-SVM is characterized by: The following steps are involved: S1, collects the soundprint signals of transformer core under different looseness conditions through no-load experiments; S2, processing the collected voiceprint signal based on variational mode decomposition VMD and wavelet threshold denoising; The signal is decomposed into 12 intrinsic mode function IMF components through VMD; The IMF component with the largest kurtosis value deviation among the IMF components is selected for wavelet threshold denoising, and the denoised IMF component is reconstructed with the remaining unprocessed IMF components to obtain the denoised signal; Extract the MFCC feature value of the voiceprint signal based on the Mel-frequency cepstral coefficient; S3, based on the MFCC eigenvalues of the voiceprint signal, the Dung Beetle Algorithm (DBO) is used to optimize the support vector machine (SVM) parameters to build a transformer core loosening diagnosis model, and the optimal kernel function and penalty factor are obtained through the optimization algorithm; S4, dividing the voiceprint signal collected in step S1 into a training set and a test set according to a fixed ratio, training the transformer core loosening diagnosis model through the training set, and testing the transformer core loosening diagnosis model through the test set, finally obtaining a trained transformer core loosening diagnosis model, and performing transformer fault voiceprint recognition through the transformer core loosening diagnosis model.
2. The transformer fault voiceprint recognition method based on voiceprint denoising and DBO-SVM according to claim 1 is characterized in that: In step S2, wavelet threshold denoising includes: Select appropriate wavelet function and decomposition level to perform wavelet decomposition on the IMF component to obtain wavelet coefficients, set the soft threshold function as the threshold function, retain the required wavelet coefficients, and reconstruct the signal by inverse transform using the processed wavelet coefficients.
3. The transformer fault voiceprint recognition method based on voiceprint denoising and DBO-SVM according to claim 1 is characterized in that: In step S2, extracting MFCC feature values of the voiceprint signal based on the Mel-frequency cepstral coefficients includes: The collected transformer voiceprint signal is preprocessed to obtain a time domain frame signal. The preprocessing includes pre-emphasis, framing and windowing. The window function used in windowing is Hamming window. The preprocessed voiceprint signal is subjected to fast Fourier transform to obtain a linear spectrum, and then the time-domain framed signal after fast Fourier transform is passed through a Mel filter bank and the logarithmic energy is calculated, and finally the required MFCC eigenvalue parameters are obtained through discrete cosine transform.
4. The transformer fault voiceprint recognition method based on voiceprint denoising and DBO-SVM according to claim 3 is characterized in that: In step S3, the dung beetle algorithm is used to optimize the support vector machine parameters to build a transformer core loosening diagnosis model, and the optimal kernel function and penalty factor are obtained by the optimization algorithm, including: Population parameter initialization: Set the population size, maximum number of iterations, upper and lower bounds, and set the proportions of the four populations: rolling dung beetles, breeding dung beetles, small dung beetles, and thieving dung beetles; Calculate the fitness of the dung beetle in each iteration, and select the error rate between the SVM prediction result and the actual result as the fitness function; Update the position of the dung beetle and make boundary judgments for each dung beetle; calculate the fitness and optimal solution of the dung beetle in each iteration, determine whether the accuracy requirements and the number of iterations are met, and update the position and speed of each particle individual; Based on the new optimal solution, the position and speed of the dung beetle are updated until the conditions meet the accuracy requirements and the number of iterations, and the optimal parameters are obtained; Based on the MFCC eigenvalues and the optimal parameter support vector machine, and trained with the training set, the transformer core loose fault diagnosis model is obtained.
5. The transformer fault voiceprint recognition system based on voiceprint denoising and DBO-SVM as claimed in any one of claims 1 to 4, characterized in that: It includes a voiceprint processing module, which is used to perform noise reduction processing on the collected on-site voiceprint to obtain a clean single voiceprint; Information extraction module, used to extract MFCC feature values of voiceprint signals; A model training module, used to build the transformer core loosening diagnosis model, and train the transformer core loosening diagnosis model according to the training set divided by the voiceprint signal data; The identification and judgment module is used to perform fault diagnosis based on the transformer core loosening fault diagnosis model.
6. The transformer fault voiceprint recognition system based on voiceprint denoising and DBO-SVM according to claim 5 is characterized in that: The voiceprint processing module extracts the transformer voiceprint signal feature value based on the Mel cepstral coefficient, specifically: The signal is decomposed into 12 intrinsic mode function IMF components through VMD; The IMF component with the largest kurtosis value deviation among the IMF components is selected for wavelet threshold denoising, and the appropriate wavelet function and decomposition layer number are selected to perform wavelet decomposition on the signal to obtain the wavelet coefficients. The soft threshold function is set as the threshold function, the required wavelet coefficients are retained, and the processed wavelet coefficients are used to reconstruct the signal through inverse transformation; finally, the denoised IMF component and the remaining unprocessed IMF components are reconstructed.
Citation Information
Cited By
Transformer fault detection method based on feature fusion
CN120086655A
