Transformer fault identification method based on voiceprint signal

Abstract

The invention relates to a transformer fault identification method based on voiceprint signals, and belongs to the technical field of cepstrum for extracting parameters in audio decoding or coding. The method comprises the following steps: setting a fault type and establishing a fault identification model for training; arranging an acoustic sensor to collect voiceprint signals of the transformer; utilizing a dream optimization algorithm to optimize the penalty factor and a successive variational mode decomposition method to decompose a plurality of mode components, and dividing the mode components into pure components and noisy components; noise reduction is carried out on the noisy component by adopting a designed threshold function in combination with wavelet threshold noise reduction, and the noisy component and the pure component after noise reduction are input into the recognition model to obtain a probability vector; and finally, fusing into a first fusion probability vector and a second fusion probability vector through fuzzy measurement, and taking the fault type corresponding to the maximum second fusion probability as the fault type of the transformer. The method can accurately capture the mapping relation between the acoustic features and the transformer fault state, and accurately identifies the transformer fault type.

Description

TECHNICAL FIELD

[0001] The present application relates to a transformer fault recognition method based on voiceprint signals, belonging to the technical field of cepstrum (G10L25 / 24) in audio decoding or parameter extraction in coding. BACKGROUND

[0002] As the core equipment of the power system, the stability of the operation state of the transformer is directly related to the safe and reliable power supply of the power grid. In particular, potential mechanical faults inside the transformer, such as winding loosening and core abnormality, if not discovered in time, can easily lead to serious damage to the equipment and even large-scale power outages.

[0003] At present, voiceprint signal recognition methods have been widely used to identify faults generated by transformers. However, the existing voiceprint recognition methods have insufficient signal noise reduction and feature separation capabilities in complex field environments, and the extraction of features related to specific mechanical fault modes embedded in acoustic signals is still insufficient, resulting in inaccurate mapping relationships between acoustic features and mechanical fault states, and often the recognized faults do not match the actual faults, seriously affecting the stable operation of the transformer. SUMMARY

[0004] The technical problem to be solved by the present application is how to accurately identify the fault type of the transformer.

[0005] The technical solution proposed by the present application to solve the above technical problem is: a transformer fault recognition method based on voiceprint signals, comprising the following steps:

[0006] Step 1: Set M types of transformer fault types and construct a transformer fault recognition model, obtain historical noise reduction signals of the transformer and input the transformer fault recognition model for training, and obtain the trained transformer fault recognition model; the input of the trained transformer fault recognition model is set as the noise reduction signal of the transformer, and the output is set as a probability vector; the probability vector contains the probability of the transformer belonging to each fault type of the M types of transformer fault types;

[0007] Step 2: Arrange N groups of acoustic sensors beside the transformer to collect real-time voiceprint signals of the transformer, each group of acoustic sensors containing n acoustic sensors;

[0008] Step 3: Preprocess the real-time voiceprint signal collected by any acoustic sensor, the preprocessing being sequentially filtering, amplifying, signal synchronization and analog-to-digital conversion of the real-time voiceprint signal;

[0009] Step 4: Set the first threshold value of the dream optimization algorithm , the iteration number threshold T and the population size B, the initial position of each individual in the population size is set by using the chaotic mapping to perform the first iteration; the initial position of each individual is a candidate value of the penalty factor in the successive variational modal decomposition method;

[0010] Step 4.1: input each initial position into the successive variational modal decomposition method respectively, and decompose the preprocessed real-time voiceprint signal by using the successive variational modal decomposition method to obtain K real-time modal components and K center frequencies of the K real-time modal components corresponding to each initial position respectively;

[0011] The energy difference E1 corresponding to each initial position is calculated according to the following formula (1)

[0012] (1);

[0013] In formula (1), is the sum of the energies of the K real-time modal components; is the energy of the preprocessed real-time voiceprint signal; A is the signal length of the preprocessed real-time voiceprint signal; a is the a-th sampling point on the signal length A; is the first amplitude of the preprocessed real-time voiceprint signal at the a-th sampling point; is the second amplitude of the k-th real-time modal component at the a-th sampling point;

[0014] Step 4.2: arrange the K center frequencies corresponding to each initial position in ascending order, denoted as , and calculate the separation degree corresponding to each initial position according to the following formula (2)

[0015] (2);

[0016] In formula (2), is the C-th center frequency;

[0017] Step 4.3: calculate the fitness function value E5 corresponding to each initial position according to the following formula (3)

[0018] (3);

[0019] In formula (3), is the first weight, and the value is ; is the second weight, and the value is ;

[0020] The initial position corresponding to the minimum fitness function value is taken as the first optimal position;

[0021] Step 4.4: calculate the diversity decay rate according to the following formula (4)

[0022] (4);

[0023] In formula (4), t is the tth iteration; is the diversity index of the tth iteration; is the diversity index before the tth iteration; is the problem dimension; is the position of the dth dimension of the bth individual in the population of the tth iteration; is the position of the bth individual in the population of the tth iteration; is the position of the bth individual in the population of the tth iteration; is the position average of the dth dimension of the tth iteration;

[0024] If t < 0.1T or t > 0.1T, the first optimal position is input into the dream optimization algorithm, and the dream optimization algorithm is controlled to enter the exploration stage to update the initial position of each individual to obtain the first updated position of each individual, complete the first iteration, and enter the second iteration; < 0.1T, If t < 0.1T or t > 0.1T, the first optimal position is input into the dream optimization algorithm, and the dream optimization algorithm is controlled to enter the exploration stage to update the initial position of each individual to obtain the first updated position of each individual, complete the first iteration, and enter the second iteration;

[0025] If t < 0.1T or t > 0.1T, the first optimal position is input into the dream optimization algorithm, and the dream optimization algorithm is controlled to enter the exploration stage to update the initial position of each individual to obtain the first updated position of each individual, complete the first iteration, and enter the second iteration; < 0.1T, If t < 0.1T or t > 0.1T, the first optimal position is input into the dream optimization algorithm, and the dream optimization algorithm is controlled to enter the exploration stage to update the initial position of each individual to obtain the first updated position of each individual, complete the first iteration, and enter the second iteration;

[0026] Step 4.5: each first updated position or second updated position is input into the successive variational modal decomposition method, respectively, and the preprocessed real-time voiceprint signal is decomposed, and so on, until the dream optimization algorithm completes the Tth iteration to stop iteration, and the Tth optimal position of the Tth iteration is obtained;

[0027] The Tth optimal position is input into the successive variational modal decomposition method, the preprocessed real-time voiceprint signal is decomposed, and T real-time modal components and T center frequencies of the T real-time modal components are obtained;

[0028] Step 5: a second threshold c, a third threshold e and a fourth threshold H are set; the T optimal real-time modal components, the T center frequencies of the T optimal real-time modal components, the T time-domain kurtosis of the T optimal real-time modal components, the T frequency-domain sparsity index of the T optimal real-time modal components and the T maximum mutual information coefficient of the T optimal real-time modal components are calculated according to formula (5);

[0029] (5);

[0030] ​​​​​​​​In equation (5), E is the mathematical expectation; It is the amplitude of the optimal real-time modal component at the h-th sampling point; The average amplitude of the optimal real-time modal components; It is the standard deviation of the amplitude of the optimal real-time modal component; The frequency domain signal obtained by fast Fourier transform of the optimal real-time modal components is the signal at the 1st... Frequency domain signal value at each frequency domain point; is the length of the frequency domain signal; I is the mutual information; X is the preprocessed real-time acoustic signature signal; Y is the optimal real-time modal component; x and y are all possible amplitudes of X and Y, respectively. and These are the marginal probability density functions of X and Y, respectively; yes and The joint probability density function; It is the maximum grid size, which is a value equal to the signal length A of the preprocessed real-time voiceprint signal. Power; It is the number of intervals into which the amplitude range of the preprocessed real-time voiceprint signal is divided; It is the number of intervals that divide the amplitude range of the optimal real-time modal component;

[0031] Will The optimal real-time modal component is taken as the clean component, and the remaining optimal real-time modal components are taken as the noisy components.

[0032] Step 6: Set the threshold function for wavelet threshold denoising according to equations (6) and (7).

[0033] (6);

[0034] (7);

[0035] In equation (7), X is the noise standard deviation; It is the median function; L is the number of wavelet decomposition levels in wavelet thresholding denoising; It is the Lth wavelet decomposition after the Lth layer. Wavelet coefficients; It is a wavelet threshold; It is the wavelet threshold of the Lth layer; It is the total number of wavelet coefficients after the Lth level decomposition; These are shape control parameters, and their values ​​are... ; sgn is the sign function;

[0036] The noisy component is denoised using wavelet threshold denoising, and the clean component and the denoised noisy component are reconstructed to obtain a real-time denoised signal.

[0037] Step 7: inputting the real-time noise reduction signal into the trained transformer fault identification model to output a probability vector corresponding to the selected acoustic sensor;

[0038] Step 8: repeating steps 3-7 to obtain a probability vector corresponding to each acoustic sensor, respectively;

[0039] Step 9: in the n probability vectors corresponding to the n acoustic sensors of any acoustic sensor group, arranging all probabilities of the same fault type in descending order;

[0040] selecting the first i acoustic sensors corresponding to the first i probabilities of the jth fault type, and calculating the first fuzzy measure of the first i acoustic sensors of the jth fault type according to formula (8)

[0041] (8);

[0042] In formula (8), is the first fuzzy density of the qth acoustic sensor in the first i acoustic sensors of the jth fault type; is the reliability value of the qth acoustic sensor, which is the proportion of the number of times that the probability of the qth acoustic sensor in the corresponding probability vector is the largest in the historical transformer fault identification to the total number of times of historical transformer fault identification, and is 0 in the first transformer fault identification; is the importance value of the qth acoustic sensor, which is ; and are the first coefficient and the second coefficient, respectively, and both are ; is the third coefficient;

[0043] Step 9.1: calculating M first fusion probabilities of M fault types according to formula (9)

[0044] (9);

[0045] In formula (9), is the i th probability of the j th fault type after arranging all probabilities of the j th fault type in descending order; when i = 1, ;

[0046] The M first fusion probabilities form a first fusion probability vector, and a first fusion probability vector of any acoustic sensor group is obtained;

[0047] Step 9.2: repeating steps 9-9.1 to obtain N first fusion probability vectors of N acoustic sensor groups;

[0048] Step 10: Among the N first fusion probability vectors, arrange all the first fusion probabilities of the same fault type in descending order;

[0049] Select the acoustic sensor groups corresponding to the first p fusion probabilities of the j-th fault type, and calculate the second fuzzy measure of the first p acoustic sensor groups of the j-th fault type according to the following formula (10).

[0050] (10);

[0051] In equation (10), It is the second fuzzy density of the r-th acoustic sensor group among the p-th acoustic sensor groups for the j-th fault type; The first one is the one mentioned The importance of each acoustic sensor group is determined by a value. ; The first one is the one mentioned The distance between the geometric center of each acoustic sensor group and the sound source point of the transformer; It is the reliable average value of the r-th acoustic sensor group, and its value is the average of the reliable values ​​of the n acoustic sensors in the r-th acoustic sensor group; and These are the fourth and fifth coefficients, each with a value of [value missing]. ; It is the sixth coefficient;

[0052] Step 10.1: Calculate the M second fusion probabilities of the M fault types according to the following formula (11).

[0053] (11);

[0054] In equation (11), It is the p-th first fusion probability after sorting all first fusion probabilities of the j-th fault type in descending order; when p=1, ;

[0055] Step 10.2: Take the fault type corresponding to the highest second fusion probability as the fault type of the transformer.

[0056] Furthermore, the transformer fault identification model is constructed by combining an attention mechanism with a dual-channel parallel fusion 2D convolutional neural network and a bidirectional gated recurrent unit network architecture.

[0057] Furthermore, the acoustic sensor adopts a multi-arm spiral arrangement, with N spiral arms.

[0058] Further, the wavelet threshold denoising selects the wavelet basis from the Daubechies wavelet family with filter length of 2-10, the nearly symmetric wavelet family with filter length of 2-10, and the Coifman wavelet family with filter length of 2-10, and sets the decomposition layer number to 4-8, selects the combination with high signal-to-noise ratio and small root mean square error from all combinations of wavelet basis and decomposition layer number, and determines the optimal wavelet basis and decomposition layer number based on the combination, and then denoises the noisy component based on the optimal wavelet basis and decomposition layer number.

[0059] The dream optimization algorithm is improved by designing the fitness function and the stage transition condition to efficiently optimize the penalty factor of the successive variational mode decomposition method. BRIEF DESCRIPTION OF DRAWINGS

[0060] The transformer fault identification method based on the voiceprint signal is further described below with reference to the drawings.

[0061] Figure 1 FIG. 1 is a schematic diagram of the arrangement of the acoustic sensors in the embodiment of the transformer fault identification method based on the voiceprint signal. DETAILED DESCRIPTION

[0062] EMBODIMENT

[0063] The transformer fault identification method based on the voiceprint signal of the embodiment includes the following steps:

[0064] Step 1: Set M types of transformer fault types and construct a transformer fault identification model, obtain historical denoised signals of the transformer, and input the transformer fault identification model for training to obtain a trained transformer fault identification model; the input of the trained transformer fault identification model is set as the denoised signal of the transformer, and the output is set as a probability vector; the probability vector contains the probability of the transformer belonging to each fault type of the M types of transformer fault types;

[0065] In the embodiment, the fault types are set to 5 types, which are core looseness, winding looseness, cooling system fault, external foreign matter, and poor sleeve connection, and a 2D convolutional neural network 2DCNN and a bidirectional gated recurrent unit network architecture BiGRU are used in double-channel parallel fusion to construct the transformer fault identification model in combination with the attention mechanism CA.

[0066] Step 2: Arrange N acoustic sensor groups around the transformer to collect real-time acoustic fingerprint signals of the transformer, each acoustic sensor group containing n acoustic sensors;

[0067] As shown in Figure 1 In this embodiment, 5 acoustic sensor groups are arranged in a multi-arm spiral pattern, with each arm representing an acoustic sensor group, and each acoustic sensor group containing 7 acoustic sensors.

[0068] Step 3: Preprocess the real-time acoustic fingerprint signal collected by any acoustic sensor, which involves sequentially filtering, amplifying, synchronizing, and analog-to-digital converting the real-time acoustic fingerprint signal;

[0069] In this embodiment, the collected real-time acoustic fingerprint signal is input into a Butterworth second-order active low-pass filter for filtering, with a cutoff frequency set to 20 kHz, to preliminarily reduce noise and eliminate high-frequency noise such as switching noise in the signal;

[0070] The filtered real-time acoustic fingerprint signal is then input into a gain amplifier for amplifying the weak voltage signal, with the gain dynamically adjusted according to the signal amplitude to ensure that the input signal of the subsequent analog-to-digital converter is within the optimal range, and the gain set to 10 dB-60 dB;

[0071] Next, the amplified real-time acoustic fingerprint signal is input into a phase-locked detection circuit to achieve signal synchronization by detecting the phase information of the signal and comparing and analyzing the phase difference between the two signals, and adjusting the signal phase;

[0072] Finally, the signal-synchronized real-time acoustic fingerprint signal is input into an analog-to-digital converter to convert the analog signal into a digital signal.

[0073] Step 4: Set the first threshold of the dream optimization algorithm =0.05, the iteration threshold T=500, and the population size B=30, and use chaotic mapping to set the initial position of each individual in the population size for the first iteration; the initial position of each individual is the candidate value of the penalty factor in the successive variable mode decomposition method;

[0074] Dream optimization algorithm is a swarm intelligence optimization algorithm that simulates the random exploration and deep thinking process in human dreams to achieve efficient solution of complex problems. It has two stages in the iteration process: exploration stage, aiming to explore the global solution space and avoid falling into local optimum; development stage, aiming to fine search in the current optimal region to improve the accuracy of the solution.

[0075] The sequential variational modal decomposition method is a self-adaptive signal decomposition technology, and belongs to the prior art in the field of time-frequency analysis. The core idea is to decompose a complex signal into a plurality of intrinsic modal components. In the decomposition process, a key super parameter, a penalty factor, is needed to control the tightness of the decomposition and prevent modal aliasing.

[0076] In this embodiment, the dream optimization algorithm is used to find the optimal penalty factor of the sequential variational modal decomposition method.

[0077] Step 4.1: input each initial position into the sequential variational modal decomposition method respectively, and use the sequential variational modal decomposition method to decompose the preprocessed real-time voiceprint signal to obtain K real-time modal components and K center frequencies of the K real-time modal components corresponding to each initial position respectively;

[0078] The energy difference E1 corresponding to each initial position is calculated according to the following formula (1)

[0079] (1);

[0080] In formula (1), is the sum of the energies of the K real-time modal components; is the energy of the preprocessed real-time voiceprint signal; A is the signal length of the preprocessed real-time voiceprint signal; a is the a-th sampling point on the signal length A; is the first amplitude of the preprocessed real-time voiceprint signal at the a-th sampling point; is the second amplitude of the k-th real-time modal component at the a-th sampling point.

[0081] Step 4.2: arrange the K center frequencies corresponding to each initial position in ascending order, and denote them as , and calculate the separation degree corresponding to each initial position according to the following formula (2)

[0082] (2);

[0083] In formula (2), is the C-th center frequency;

[0084] Step 4.3: calculate the fitness function value E5 corresponding to each initial position according to the following formula (3)

[0085] (3);

[0086] In formula (3), is the first weight, and the value is ; is the second weight, and the value is ;

[0087] In this embodiment, Take 5. Take 0.6;

[0088] The initial position with the smallest corresponding fitness function value is taken as the first optimal position. This is the optimal penalty factor found in the first iteration using the dream optimization algorithm.

[0089] Step 4.4: Calculate the diversity decay rate using the following formula (4).

[0090] (4);

[0091] In equation (4), t is the t-th iteration; It is the diversity index of the t-th iteration; It is the diversity index before ∆t iterations; This refers to the problem dimension, specifically the number of hyperparameters that need to be optimized. It is the position of the b-th individual in the population during the t-th iteration, in the d-th dimension. It is the first The iteration of the ... Average positional value in each dimension;

[0092] like < If t≤0.1T, input the first optimal position into the dream optimization algorithm, and control the dream optimization algorithm to enter the exploration phase to update the initial position of each individual to obtain the first updated position of each individual, complete the first iteration, and enter the second iteration;

[0093] like ≥ And if t > 0.1T, input the first optimal position into the dream optimization algorithm, and control the dream optimization algorithm to enter the development stage to update the initial position of each individual to obtain the second updated position of each individual, complete the first iteration, and enter the second iteration;

[0094] The dream optimization algorithm updates the position of individuals in the population based on the best position found at different stages. In contrast, the traditional dream optimization algorithm transitions to a new stage after a certain number of iterations. In this embodiment, the population's exploration ability is quantified by the diversity decay rate to determine the appropriate stage. This significantly improves local development efficiency while ensuring global search capabilities, making it more suitable for multimodal component processing scenarios.

[0095] Step 4.5: Input each first update position or second update position into the successive variational mode decomposition method to decompose the preprocessed real-time voiceprint signal. Continue in this manner until the dream optimization algorithm completes the Tth iteration and stops iterating, obtaining the Tth best position of the Tth iteration, which is the best penalty factor of the successive variational mode decomposition method.

[0096] The optimal T-th position is input into the successive variational mode decomposition method to decompose the preprocessed real-time voiceprint signal, resulting in... One real-time modal component and One real-time modal component One center frequency;

[0097] Step 5: Set the second threshold c, the third threshold e, and the fourth threshold H; calculate according to the following formula (5). The optimal real-time modal components temporal kurtosis Frequency domain sparsity index and maximum mutual information coefficient ;

[0098] (5);

[0099] In equation (5), E is the mathematical expectation; It is the amplitude of the optimal real-time modal component at the h-th sampling point; The average amplitude of the optimal real-time modal components; It is the standard deviation of the amplitude of the optimal real-time modal component; The frequency domain signal obtained by fast Fourier transform of the optimal real-time modal components is the signal at the 1st... Frequency domain signal value at each frequency domain point; is the length of the frequency domain signal; I is the mutual information; X is the preprocessed real-time acoustic signature signal; Y is the optimal real-time modal component; x and y are all possible amplitudes of X and Y, respectively. and These are the marginal probability density functions of X and Y, respectively; yes and The joint probability density function; It is the maximum grid size, which is a value equal to the signal length A of the preprocessed real-time voiceprint signal. Power; It is the number of intervals into which the amplitude range of the preprocessed real-time voiceprint signal is divided; It is the number of intervals that divide the amplitude range of the optimal real-time modal component;

[0100] Will The optimal real-time modal component is taken as the clean component, and the remaining optimal real-time modal components are taken as the noisy components.

[0101] In this embodiment, temporal kurtosis is used. Measuring the "sharpness" of a signal's time-domain distribution helps distinguish between low-kurtosis stationary noise and high-kurtosis effective sound source components. Effective sound source components typically have concentrated energy in a specific frequency band, while noise components have a more dispersed energy distribution; therefore, the frequency domain sparsity index is used. The maximum mutual information coefficient is used to measure the concentration of signal energy in the frequency domain. Effective components contain more information related to the original signal, while noise components have lower correlation; therefore, the maximum mutual information coefficient is employed. The nonlinear correlation between the decomposed modal components and the original signal is measured.

[0102] Step 6: Set the threshold function for wavelet threshold denoising according to equations (6) and (7).

[0103] (6);

[0104] (7);

[0105] In equation (7), X is the noise standard deviation; It is the median function; L is the number of wavelet decomposition levels in wavelet thresholding denoising; It is the Lth wavelet decomposition after the Lth layer. Wavelet coefficients; It is a wavelet threshold; It is the wavelet threshold of the Lth layer; It is the total number of wavelet coefficients after the Lth level decomposition; These are shape control parameters, and their values ​​are... ; sgn is the sign function;

[0106] Wavelet threshold denoising is used to denoise the noisy components, and the clean components and the denoised noisy components are reconstructed to obtain the real-time denoised signal.

[0107] In this embodiment, the wavelet bases selected for wavelet threshold denoising include Dobesi wavelets with filter lengths of 2 to 10 in the Dobesi wavelet family dbN, approximately symmetric wavelets with filter lengths of 2 to 10 in the approximately symmetric wavelet family symN, and Koifman wavelets with filter lengths of 2 to 10 in the Koifman wavelet family coifN. The number of decomposition layers is set to 4 to 8. By traversing all wavelet bases and merging the decomposition layer arrays, the signal-to-noise ratio and root mean square error under different combinations are compared. The combination with a large signal-to-noise ratio and a small root mean square error is selected to determine the optimal wavelet base and the number of decomposition layers. The noisy components are denoised based on the optimal wavelet base and the number of decomposition layers.

[0108] Step 7: input the real-time noise reduction signal into the trained transformer fault identification model, and output the probability vector corresponding to the selected acoustic sensor.

[0109] In this embodiment, the 2DCNN channel in the transformer fault identification model will first extract the two-dimensional log-mel spectrum graph from the one-dimensional voiceprint time sequence data in the noise reduction signal through the mel filter bank, and then extract the image features using the two-dimensional convolutional neural network, while the BiGRU channel directly extracts the voiceprint signal features using the bidirectional gated recurrent unit, the two channels are parallel, then the cross attention CA mechanism fuses the features extracted by the two channels, and finally the Softmax classifier outputs the probability distribution of each fault type, i.e. the probability vector.

[0110] Step 8: repeat steps 3-7 to obtain the probability vector corresponding to each acoustic sensor.

[0111] Step 9: in the 7 probability vectors corresponding to the 7 acoustic sensors of any acoustic sensor group, arrange all the probabilities of the same fault type in descending order;

[0112] Select the acoustic sensors corresponding to the first i probabilities of the jth fault type, and calculate the first fuzzy measure of the first i acoustic sensors of the jth fault type according to the following formula (8)

[0113] (8);

[0114] In formula (8), is the first fuzzy density of the qth acoustic sensor in the first i acoustic sensors of the jth fault type; is the reliability value of the qth acoustic sensor, which is the proportion of the number of times that the probability of the qth acoustic sensor in the corresponding probability vector in the historical transformer fault identification is the largest and the transformer historical fault type is consistent to the total number of times of the historical transformer fault identification, and is 0 in the first transformer fault identification; is the importance value of the qth acoustic sensor, which is determined by the accuracy, sensitivity and noise level of the qth acoustic sensor, and the value is ; and are the first coefficient and the second coefficient, respectively, and the values are both ; is the third coefficient;

[0115] In this embodiment, 0.5, 0.7.

[0116] Step 9.1: calculate the M first fusion probabilities of the M fault types according to the following formula (9)

[0117] (9);

[0118] In formula (9), is the i-th probability of all the probabilities of the j-th fault type arranged in descending order; when i=1, ;

[0119] The M first fusion probabilities form a first fusion probability vector to obtain the first fusion probability vector of the sound sensor group;

[0120] Step 9.2: Repeat step 9~9.1 to obtain five first fusion probability vectors of five sound sensor groups;

[0121] Step 10: In the five first fusion probability vectors, all the first fusion probabilities of the same fault type are arranged in descending order;

[0122] The sound sensor groups corresponding to the first p first fusion probabilities of the j-th fault type are selected, and the second fuzzy measure of the first p sound sensor groups of the j-th fault type is calculated according to formula (10)

[0123] (10);

[0124] In formula (10), is the second fuzzy density of the r-th sound sensor group in the first p sound sensor groups of the j-th fault type; is the importance of the r-th sound sensor group, and the value is ; is the distance between the geometric center point of the r-th sound sensor group and the sound source point of the transformer; is the reliable average value of the r-th sound sensor group, and the value is the average value of the reliable values of n sound sensors in the r-th sound sensor group; and are the fourth coefficient and the fifth coefficient, and the values are both ; is the sixth coefficient; In this embodiment,

[0125] 0.5, 0.7.

[0126] Step 10.1: Calculate M second fusion probabilities of M fault types according to formula (11)

[0127] (11);

[0128] In formula (11),​​ is the pth first fusion probability after all the first fusion probabilities of the jth fault type are sorted in descending order; when p = 1,

[0129] Step 10.2: taking the fault type corresponding to the maximum second fusion probability as the fault type of the transformer.

[0130] In this embodiment, the second fusion probabilities corresponding to the five fault types of core looseness, winding looseness, cooling system fault, external foreign matter and poor bushing connection obtained according to formula (11) are 0.2, 0.1, 0.4, 0.1 and 0.2 respectively, so it is identified that the fault type of the transformer is the cooling system fault.

[0131] The above merely describes preferred embodiments of the present application, but the present application is not limited thereto, and all equivalent replacements or equivalent changes according to the concept and technical solutions of the present application should be covered within the protection scope of the present application.​

Claims

1. A transformer fault identification method based on voiceprint signals, characterized in that... The steps include the following: Step 1: Define M types of transformer faults and construct a transformer fault identification model. Obtain historical denoised signals of the transformer and input them into the transformer fault identification model for training to obtain a trained transformer fault identification model. The input of the trained transformer fault identification model is set as the denoised signal of the transformer, and the output is set as a probability vector. The probability vector contains the probability that the transformer belongs to each of the M types of transformer faults. Step 2: Set up N acoustic sensor groups next to the transformer to collect the real-time acoustic signature signal of the transformer. Each acoustic sensor group contains n acoustic sensors. Step 3: Select any real-time voiceprint signal collected by a sound sensor for preprocessing. The preprocessing involves sequentially filtering, amplifying, synchronizing, and converting the real-time voiceprint signal to digital. Step 4: Set the first threshold for the dream optimization algorithm The iteration number threshold T and the population size B are used to set the initial position of each individual in the population size using chaotic mapping for the first iteration; the initial position of each individual is a candidate value of the penalty factor in the successive variational mode decomposition method. Step 4.1: Input each initial position into the successive variational mode decomposition method, and use the successive variational mode decomposition method to decompose the preprocessed real-time acoustic signal to obtain K real-time modal components and K center frequencies of the K real-time modal components corresponding to each initial position. Calculate the energy difference E1 corresponding to each initial position using the following formula (1) （1）； In equation (1), It is the sum of the energies of the K real-time modal components; A is the energy of the preprocessed real-time voiceprint signal; A is the signal length of the preprocessed real-time voiceprint signal; a is the a-th sampling point on the signal length A. It is the first amplitude of the preprocessed real-time voiceprint signal at the a-th sampling point; It is the second amplitude of the k-th real-time modal component at the a-th sampling point; Step 4.2: Arrange the K center frequencies corresponding to each initial position in ascending order, and denote them as follows: And calculate the degree of separation corresponding to each initial position according to the following formula (2). （2）； In equation (2), It is the Cth center frequency; Step 4.3: Calculate the fitness function value E5 corresponding to each initial position using the following formula (3). （3）； In equation (3), It is the first weight, and its value is... ; It is the second weight, and its value is... ; The initial position with the smallest corresponding fitness function value is taken as the first optimal position; Step 4.4: Calculate the diversity decay rate using the following formula (4). （4）； In equation (4), t is the t-th iteration; It is the diversity index of the t-th iteration; It is the diversity index before ∆t iterations; It is a problem dimension; It is the position of the b-th individual in the population during the t-th iteration, in the d-th dimension. It is the first The iteration of the ... Average positional value in each dimension; like < If t≤0.1T, input the first optimal position into the dream optimization algorithm, and control the dream optimization algorithm to enter the exploration phase to update the initial position of each individual to obtain the first updated position of each individual, complete the first iteration, and enter the second iteration; like ≥ And if t > 0.1T, input the first optimal position into the dream optimization algorithm, and control the dream optimization algorithm to enter the development stage to update the initial position of each individual to obtain the second updated position of each individual, complete the first iteration, and enter the second iteration; Step 4.5: Input each first update position or second update position into the successive variational mode decomposition method to decompose the preprocessed real-time voiceprint signal, and so on, until the dream optimization algorithm completes the Tth iteration and stops iterating, to obtain the Tth best position of the Tth iteration; The T-th optimal position is input into the successive variational mode decomposition method to decompose the preprocessed real-time voiceprint signal, resulting in... One real-time modal component and One real-time modal component One center frequency; Step 5: Set the second threshold c, the third threshold e, and the fourth threshold H; calculate according to the following formula (5). The optimal real-time modal components temporal kurtosis Frequency domain sparsity index and maximum mutual information coefficient ; （5）； In equation (5), E is the mathematical expectation; It is the amplitude of the optimal real-time modal component at the h-th sampling point; The average amplitude of the optimal real-time modal components; It is the standard deviation of the amplitude of the optimal real-time modal component; The frequency domain signal obtained by fast Fourier transform of the optimal real-time modal components is the signal at the 1st... Frequency domain signal value at each frequency domain point; is the length of the frequency domain signal; I is the mutual information; X is the preprocessed real-time acoustic signature signal; Y is the optimal real-time modal component; x and y are all possible amplitudes of X and Y, respectively. and These are the marginal probability density functions of X and Y, respectively; yes and The joint probability density function; It is the maximum grid size, which is a value equal to the signal length A of the preprocessed real-time voiceprint signal. Power; It is the number of intervals into which the amplitude range of the preprocessed real-time voiceprint signal is divided; It is the number of intervals that divide the amplitude range of the optimal real-time modal component; Will The optimal real-time modal component is taken as the clean component, and the remaining optimal real-time modal components are taken as the noisy components. Step 6: Set the threshold function for wavelet threshold denoising according to equations (6) and (7). （6）； （7）； In equation (7), X is the noise standard deviation; It is the median function; L is the number of wavelet decomposition levels in wavelet thresholding denoising; It is the Lth wavelet decomposition after the Lth layer. Wavelet coefficients; It is a wavelet threshold; It is the wavelet threshold of the Lth layer; It is the total number of wavelet coefficients after the Lth level decomposition; These are shape control parameters, and their values ​​are... ; sgn is a sign function; The noisy component is denoised using wavelet threshold denoising, and the clean component and the denoised noisy component are reconstructed to obtain a real-time denoised signal. Step 7: Input the real-time noise reduction signal into the trained transformer fault identification model, and output the probability vector corresponding to the randomly selected acoustic sensor; Step 8: Repeat steps 3-7 to obtain the probability vector corresponding to each acoustic sensor; Step 9: In any group of n acoustic sensors, arrange all probabilities of the same fault type in descending order from the n probability vectors corresponding to the n acoustic sensors. Select the acoustic sensors corresponding to the first i probabilities of the j-th fault type, and calculate the first fuzzy measure of the first i acoustic sensors for the j-th fault type according to the following formula (8). （8）； In equation (8), It is the first fuzzy density of the qth acoustic sensor among the i acoustic sensors before the jth fault type; It is the reliability value of the qth acoustic sensor, which is the proportion of the number of times the fault type with the highest probability in the probability vector of the qth acoustic sensor in the historical transformer fault identification is consistent with the historical transformer fault type to the total number of historical transformer fault identifications. It is 0 in the first transformer fault identification. This is the importance value of the q-th acoustic sensor, and its value is [value missing]. ; and These are the first and second coefficients, both with values ​​of 1. ; It is the third coefficient; Step 9.1: Calculate the M first fusion probabilities of the M fault types according to the following formula (9). （9）； In equation (9), It is the i-th probability after all probabilities of the j-th fault type are arranged in descending order; when i=1, ; The M first fusion probabilities are combined to form a first fusion probability vector to obtain the first fusion probability vector of any one acoustic sensor group; Step 9.2: Repeat steps 9 to 9.1 to obtain N first fusion probability vectors for the N acoustic sensor groups; Step 10: Among the N first fusion probability vectors, arrange all the first fusion probabilities of the same fault type in descending order; Select the acoustic sensor groups corresponding to the first p fusion probabilities of the j-th fault type, and calculate the second fuzzy measure of the first p acoustic sensor groups of the j-th fault type according to the following formula (10). （10）； In equation (10), It is the second fuzzy density of the r-th acoustic sensor group among the p-th acoustic sensor groups for the j-th fault type; The first one is the one mentioned The importance of each acoustic sensor group is determined by a value. ; The first one is the one mentioned The distance between the geometric center of each acoustic sensor group and the sound source point of the transformer; It is the reliable average value of the r-th acoustic sensor group, and its value is the average of the reliable values ​​of the n acoustic sensors in the r-th acoustic sensor group; and These are the fourth and fifth coefficients, each with a value of [value missing]. ; It is the sixth coefficient; Step 10.1: Calculate the M second fusion probabilities of the M fault types according to the following formula (11). （11）； In equation (11), It is the p-th first fusion probability after sorting all first fusion probabilities of the j-th fault type in descending order; when p=1, ; Step 10.2: Take the fault type corresponding to the highest second fusion probability as the fault type of the transformer.

2. The transformer fault identification method based on voiceprint signals according to claim 1, characterized in that: The transformer fault identification model is constructed by combining a dual-channel parallel fusion 2D convolutional neural network and a bidirectional gated recurrent unit network architecture with an attention mechanism.

3. The transformer fault identification method based on voiceprint signals according to claim 1, characterized in that: The acoustic sensor is arranged in a multi-arm spiral pattern, with N spiral arms.

4. The transformer fault identification method based on acoustic signature signal according to claim 1, characterized in that: The wavelet basis selected for wavelet threshold denoising includes Dobesi wavelets with filter lengths of 2 to 10 from the Dobesi wavelet family, approximately symmetric wavelets with filter lengths of 2 to 10 from the approximately symmetric wavelet family, and Koifman wavelets with filter lengths of 2 to 10 from the Koifman wavelet family. The number of decomposition layers is set to 4 to 8. By traversing all wavelet basis and merging the decomposition layer array, the signal-to-noise ratio and root mean square error under different combinations are compared. The combination with a large signal-to-noise ratio and a small root mean square error is selected to determine the optimal wavelet basis and the number of decomposition layers. The noisy components are denoised based on the optimal wavelet basis and the number of decomposition layers.

Images