Transformer fault identification method, system, equipment and medium

Abstract

The invention discloses a transformer fault identification method, system and device and a medium, and relates to the technical field of transformer fault identification, and the method comprises the steps: collecting an audio signal of a transformer, and extracting a chaotic domain multi-scale parting dimension feature, a time domain zero crossing rate feature and a frequency domain fractional order spectrum entropy feature of the audio signal; constructing a biological heuristic multi-domain neural network model; the biological heuristic multi-domain neural network model comprises a spiking neural network, a bionic convolutional layer and a chaotic neural network; inputting the extracted audio signal features into a biological heuristic multi-domain neural network model, and identifying the fault category of the transformer; according to the method, the accuracy of fault identification is improved, and fault diagnosis can be stably carried out in a complex and changeable operation environment.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment fault diagnosis, and in particular to a transformer fault identification method, system, equipment and medium. Background Art

[0002] With the continuous development of power systems, transformers, as important equipment in power transmission and distribution systems, have a direct impact on the safety and stability of the entire power system. Therefore, how to effectively monitor and diagnose the operating status of transformers, detect faults in a timely manner and take corresponding maintenance measures has become a key issue in power equipment management.

[0003] Traditional transformer fault diagnosis methods mainly rely on electrical measurements and oil gas analysis, but these methods often have certain limitations when facing complex fault types and changing operating environments. In recent years, with the development of information technology and intelligent algorithms, fault diagnosis methods based on audio signals have gradually attracted the attention of researchers. During the operation of the transformer, various sound signals are generated, and these sound signals contain rich fault information. By analyzing and processing the transformer audio signals, the transformer fault can be effectively identified. Researchers have proposed a variety of transformer fault diagnosis technologies based on audio signals, which mainly focus on the extraction of signal features and the design of classification models. In terms of signal feature extraction, researchers have developed a variety of feature extraction methods in the time domain, frequency domain and chaos domain. For example, short-time energy, zero crossing rate and other indicators commonly used in time domain features can capture the transient information of audio signals; in frequency domain features, methods such as power spectrum and spectral entropy can effectively describe the spectral distribution of signals; and chaotic domain features such as multi-scale fractal dimension further capture the nonlinear dynamic characteristics hidden in the signal.

[0004] However, in the existing transformer fault diagnosis technology, only relying on time domain features such as short-time energy, zero crossing rate, etc., although it can capture the transient information of the signal, it cannot effectively characterize the frequency domain information brought by the spectrum change. Similarly, using only frequency domain features such as spectral entropy, power spectral density, etc., cannot fully reflect the nonlinear and non-stationary characteristics of the signal, and its diagnostic performance is still insufficient when facing variable working conditions. Summary of the invention

[0005] In view of the deficiency that the prior art only uses a single feature for analysis, which is difficult to fully reflect the fault characteristics of the transformer and leads to a decrease in diagnostic accuracy, the present invention proposes a transformer fault identification method, system, device and medium. By utilizing audio signal processing technology and deep learning models, combined with feature extraction and optimization algorithms in the chaotic domain, time domain and frequency domain, high-precision identification of transformer faults is achieved, thereby solving the problems existing in the prior art.

[0006] A transformer fault identification method comprises the following steps:

[0007] Collect the audio signal of the transformer, and extract the chaotic domain multi-scale fractal dimension features, time domain zero crossing rate features, and frequency domain fractional-order spectrum entropy features of the audio signal;

[0008] The extracted audio signal features are input into a bio-inspired multi-domain neural network model to identify the fault type of the transformer; the bio-inspired multi-domain neural network model includes a pulse neural network, a bionic convolutional layer and a chaotic neural network. The identification of the fault type of the transformer by the bio-inspired multi-domain neural network model specifically includes the following steps:

[0009] A chaotic neural network is used to extract the multi-scale classification dimension features of the chaotic domain to obtain the chaotic domain features; the chaotic domain features are input into the bionic convolution layer as feedback signals to adjust the weights of the bionic convolution layer, and the frequency domain fractional-order spectrum entropy features are processed through the adjusted bionic convolution layer to obtain the frequency domain features; the frequency domain features and the chaotic domain features are introduced into the pulse neural network to adjust the membrane potential update rule of the pulse neural network, and the adjusted pulse neural network is used to process the time domain zero crossing rate features to obtain the time domain features;

[0010] The time domain features, frequency domain features and chaos domain features are fused and input into the fully connected layer to identify the fault category.

[0011] Furthermore, the extraction of the chaotic domain multi-scale fractal dimension features specifically includes the following steps:

[0012] The audio signal is segmented into fixed length L, each segment is denoted by x k (t), where K represents the segment number;

[0013] For each segment signal x k (t) Perform discrete wavelet transform, select Daubechies wavelet basis function, and set different scale factors τ to τ1, τ2, ..., τ n ; Perform j-layer wavelet decomposition on the segmented signal:

[0014]

[0015] Among them A j is the approximation coefficient at different scales, D j is the detail coefficient;

[0016] Calculate the Euclidean distance matrix D between points in the reconstructed phase space:

[0017]

[0018] i and j are the element subscript indices in the matrix D;

[0019] Calculate the box count N(r) at different radii r in the distance matrix, and then find the classification dimension D f :

[0020]

[0021] in is a step function, when rD ij The value is 1 when ≥0, otherwise it is 0;

[0022]

[0023] The classification dimension D calculated under each scale factor τ f Arrange the scale factors in order to construct a multi-scale feature matrix, and standardize the constructed multi-scale feature matrix to ensure that the eigenvalues ​​are at the same order of magnitude;

[0024]

[0025] Among them, u is the mean of the feature matrix, σ is the standard deviation of the feature matrix;

[0026] The final chaotic domain characteristic matrix F is obtained, which is expressed as:

[0027]

[0028] in, Indicated by the scale factor τ j Next, the classification dimension value of the i-th signal segment.

[0029] Furthermore, the extraction of the time domain zero crossing rate feature comprises the following steps:

[0030] Select the window length L and overlap length O to segment the signal, and segment the signal x(t) into multiple windows, each window contains L samples, and overlaps each other by O samples. The segmentation formula is expressed as:

[0031] x k (n) = x(n+k(LO)),

[0032] Where k is the index of the window, n=0,1,2,3,4,5...,L-1;

[0033] Calculate the window function with a window length of L:

[0034]

[0035] x k (n) = x (n + k (LO))

[0036] For the short-time zero-crossing rate Z of the i-th window i Calculated as:

[0037]

[0038] where sgn(x) is the sign function, which is expressed as:

[0039]

[0040] The short-time zero-crossing rate Z of each window i Mapped onto the time axis, the required short-time zero-crossing rate feature image is formed, and the short-time zero-crossing rate feature representation in the time domain is obtained:

[0041] Z=[Z1 Z2 ... Z T ],

[0042] Where T is the total number of windows.

[0043] Furthermore, the extraction of the frequency domain fractional-order spectrum entropy feature specifically includes the following steps:

[0044] Decompose the audio signal into a set of intrinsic mode functions, detect local extreme points of the audio signal, interpolate and calculate the mean of the upper and lower envelopes; calculate the details of the signal, and repeat this process until the IMF conditions are met; the IMF conditions include:

[0045] The number of zero crossings and extrema is equal or differs by at most one, and the local mean is zero:

[0046]

[0047] IMF i (t) is the ith eigenmode function, r n (t) is the remaining signal;

[0048] IMF for the IMF i (t), and perform S transformation S{IMF i (t)}:

[0049]

[0050] in

[0051] Get the parsed signal:

[0052] z i (t) = IMF i (t)+iS{IMFj i (t)}(f,t),

[0053] The details of the final signal D i (t) is expressed as:

[0054] D i (t) = Re{z i (t)}=IMF i (t),

[0055] Calculate the instantaneous amplitude A i (t) and instantaneous phase θ i (t) is:

[0056]

[0057] Calculate the instantaneous frequency ω i (t) is:

[0058]

[0059] Construct the S-transform time-frequency spectrum S(f,τ) to obtain the time-frequency representation of the signal:

[0060]

[0061] Where S(ω,t) represents the instantaneous spectrum value at frequency ω and time t;

[0062] Under the obtained fractional order α, the spectrum H(ω,t) is subjected to fractional Fourier transform:

[0063]

[0064] Where K α (t,u) is the fractional Fourier kernel function;

[0065] Calculate the spectrum after fractional Fourier transform FRFT α The energy spectral density of (ω) is:

[0066] E(ω)=|FRFT α (ω)| 2 ,

[0067] Calculate the probability distribution P for each frequency band α (ω):

[0068]

[0069] Calculate the fractional spectral entropy S(α):

[0070]

[0071] According to the frequency ω, the fractional order α and the spectrum entropy value S(α), the fractional spectral entropy feature image is constructed, and the feature image is converted into a feature matrix M. Each element of the matrix corresponds to a pixel value in the feature image, and the final frequency domain matrix is ​​obtained, which is specifically expressed as:

[0072]

[0073] Where M(i,j) = S(α i ,ω j ), indicating that the element in the i-th row and j-th column of the feature matrix corresponds to the fractional order α i and frequency ω i The spectrum entropy value S(α i ,ω j ).

[0074] Furthermore, the frequency domain features and the chaotic domain features are introduced into the spiking neural network to adjust the membrane potential update rule of the spiking neural network, and the adjusted spiking neural network is used to simulate the pulse emission mechanism of biological neurons to process the time domain zero crossing rate features to obtain the time domain features, which specifically includes the following steps:

[0075] The pulse firing mechanism of the spiking neural network simulating biological neurons is expressed as:

[0076]

[0077] Among them, S(t) * is a pulse emission signal, a pulse emitted at time t; V(t) represents the membrane potential of the neuron at time t, and the membrane potential represents the total response of the neuron to the input; the membrane potential update formula is expressed as:

[0078]

[0079] Where V(t-1) is the membrane potential of the neuron at the previous time step t-1; ω i Input x i The weight of (t) indicates the connection strength; x i (t) input signal at time t;

[0080] The bionic convolutional layer simulates the processing mechanism of the biological visual cortex, and its processing process is expressed as:

[0081]

[0082] Among them, Y ij is the output after the convolution operation, indicating the value of position (i, j) in the feature map; X (i+m)(i+m) is the value of the input feature map at position (i+m,j+m); K mn is the weight of the convolution kernel;

[0083] Chaotic neural network simulates the behavior mechanism in biological chaotic system, which is expressed as:

[0084] X(t+1)=αX(t)(1-X(t)),

[0085] X(t+1) is the state at time t+1, X(t) is the state at time t, and α is the adjustment coefficient;

[0086] The frequency domain feature Y extracted by the bionic convolution layer ij After the introduction of the pulse neural network, its membrane potential is expressed as:

[0087]

[0088] V(t) is the membrane potential at time t, V(t+1) is the membrane potential at time t+1, ω i is the input signal x i (t), β is the adjustment coefficient; X(t+1) is the frequency domain feature input signal of the bionic convolution layer;

[0089] The output X of the chaotic neural network * (t+1) is used as a feedback signal to adjust the membrane potential update rule V(t) of the spiking neural network and the weight K of the bionic convolutional layer mn , specifically expressed as:

[0090] K mn (t+1)=K mn (t)·(1+γ·X * (t+1)),

[0091] Among them, β and γ are adjustment coefficients;

[0092] By introducing the chaotic domain characteristics into the spiking neural network, the adjusted membrane potential update rule is expressed as:

[0093]

[0094] The pulse neural network adjusted based on the membrane potential update rule is used to process the time domain zero crossing rate feature to obtain the time domain feature S(t).

[0095] Furthermore, the fusion of the time domain features, the frequency domain features and the chaos domain features specifically includes the following steps:

[0096] The time domain feature S(t) and frequency domain feature Y ij And the chaotic domain features X(t+1) are converged to form a feature matrix F:

[0097] F=[S(t)Y ijX(t+1)],

[0098] Adjust the synaptic weights of the feature matrix F:

[0099] W=W ij +η·ΔW ij ,

[0100] Among them, W ij is the weight of the position (i, j) of the fusion weight matrix; η is the learning rate, which controls the amplitude of weight adjustment, ΔW ij is the volume weight adjustment amount;

[0101] Introduce the neural modulation mechanism to adjust the weights in the feature fusion process:

[0102] W ij (t+1)=W ij (t)·(1+γ·D(t)),

[0103] γ is the modulation coefficient, D(t) is the modulation signal;

[0104] Apply synaptic weight adjustment and neural modulation mechanism to the feature matrix F to obtain the fused feature F fused :

[0105] F fused =F·W.

[0106] Furthermore, it also includes optimizing the biologically inspired multi-domain neural model using a global optimal firefly optimization algorithm; wherein the hyperparameters optimized in the biologically inspired multi-domain neural model include: a learning rate lr, a hyperparameter tls of a time domain feature learning module, a hyperparameter ffs of a frequency domain feature learning module, a hyperparameter chs of a chaos domain feature learning module, a hyperparameter fms of a fusion module, and a hyperparameter cls of a classification head; optimizing the hyperparameters using the global optimal firefly algorithm includes the following steps:

[0107] Set the value range of the hyperparameter to be optimized:

[0108] lr: [0.0001,0.01]; tls: [32,256]; ffs: [64,512]; chs: [64,512]; fms: [32,256]; cls: [32,256];

[0109] Initialize firefly individuals: Randomly initialize firefly individuals within the value range of the hyperparameter to be optimized. Each individual represents a hyperparameter combination. Assuming the population size is n, each firefly individual x i Contains six parameters:

[0110] X i =[lr itls i ffs i chs i fms i cls i ],

[0111] For each firefly individual, a biologically inspired multi-domain neural model is trained according to its hyperparameter combination; the specified hyperparameters are used during training and the performance metrics of the model are recorded;

[0112] The fitness value is calculated according to the performance indicators of the biologically inspired multi-domain neural model, and the validation error is used as the fitness function. The fitness value X i Represents the fitness of the i-th firefly individual:

[0113] f(X i )=validation loss(X i ),

[0114] Among them, validation loss (X i ) represents the loss function value of the i-th firefly individual on the validation set;

[0115] f(X i )=validation loss(X i ) is structured as follows: the input data is propagated forward through the model to obtain the predicted value The predicted value is calculated using the loss function mean square error loss The loss between the true value Y is:

[0116]

[0117] Among them, m is the number of samples in the validation set, is the predicted value of the jth sample, Y j is the true value of the jth sample;

[0118] Update the model parameters according to the gradient of the loss function;

[0119] Make predictions on the validation set and calculate the validation error. Assuming the number of samples in the validation set is m, the validation error is expressed as:

[0120]

[0121] in, is the predicted value of the jth sample in the validation set, Y val,j is the true value of the jth sample;

[0122] Calculate the light intensity of each firefly individual, assuming I iRepresents the light intensity of the i-th firefly individual:

[0123]

[0124] Update the position of individual fireflies according to the light intensity, and update the position of firefly i so that it moves towards firefly j with higher brightness:

[0125] x i =x i +β ij (x j -x i )+α(rand-0.5),

[0126]

[0127] Among them, α is the random movement coefficient, rand is a random number that obeys uniform distribution, γ is the light intensity absorption coefficient, and r ij is the distance between fireflies i and j. Each firefly adjusts its position according to the attraction and random disturbance it receives. Firefly i will move towards firefly j with higher brightness, and the step size is determined by the attraction and a random factor.

[0128] After multiple iterative updates, the optimal fitness value and the corresponding hyperparameter combination are recorded. When the maximum number of iterations is reached or the fitness value is no longer updated, the movement of the firefly population is terminated, and the firefly individual with the smallest fitness value is selected as the optimal hyperparameter combination.

[0129] The present invention also includes a transformer fault identification system, comprising:

[0130] Collect the audio signal of the transformer, and extract the chaotic domain multi-scale fractal dimension features, time domain zero crossing rate features, and frequency domain fractional-order spectrum entropy features of the audio signal;

[0131] The recognition module is used to input the extracted audio signal features into a bio-inspired multi-domain neural network model to identify the fault type of the transformer; the bio-inspired multi-domain neural network model includes a pulse neural network, a bionic convolutional layer and a chaotic neural network; identifying the fault type of the transformer by the bio-inspired multi-domain neural network model specifically includes:

[0132] The feature extraction unit is used to extract the multi-scale classification dimension features of the chaotic domain by using a chaotic neural network to obtain the chaotic domain features; the chaotic domain features are input into the bionic convolution layer as feedback signals to adjust the weight of the bionic convolution layer, and the frequency domain fractional-order spectrum entropy features are processed by the adjusted bionic convolution layer to obtain the frequency domain features; the frequency domain features and the chaotic domain features are introduced into the pulse neural network to adjust the membrane potential update rule of the pulse neural network, and the time domain zero crossing rate features are processed by the adjusted pulse neural network to obtain the time domain features;

[0133] The feature fusion unit is used to fuse the time domain features, frequency domain features and chaos domain features and input them into the fully connected layer to identify the fault category.

[0134] The present invention also includes a transformer fault identification computer device, including: a memory, a processor, and a computer program stored in the memory, and the processor implements the steps of the transformer fault identification method when executing the computer program.

[0135] The present invention also includes a readable storage medium, wherein the readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, they are used to execute the steps of the transformer fault identification method.

[0136] The present invention provides a transformer fault identification method, system, device and medium, which have the following beneficial effects:

[0137] The present invention extracts the chaotic domain multi-scale fractal dimension features, time domain zero crossing rate features and frequency domain fractional spectral entropy features of the transformer audio signal, introduces the frequency domain fractional spectral entropy features into the membrane potential update formula, enhances the responsiveness of the pulse neural network to the input signal, realizes the comprehensive analysis of the frequency domain fractional spectral entropy features and the time domain zero crossing rate features, uses the output of the chaotic neural network as a feedback signal, adjusts the membrane potential update rule of the pulse neural network and the weight of the bionic convolution layer, realizes the organic fusion between different feature domains, and realizes the diagnosis of the fault category. This method extracts the multi-scale fractal dimension features of the chaotic domain, combines the zero crossing rate of the time domain and the fractional spectral entropy features of the frequency domain, thereby realizing a comprehensive characterization of the transformer operating state. Compared with the existing feature fusion scheme, this method can not only capture the basic characteristics of the signal, but also fully reflect the fault characteristics of the transformer when facing variable working conditions, and greatly improve the accuracy and robustness of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0138] Figure 1 A schematic diagram of a transformer fault identification process in an embodiment of the present invention;

[0139] Figure 2It is a multi-scale fractal dimension feature heat scatter diagram in an embodiment of the present invention;

[0140] Figure 3 It is a heat scatter diagram of the short-time zero crossing rate characteristics in an embodiment of the present invention;

[0141] Figure 4 It is a heat scatter plot of the fractional-order spectral entropy feature in an embodiment of the present invention. DETAILED DESCRIPTION

[0142] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0143] The present invention proposes a transformer fault identification method, which realizes high-precision identification of transformer fault types by extracting multi-domain features of audio signals, fusion of features and application of deep learning models, such as Figure 1 As shown; the method specifically comprises the following steps:

[0144] S1. Extract multi-scale classification dimension features in the chaotic domain.

[0145] S11, audio segmentation processing and wavelet transformation.

[0146] The audio signal is segmented into fixed length L, each segment is denoted by x k (t), where K is the segment number, for each segment signal x k (t) Perform discrete wavelet transform, select Daubechies wavelet basis function, and set different scale factors τ to τ1, τ2, ..., τ n ; Perform j-layer wavelet decomposition on the segmented signal:

[0147]

[0148] Among them A j is the approximation coefficient at different scales, D j is the detail factor.

[0149] S1.2. Calculate the classification dimension of the audio.

[0150] Calculate the Euclidean distance matrix D between points in the reconstructed phase space:

[0151]

[0152] i and j are the element subscript indices in the matrix D.

[0153] Calculate the box count N(r) at different radii r in the distance matrix, and then find the classification dimension D f:

[0154]

[0155] in is a step function, when rD ij The value is 1 when ≥0, otherwise it is 0.

[0156]

[0157] S1.3. Construct a multi-scale classification dimension feature matrix representation. The extracted multi-scale classification dimension feature heat scatter plot is as follows: Figure 2 shown.

[0158] The classification dimension D calculated under each scale factor τ f Arrange in order of scale factors to construct a multi-scale classification dimension feature matrix, and standardize the constructed multi-scale classification dimension feature matrix to ensure that the eigenvalues ​​are at the same order of magnitude.

[0159]

[0160] Among them, u is the mean of the feature matrix, and σ is the standard deviation of the feature matrix.

[0161] The final chaotic domain characteristic matrix F is expressed as:

[0162]

[0163] in, Indicated by the scale factor τ j Next, the classification dimension value of the i-th signal segment.

[0164] S2. Extract the time domain zero crossing rate features.

[0165] S2.1. Divide the audio into frames and add windows.

[0166] Select the window length L and overlap length O to split the signal, and split the signal x(t) into multiple windows, each containing L samples, and overlapping each other by O samples. The segmentation formula is as follows:

[0167] x k (n) = x (n + k (LO)) (6)

[0168] Where k is the index of the window, n=0, 1, 2, 3, 4, 5, ..., L-1.

[0169] Window function calculation, for a window function of length L:

[0170]

[0171] Where n=0,1,2,3,4,5......,L-1.

[0172] x k (n) = x (n + k (LO)) (8)

[0173] Where k is the index of the window, n=0, 1, 2, 3, 4, 5, ..., L-1.

[0174] S2.2. Calculate the zero crossing rate for each window and obtain the short-time zero crossing rate feature representation. The extracted short-time zero crossing rate feature heat scatter plot is as follows: Figure 3 shown.

[0175] For the i-th window, the short-time energy Z i Calculated as:

[0176]

[0177] where sgn(x) is the sign function, defined as:

[0178]

[0179] The short-time zero-crossing rate Z of each window i Mapped onto the time axis, the required short-time zero-crossing rate feature image is formed, and the short-time zero-crossing rate feature representation in the time domain is obtained:

[0180] Z=[Z1 Z2…Z T ]

[0181] Where Twindow is the total number of windows.

[0182] S3, frequency domain fractional-order spectrum entropy characteristics.

[0183] S3.1. Empirical mode decomposition and construction of S-transform spectrum.

[0184] Decompose the signal into a set of intrinsic mode functions. Detect local extreme points of the signal and interpolate to calculate the mean of the upper and lower envelopes. Calculate the details of the signal and repeat this process until the IMF conditions are met:

[0185] The number of zero crossings and extreme points is equal or differs by at most one; the local mean is zero.

[0186]

[0187] IMF i (t) is the ith eigenmode function, r n (t) is the remaining signal.

[0188] IMF for the IMF i(t), and perform S transformation S{IMF i (t)}:

[0189]

[0190] in

[0191] Get the parsed signal:

[0192] z i (t) = IMF i (t)+iS{IMFj i (t)}(f,t) (12)

[0193] Calculate the instantaneous amplitude A i (t) and instantaneous phase θ i (t) is:

[0194]

[0195] Calculate the instantaneous frequency ω i (t) is:

[0196]

[0197] Construct the S-transform time-frequency spectrum S(f,τ) to obtain the time-frequency representation of the signal:

[0198]

[0199] Where S(ω,t) represents the instantaneous spectrum value at frequency ω and time t.

[0200] S3.2, fractional Fourier transform and spectrum entropy calculation;

[0201] Under the obtained fractional order α, the spectrum H(ω,t) is subjected to fractional Fourier transform:

[0202]

[0203] Where K α (t,u) is the fractional Fourier kernel function.

[0204] Calculate the spectrum after fractional Fourier transform FRFT α Energy spectral density of (ω):

[0205] E(ω)=|FRFT α (ω)| 2 (18)

[0206] Calculate the probability distribution P for each frequency band α (ω):

[0207]

[0208] Calculate the fractional spectral entropy S(α):

[0209]

[0210] Where S(α) describes the complexity and uncertainty of the signal in different frequency components.

[0211] S3.3, construct the fractional-order spectral entropy feature image and feature matrix, and the extracted fractional-order spectral entropy feature heat scatter plot is as follows Figure 4 shown.

[0212] The horizontal axis represents the frequency ω, the vertical axis represents the fractional order α, and the color represents the spectrum entropy value S(α), and a two-dimensional feature image is drawn. The feature image is converted into a feature matrix M, where each element of the matrix corresponds to a pixel value in the feature image. The final frequency domain matrix is ​​specifically expressed as:

[0213]

[0214] Where M(i,j) = S(α i ,ω j ), which represents the element in the i-th row and j-th column of the feature matrix, corresponding to the fractional order α i and frequency ω i The spectrum entropy value S(α i ,ω j ).

[0215] S4. Build a biologically inspired multi-domain neural model.

[0216] S4.1. Time domain feature learning.

[0217] The spiking neural network (SNN) is used to process the time domain feature map. SNN simulates the pulse firing mechanism of biological neurons and captures the time dependency.

[0218] Membrane potential update formula:

[0219]

[0220] Where V(t) is the membrane potential of the neuron at time t, the membrane potential represents the total response of the neuron to the input, and V(t-1) is the membrane potential of the neuron at the previous time step t-1. i Input x i The weight of (t) represents the connection strength, x i (t) Input signal at time t.

[0221] Pulse emission formula:

[0222]

[0223] Among them, S(t) * is the pulse emission signal, the pulse emitted at time t.

[0224] Threshold is the pulse emission threshold, which determines whether to emit a pulse. If the membrane potential V(t) exceeds the threshold, a pulse is emitted.

[0225] S4.2. Frequency domain feature learning.

[0226] Use biomimetic convolutional layers to process frequency domain feature maps. The biomimetic convolutional layer simulates the processing mechanism of biological visual cortex and captures frequency dependencies.

[0227]

[0228] Y ij is the output after the convolution operation, representing the value of position (i, j) in the feature map, X (i+m)(i+m) is the value of the input feature map at position (i+m,j+m), K mn is the weight of the convolution kernel and the filter of the convolution operation.

[0229] S4.3. Chaotic Domain Feature Learning

[0230] Use chaotic neural network to process chaotic domain feature map. Chaotic neural network simulates the behavior in biological chaotic system and extracts chaotic domain features.

[0231] Chaotic system formula:

[0232] X(t+1)=αX(t)(1-X(t)) (25)

[0233] X(t+1) is the state at time t+1, X(t) is the state at time t, and α is the parameter of the chaotic system, which determines the chaotic characteristics of the system.

[0234] S4.4. Constructing feature fusion module

[0235] (1) Feature aggregation: The time domain features S(t) and frequency domain features Y ij It is then converged with the chaotic domain features X(t+1) to form a feature matrix F.

[0236] F=[S(t)Y ij X(t+1)]

[0237] (2) Synaptic weight adjustment: The synaptic weight of the feature matrix F is adjusted to simulate the plasticity of biological synapses. Dynamic feature fusion is achieved by adaptively adjusting the weights of each feature domain.

[0238] Synaptic weight adjustment formula:

[0239] W ij =W ij +η·ΔW ij (26)

[0240] W ij is the weight of the position (i, j) of the fusion weight matrix, η is the learning rate, which controls the amplitude of weight adjustment, ΔW ij is the volume weight adjustment amount, which is calculated based on the model's training error. Through the above formula, the weight is adaptively adjusted to achieve feature fusion.

[0241] (3) Neural modulation mechanism: Introduce neural modulation mechanism (such as dopamine modulation) to dynamically adjust the weights in the feature fusion process and enhance the relevance of features and information transmission.

[0242] W ij (t+1)=W ij (t)·(1+γ·D(t)) (27)

[0243] γ is the modulation coefficient, D(t) is the modulation signal. Through the above formula, the weight is dynamically adjusted to improve the feature fusion effect.

[0244] (4) Fusion feature calculation: Apply synaptic weight adjustment and neural modulation mechanism to the feature matrix F to obtain the fusion feature F fused :

[0245] F fused =F·W

[0246] S4.5. Regularization and Optimization

[0247] Bio-inspired regularization: Synaptic pruning and neuron inhibition mechanisms are introduced to ensure the stability and efficiency of feature fusion.

[0248] Synaptic pruning formula:

[0249]

[0250] Here, θ is the pruning threshold.

[0251] S5. Use the global optimal firefly optimization algorithm to optimize the parameters of the biologically inspired multi-domain god model

[0252] S5.1. Initialize the firefly population

[0253] In the biologically inspired multi-domain neural model, the hyperparameters that need to be optimized are: learning rate (lr), hyperparameters of the time domain feature learning module (tls), hyperparameters of the frequency domain feature learning module (ffs), hyperparameters of the chaos domain feature learning module (chs), hyperparameters of the fusion module (fms) and hyperparameters of the classification head (cls). The global optimal firefly algorithm is used to optimize the hyperparameters, including the following steps:

[0254] (1) Set the value range of the hyperparameter to be optimized:

[0255] (lr): [0.0001,0.01]; (tls): [32,256]; (ffs): [64,512]; (chs): [64,512]; (fms): [32,256]; (cls): [32,256].

[0256] (2) Initialize firefly individuals: Initialize firefly individuals randomly within the above range, and each individual represents a hyperparameter combination. Assuming the population size is n, each firefly individual x i Contains six parameters:

[0257] X i =[lr i tls i ffs i chs i fms i cls i ]

[0258] S5.2. Calculate the fitness function.

[0259] (1) Training model: For each firefly individual, train the biologically inspired multi-domain neural model according to its hyperparameter combination. Use the specified hyperparameters during training and record the performance indicators of the model.

[0260] (2) Calculate the fitness function: Calculate the fitness value based on the performance index of the model, and use the validation error as the fitness function. Then the fitness value X i represents the fitness of the i-th firefly individual.

[0261] f(X i )=validation loss(X i )

[0262] Among them, validation loss (X i ) represents the loss function value of the i-th firefly individual on the validation set.

[0263] f(X i)=validation loss(X i ) has the structure:

[0264] (1) Input data is forward propagated through the model to obtain the predicted value The predicted value is calculated using the loss function mean square error loss The loss between the true value Y is:

[0265]

[0266] Among them, m is the number of samples in the validation set, is the predicted value of the jth sample, Y j is the true value of the jth sample.

[0267] Update the model parameters according to the gradient of the loss function;

[0268] Make predictions on the validation set and calculate the validation error. Assuming the number of samples in the validation set is m, the validation error can be calculated using the following formula:

[0269]

[0270] in, is the predicted value of the jth sample in the validation set, Y val,j is the true value of the jth sample.

[0271] S5.3. Update the firefly position.

[0272] Calculate the light intensity of each firefly individual, assuming I i Represents the light intensity of the i-th firefly individual:

[0273]

[0274] Update the position of individual fireflies according to the light intensity, and update the position of firefly i so that it moves towards firefly j with higher brightness:

[0275] x i =x i +β ij (x j -x i )+α(rand-0.5) (31)

[0276]

[0277] Among them, α is the random movement coefficient, rand is a random number that obeys uniform distribution, γ is the light intensity absorption coefficient, and r ijis the distance between fireflies i and j. Each firefly adjusts its position according to the attraction and random disturbance it receives. Firefly i will move towards firefly j with higher brightness, and the step length of the movement is determined by attraction and a random factor.

[0278] S5.4, iterative optimization to obtain the optimal solution.

[0279] Check the stop condition: Repeat steps S82 and S83, record the optimal fitness value and the corresponding hyperparameter combination, and terminate the firefly population movement when the maximum number of iterations is reached or the fitness value is no longer significantly improved. Select the firefly individual with the smallest fitness value (maximum light intensity) as the optimal hyperparameter combination.

[0280] S6. Use the model with the optimal hyperparameters to train and classify transformer audio.

[0281] The optimal hyperparameter combination recorded in step S54 is output to the bio-inspired multi-domain neural model for training, and the fused features obtained in step S44 are input to the fully connected layer for classification.

[0282] The fully connected layer is used to map the high-dimensional feature space to the category space. The output of the fully connected layer is the score Z of each category:

[0283] Z=W fc ·F fused +b

[0284] Among them, W fc and b are the weight matrix and bias term of the fully connected layer.

[0285] The softmax layer converts the output Z of the fully connected layer into a probability distribution. The formula of the softmax function is:

[0286]

[0287] Among them, P(y=i|F fused ) represents the input feature F fused The probability of belonging to class i.

[0288] The output classification result is the category with the highest probability:

[0289]

[0290] Finally, the model is trained on the complete training set, and its performance indicators such as accuracy, recall, and confusion matrix on the validation set and test set are recorded to obtain the optimal prediction model.

[0291] Based on the same inventive concept, the present invention also proposes a transformer fault identification system, comprising:

[0292] The audio signal of the transformer is collected, and the chaotic domain multi-scale fractal dimension characteristics, time domain zero crossing rate characteristics and frequency domain fractional-order spectrum entropy characteristics of the audio signal are extracted.

[0293] The recognition module is used to input the extracted audio signal features into the bio-inspired multi-domain neural network model to identify the fault type of the transformer; the bio-inspired multi-domain neural network model includes a pulse neural network, a bionic convolutional layer and a chaotic neural network; the fault type of the transformer identified by the bio-inspired multi-domain neural network model specifically includes:

[0294] The feature extraction unit is used to extract the multi-scale classification dimension features of the chaotic domain by using a chaotic neural network to obtain the chaotic domain features; the chaotic domain features are input into the bionic convolution layer as feedback signals to adjust the weights of the bionic convolution layer, and the frequency domain fractional-order spectrum entropy features are processed by the adjusted bionic convolution layer to obtain the frequency domain features; the frequency domain features and the chaotic domain features are introduced into the pulse neural network to adjust the membrane potential update rule of the pulse neural network, and the time domain zero crossing rate features are processed by the adjusted pulse neural network to obtain the time domain features.

[0295] The feature fusion unit is used to fuse the time domain features, frequency domain features and chaos domain features and input them into the fully connected layer to identify the fault category.

[0296] The present invention also proposes a transformer fault identification computer device, comprising: a memory, a processor, and a computer program stored in the memory, and the processor implements the steps of the transformer fault identification method when executing the computer program.

[0297] The present invention further provides a readable storage medium, which stores a computer program. The computer program includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the transformer fault identification method.

[0298] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A transformer fault identification method, characterized in that: The following steps are involved: Collect the audio signal of the transformer, and extract the chaotic domain multi-scale fractal dimension features, time domain zero crossing rate features, and frequency domain fractional-order spectrum entropy features of the audio signal; The extracted audio signal features are input into a bio-inspired multi-domain neural network model to identify the fault type of the transformer; the bio-inspired multi-domain neural network model includes a pulse neural network, a bionic convolutional layer and a chaotic neural network. The identification of the fault type of the transformer by the bio-inspired multi-domain neural network model specifically includes the following steps: A chaotic neural network is used to extract the multi-scale classification dimension features of the chaotic domain to obtain the chaotic domain features; the chaotic domain features are input into the bionic convolution layer as feedback signals to adjust the weights of the bionic convolution layer, and the frequency domain fractional-order spectrum entropy features are processed through the adjusted bionic convolution layer to obtain the frequency domain features; the frequency domain features and the chaotic domain features are introduced into the pulse neural network to adjust the membrane potential update rule of the pulse neural network, and the adjusted pulse neural network is used to process the time domain zero crossing rate features to obtain the time domain features; The time domain features, frequency domain features and chaos domain features are fused and input into the fully connected layer to identify the fault category.

2. A transformer fault identification method according to claim, characterized in that: The extraction of the chaotic domain multi-scale classification dimensional features specifically includes the following steps: The audio signal is segmented into fixed length L, each segment is denoted by x k (t), where K represents the segment number; For each segment signal x k (t) Perform discrete wavelet transform, select Daubechies wavelet basis function, and set different scale factors τ to τ1, τ2, ..., τ n ; Perform j-layer wavelet decomposition on the segmented signal: Among them A j is the approximation coefficient at different scales, D j is the detail coefficient; Calculate the Euclidean distance matrix D between points in the reconstructed phase space: i and j are the element subscript indices in the matrix D; Calculate the box count N(r) at different radii r in the distance matrix, and then find the classification dimension D f : in is a step function, when rD ij The value is 1 when ≥0, otherwise it is 0; The classification dimension D calculated under each scale factor τ f Arrange the scale factors in order, construct a multi-scale classification dimension feature matrix, and standardize the constructed multi-scale classification dimension feature matrix to ensure that the eigenvalues ​​are at the same order of magnitude; Among them, μ is the mean of the feature matrix, σ is the standard deviation of the feature matrix; The final chaotic domain characteristic matrix F is obtained, which is expressed as: in, Indicated by the scale factor τ j Next, the classification dimension value of the i-th signal segment.

3. A transformer fault identification method according to claim 1, characterized in that: The extraction of the time domain zero crossing rate feature comprises the following steps: Select the window length L and overlap length O to segment the signal, and segment the signal x(t) into multiple windows, each window contains L samples, and overlaps each other by O samples. The segmentation formula is expressed as: x k (n)=x(n+k(L-O)), Where k is the index of the window, n=0,1,2,3,4,5...,L-1; Calculate the window function with a window length of L: x k (n)=x(n+k(L-O)), For the short-time zero-crossing rate Z of the i-th window i Calculated as: where sgn(x) is the sign function, which is expressed as: The short-time zero-crossing rate Z of each window i Mapped onto the time axis, the required short-time zero-crossing rate feature image is formed, and the short-time zero-crossing rate feature representation in the time domain is obtained: Z=[Z1 Z2...Z T ], Where T is the total number of windows.

4. A transformer fault identification method according to claim 1, characterized in that: The extraction of the frequency domain fractional-order spectrum entropy feature specifically includes the following steps: Decompose the audio signal into a set of intrinsic mode functions, detect local extreme points of the audio signal, interpolate and calculate the mean of the upper and lower envelopes; calculate the details of the signal, and repeat this process until the IMF conditions are met; the IMF conditions include: The number of zero crossings and extrema is equal or differs by at most one, and the local mean is zero: IMF i (t) is the i-th eigenmode function, r n (t) is the remaining signal; IMF for the IMF i (t), and perform S transformation S{IMF i (t)}: in Get the parsed signal: z i (t)=IMF i (t)+iS{IMFj i (t)}(f,t), The details of the final signal D i (t) is expressed as: D i (t)=Re{z i (t)}=IMF i (t), Calculate the instantaneous amplitude A i (t) and instantaneous phase θ i (t) is: Calculate the instantaneous frequency ω i (t) is: Construct the S-transform time-frequency spectrum S(f,τ) to obtain the time-frequency representation of the signal: Where S(ω,t) represents the instantaneous spectrum value at frequency ω and time t; Under the obtained fractional order α, the spectrum H(ω,t) is subjected to fractional Fourier transform: Where K α (t,u) is the fractional Fourier kernel function; Calculate the spectrum FRFT after fractional Fourier transform α The energy spectral density of (ω) is: E(ω)=|FRFT α (oh)| 2 , Calculate the probability distribution P for each frequency band α (ω): Calculate the fractional spectral entropy S(α): According to the frequency ω, the fractional order α and the spectrum entropy value S(α), the fractional spectral entropy feature image is constructed, and the feature image is converted into a feature matrix M. Each element of the matrix corresponds to a pixel value in the feature image, and the final frequency domain matrix is ​​obtained, which is specifically expressed as: Where M(i,j) = S(α i ,ω j ), indicating that the element in the i-th row and j-th column of the feature matrix corresponds to the fractional order α i and frequency ω i The spectrum entropy value S(α i ,ω j ).

5. A transformer fault identification method according to claim 1, characterized in that: The method of introducing the frequency domain features and the chaotic domain features into the spiking neural network to adjust the membrane potential update rule of the spiking neural network, using the adjusted spiking neural network to simulate the pulse emission mechanism of biological neurons to process the time domain zero crossing rate features, and obtaining the time domain features, specifically includes the following steps: The pulse firing mechanism of the spiking neural network simulating biological neurons is expressed as: Among them, S(t) * is a pulse emission signal, a pulse emitted at time t; V(t) represents the membrane potential of the neuron at time t, and the membrane potential represents the total response of the neuron to the input; the membrane potential update formula is expressed as: Where V(t-1) is the membrane potential of the neuron at the previous time step t-1; ω i Input x i The weight of (t) indicates the connection strength; x i (t) input signal at time t; The bionic convolutional layer simulates the processing mechanism of the biological visual cortex, and its processing process is expressed as: Among them, Y ij is the output after the convolution operation, indicating the value of position (i, j) in the feature map; X (i+m)(i+m) is the value of the input feature map at position (i+m,j+m); K mn is the weight of the convolution kernel; Chaotic neural network simulates the behavior mechanism in biological chaotic system, which is expressed as: X(t+1)=αX(t)(1-X(t)), X(t+1) is the state at time t+1, X(t) is the state at time t, and α is the adjustment coefficient; The frequency domain feature Y extracted by the bionic convolution layer ij After the introduction of the pulse neural network, its membrane potential is expressed as: V(t) is the membrane potential at time t, V(t+1) is the membrane potential at time t+1, ω i is the input signal x i (t), β is the adjustment coefficient; X(t+1) is the frequency domain feature input signal of the bionic convolution layer; The output X of the chaotic neural network * (t+1) is used as a feedback signal to adjust the membrane potential update rule V(t) of the spiking neural network and the weight K of the bionic convolutional layer mn , specifically expressed as: K mn (t+1)=K mn (t)·(1+γ·X * (t+1)), Among them, β and γ are adjustment coefficients; By introducing the chaotic domain characteristics into the spiking neural network, the adjusted membrane potential update rule is expressed as: The pulse neural network adjusted based on the membrane potential update rule is used to process the time domain zero crossing rate feature to obtain the time domain feature S(t).

6. A transformer fault identification method according to claim 5, characterized in that: The fusion of the time domain features, the frequency domain features and the chaos domain features specifically includes the following steps: The time domain feature S(t) and frequency domain feature Y ij And the chaotic domain features X(t+1) are converged to form a feature matrix F: F=[S(t) Y ij X(t+1)], Adjust the synaptic weights of the feature matrix F: W=W ij +η ΔW ij , Among them, W ij is the weight of the position (i, j) of the fusion weight matrix; η is the learning rate, which controls the amplitude of weight adjustment, ΔW ij is the volume weight adjustment amount; Introduce the neural modulation mechanism to adjust the weights in the feature fusion process: W ij (t+1)=W ij (t)·(1+γ·D(t)), γ is the modulation coefficient, D(t) is the modulation signal; Apply synaptic weight adjustment and neural modulation mechanism to feature matrix F to obtain fused feature F fused : F fused =F·W。 7. A transformer fault identification method according to claim 1, characterized in that: The method also includes optimizing the biologically inspired multi-domain neural model using a global optimal firefly optimization algorithm; wherein the hyperparameters optimized in the biologically inspired multi-domain neural model include: a learning rate lr, a hyperparameter tls of a time domain feature learning module, a hyperparameter ffs of a frequency domain feature learning module, a hyperparameter chs of a chaos domain feature learning module, a hyperparameter fms of a fusion module, and a hyperparameter cls of a classification head; and optimizing the hyperparameters using a global optimal firefly algorithm includes the following steps: Set the value range of the hyperparameter to be optimized: lr: [0.0001,0.01]; tls: [32,256]; ffs: [64,512]; chs: [64,512]; fms: [32,256]; cls: [32,256]; Initialize firefly individuals: Randomly initialize firefly individuals within the value range of the hyperparameter to be optimized. Each individual represents a hyperparameter combination. Assuming the population size is n, each firefly individual x i Contains six parameters: X i =[lr i tls i ffs i chs i fms i cls i ], For each firefly individual, a biologically inspired multi-domain neural model is trained according to its hyperparameter combination; the specified hyperparameters are used during training and the performance metrics of the model are recorded; The fitness value is calculated according to the performance indicators of the biologically inspired multi-domain neural model, and the validation error is used as the fitness function. The fitness value X i Represents the fitness of the i-th firefly individual: f(X i )=validation loss(X i ), Among them, validation loss (X i ) represents the loss function value of the i-th firefly individual on the validation set; f(X i )=validation loss(X i ) is structured as follows: the input data is propagated forward through the model to obtain the predicted value The predicted value is calculated using the loss function mean square error loss The loss between the true value Y: Among them, m is the number of samples in the validation set, is the predicted value of the jth sample, Y j is the true value of the jth sample; Update the model parameters according to the gradient of the loss function; Make predictions on the validation set and calculate the validation error. Assuming the number of samples in the validation set is m, the validation error is expressed as: in, is the predicted value of the jth sample in the validation set, Y val,j is the true value of the jth sample; Calculate the light intensity of each firefly individual, assuming I i Represents the light intensity of the i-th firefly individual: Update the position of individual fireflies according to the light intensity, and update the position of firefly i so that it moves towards firefly j with higher brightness: x i =x i +b ij (x j -x i )+α(rand-0.5), Among them, α is the random movement coefficient, rand is a random number that obeys uniform distribution, γ is the light intensity absorption coefficient, and r ij is the distance between fireflies i and j. Each firefly adjusts its position according to the attraction and random disturbance it receives. Firefly i will move towards firefly j with higher brightness, and the step length of the movement is determined by the attraction and a random factor. After multiple iterative updates, the optimal fitness value and the corresponding hyperparameter combination are recorded. When the maximum number of iterations is reached or the fitness value is no longer updated, the movement of the firefly population is terminated, and the firefly individual with the smallest fitness value is selected as the optimal hyperparameter combination.

8. A transformer fault identification system, characterized in that: include: Collect the audio signal of the transformer, and extract the chaotic domain multi-scale fractal dimension features, time domain zero crossing rate features, and frequency domain fractional-order spectrum entropy features of the audio signal; The recognition module is used to input the extracted audio signal features into a bio-inspired multi-domain neural network model to identify the fault type of the transformer; the bio-inspired multi-domain neural network model includes a pulse neural network, a bionic convolutional layer and a chaotic neural network; identifying the fault type of the transformer by the bio-inspired multi-domain neural network model specifically includes: The feature extraction unit is used to extract the multi-scale classification dimension features of the chaotic domain by using a chaotic neural network to obtain the chaotic domain features; the chaotic domain features are input into the bionic convolution layer as feedback signals to adjust the weight of the bionic convolution layer, and the frequency domain fractional-order spectrum entropy features are processed by the adjusted bionic convolution layer to obtain the frequency domain features; the frequency domain features and the chaotic domain features are introduced into the pulse neural network to adjust the membrane potential update rule of the pulse neural network, and the time domain zero crossing rate features are processed by the adjusted pulse neural network to obtain the time domain features; The feature fusion unit is used to fuse the time domain features, frequency domain features and chaos domain features and input them into the fully connected layer to identify the fault category.

9. A transformer fault identification computer device, characterized in that: include: A memory, a processor, and a computer program stored in the memory, wherein the processor implements the steps of the transformer fault identification method according to any one of claims 1 to 7 when executing the computer program.

10. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and the computer program includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the transformer fault identification method according to any one of claims 1 to 7.

Images