IPO-based multi-modal fusion power equipment fault identification method and system

Abstract

The invention discloses an IPO-based multi-mode fusion power equipment fault identification method and system, and the method comprises the steps: carrying out the variational mode decomposition of an original vibration signal and a voiceprint signal, and obtaining a denoised vibration signal and a denoised voiceprint signal; processing the denoised vibration signals and voiceprint signals through a Mel filter bank, a Barker filter bank and a Gammatone filter bank to obtain a cepstrum group; performing Gaussian filtering and image segmentation on the infrared image to obtain an infrared target image; and inputting the obtained multi-modal data into the multi-modal fusion power transformer fault identification model after the hyper-parameters are optimized by the improved Meland spar optimization algorithm, and carrying out power equipment fault identification. According to the method, the limitation of a traditional single-mode diagnosis method is overcome, and the accuracy and efficiency of power transformer fault identification are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment fault identification, and particularly to a multi-modal fusion power equipment fault identification method and system based on IPO (Improved Puma Optimization Algorithm). Background Art

[0002] As the core equipment of the power system, the reliability of the operation state of a power transformer directly affects the stable operation of the power system. Traditional single-modal diagnosis methods, such as fault diagnosis relying only on single-modal data such as vibration signals, voiceprint signals, or infrared images, have limitations. The multi-modal data of power transformers has the characteristics of multi-source, heterogeneity, and temporality, and it is difficult for single-modal diagnosis methods to comprehensively and accurately identify faults. In addition, traditional optimization algorithms also have problems such as poor global search ability and being easily trapped in local optimal solutions when dealing with the optimization of model hyperparameters, which affect the accuracy and efficiency of fault identification. Summary of the Invention

[0003] The present invention aims to provide a multi-modal fusion power equipment fault identification method and system based on IPO, which overcomes the limitations of traditional single-modal diagnosis methods and improves the accuracy and efficiency of power transformer fault identification.

[0004] The present invention is realized through the following technical solutions. The multi-modal fusion power equipment fault identification method based on IPO comprises the following steps: S1: Perform variational mode decomposition (VMD) on the acquired original vibration signal and voiceprint signal to obtain the denoised vibration signal and voiceprint signal. Process the denoised vibration signal and voiceprint signal through a Mel filter bank, a Bark filter bank, and a gamma-pass filter bank to obtain a cepstrum map group; perform Gaussian filtering and image segmentation on the infrared image to obtain an infrared target image; the denoised vibration signal and voiceprint signal, the cepstrum map group, and the infrared target image form multi-modal data; S2: Construct a multi-modal fusion power transformer fault identification model, which includes three branches. The first branch inputs the denoised vibration signal and voiceprint signal, the second branch inputs the cepstrum map group, and the third branch inputs the infrared target image. The first branch uses one-dimensional convolution to extract features from the denoised vibration signal and voiceprint signal, and then uses a long short-term memory network for processing; the second branch uses depthwise separable convolution (DSCNN) to extract features from the cepstrum map group; the third branch uses two-dimensional convolution (2D CNN) to extract features from the infrared target image; use bidirectional cross-attention (BCA) to obtain the cross-attention information of different modal data from the features extracted from the three branches, and then perform feature fusion through a fusion layer, and obtain the fault type through a fully connected layer; S3: Use the Improved Puma Optimisation (IPO) algorithm to optimize the hyperparameters of the multimodal fusion power transformer fault recognition model, and input the multimodal data into the optimized multimodal fusion power transformer fault recognition model for power transformer fault recognition; in the improved puma optimization algorithm, tent chaotic mapping is used to improve population initialization, and the t-distribution mutation improvement strategy is used to improve the original puma optimization algorithm.

[0005] Further preferably, use the Mel filter bank, Bark filter bank, and gammatone filter bank to extract Mel-frequency cepstral coefficients (MFCC), Bark-frequency cepstral coefficients (BFCC), and gammatone-frequency cepstral coefficients (GFCC) respectively to obtain a set of cepstral diagrams.

[0006] Further preferably, convert between frequency and Mel frequency through the Mel scale, and construct the Mel filter bank according to the Mel frequency; convert between frequency and Bark frequency through the Bark scale, and construct the Bark filter bank according to the Bark frequency; convert between frequency and gammatone frequency through the gammatone scale, and construct the gammatone filter bank according to the gammatone frequency; convert the frequency domain of the filter bank to the cepstral domain through the discrete cosine transform.

[0007] Further preferably, use the region growing method for image segmentation.

[0008] Further preferably, use the improved puma optimization algorithm (IPO) to optimize the hyperparameters of the multimodal fusion power transformer fault recognition model. The specific steps are as follows: Step 1: Obtain and label the multimodal data, and then divide it into a training set and a test set; Step 2: Set the maximum number of iterations, population size, and hyperparameters; Step 3: Create an initial random population with better uniformity and ergodicity through tent chaotic mapping; Step 4: Calculate the population fitness value; Step 5: Update the population position and the optimal solution through the exploration and exploitation phases of the algorithm, apply t-distribution mutation for position mutation, and update the population position and the optimal solution again; Step 6: Determine whether the termination condition is met (reaching the maximum number of iterations or meeting the accuracy requirement). If the termination condition is met, output the optimal hyperparameters to obtain the optimized multimodal fusion power transformer fault recognition model. Otherwise, return to Step 4.

[0009] Further preferably, the hyperparameters include the learning rate, the size and number of convolutional kernels, and the number of neurons in the fully connected layer.

[0010] The present invention also provides a multimodal fusion power equipment fault recognition system based on IPO, including: Data acquisition module, used to obtain original vibration signals, voiceprint signals, and infrared images; The feature engineering module performs variational modal decomposition on the original vibration signal and voiceprint signal to obtain the denoised vibration signal and voiceprint signal, processes the denoised vibration signal and voiceprint signal through a Mel filter group, a Bark filter group and a gammatone filter group to obtain a cepstrum group; performs Gaussian filtering and image segmentation on the infrared image to obtain an infrared target image; the denoised vibration signal and voiceprint signal, the cepstrum group and the infrared target image constitute multimodal data; The fault identification module has a built-in multi-modal fusion power transformer fault identification model, which includes three branches. The first branch inputs the denoised vibration signal and voiceprint signal, the second branch inputs the cepstral group, and the third branch inputs the infrared target image. The first branch uses one-dimensional convolution to extract features from the denoised vibration signal and voiceprint signal, and then uses a long short-term memory network for processing; the second branch uses deep separable convolution to extract features from the cepstral group; the third branch uses two-dimensional convolution to extract features from the infrared target image; bidirectional cross attention is used to obtain the cross attention information of different modal data from the features extracted by the three branches, and then feature fusion is performed through the fusion layer, and the fault type is obtained through the fully connected layer; The model optimization module adopts the improved Puma optimization algorithm to optimize the hyperparameters of the multi-modal fusion power transformer fault identification model. The improved Puma optimization algorithm adopts tent chaotic mapping to improve population initialization and adopts t distribution variation improvement strategy to improve the original Puma optimization algorithm.

[0011] Beneficial effects of the present invention: 1. The present invention integrates multi-modal data such as vibration signals, voiceprint signals and infrared images for fault identification, fully utilizes multi-source data information, overcomes the limitations of traditional single-modal diagnosis methods, and improves the accuracy of fault identification.

[0012] 2. The constructed multi-modal fusion power transformer fault identification model performs effective feature extraction and fusion according to the characteristics of different modal data, which can better mine the potential information in multi-modal data and improve the fault identification ability.

[0013] 3. The IPO algorithm is used to optimize the model hyperparameters, which improves the shortcomings of the traditional PO algorithm, improves the global search capability and local development capability of the algorithm, enables the model to converge to the optimal solution faster, and improves the efficiency of fault identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 The figure is a flow chart of the method of the present invention.

[0015] Figure 2Schematic diagram of the process for optimizing the hyperparameters of a multi-modal fusion power transformer fault identification model using the cougar optimization algorithm.

[0016] Figure 3 Average optimal fitness value curves of each optimization algorithm when testing function F1; Figure 4 Average optimal fitness value curves of each optimization algorithm when testing function F2; Figure 5 Average optimal fitness value curves of each optimization algorithm when testing function F3; Figure 6 Average optimal fitness value curves of each optimization algorithm when testing function F4; Figure 7 Comparison of the convergence curves of the improved cougar optimization algorithm (IPO) and the original cougar optimization algorithm (PO); Figure 8 Transformer fault diagnosis results of the multi-modal fusion power transformer fault identification model with optimal hyperparameters. Specific implementation manner

[0017] The following will refer to the accompanying drawings to describe in detail the specific embodiments of the present application. According to these detailed descriptions, those skilled in the art can clearly understand the present application. Without departing from the principle of the present application, the features in different embodiments can be combined to obtain new implementation manners, or some features in certain embodiments can be replaced to obtain other preferred implementation manners.

[0018] Refer to Figure 1 , the multi-modal fusion power equipment fault identification method based on IPO is as follows: S1: Perform variational mode decomposition (VMD) on the acquired original vibration signal and voiceprint signal to obtain the denoised vibration signal and voiceprint signal. Process the denoised vibration signal and voiceprint signal through a Mel filter bank, a Bark filter bank, and a gamma-pass filter bank to obtain a cepstrum map group; perform Gaussian filtering and image segmentation on the infrared image to obtain an infrared target image; the denoised vibration signal and voiceprint signal, the cepstrum map group, and the infrared target image form multi-modal data; S2: Construct a multi-modal fusion power transformer fault identification model, which includes three branches. The first branch inputs the denoised vibration signal and voiceprint signal, the second branch inputs the cepstrum diagram group, and the third branch inputs the infrared target image. The first branch uses one-dimensional convolution to extract features from the denoised vibration signal and voiceprint signal, and then uses a long short-term memory network for processing; the second branch uses depthwise separable convolution (DSCNN) to extract features from the cepstrum diagram group; the third branch uses two-dimensional convolution (2D CNN) to extract features from the infrared target image; bidirectional cross-attention (BCA) is used to obtain the cross-attention information of different modal data from the features extracted by the three branches, and then feature fusion is performed through a fusion layer, and the fault type is obtained through a fully connected layer. S3: Use the Improved Puma Optimisation (IPO) algorithm to optimize the hyperparameters of the multi-modal fusion power transformer fault identification model, and input the multi-modal data into the optimized multi-modal fusion power transformer fault identification model for power transformer fault identification.

[0019] In step S1, the vibration signal and voiceprint data are preprocessed by variational mode decomposition (VMD), specifically as follows: Variational mode decomposition (VMD) is a noise reduction processing technology that can eliminate the influence of noise collected by sensors on fault diagnosis. The signal decomposition process of the VMD algorithm is as follows: Construct a variational problem: ; where: represents the k-th modal component, represents the center frequency of the k-th mode, represents the partial derivative with respect to time t, represents the Dirac function, is the convolution operator, t is time, j is the imaginary unit, is the original signal, and e is the natural constant.

[0020] Solve the variational problem: ; where: L(•) is the augmented Lagrangian function; is the Lagrange multiplier; is the quadratic penalty factor, . Introduce the quadratic penalty factor and the Lagrange multiplier , under the condition of maintaining the strictness of constraints and meeting the accuracy of the reconstructed signal, the constrained variational problem is transformed into an unconstrained optimization problem. Using the Alternating Direction Method of Multipliers (ADMM), each mode and its respective center frequency are adjusted to obtain the optimal , and .

[0021] ; where: " " represents the Fourier transform, represents the value of the Fourier transform of the i-th mode component at the (n + 1)-th iteration, is the Fourier transform of the original signal, is the value of the Fourier transform of the Lagrange multiplier at the n-th iteration at the frequency domain ω, is the value of the Fourier transform of the k-th mode component at the (n + 1)-th iteration, is the center frequency of the k-th mode component at the (n + 1)-th iteration, The k-th mode component at the -th iteration of the Fourier transform at the angular frequency ω, is the number of iterations, is the noise tolerance to ensure the fidelity of signal decomposition.

[0022] The denoised vibration signal and voiceprint signal are processed by the Mel filter bank, Bark filter bank and gammatone filter bank to obtain a group of cepstrum diagrams, specifically: The original vibration signal and voiceprint signal are a non-stationary and non-linear time series. In order to better capture the vibration and voiceprint characteristics of transformer operation, and also considering the high adaptability and effectiveness of the spectral transformation method in feature engineering technology for extracting the characteristics of time series data, the Mel filter bank, Bark filter bank and gammatone filter bank are used to extract the Mel cepstral coefficients (MFCC), Bark cepstral coefficients (BFCC) and gammatone cepstral coefficients (GFCC) respectively to obtain a group of cepstrum diagrams.

[0023] The methods for obtaining MFCC, BFCC and GFCC are: Pre - enhance the signal, apply windowing, and perform Fast Fourier Transform (FFT); construct Mel filter banks, Bark filter banks, and gammatone filter banks; apply Discrete Cosine Transform (DCT) to decorrelate the filter bank coefficients. By enhancing the high - frequency components, the signal - to - noise ratio is increased and signal distortion is reduced; by windowing, spectral leakage is reduced; by Fast Fourier Transform, the signal is converted into the energy distribution in the frequency domain. By using Mel scale, Bark scale, and gammatone scale respectively to construct Mel filter banks, Bark filter banks, and gammatone filter banks to extract the vibration signals and voiceprint features of different frequency components. By applying Discrete Cosine Transform to decorrelate the filter bank coefficients, an cepstrum map group containing MFCC, BFCC, and GFCC is obtained.

[0024] The pre - enhancement operation is as shown in the following formula: ; Where: is the signal at time t after variational mode decomposition and noise reduction, is the signal at time t - 1 after variational mode decomposition and noise reduction, is the signal at time t after pre - enhancement, t is time, and β is the pre - enhancement coefficient.

[0025] The windowing operation is as shown in the following formula: ; Where: is the windowed signal, L is the window length, W s is the window step length, m is the window index, is the Hamming window function.

[0026] The Fast Fourier Transform operation is as shown in the following formula: ; Where: is the frequency - domain signal, u is the frequency - component index, is the sample value of the time - domain signal at the discrete time point i, j is the imaginary unit, and N is the total number of discrete time points.

[0027] Multiple filter banks are used to extract the vibration signals and voiceprint features of different frequency components, specifically: Convert between frequency and Mel frequency through Mel scale, and construct Mel filter banks according to Mel frequency. The conversion formula is as follows: ; Where: is the Mel frequency, is the frequency.

[0028] Convert between frequency and Bark frequency through Bark scale, and construct Bark filter banks according to Bark frequency. The conversion formula is as follows: ; Where: is the Barker frequency.

[0029] Convert between the frequency and the gammaton frequency through gammaton scaling, and construct a gammaton filter bank according to the gammaton frequency. The conversion formula is as follows: ; Where: is the gammaton frequency.

[0030] Convert the frequency domain of the filter bank to the cepstrum domain through discrete cosine transform. The formula is as follows: ; Where: is the discrete sequence of the filter bank frequencies, n is the discrete index in the frequency domain, is the cepstrum coefficient of the filter bank, is the scaling factor to keep the discrete cosine transform orthogonal and with unit energy, is the value of the d-th cosine basis function at the p-th point, M is the number of filter banks, d is the discrete index in the cepstrum domain, and p is a variable related to the discrete sequence index in the frequency domain, is the auxiliary variable defining the scaling factor .

[0031] In step S1, perform Gaussian filtering on the infrared image and use the region growing method for image segmentation to accurately segment the continuous high-temperature area, avoid the influence of background noise and background temperature, and obtain the infrared target image.

[0032] In step S2, the multi-modal fusion power transformer fault identification model includes three branches. The first branch inputs the denoised vibration signal and voiceprint signal, the second branch inputs the cepstrum diagram group, and the third branch inputs the infrared target image. Considering the timeliness and continuity of the one-dimensional vibration signal and voiceprint signal, the first branch uses one-dimensional convolution (1DCNN) to extract features from the denoised vibration signal and voiceprint signal, and then uses the long short-term memory network (LSTM) to mine the potential time-dependent relationships in the time-series data. Considering the discreteness of the cepstrum diagram group of the vibration signal and voiceprint signal, and the characteristics that the cepstrum diagrams extracted by different filter banks are decoupled from each other and source-related, the second branch uses depthwise separable convolution (DSCNN) separable in channels and space to extract features from the cepstrum diagram group. Considering the low resolution and ambiguity of the infrared image, the third branch uses two-dimensional convolution (2D CNN) to extract features from the infrared target image. Considering the data source correlation and mutual learnability between different modal data, bidirectional cross-attention (BCA) is used to obtain the cross-attention information of different modal data from the features extracted by the three branches, and then feature fusion is performed through the fusion layer, and the fault type is obtained through the fully connected layer.

[0033] Considering the original Puma Optimization (PO) algorithm, since the initial positions of the population are random, it may reduce the population diversity, resulting in poor global search ability and low convergence efficiency. At the same time, the PO algorithm is prone to falling into local optimal solutions in the later stage. The original Puma Optimization algorithm is improved by using tent chaotic mapping to improve population initialization and t-distribution mutation improvement strategy, which improves the diversity of the initial population and enhances the global search ability and local development ability of the algorithm.

[0034] Compared with the random function initialization of the original Puma Optimization algorithm, the tent chaotic mapping has better ergodicity and uniformity, and the formula is as follows: ; where: T q is the value of the tent chaotic mapping at the q-th iteration, T q+1 is the value of the tent chaotic mapping at the (q + 1)-th iteration, R is the control parameter, is the initial population position, is the lower limit of the population value, is the upper limit of the population value, is the tent chaotic mapping value matrix used to generate the population.

[0035] The t-distribution mutation combines the advantages of Cauchy mutation and Gaussian mutation, and can enhance the global search ability and local development ability of the Puma Optimization algorithm. The formula is as follows: ; where: is the position of the i-th population, is the position of the i-th population after mutation, represents the t-distribution with degrees of freedom, and the degrees of freedom is equal to the number of iterations, which determines the intensity and characteristics of the mutation.

[0036] Considering that the multi-modal fusion power transformer fault identification model needs to set hyperparameters such as the learning rate, the size and number of convolutional kernels, and the number of neurons in the fully connected layer, which have an important impact on the diagnostic accuracy and convergence speed. Therefore, the improved puma optimization algorithm (IPO) is used to optimize the hyperparameters of the multi-modal fusion power transformer fault identification model to obtain the optimal parameters. The optimization process is as Figure 2 shown, and the specific steps are as follows: Step 1: Obtain multi-modal data and label it, and then divide it into a training set and a test set; Step 2: Set the maximum number of iterations, population size, and hyperparameters; Step 3: Create an initial random population with better uniformity and ergodicity through the tent chaotic map; Step 4: Calculate the population fitness value.

[0037] Step 5: Update the population position and the optimal solution through the exploration and exploitation phases of the algorithm, apply t-distribution mutation for position mutation, and update the population position and the optimal solution again; Step 6: Determine whether the termination condition is met (reaching the maximum number of iterations or meeting the accuracy requirement). If the termination condition is met, output the optimal hyperparameters to obtain the optimized multi-modal fusion power transformer fault identification model. Otherwise, return to Step 4.

[0038] Input the multi-modal data into the multi-modal fusion power transformer fault identification model with the optimal hyperparameters to achieve accurate identification of power transformer faults. Specifically: Improve the traditional PO algorithm by fusing the tent chaotic map and t-distribution mutation to obtain the IPO algorithm; divide the multi-modal data into a training set and a test set; select the training set to optimize the hyperparameters of the multi-modal fusion power transformer fault diagnosis model using the IPO algorithm. Based on the trained multi-modal fusion power transformer fault identification model, perform fault diagnosis on the test set and output performance indicators such as the diagnostic result and the accuracy rate.

[0039] To verify the performance of the IPO algorithm, four test functions F1 - F4 in CEC17 were selected. Among them, F1 - F2 are unimodal functions, F3 is a simple multimodal function, and F4 is a composite function composed of multiple benchmark functions. And it was compared with WOA (Whale Optimization Algorithm), MSO (Mirage Search Optimization Algorithm), and PO under the same settings of other parameters. The average optimal fitness value curves of each optimization algorithm are as Figures 3 - 6 shown. Compared with the WOA algorithm, MSO algorithm, and PO algorithm, the IPO algorithm designed in the present invention has a faster convergence speed and higher accuracy for the four test functions, proving the effectiveness of the IPO algorithm improvement strategy.

[0040] To test the effect of the IPO algorithm on optimizing the hyperparameters of the fault diagnosis model, it was compared with the PO algorithm. The initial population size was set to 10, and the maximum number of iterations was 20. It was tested on the training set, and the test results are shown in Figure 7 . From Figure 7 it can be seen that due to the introduction of the tent chaotic map to improve the population initialization method and the population update strategy based on t - distribution mutation, in the early stage of iteration, the IPO algorithm obtained a higher accuracy rate with a more uniform initial population; in the middle stage of iteration, compared with the PO algorithm (the population update strategy is greedy selection based on fitness value), the IPO algorithm has a faster convergence speed; but in the end, due to a more traversable initial population and a stronger global search ability, the IPO algorithm obtained network hyperparameters with a higher diagnostic accuracy, meeting the high requirements for diagnostic accuracy in transformer fault diagnosis.

[0041] After the optimal hyperparameters were obtained through the IPO algorithm, a multi - modal fusion power transformer fault identification model with the optimal hyperparameters was used for transformer fault diagnosis, and the results are as Figure 8 shown. From Figure 8 it can be seen that the difference between the true fault type and the fault type is small. This method can achieve high - precision fault diagnosis and provide a strong guarantee for the safe and stable operation of the transformer.

[0042] Another embodiment of the present invention provides a multi - modal fusion power equipment fault identification system based on IPO, including: A data acquisition module for acquiring original vibration signals, voiceprint signals, and infrared images; A feature engineering module that performs variational mode decomposition on the acquired original vibration signals and voiceprint signals to obtain denoised vibration signals and voiceprint signals, processes the denoised vibration signals and voiceprint signals through Mel filter banks, Bark filter banks, and gamma - pass filter banks to obtain cepstrum map groups; performs Gaussian filtering and image segmentation on the infrared images to obtain infrared target images; the denoised vibration signals and voiceprint signals, cepstrum map groups, and infrared target images form multi - modal data; The fault identification module is built with a multi-modal fusion power transformer fault identification model, which includes three branches. The first branch inputs the denoised vibration signal and voiceprint signal, the second branch inputs the cepstrum diagram group, and the third branch inputs the infrared target image. The first branch uses one-dimensional convolution to extract features from the denoised vibration signal and voiceprint signal, and then uses a long short-term memory network for processing; the second branch uses depthwise separable convolution to extract features from the cepstrum diagram group; the third branch uses two-dimensional convolution to extract features from the infrared target image; bidirectional cross-attention is used to obtain the cross-attention information of different modal data from the features extracted by the three branches, and then feature fusion is performed through a fusion layer, and the fault type is obtained through a fully connected layer. The model optimization module uses an improved puma optimization algorithm to optimize the hyperparameters of the multi-modal fusion power transformer fault identification model. The improved puma optimization algorithm uses tent chaotic mapping to improve population initialization and uses a t-distribution mutation improvement strategy to improve the original puma optimization algorithm.

[0043] The above only expresses the preferred embodiments of the present invention and does not limit the present invention in other forms. Any person skilled in the art may use the disclosed content to make changes or modifications into equivalent embodiments with equivalent changes. However, any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A multi-modal fusion power equipment fault identification method based on IPO, characterized in that The steps are as follows: S1: Perform variational mode decomposition on the acquired original vibration signal and voiceprint signal to obtain the vibration signal and voiceprint signal after noise reduction. Process the vibration signal and voiceprint signal after noise reduction through a Mel filter bank, a Bark filter bank, and a gammatone filter bank to obtain a cepstrum map group; perform Gaussian filtering and image segmentation on the infrared image to obtain an infrared target image; the vibration signal and voiceprint signal after noise reduction, the cepstrum map group, and the infrared target image constitute multi-modal data; S2: Construct a multi-modal fusion power transformer fault identification model, including three branches. The first branch inputs the vibration signal and voiceprint signal after noise reduction, the second branch inputs the cepstrum map group, and the third branch inputs the infrared target image. The first branch uses one-dimensional convolution to extract features from the vibration signal and voiceprint signal after noise reduction, and then uses a long short-term memory network for processing; the second branch uses depthwise separable convolution to extract features from the cepstrum map group; the third branch uses two-dimensional convolution to extract features from the infrared target image; use bidirectional cross-attention to obtain cross-attention information of different modal data from the features extracted by the three branches, then perform feature fusion through a fusion layer, and obtain the fault type through a fully connected layer; S3: Use an improved puma optimization algorithm to optimize the hyperparameters of the multi-modal fusion power transformer fault identification model, and input the multi-modal data into the optimized multi-modal fusion power transformer fault identification model for power transformer fault identification; for the improved puma optimization algorithm, use tent chaotic mapping to improve population initialization, and use the t-distribution mutation improvement strategy to improve the original puma optimization algorithm.

2. The multimodal fusion power equipment fault identification method according to claim 1, wherein Use a Mel filter bank, a Bark filter bank, and a gammatone filter bank to extract Mel cepstral coefficients, Bark cepstral coefficients, and gammatone cepstral coefficients respectively to obtain a cepstrum map group.

3. The multimodal fusion power equipment fault identification method according to claim 1, wherein, Convert between frequency and Mel frequency through the Mel scale, and construct a Mel filter bank according to the Mel frequency; convert between frequency and Bark frequency through the Bark scale, and construct a Bark filter bank according to the Bark frequency; convert between frequency and gammatone frequency through the gammatone scale, and construct a gammatone filter bank according to the gammatone frequency; convert the frequency domain of the filter bank to the cepstrum domain through discrete cosine transform.

4. The multi-modal fusion power equipment fault identification method according to claim 1, characterized in that Use the region growing method for image segmentation.

5. The multimodal fusion power equipment fault identification method according to claim 1, wherein Use an improved puma optimization algorithm to optimize the hyperparameters of the multi-modal fusion power transformer fault identification model. The steps are as follows: Step 1: Acquire and label multi-modal data, and then divide it into a training set and a test set; Step 2: Set the maximum number of iterations, population size, and hyperparameters; Step 3: Create an initial random population with better uniformity and ergodicity through tent chaotic mapping; Step 4: Calculate the population fitness value; Step 5: Update the population position and the optimal solution through the exploration and exploitation phases of the algorithm, apply t-distribution mutation for position mutation, and update the population position and the optimal solution again; Step 6: Judge whether the termination condition is satisfied. If the termination condition is satisfied, output the optimal hyperparameters to obtain the optimized multi-modal fusion power transformer fault identification model. Otherwise, return to Step 4.

6. The multimodal fusion power equipment fault identification method according to claim 4, wherein, The hyperparameters include the learning rate, the size and number of convolutional kernels, and the number of neurons in the fully connected layer.

7. A multi-modal fusion power equipment fault identification system based on IPO, characterized in that, It includes: A data acquisition module for acquiring original vibration signals, voiceprint signals, and infrared images; A feature engineering module that performs variational mode decomposition on the acquired original vibration signals and voiceprint signals to obtain denoised vibration signals and voiceprint signals, processes the denoised vibration signals and voiceprint signals through a Mel filter bank, a Bark filter bank, and a gamma-pass filter bank to obtain a group of cepstrum diagrams; performs Gaussian filtering and image segmentation on the infrared images to obtain infrared target images; the denoised vibration signals and voiceprint signals, the group of cepstrum diagrams, and the infrared target images form multimodal data; A fault identification module with a multimodal fusion power transformer fault identification model built in, including three branches. The first branch inputs the denoised vibration signals and voiceprint signals, the second branch inputs the group of cepstrum diagrams, and the third branch inputs the infrared target images. The first branch uses one-dimensional convolution to extract features from the denoised vibration signals and voiceprint signals, and then uses a long short-term memory network for processing; the second branch uses depthwise separable convolution to extract features from the group of cepstrum diagrams; the third branch uses two-dimensional convolution to extract features from the infrared target images; bidirectional cross-attention is used to obtain cross-attention information of different modal data from the features extracted from the three branches, and then feature fusion is performed through a fusion layer, and the fault type is obtained through a fully connected layer; A model optimization module that uses an improved cougar optimization algorithm to optimize the hyperparameters of the multimodal fusion power transformer fault identification model. The improved cougar optimization algorithm improves the population initialization using tent chaotic mapping and improves the original cougar optimization algorithm using a t-distribution mutation improvement strategy.

Images