Transformer fault diagnosis method based on golden vehicle optimization algorithm

Abstract

The invention relates to the technical field of transformer fault diagnosis, and particularly provides a transformer fault diagnosis method based on a golden vehicle optimization algorithm. The method comprises the steps that a GJO-SDAE joint optimization network model is constructed, classification accuracy and reconstruction errors serve as fitness functions, and network structure parameters and hyper-parameter combinations of SDAE are synchronously optimized; designing an anti-noise enhancement mechanism through a GJO-SDAE combined optimization network model, and meanwhile, designing a weighted fusion strategy of multi-level noise reduction features; according to the method, the optimized network model is obtained through GJO, meanwhile, an interpretability module based on feature importance is introduced, a diagnosis decision basis with semantic tags is generated by analyzing a weight matrix optimized through GJO, and the transformer fault diagnosis performance is improved through the method.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer fault diagnosis, and in particular to a transformer fault diagnosis method based on a golden jackal optimization algorithm. Background Art

[0002] Power transformers, as key components of power grid energy transmission, have a direct impact on the stability of the power system. Internal faults can trigger cascading power outages, resulting in tens of billions of yuan in annual economic losses and significant safety risks. Dissolved Gas Analysis (DGA) technology detects the concentration and generation rate of characteristic gases such as H2, CH4, and C2H2 in transformer oil, creating a unique gas fingerprint that provides key evidence for diagnosing latent transformer faults. For example, a C2H2 / C2H4 ratio exceeding 1 typically indicates a discharge fault, while a change in the CH4 / H2 ratio is closely related to the severity of localized overheating. However, DGA monitoring data has inherent flaws. Sensor noise, variations in oil sample storage conditions, and coupled fault modes result in high-dimensional, nonlinear data, and a significant imbalance in the ratio of normal to faulty samples. Traditional diagnostic methods, such as the International Electrotechnical Commission (IEC) three-ratio method, rely heavily on manually encoded rules. Shallow models like support vector machines (SVMs) struggle to capture complex gas correlation patterns and exhibit significant performance bottlenecks in the presence of noise and edge case detection.

[0003] The rise of deep learning technology has provided a new paradigm for DGA diagnosis. The Stacked Denoising Autoencoder (SDAE) constructs deep feature representations through layer-by-layer greedy training. Its unique denoising pre-training mechanism, such as the use of masked or additive noise, gives the model dual advantages: automatic filtering of high-frequency noise interference at the encoding layer; and the ability to capture the nonlinear manifold structure between gas parameters, avoiding the artifacts of traditional feature engineering. However, the performance of SDAE is deeply coupled to structural parameters such as network architecture depth and neuron size, as well as hyperparameters such as learning rate and sparsity constraints. Existing research has largely relied on empirically driven trial-and-error tuning or grid search, resulting in exponentially increasing computational costs, prone to falling into local optimal solutions, and unable to adapt to the dynamically changing DGA data distribution. Summary of the Invention

[0004] In view of this, the present invention provides a transformer fault diagnosis method based on the Golden Jackal optimization algorithm to improve the transformer fault diagnosis performance.

[0005] In a first aspect, the present invention provides a transformer fault diagnosis method based on the Golden Jackal optimization algorithm, the method comprising:

[0006] Step 1: Construct a joint optimization network model of the Golden Jackal Optimization Algorithm-Stacked Denoising Autoencoder (GJO) and the SDAE. The network structure parameters and hyperparameter combinations of the SDAE are optimized simultaneously using the classification accuracy and reconstruction error as the fitness function.

[0007] Step 2: Through the GJO-SDAE joint optimization network model, an anti-noise enhancement mechanism is designed, and a weighted fusion strategy of multi-level noise reduction features is designed;

[0008] Step 3: Obtain the optimized network model through GJO, and introduce the interpretability module based on feature importance. By analyzing the weight matrix optimized by GJO, a diagnostic decision basis with semantic labels is generated.

[0009] Optionally, constructing the GJO-SDAE joint optimization network model in step 1 includes:

[0010] Initialize the golden jackal population, the expression is:

[0011] Y0=Y min +rand×(Y max -Y min );

[0012] Among them, Y0 represents the position of the initial golden jackal population; rand is a random number in the range of 0 to 1; Y min and Y max Represent the upper and lower bounds of the hunting problem respectively;

[0013] The prey matrix Prey is expressed as:

[0014]

[0015] Among them, Yi,j represents the j-th dimension position of the i-th prey, n represents the number of prey, and d represents the dimension of the hunting problem. The fitness matrix of the prey is established according to the objective function, and its expression is:

[0016]

[0017] Where f(·) is a multi-objective fitness function, which assigns the best and second-best fitness values to male and female golden jackals, respectively, to form a golden jackal pair to capture prey;

[0018] Assuming that the prey's escape energy factor is E, when |E|>1, the golden jackal pair is in the prey tracking stage; when |E|≤1, the golden jackal pair is in the encirclement and attack stage. The expression of the escape energy factor is:

[0019] E=E1×E0;

[0020] Among them, E1 is the process of reducing the prey's escape energy, and E0 represents the prey's initial escape energy, which is expressed as:

[0021]

[0022] E0=2*rand-1;

[0023] Where c1 is a constant with a value of 1.5; t represents the current number of iterations, T represents the maximum number of iterations, and rand is a random number in the range of 0 to 1.

[0024] The golden jackal's position update process is divided into two stages: tracking and encirclement and attack. When tracking prey, assuming that the male golden jackal and the female golden jackal approach the prey together, the position update formula is:

[0025]

[0026] Among them, Y M (t) represents the position of the male golden jackal at the tth iteration; Y FM (t) represents the position of the female golden jackal at the tth iteration; Y1(t) represents the population position obtained under the guidance of the male golden jackal, and Y2(t) represents the population position obtained under the guidance of the female golden jackal; Y(t) represents the population position after t iterations, and l is a random number;

[0027] Based on Levy distribution, its expression is:

[0028] l=0.05×Levy

[0029]

[0030] Among them, β is 1.5; η and α obey the normal distribution:

[0031] σ η and σ α satisfy: σ α =1;

[0032] Where Γ(·) is the gamma function;

[0033] SDAE is composed of multiple stacked denoising autoencoders (DAEs). It extracts robust features of the data through unsupervised pre-training and supervised fine-tuning, reconstructs the original data from noisy input, and learns noise-invariant features. Multiple DAEs are cascaded layer by layer, with the output of the previous layer serving as the input of the next layer. Each layer is trained separately to minimize the reconstruction error of the current layer. After stacking all layers, a Softmax classifier is added, and the entire network is fine-tuned using labeled data. The Golden Jackal Optimization (GJO) algorithm encodes the parameters of the SDAE's number of hidden layers, number of nodes, and learning rate as optimization variables, and uses classification accuracy and reconstruction error as fitness functions to search for the globally optimal parameter combination. GJO is used to optimize the noise injection intensity to match the network capacity.

[0034] Optionally, the fitness function using classification accuracy and reconstruction error includes:

[0035] Design a multi-objective fitness function, the objective function fitness, whose expression is:

[0036]

[0037] Accuracy represents the classification accuracy. SDAE encoding features are used to train the support vector machine (SVM) classifier. The average value is obtained from the five-fold cross validation.

[0038] The reconstruction error MSE is expressed as:

[0039]

[0040] Where N is the number of samples, x i is the actual value, is the predicted value.

[0041] Optionally, step 2 includes:

[0042] By embedding an adaptive noise injection module in the input layer of the stacked denoising autoencoder (SDAE), the noise ratio parameters are optimized in real time by the Golden Jackal Optimization (GJO) algorithm, and the noise intensity distribution is dynamically adjusted. At the same time, a hierarchical feature fusion strategy is designed to perform weighted splicing of the denoising features of multiple hidden layers, and the weight coefficients are determined by GJO optimization.

[0043] Optionally include:

[0044] a. Noise simulation: add Gaussian white noise to the original data. The noise has a mean of 0 and a variance of σ. 2 Normal distribution: ε~Ν(0,σ 2 ), data after adding noise: x noisy =x+ε; where the variance σ 2 Controlling the noise intensity is defined by the signal-to-noise ratio (SNR), which is expressed as:

[0045]

[0046] in, represents the original signal variance, represents the variance after adding noise;

[0047] b. Design hierarchical feature fusion, using a three-level progressive feature extraction architecture. The first layer focuses on local detail features and retains nonlinear features through the activation function ReLU; the second layer constructs combined correlation features and combines batch normalization with a 20% dropout layer to enhance generalization; the third layer extracts global abstract features and uses the activation function Sigmoid to output probability distribution features. Design a learnable feature fusion, which is expressed as:

[0048]

[0049] Among them, w i is the weight coefficient, F i is the distribution characteristic value;

[0050] The Golden Jackal Optimization (GJO) algorithm is used to automatically search for the optimal weight combination in the interval [0, 1], and the Fisher criterion is used as the optimization objective to maximize the inter-class discrimination. The weight search process is carried out simultaneously with model training to form an end-to-end joint optimization framework.

[0051] Optionally, step 3 includes:

[0052] Through multiple rounds of GJO iterative search, the network model automatically streamlined the network structure to a 3-hidden layer configuration, compressing the network model parameters to less than 40% while maintaining a diagnostic accuracy of ≥98.5%;

[0053] An interpretability module based on feature importance is constructed, and the encoder weight matrix optimized by parsing GJO is used to reversely map the association strength between key gas components and fault types to generate diagnostic decision-making basis with semantic labels.

[0054] Optionally include:

[0055] Optimize the network model structure. During the initialization phase, a flexible structure with a variable number of neurons in each layer is set based on a predefined three-hidden layer network model. The structural parameters of the network depth, number of neurons in each layer, and activation function type are encoded as optimization variables of the GJO algorithm. In designing the optimized fitness function, the two conflicting goals of classification accuracy and model compression rate are comprehensively considered.

[0056] Using an optimization goal priority design, the optimization is divided into two stages: Stage 1: Rapid Convergence: Optimizing MSE and Compression Rate with only a classification accuracy of ≥ 98.5% as the constraint; Stage 2: Fine-tuning: After the fixed classification accuracy is achieved, further minimizing MSE with a model compression rate of ≤ 40% as the hard constraint;

[0057] The construction of the interpretability module first extracts feature importance. Based on the optimized encoder weight matrix, a layer-by-layer backpropagation algorithm is used to quantify the contribution of each gas component to the encoded output. The gradient-weighted class activation mapping (Grad-CAM) technology is introduced to generate a feature importance heat map. Then, association rules are generated. By constructing a Bayesian network model of gas composition and fault type, the encoder weights are converted into conditional probability parameters. Markov blanket analysis is used to identify the causal boundaries of key gas components and automatically generate diagnostic rules with physical significance. Finally, the decision basis is visualized and an interactive diagnostic interface is developed to present four-dimensional information in real time.

[0058] Optionally, for dynamic structure search, the GJO algorithm performs global exploration in the solution space by simulating the hunting behavior of jackal packs; each iteration generates a new candidate solution for the network structure, and uses the escape energy factor E to control population diversity to avoid falling into local optimality; when the classification accuracy of the candidate structure is less than a threshold, the structural reorganization mechanism is automatically triggered.

[0059] In a second aspect, an embodiment of the present invention provides a computer-readable storage medium, which includes a stored program, wherein when the program is running, the device where the computer-readable storage medium is located is controlled to execute the transformer fault diagnosis method based on the Golden Jackal optimization algorithm in the first aspect or any possible implementation of the first aspect.

[0060] In a third aspect, an embodiment of the present invention provides an electronic device, comprising: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory, and the one or more computer programs include instructions, which, when executed by the device, enable the device to execute the transformer fault diagnosis method based on the Golden Jackal optimization algorithm in the first aspect or any possible implementation of the first aspect.

[0061] In the technical solution provided by the present invention, the method includes constructing a Golden Jackal Optimization Algorithm-Stacked Noise Reduction Autoencoder GJO-SDAE joint optimization network model, taking classification accuracy and reconstruction error as fitness functions, and simultaneously optimizing the network structure parameters and hyperparameter combinations of SDAE; through the GJO-SDAE joint optimization network model, designing an anti-noise enhancement mechanism, and at the same time designing a weighted fusion strategy of multi-level noise reduction features; obtaining the optimized network model through GJO, and at the same time introducing an interpretability module based on feature importance, and generating a diagnostic decision basis with semantic labels by parsing the weight matrix optimized by GJO. This method improves the transformer fault diagnosis performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0063] Figure 1 A flowchart of a transformer fault diagnosis method provided by an embodiment of the present invention;

[0064] Figure 2 A schematic diagram of fault probability judgment provided by an embodiment of the present invention;

[0065] Figure 3 A schematic diagram of four-dimensional information provided by an embodiment of the present invention; Figure 3 (a) is the gas content distribution histogram; Figure 3 (b) is a schematic diagram of feature contribution ranking; Figure 3 (c) Schematic diagram of the diagnostic rule with highlighted activation in the association rule matching results; Figure 3 (d) is a three-dimensional scatter matrix diagram;

[0066] Figure 4 A schematic diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0067] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0068] It should be understood that the embodiments described are only a portion of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative work are within the scope of protection of the present invention.

[0069] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The singular forms "a", "the" and "the" used in the embodiments of the present invention are also intended to include plural forms, unless the context clearly indicates other meanings.

[0070] It should be understood that the term "and / or" as used herein simply describes a relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A alone, A and B together, or B alone. Furthermore, the character " / " in this document generally indicates an "or" relationship between the associated objects.

[0071] The word "if," as used herein, may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to the determination" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)," depending on the context.

[0072] Swarm intelligence optimization algorithms offer an effective path to overcoming this dilemma. Compared to traditional gradient descent methods, the Golden Jackal Optimization (GJO) algorithm exhibits unique optimization characteristics by simulating the group hunting behavior of male golden jackals. The dynamic hunting mechanism introduces an escape energy factor E and a collaboration coefficient C to control population diversity and information interaction intensity, respectively, achieving an adaptive balance between global exploration and local development. Experiments have shown that GJO converges faster than classic algorithms such as particle swarm optimization (PSO) and genetic algorithms (GA) when solving high-dimensional, non-convex optimization problems. Although previous studies have validated GJO's superiority in tasks such as feature selection and neural network pruning, its systematic application in SDAE architecture search and DGA diagnosis remains unexplored. In particular, establishing a co-evolutionary framework between algorithms and deep learning models has become a key breakthrough in improving the performance of intelligent transformer diagnosis.

[0073] Figure 1 A flowchart of a transformer fault diagnosis method provided by an embodiment of the present invention is shown in FIG. Figure 1 As shown, the method includes:

[0074] Step 1: Construct a joint optimization network model of the Golden Jackal optimization algorithm-stacked denoising autoencoder GJO-SDAE, use classification accuracy and reconstruction error as fitness functions, and simultaneously optimize the network structure parameters and hyperparameter combinations of SDAE.

[0075] The Golden Jackal Optimization (GJO) algorithm is a metaheuristic optimization algorithm based on swarm intelligence, inspired by the collaborative hunting behavior of golden jackals. The algorithm simulates the process of male golden jackals (leaders) and female golden jackals (followers) collaborating to capture prey. It balances global exploration with local exploitation by dynamically adjusting the search strategy.

[0076] In the embodiment of the present invention, constructing the GJO-SDAE joint optimization network model in step 1 includes:

[0077] Initialize the golden jackal population, the expression is:

[0078] Y0=Y min +rand×(Y max -Y min );

[0079] Among them, Y0 represents the position of the initial golden jackal population; rand is a random number in the range of 0 to 1; Y min and Y max Represent the upper and lower bounds of the hunting problem respectively;

[0080] The prey matrix Prey is expressed as:

[0081]

[0082] Among them, Yi,j represents the j-th dimension position of the i-th prey, n represents the number of prey, and d represents the dimension of the hunting problem. The fitness matrix of the prey is established according to the objective function, and its expression is:

[0083]

[0084] Where f(·) is a multi-objective fitness function, which assigns the best and second-best fitness values to male and female golden jackals, respectively, to form a golden jackal pair to capture prey;

[0085] Assuming that the prey's escape energy factor is E, when |E|>1, the golden jackal pair is in the prey tracking stage; when |E|≤1, the golden jackal pair is in the encirclement and attack stage. The expression of the escape energy factor is:

[0086] E=E1×E0;

[0087] Among them, E1 is the process of reducing the prey's escape energy, and E0 represents the prey's initial escape energy, which is expressed as:

[0088]

[0089] E0=2*rand-1;

[0090] Where c1 is a constant with a value of 1.5; t represents the current number of iterations, T represents the maximum number of iterations, and rand is a random number in the range of 0 to 1.

[0091] The golden jackal's position update process is divided into two stages: tracking and encirclement and attack. When tracking prey, assuming that the male golden jackal and the female golden jackal approach the prey together, the position update formula is:

[0092]

[0093] Among them, Y M (t) represents the position of the male golden jackal at the tth iteration; Y FM (t) represents the position of the female golden jackal at the tth iteration; Y1(t) represents the population position obtained under the guidance of the male golden jackal, and Y2(t) represents the population position obtained under the guidance of the female golden jackal; Y(t) represents the population position after t iterations, and l is a random number;

[0094] Based on Levy distribution, its expression is:

[0095] l=0.05×Levy

[0096]

[0097] Among them, β is 1.5; η and α obey the normal distribution:

[0098] σ η and σ α satisfy: σ α =1;

[0099] Where Γ(·) is the gamma function;

[0100] SDAE is a deep generative model composed of multiple stacked denoising autoencoders (DAEs). It extracts robust features from data through unsupervised pre-training and supervised fine-tuning, reconstructs original data from noisy inputs, and learns noise-invariant features. Multiple DAEs are cascaded layer by layer, with the output of the previous layer serving as the input to the next. For example, the sequence is: input layer → DAE1 (hidden layer 1) → DAE2 (hidden layer 2) → ... Each layer is trained independently to minimize the reconstruction error of the current layer. After stacking all layers, a softmax classifier is added, and the entire network is fine-tuned using labeled data. SDAE's performance is significantly affected by the network structure (number of layers and nodes) and hyperparameters (learning rate, noise ratio), making traditional trial-and-error methods inefficient. The Golden Jackal Optimized (GJO) algorithm encodes the SDAE's parameters (number of hidden layers, number of nodes, and learning rate) as optimization variables. It uses classification accuracy and reconstruction error as fitness functions to search for the globally optimal parameter combination. GJO optimizes the noise injection intensity to match the network capacity, improving noise resistance.

[0101] In the embodiment of the present invention, the fitness function using the classification accuracy and reconstruction error includes:

[0102] Design a multi-objective fitness function, the objective function fitness, whose expression is:

[0103]

[0104] Where, accuracy represents the classification accuracy. SDAE encoding features are used to train the support vector machine (SVM) classifier (RBF kernel, C = 1.0); the mean value is taken from the five-fold cross validation.

[0105] The reconstruction error MSE is expressed as:

[0106]

[0107] Where N is the number of samples, x i is the actual value, is the predicted value.

[0108] In the embodiment of the present invention, before step 1, the following steps are further included:

[0109] ① Data acquisition: 10,000 sets of historical DGA data were obtained from the power system monitoring platform. Each set contained five gas contents (H2, CH4, C2H2, C2H4, C2H6) and their derived ratio characteristics (IEC three ratios and Dornenburg ratio).

[0110] ②Z-score standardization:

[0111]

[0112] Where μ is the mean and ρ is the variance;

[0113] ③ Dataset division: divide the training set, validation set, and test set into 7:2:1 to maintain category balance.

[0114] Step 2: Through the GJO-SDAE joint optimization network model, an anti-noise enhancement mechanism is designed, and a weighted fusion strategy of multi-level noise reduction features is designed.

[0115] In this embodiment of the present invention, step 2 includes:

[0116] By embedding an adaptive noise injection module in the input layer of the stacked denoising autoencoder (SDAE), the noise ratio parameters are optimized in real time by the Golden Jackal Optimization (GJO) algorithm, and the noise intensity distribution is dynamically adjusted. At the same time, a hierarchical feature fusion strategy is designed to perform weighted splicing of the denoising features of multiple hidden layers, and the weight coefficients are determined by GJO optimization.

[0117] In an embodiment of the present invention, the present invention includes:

[0118] a. Noise simulation: add Gaussian white noise to the original data. The noise has a mean of 0 and a variance of σ. 2 Normal distribution: ε~Ν(0,σ 2 ), data after adding noise: x noisy =x+ε; where the variance σ 2 Controlling the noise intensity is defined by the signal-to-noise ratio (SNR), which is expressed as:

[0119]

[0120] in, represents the original signal variance, represents the variance after adding noise;

[0121] For example, SNR = 20dB corresponds to σ noise =0.1×σ signal .

[0122] b. Design hierarchical feature fusion, using a three-level progressive feature extraction architecture. The first layer focuses on local detail features and retains nonlinear features through the activation function ReLU; the second layer constructs combined correlation features and combines batch normalization with a 20% dropout layer to enhance generalization; the third layer extracts global abstract features and uses the activation function Sigmoid to output probability distribution features. Design a learnable feature fusion, which is expressed as:

[0123]

[0124] Among them, w i is the weight coefficient, F iis the distribution eigenvalue; the Golden Jackal Optimization algorithm (GJO) is used to automatically search for the optimal weight combination in the interval [0, 1], and the Fisher criterion is used as the optimization objective to maximize the inter-class discrimination; the weight search process is carried out simultaneously with model training to form an end-to-end joint optimization framework.

[0125] Step 3: Obtain the optimized network model through GJO, and introduce the interpretability module based on feature importance. By analyzing the weight matrix optimized by GJO, generate the diagnostic decision basis with semantic labels.

[0126] In this embodiment of the present invention, step 3 includes:

[0127] Through multiple rounds of GJO iterative search, the network model automatically streamlined the network structure to a 3-hidden layer configuration, compressing the network model parameters to less than 40% while maintaining a diagnostic accuracy of ≥98.5%;

[0128] An interpretability module based on feature importance is constructed. The encoder weight matrix optimized by parsing GJO is used to reversely map the correlation strength between key gas components and fault types, and generate diagnostic decision-making basis with semantic labels, achieving a dual improvement in high-precision diagnosis and model interpretability.

[0129] In an embodiment of the present invention, the present invention includes:

[0130] To optimize the network model structure, during the initialization phase, a flexible structure with a variable number of neurons in each layer was set based on a predefined three-hidden layer network model. Structural parameters such as network depth, number of neurons per layer, and activation function type were encoded as optimization variables for the GJO algorithm. For the design of the optimized fitness function, two conflicting objectives were considered: classification accuracy (target value ≥ 98.5%) and model compression rate (target value ≤ 40%).

[0131] Using an optimization goal priority design, the optimization is divided into two stages: Stage 1: Rapid Convergence: Optimizing MSE and Compression Rate with only a classification accuracy of ≥ 98.5% as the constraint; Stage 2: Fine-tuning: After the fixed classification accuracy is achieved, further minimizing MSE with a model compression rate of ≤ 40% as the hard constraint;

[0132] For dynamic structure search, the GJO algorithm simulates the hunting behavior of jackal packs to conduct global exploration in the solution space. Each iteration generates new candidate network structure solutions and uses the escape energy factor E to control population diversity and avoid falling into local optimality. When the classification accuracy of the candidate structure is less than the threshold, the structural reorganization mechanism is automatically triggered.

[0133] The construction of the interpretability module first extracts the importance of features. Based on the optimized encoder weight matrix, a layer-by-layer backpropagation algorithm is used to quantify the contribution of each gas component to the encoded output. The gradient-weighted class activation mapping (Grad-CAM) technology is introduced to generate a feature importance heat map. Then, association rules are generated. By building a Bayesian network model of gas components and fault types, the encoder weights are converted into conditional probability parameters. Markov blanket analysis is used to identify the causal boundaries of key gas components, and diagnostic rules with physical significance are automatically generated. Eight strong association rules are generated to judge the fault probability and determine the rule confidence, such as Figure 2 As shown, the gas content types include hydrogen H2, methane CH4, ethane C2H6, ethylene C2H4, and acetylene C2H2; the association rules are 1-8; and the fault types include partial discharge, arc discharge, spark discharge, high temperature overheating, medium temperature overheating, and low temperature overheating. For example, "When the H2 concentration is >150μL / L and the CH4 concentration is >80μL / L, the probability of partial discharge is >90%." The rule confidence level exceeds 95%. Finally, the decision basis is visualized, as shown in the figure below. Figure 3 As shown, an interactive diagnostic interface is developed to present four-dimensional information in real time: Figure 3 As shown in (a), the gas content distribution histogram shows the concentration level of each component in the current sample; Figure 3 As shown in (b), the feature importance ranking dynamically updates the contribution ranking of each gas; Figure 3 As shown in (c), the association rule matching results highlight the activated diagnostic rules; Figure 3 As shown in (d), the 3D scatter matrix reveals the combined effects of multiple gas components. For example, the H₂-CH₄-C₂H₂ concentration combination forms distinct fault type clusters in 3D space. This visualization interface supports threshold parameter adjustment, allowing users to intuitively explore the boundary effects of different gas combinations on fault identification.

[0134] In the technical solution provided by the present invention, the method includes constructing a Golden Jackal Optimization Algorithm-Stacked Noise Reduction Autoencoder GJO-SDAE joint optimization network model, taking classification accuracy and reconstruction error as fitness functions, and simultaneously optimizing the network structure parameters and hyperparameter combinations of SDAE; through the GJO-SDAE joint optimization network model, designing an anti-noise enhancement mechanism, and at the same time designing a weighted fusion strategy of multi-level noise reduction features; obtaining the optimized network model through GJO, and at the same time introducing an interpretability module based on feature importance, and generating a diagnostic decision basis with semantic labels by parsing the weight matrix optimized by GJO. This method improves the transformer fault diagnosis performance.

[0135] Each step of the embodiment of the present invention may be performed by an electronic device, including but not limited to a tablet computer, a portable PC, a desktop computer, etc.

[0136] An embodiment of the present invention provides a computer-readable storage medium, which includes a stored program. When the program runs, the electronic device where the computer-readable storage medium is located is controlled to execute the above-mentioned embodiment of the transformer fault diagnosis method based on the Golden Jackal optimization algorithm.

[0137] Figure 4 A schematic diagram of an electronic device provided by an embodiment of the present invention is shown in FIG. Figure 4 As shown, the electronic device 21 includes: a processor 211, a memory 212, and a computer program 213 stored in the memory 212 and executable on the processor 211. When the computer program 213 is executed by the processor 211, the transformer fault diagnosis method based on the Golden Jackal optimization algorithm in the embodiment is implemented. To avoid repetition, they are not described here one by one.

[0138] The electronic device 21 includes, but is not limited to, a processor 211 and a memory 212. Those skilled in the art will understand that Figure 4 It is only an example of the electronic device 21 and does not constitute a limitation of the electronic device 21. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the electronic device may also include input and output devices, network access devices, buses, etc.

[0139] The processor 211 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0140] The memory 212 can be an internal storage unit of the electronic device 21, such as a hard disk or memory of the electronic device 21. The memory 212 can also be an external storage device of the electronic device 21, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash memory card, etc. equipped on the electronic device 21. Furthermore, the memory 212 can also include both an internal storage unit of the electronic device 21 and an external storage device. The memory 212 is used to store computer programs and other programs and data required by the network device. The memory 212 can also be used to temporarily store data that has been output or is about to be output.

[0141] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0142] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A transformer fault diagnosis method based on the Golden Jackal optimization algorithm, characterized in that: The method comprises: Step 1: Construct a joint optimization network model of the Golden Jackal Optimization Algorithm-Stacked Denoising Autoencoder (GJO) and the SDAE. The network structure parameters and hyperparameter combinations of the SDAE are optimized simultaneously using the classification accuracy and reconstruction error as the fitness function. Step 2: Through the GJO-SDAE joint optimization network model, an anti-noise enhancement mechanism is designed, and a weighted fusion strategy of multi-level noise reduction features is designed; Step 3: Obtain the optimized network model through GJO, and introduce the interpretability module based on feature importance. By analyzing the weight matrix optimized by GJO, a diagnostic decision basis with semantic labels is generated.

2. The method according to claim 1, characterized in that The step 1 of constructing the GJO-SDAE joint optimization network model includes: Initialize the golden jackal population, the expression is: Y0=Y min +rand×(Y max -AND min ); Among them, Y0 represents the position of the initial golden jackal population; rand is a random number in the range of 0 to 1; Y min and Y max Represent the upper and lower bounds of the hunting problem respectively; The prey matrix Prey is expressed as: Among them, Yi,j represents the j-th dimension position of the i-th prey, n represents the number of prey, and d represents the dimension of the hunting problem. The fitness matrix of the prey is established according to the objective function, and its expression is: Where f(·) is a multi-objective fitness function, which assigns the best and second-best fitness values to male and female golden jackals, respectively, to form a golden jackal pair to capture prey; Assuming that the prey's escape energy factor is E, when |E|>1, the golden jackal pair is in the prey tracking stage; when |E|≤1, the golden jackal pair is in the encirclement and attack stage. The expression of the escape energy factor is: E=E1×E0; Among them, E1 is the process of reducing the prey's escape energy, and E0 represents the prey's initial escape energy, which is expressed as: E0=2*rand-1; Where c1 is a constant with a value of 1.5; t represents the current number of iterations, T represents the maximum number of iterations, and rand is a random number in the range of 0 to 1. The golden jackal's position update process is divided into two stages: tracking and encirclement and attack. When tracking prey, assuming that the male golden jackal and the female golden jackal approach the prey together, the position update formula is: Among them, Y M (t) represents the position of the male golden jackal at the tth iteration; Y FM (t) represents the position of the female golden jackal at the tth iteration; Y1(t) represents the population position obtained under the guidance of the male golden jackal, and Y2(t) represents the population position obtained under the guidance of the female golden jackal; Y(t) represents the population position after t iterations, and l is a random number; Based on Levy distribution, its expression is: l=0.05×Levy Among them, β is 1.5; η and α obey the normal distribution: σ η and σ α satisfy: σ α =1; Where Γ(·) is the gamma function; SDAE is composed of multiple stacked denoising autoencoders (DAEs). It extracts robust features of the data through unsupervised pre-training and supervised fine-tuning, reconstructs the original data from noisy input, and learns noise-invariant features. Multiple DAEs are cascaded layer by layer, with the output of the previous layer serving as the input of the next layer. Each layer is trained separately to minimize the reconstruction error of the current layer. After stacking all layers, a Softmax classifier is added, and the entire network is fine-tuned using labeled data. The Golden Jackal Optimization (GJO) algorithm encodes the parameters of the SDAE's number of hidden layers, number of nodes, and learning rate as optimization variables, and uses classification accuracy and reconstruction error as fitness functions to search for the globally optimal parameter combination. GJO is used to optimize the noise injection intensity to match the network capacity.

3. The method according to claim 1 or 2, characterized in that The fitness function using classification accuracy and reconstruction error includes: Design a multi-objective fitness function, the objective function fitness, whose expression is: Accuracy represents the classification accuracy. SDAE encoding features are used to train the support vector machine (SVM) classifier. The average value is obtained from the five-fold cross validation. The reconstruction error MSE is expressed as: Where N is the number of samples, x i is the actual value, is the predicted value.

4. The method according to claim 1, wherein The step 2 includes: By embedding an adaptive noise injection module in the input layer of the stacked denoising autoencoder (SDAE), the noise ratio parameters are optimized in real time by the Golden Jackal Optimization (GJO) algorithm, and the noise intensity distribution is dynamically adjusted. At the same time, a hierarchical feature fusion strategy is designed to perform weighted splicing of the denoising features of multiple hidden layers, and the weight coefficients are determined by GJO optimization.

5. The method according to claim 1 or 4, characterized in that include: a. Noise simulation: add Gaussian white noise to the original data. The noise has a mean of 0 and a variance of σ. 2 Normal distribution: ε~Ν(0,σ 2 ), data after adding noise: x noisy =x+ε; where the variance σ 2 Controlling the noise intensity is defined by the signal-to-noise ratio (SNR), which is expressed as: in, represents the original signal variance, represents the variance after adding noise; b. Design hierarchical feature fusion, using a three-level progressive feature extraction architecture. The first layer focuses on local detail features and retains nonlinear features through the activation function ReLU; the second layer constructs combined correlation features and combines batch normalization with a 20% dropout layer to enhance generalization; the third layer extracts global abstract features and uses the activation function Sigmoid to output probability distribution features. Design a learnable feature fusion, which is expressed as: Among them, w i is the weight coefficient, F i is the distribution characteristic value; The Golden Jackal Optimization (GJO) algorithm is used to automatically search for the optimal weight combination in the interval [0, 1], and the Fisher criterion is used as the optimization objective to maximize the inter-class discrimination. The weight search process is carried out simultaneously with model training to form an end-to-end joint optimization framework.

6. The method according to claim 1, characterized in that The step 3 comprises: Through multiple rounds of GJO iterative search, the network model automatically streamlined the network structure to a 3-hidden layer configuration, compressing the network model parameters to less than 40% while maintaining a diagnostic accuracy of ≥98.5%; An interpretability module based on feature importance is constructed, and the encoder weight matrix optimized by parsing GJO is used to reversely map the association strength between key gas components and fault types to generate diagnostic decision-making basis with semantic labels.

7. The method according to claim 1 or 6, characterized in that include: Optimize the network model structure. During the initialization phase, a flexible structure with a variable number of neurons in each layer is set based on a predefined three-hidden layer network model. The structural parameters of the network depth, number of neurons in each layer, and activation function type are encoded as optimization variables of the GJO algorithm. In designing the optimized fitness function, the two conflicting goals of classification accuracy and model compression rate are comprehensively considered. Using an optimization goal priority design, the optimization is divided into two stages: Stage 1: Rapid Convergence: Optimizing MSE and Compression Rate with only a classification accuracy of ≥ 98.5% as the constraint; Stage 2: Fine-tuning: After the fixed classification accuracy is achieved, further minimizing MSE with a model compression rate of ≤ 40% as the hard constraint; The construction of the interpretability module first extracts feature importance. Based on the optimized encoder weight matrix, a layer-by-layer backpropagation algorithm is used to quantify the contribution of each gas component to the encoded output. The gradient-weighted class activation mapping (Grad-CAM) technology is introduced to generate a feature importance heat map. Then, association rules are generated. By constructing a Bayesian network model of gas composition and fault type, the encoder weights are converted into conditional probability parameters. Markov blanket analysis is used to identify the causal boundaries of key gas components and automatically generate diagnostic rules with physical significance. Finally, the decision basis is visualized and an interactive diagnostic interface is developed to present four-dimensional information in real time.

8. The method according to claim 7, characterized in that include: For dynamic structure search, the GJO algorithm performs global exploration in the solution space by simulating the hunting behavior of a jackal group; Each iteration generates a new candidate solution for the network structure, and uses the escape energy factor E to control population diversity and avoid falling into local optimality. When the classification accuracy of the candidate structure is less than the threshold, the structural reorganization mechanism is automatically triggered.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a stored program, wherein when the program is running, the device where the computer-readable storage medium is located is controlled to execute the transformer fault diagnosis method based on the Golden Jackal optimization algorithm according to any one of claims 1 to 8.

10. An electronic device, characterized in that: include: one or more processors; Memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory, and the one or more computer programs include instructions, which, when executed by the device, enable the device to execute the transformer fault diagnosis method based on the Golden Jackal optimization algorithm according to any one of claims 1 to 8.

Images