Distribution transformer fault diagnosis method based on full-band multi-modal data fusion

Abstract

The invention discloses a distribution transformer fault diagnosis method based on full-band multi-modal data fusion. The method comprises the following steps: collecting different feature maps of a transformer; performing feature extraction and weighted summation on the collected different feature maps to obtain a multi-source fusion feature vector; performing data preprocessing on the multi-source fusion feature vector, and performing normalization processing after data preprocessing; training an improved probabilistic neural network by using the normalized multi-source fusion features; optimizing the polynomial order of the trained improved probabilistic neural network by adopting an improved parrot optimization algorithm; and carrying out transformer fault diagnosis by adopting the optimized improved probabilistic neural network. According to the method, the improved parrot-shaped optimization algorithm is adopted, the foraging behavior of the parrot-shaped algorithm is simulated, the polynomial order of the improved probabilistic neural network is optimized by effectively combining global search and local search strategies, and the accuracy of transformer data fusion diagnosis and the generalization ability of the probabilistic neural network are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer fault diagnosis, and in particular to a distribution transformer fault diagnosis method based on full-band multi-modal data fusion. Background Art

[0002] In the power system, the stable operation of distribution transformers is crucial, and the accuracy of fault diagnosis directly affects the reliability of power supply. Traditional fault diagnosis methods often rely on a single data source, which limits the comprehensiveness and accuracy of diagnosis.

[0003] The use of probabilistic neural networks (PNNs) for distribution transformer fault diagnosis can identify and classify fault types based on existing labeled data. However, this approach may face challenges when dealing with unknown or rare fault modes because they usually require large amounts of data or more advanced models to achieve good generalization. In addition, complex neural networks have high requirements for computing resources during training and evaluation, which may limit their applicability in resource-constrained environments. At the same time, for real-time application scenarios that require rapid response, the inference time of these models may be too long to meet the needs of immediacy. Therefore, studying how to optimize the PNN model to improve its generalization ability, reduce resource consumption, and speed up processing speed is important for improving the efficiency and accuracy of transformer fault diagnosis under full-band multimodal data. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a distribution transformer fault diagnosis method based on full-band multimodal data fusion, which aims to solve the problems in the background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a distribution transformer fault diagnosis method based on full-band multimodal data fusion, comprising the following steps: Step S1: collecting different characteristic graphs of the transformer, including a voiceprint spectrum graph, an ultra-high frequency partial discharge spectrum, a high frequency partial discharge spectrum, and an ultrasonic spectrum; Step S2: using a convolutional neural network to extract features from different collected feature maps to obtain a one-dimensional feature vector, performing weighted summation on the one-dimensional feature vector to obtain a multi-source fusion feature vector; Step S3: performing data preprocessing on the multi-source fusion feature vector, and then performing normalization processing after the data preprocessing; Step S4: constructing an improved probabilistic neural network PNN; initializing the polynomial order of the improved probabilistic neural network; inputting the normalized multi-source fusion feature vector into the improved probabilistic neural network after the polynomial order is initialized for training, and obtaining a trained improved probabilistic neural network; Step S5: using the improved Arctic Puffin optimization algorithm to optimize the polynomial order of the trained improved probabilistic neural network; using the optimized improved probabilistic neural network to perform transformer fault diagnosis; The specific process of optimizing the polynomial order of the improved probabilistic neural network using the improved Arctic Puffin optimization algorithm is as follows: Step S5.11: setting the parameters of the improved Arctic puffin optimization algorithm, including the number of puffin individuals, search space, total number of iterations, and introducing Logistic chaos mapping to initialize the initial population position; Step S5.12: Initialize the current iteration number t=1 until t>T is satisfied, where T represents the total number of iterations. Perform fitness evaluation on each puffin individual, calculate the puffin individual with the highest fitness, and record its position and fitness value. Step S5.13: Dynamically adjust the behavior coefficient according to the current number of iterations and the total number of iterations ; When the behavior coefficient When it is greater than the set threshold C, it is the global search stage; when the behavior coefficient is less than or equal to the set threshold C, it is the local search stage; Step S5.14: When in the global search phase, the puffin individuals update their positions by simulating aerial search strategies, diving predation strategies, and population merging strategies; Step S5.15: When in the local search stage, the puffin individuals update their own positions by simulating collective foraging strategies, enhanced search strategies, strategies for escaping natural enemies, and population integration strategies; Step S5.16: Determine whether the current number of iterations is the maximum number of iterations. If yes, output the position of the puffin individual with the highest fitness as the optimal polynomial order of the improved probabilistic neural network. If no, return to step S5.12 and iterate again.

[0006] Further, improve the kernel function of probabilistic neural network It is expressed as: ; In the formula, represents the improved polynomial kernel function; represents the Gaussian kernel function; represents the weight parameter; Indicates adjustable parameters; , It represents the vector calculated by decomposing the multi-source fusion feature vector, which is used to measure the correlation between the diameters of different feature vectors in the multi-source fusion feature vector; Improved polynomial kernel function It is expressed as: ; In the formula, represents the scaling parameter; represents the bias parameter; Indicates the order; express The transpose of Gaussian kernel function It is expressed as: ; In the formula, Indicates the bandwidth parameter.

[0007] Furthermore, the parameters of the improved Arctic puffin optimization algorithm are set, including the number of puffin individuals, search space, and total number of iterations, and the Logistic chaotic map is introduced to initialize the initial population position, which is expressed as: ; In the formula, Indicates The chaotic value generated by the iteration; represents the control parameter, and its value is [3.57,4]; Indicates The chaotic value generated by the iteration; Indicates After the iteration The location of individual puffins; and Respectively represent the upper and lower bounds of the search space; represents the total number of puffin individuals; Behavior coefficient It is expressed as: ; In the formula, Represents a random number in (0,1).

[0008] Furthermore, the aerial search strategy is simulated and improved by dynamically adjusting the Levy flight parameters, which is expressed as: ; In the formula, represents a puffin individual randomly selected from the puffin population, is a random integer between 1 and N-1; Indicates the dynamic adjustment of the step size of Levy flight; Represents the dimension; Represents random numbers that conform to the standard normal distribution; Indicates that after the air search strategy is updated The location of individual puffins; represents the random perturbation adjusted according to the current number of iterations; It is expressed as: ; In the formula, It is to dynamically adjust the maximum value of Levy flight. It is the minimum value for dynamically adjusting Levy flight.

[0009] Furthermore, the diving predation strategy is simulated and a dynamic speed adjustment factor is introduced for improvement, which is expressed as: ; In the formula, Indicates dynamic adjustment speed factor; It means that after the diving predation strategy is updated, The location of individual puffins.

[0010] Furthermore, the population merging strategy is simulated, which is expressed as: ; In the formula, It means sorting the fitness values ​​of the puffins from small to large. Indicates The position of the i-th puffin individual after iterations; represents the set of individual puffin positions after being updated with the aerial search strategy and the diving predation strategy; According to The fitness of individual puffin The sorting operation is performed on the individuals of the Chinese puffin.

[0011] Furthermore, the collective foraging strategy is simulated as follows: ; In the formula, represents the collaboration parameters, Indicates the maximum value of the collaboration parameter; , and All represent puffin individuals randomly selected from the current puffin population; , and is a random integer between 1 and N-1, excluding i; Represents the position of the i-th puffin individual after the collective foraging strategy is updated.

[0012] Furthermore, the simulation strengthens the search strategy, expressed as: ; In the formula, represents the adaptive factor; Represents the position of the i-th puffin individual after the enhanced search strategy update.

[0013] Furthermore, the strategy of escaping from natural enemies is simulated and the distance-based predator perception model is introduced for improvement, which is expressed as: ; ; In the formula, represents the predator position and The distance between the positions of individual puffins; Indicates The location of individual puffins; Indicates the location of the predator; Indicates the set threshold; is a randomly generated number between 0 and 1; represents the position of the i-th puffin individual after the strategy of escaping from natural enemies is updated; when Less than , the puffin individuals take evasive action, when Greater than or equal to , puffin individuals actively avoid predators.

[0014] Furthermore, the population fusion strategy is simulated, which is expressed as: ; In the formula, It represents the individual position set of puffins after the collective foraging strategy, enhanced search strategy and escape from natural enemies strategy are updated; According to The fitness of individual puffin The sorting operation is performed on the individuals of the Chinese puffin.

[0015] Compared with the existing technology, the present invention has the following beneficial effects:

[0016] (1) The present invention adopts an improved Arctic puffin optimization algorithm, simulates the foraging behavior of the Arctic puffin, effectively combines global search and local search, optimizes the polynomial order of the improved probabilistic neural network PNN, and improves the accuracy of transformer data fusion diagnosis and the generalization ability of the probabilistic neural network. The present invention can quickly process and analyze a large amount of data, meet the needs of real-time monitoring and rapid response, and thus provide a solid guarantee for the stable operation of the power system.

[0017] (2) The present invention can comprehensively capture the behavioral characteristics of the transformer in the entire frequency band and extract relevant information from multi-dimensional data such as temperature, current, sound and vibration. Through comprehensive analysis, the accuracy of fault detection is significantly improved, the risk of false alarms and missed alarms caused by a single data source is reduced, and the diagnostic efficiency of complex fault conditions is enhanced. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0019] like Figure 1 As shown, the present invention provides a technical solution: a distribution transformer fault diagnosis method based on full-band multimodal data fusion, comprising the following steps:

[0020] Step S1: collecting different characteristic graphs of the transformer, including a voiceprint spectrum graph, an ultra-high frequency partial discharge spectrum, a high frequency partial discharge spectrum, and an ultrasonic spectrum.

[0021] Step S2: Use a convolutional neural network (CNN) to extract features from different collected feature maps to obtain a one-dimensional feature vector, perform weighted summation on the one-dimensional feature vector, and obtain a multi-source fusion feature vector.

[0022] Step S3: Perform data preprocessing on the multi-source fusion feature vector, and then perform normalization processing after the data preprocessing.

[0023] Step S4: construct an improved probabilistic neural network (PNN); initialize the polynomial order of the improved probabilistic neural network; input the normalized multi-source fusion feature vector into the improved probabilistic neural network after the polynomial order is initialized for training, and obtain a trained improved probabilistic neural network.

[0024] Among them, the kernel function of the improved probabilistic neural network (PNN) It is expressed as: (1); In the formula, represents the improved polynomial kernel function; represents the Gaussian kernel function; represents the weight parameter; Indicates adjustable parameters; , Represents the vector calculated from the decomposition of the multi-source fusion feature vector, which is used to measure the correlation between the diameters of different feature vectors in the multi-source fusion feature vector.

[0025] Improved polynomial kernel function It is expressed as: (2); In the formula, Represents the scaling parameter, which is used to adjust the scaling of the kernel function on the inner product of the input vector; represents the bias parameter; Indicates the order; express The transpose of .

[0026] Gaussian kernel function It is expressed as: (3); In the formula, Indicates the bandwidth parameter.

[0027] Step S5: using the improved Arctic Puffin optimization algorithm to optimize the polynomial order of the trained improved probabilistic neural network; using the optimized improved probabilistic neural network (PNN) to perform transformer fault diagnosis.

[0028] Among them, the specific process of optimizing the polynomial order of the improved probabilistic neural network (PNN) using the improved Arctic Puffin optimization algorithm is as follows: Step S5.11: Set the parameters of the improved Arctic puffin optimization algorithm, including the number of puffin individuals, search space, and total number of iterations, and introduce the Logistic chaotic map to initialize the initial population position, which can make the initial population more evenly distributed in the search space and improve the search efficiency and accuracy of the algorithm, expressed as: (4); In the formula, Indicates The chaotic value generated by the iteration; represents the control parameter, and its value is [3.57,4]; Indicates The chaotic value generated by the iteration; Indicates After the iteration The location of individual puffins; and Respectively represent the upper and lower bounds of the search space; Represents the total number of puffin individuals.

[0029] Step S5.12: Initialize the current iteration number t=1 until t>T is satisfied, where T represents the total number of iterations. Perform fitness evaluation on each puffin individual, calculate the puffin individual with the highest fitness, and record its position and fitness value.

[0030] Step S5.13: Dynamically adjust the behavior coefficient according to the current number of iterations and the total number of iterations ; Behavior coefficient It is expressed as: (5); In the formula, Represents a random number in (0,1).

[0031] When the behavior coefficient When it is greater than the set threshold C, it means that the current stage is the global search stage. The puffin individuals will conduct more extensive exploration to find new solution space areas and update their own positions. The puffin individuals with the highest fitness will be selected as the new puffin population.

[0032] When the behavior coefficient is less than or equal to the set threshold C, it means that the current stage is the local search stage, and the puffin individuals will concentrate on conducting a more detailed search near the currently found optimal solution.

[0033] Step S5.14: When in the global search phase, the puffin individuals update their own positions by simulating aerial search strategies, diving predation strategies, and population merging strategies.

[0034] Step S5.141: simulate the aerial search strategy, and introduce the dynamic adjustment of Levy flight parameters to improve the aerial search strategy; puffins often fly in groups to improve flight efficiency and facilitate cooperative hunting. They fly at low altitudes to catch underwater food. When hunting, they not only look for prey, but also alert predators. When conditions are suitable, they will move quickly to the water surface to catch prey, which is expressed as: (6); In the formula, represents a puffin individual randomly selected from the puffin population, is a random integer between 1 and N-1; Indicates the dynamic adjustment of the step size of Levy flight; Represents the dimension; Represents random numbers that conform to the standard normal distribution; Indicates that after the air search strategy is updated The location of individual puffins; represents a random perturbation adjusted according to the current iteration number.

[0035] It is expressed as: (7); In the formula, It is the maximum value of Levy flight that is dynamically adjusted at the beginning of the iteration. It is the minimum value for dynamically adjusting Levy flight in the later stages of iteration.

[0036] Step S5.142: simulate the diving predation strategy and introduce a dynamic adjustment speed factor for improvement; puffins adopt a fast diving strategy when hunting, which enables them to quickly adjust the flight path to improve the efficiency of hunting. When competing for food with other competitors, this fast diving ability is crucial because they need to ensure that they can quickly and effectively capture prey to survive, which is expressed as: (8); In the formula, Indicates dynamic adjustment speed factor; It means that after the diving predation strategy is updated, The location of individual puffins.

[0037] Step S5.143: Simulate population merging strategy; in order to adapt to different environmental conditions and seek the best solution, the puffins generated by the aerial search strategy and the simulated diving predation strategy are aggregated. Then, the fitness of each puffin individual is sorted, and the top N puffins with the highest fitness are selected to form a new generation of puffins, which is expressed as: (9); In the formula, It means sorting the fitness values ​​of the puffins from small to large. Indicates The position of the i-th puffin individual after iterations; represents the set of individual puffin positions after being updated with the aerial search strategy and the diving predation strategy; According to The fitness of individual puffin The sorting operation is performed on the individuals of the Chinese puffin.

[0038] Step S5.15: When in the local search stage, puffins update their own positions by simulating collective foraging strategies, enhanced search strategies, strategies to escape from natural enemies, and population integration strategies.

[0039] Step S5.151: Simulate collective foraging strategy; puffins tend to gather around schools of fish near the water surface, and improve the efficiency and success rate of predation through collective action. They will also stay on the sea surface and observe the behavior of their companions to determine the location of food resources, which is expressed as: (10); In the formula, represents the collaboration parameters, Indicates the maximum value of the collaboration parameter; , and All represent puffin individuals randomly selected from the current puffin population; , and is a random integer between 1 and N-1 (excluding i); Represents the position of the i-th puffin individual after the collective foraging strategy is updated.

[0040] Step S5.152: Simulate the enhanced search strategy; after collective foraging, puffins may perceive that the food supply in the current foraging area is decreasing. In order to continue to obtain food and meet nutritional needs, they need to change their position or direction and conduct more active searches to find more food resources, which is expressed as: (11); In the formula, represents the adaptive factor, which is used to adjust the position of individual puffin in the water; Represents the position of the i-th puffin individual after the enhanced search strategy update.

[0041] Step S5.153: Simulate the strategy of escaping from natural enemies and introduce a distance-based predator perception model for improvement; when puffins sense that there are predators nearby, they warn their companions by making specific calls to alert them to potential threats. When puffins perceive that predators are approaching, they will quickly change their positions and take a circuitous route to swim to a safe area to avoid being preyed on, which is expressed as: (12); (13); In the formula, represents the predator position and The distance between the positions of individual puffins; Indicates The location of individual puffins; Indicates the location of the predator; Indicates the set threshold; is a randomly generated number between 0 and 1; Represents the position of the i-th puffin individual after the strategy of escaping from natural enemies is updated.

[0042] when Less than , indicating that the predator is approaching, the puffin will quickly take evasive action and make a large change of position. Greater than or equal to , puffin individuals actively avoid predators.

[0043] Step S5.154: Simulate the population fusion strategy; generate puffin individuals by integrating the collective foraging strategy, enhanced search strategy and escaping natural enemy strategy, and sort these puffin individuals according to fitness, and select the N puffin individuals with the highest fitness, which can be expressed as: (14); In the formula, It represents the individual position set of puffins after the collective foraging strategy, enhanced search strategy and escape from natural enemies strategy are updated; According to The fitness of individual puffin The sorting operation is performed on the individuals of the Chinese puffin.

[0044] Step S5.16: Determine whether the current number of iterations is the maximum number of iterations. If yes, output the position of the puffin individual with the highest fitness as the optimal polynomial order of the improved probabilistic neural network (PNN). If no, return to step S5.12 and iterate again.

[0045] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A distribution transformer fault diagnosis method based on full-band multimodal data fusion, characterized in that: The steps include: Step S1: collecting different characteristic graphs of the transformer, including a voiceprint spectrum graph, an ultra-high frequency partial discharge spectrum, a high frequency partial discharge spectrum, and an ultrasonic spectrum; Step S2: using a convolutional neural network to extract features from different collected feature maps to obtain a one-dimensional feature vector, performing weighted summation on the one-dimensional feature vector to obtain a multi-source fusion feature vector; Step S3: performing data preprocessing on the multi-source fusion feature vector, and then performing normalization processing after the data preprocessing; Step S4: construct an improved probabilistic neural network PNN; Initializing the polynomial order of the improved probabilistic neural network; inputting the normalized multi-source fusion feature vector into the improved probabilistic neural network after the polynomial order is initialized for training, and obtaining a trained improved probabilistic neural network; Step S5: using the improved Arctic Puffin optimization algorithm to optimize the polynomial order of the trained improved probabilistic neural network; using the optimized improved probabilistic neural network to perform transformer fault diagnosis; The specific process of optimizing the polynomial order of the improved probabilistic neural network using the improved Arctic Puffin optimization algorithm is as follows: Step S5.11: setting the parameters of the improved Arctic puffin optimization algorithm, including the number of puffin individuals, search space, total number of iterations, and introducing Logistic chaos mapping to initialize the initial population position; Step S5.12: Initialize the current iteration number t=1 until t>T is satisfied, where T represents the total number of iterations. Perform fitness evaluation on each puffin individual, calculate the puffin individual with the highest fitness, and record its position and fitness value. Step S5.13: Dynamically adjust the behavior coefficient according to the current number of iterations and the total number of iterations ; When the behavior coefficient When it is greater than the set threshold C, it is the global search stage; when the behavior coefficient is less than or equal to the set threshold C, it is the local search stage; Step S5.14: When in the global search phase, the puffin individuals update their positions by simulating aerial search strategies, diving predation strategies, and population merging strategies; Step S5.15: When in the local search stage, the puffin individuals update their own positions by simulating collective foraging strategies, enhanced search strategies, strategies for escaping natural enemies, and population integration strategies; Step S5.16: Determine whether the current number of iterations is the maximum number of iterations. If yes, output the position of the puffin individual with the highest fitness as the optimal polynomial order of the improved probabilistic neural network. If no, return to step S5.12 and iterate again.

2. The distribution transformer fault diagnosis method based on full-band multimodal data fusion according to claim 1 is characterized in that: Improving the kernel function of probabilistic neural networks It is expressed as: ； In the formula, represents the improved polynomial kernel function; represents the Gaussian kernel function; represents the weight parameter; Indicates adjustable parameters; , It represents the vector calculated by decomposing the multi-source fusion feature vector, which is used to measure the correlation between the diameters of different feature vectors in the multi-source fusion feature vector; Improved polynomial kernel function It is expressed as: ； In the formula, represents the scaling parameter; represents the bias parameter; Indicates the order; express The transpose of Gaussian kernel function It is expressed as: ； In the formula, Indicates the bandwidth parameter.

3. The distribution transformer fault diagnosis method based on full-band multimodal data fusion according to claim 2 is characterized in that: The parameters of the improved Arctic puffin optimization algorithm are set, including the number of puffin individuals, search space, and total number of iterations. Logistic chaotic mapping is introduced to initialize the initial population position, which is expressed as: ； In the formula, Indicates The chaotic value generated by the iteration; Indicates control parameters; Indicates The chaotic value generated by the iteration; Indicates After the iteration The location of individual puffins; and Respectively represent the upper and lower bounds of the search space; represents the total number of puffin individuals; Behavior coefficient It is expressed as: ； In the formula, Represents a random number in (0,1).

4. The distribution transformer fault diagnosis method based on full-band multimodal data fusion according to claim 3 is characterized in that: The air search strategy is simulated and improved by dynamically adjusting the Levy flight parameters, which is expressed as: ； In the formula, represents a puffin individual randomly selected from the puffin population, is a random integer between 1 and N-1; Indicates the dynamic adjustment of the step size of Levy flight; Represents the dimension; Represents random numbers that conform to the standard normal distribution; Indicates that after the air search strategy is updated The location of individual puffins; represents the random perturbation adjusted according to the current number of iterations; It is expressed as: ； In the formula, It is to dynamically adjust the maximum value of Levy flight. It is the minimum value for dynamically adjusting Levy flight.

5. The distribution transformer fault diagnosis method based on full-band multimodal data fusion according to claim 4 is characterized in that: The dive predation strategy is simulated and improved by introducing a dynamic adjustment speed factor, which is expressed as: ； In the formula, Indicates dynamic adjustment speed factor; It means that after the diving predation strategy is updated, The location of individual puffins.

6. The distribution transformer fault diagnosis method based on full-band multimodal data fusion according to claim 5 is characterized in that: The simulated population merging strategy is expressed as: ； In the formula, It means sorting the fitness values ​​of the puffins from small to large. Indicates The position of the i-th puffin individual after iterations; represents the set of individual puffin positions after being updated with the aerial search strategy and the diving predation strategy; Indicates based on The fitness of individual puffins The sorting operation is performed on the individuals of the Chinese puffin.

7. The distribution transformer fault diagnosis method based on full-band multimodal data fusion according to claim 6 is characterized in that: Simulate the collective foraging strategy, expressed as: ； In the formula, represents the collaboration parameters, Indicates the maximum value of the collaboration parameter; , and All represent puffin individuals randomly selected from the current puffin population; , and is a random integer between 1 and N-1, excluding i; Represents the position of the i-th puffin individual after the collective foraging strategy is updated.

8. The distribution transformer fault diagnosis method based on full-band multimodal data fusion according to claim 7 is characterized in that: Simulate the enhanced search strategy, expressed as: ； In the formula, represents the adaptive factor; Represents the position of the i-th puffin individual after the enhanced search strategy update.

9. The distribution transformer fault diagnosis method based on full-band multimodal data fusion according to claim 8 is characterized in that: The strategy of escaping from natural enemies is simulated and the distance-based predator perception model is introduced for improvement, which is expressed as: ； ； In the formula, represents the predator position and The distance between the positions of individual puffins; Indicates The location of individual puffins; Indicates the location of the predator; Indicates the set threshold; is a randomly generated number between 0 and 1; represents the position of the i-th puffin individual after the strategy of escaping from natural enemies is updated; when Less than , the puffin individuals take evasive action, when Greater than or equal to , puffin individuals actively avoid predators.

10. The distribution transformer fault diagnosis method based on full-band multimodal data fusion according to claim 9 is characterized in that: The simulated population fusion strategy is expressed as: ； In the formula, It represents the individual position set of puffins after the collective foraging strategy, enhanced search strategy and escape from natural enemies strategy are updated; Indicates based on The fitness of individual puffins The sorting operation is performed on the individuals of the Chinese puffin.

Images