A Transformer Partial Discharge Pattern Recognition Method and System Based on Optimized Probabilistic Neural Network

Abstract

This invention discloses a method and system for transformer partial discharge pattern recognition based on an optimized probabilistic neural network. The method first acquires the transformer partial discharge signal; then, it establishes a two-dimensional spectrum of the partial discharge phase distribution pattern based on the partial discharge signal and extracts discharge statistical features from this spectrum; finally, it inputs the discharge statistical features into the optimized probabilistic neural network for pattern recognition to obtain the partial discharge pattern. The optimized probabilistic neural network uses a pollination algorithm to optimize the smoothing factor, and the switching probability in the pollination algorithm is a nonlinear function that decreases with the number of iterations. This invention can more accurately and efficiently classify and identify different partial discharge patterns of transformers, providing a data foundation for transformer fault diagnosis and resolution.

Description

Technical Field

[0001] This invention relates to the field of transformer partial discharge pattern recognition technology, and in particular to a transformer partial discharge pattern recognition method based on an optimized probabilistic neural network. Background Technology

[0002] Power transformers are among the most critical electrical devices in the power grid. Transformer failures can cause widespread power outages, impacting production and daily life, and resulting in significant economic losses. Statistics show that over 80% of transformer failures in actual operation are caused by insulation problems, with partial discharge being a major cause of transformer insulation degradation. Different modes of partial discharge exhibit varying characteristics and cause different degrees of damage to the internal insulation materials of the transformer. Therefore, quickly and accurately identifying the specific mode of partial discharge in a transformer is crucial for accurately assessing the internal insulation condition and promptly eliminating transformer faults.

[0003] Currently, one of the technical challenges in transformer partial discharge pattern recognition is that, due to the complexity of the internal structure of power transformers, the collected transformer discharge signals exhibit problems such as high noise, multimodal aliasing, high dimensionality of feature parameters, and high randomness. Existing partial discharge pattern recognition methods suffer from problems such as high computational cost of feature extraction and low recognition efficiency of classification models, which seriously reduces the efficiency and accuracy of transformer partial discharge pattern recognition. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a transformer partial discharge pattern recognition method and system based on an optimized probabilistic neural network with high accuracy and high computational efficiency.

[0005] Technical solution: The present invention provides a transformer partial discharge pattern recognition method based on an optimized probabilistic neural network, comprising the following steps:

[0006] Acquire the partial discharge signal of the transformer;

[0007] A two-dimensional spectrum of the partial discharge phase distribution mode is established based on the partial discharge signal, and discharge statistical features are extracted from the spectrum.

[0008] The discharge statistical features are input into an optimized probabilistic neural network for pattern recognition to obtain partial discharge patterns.

[0009] The optimized probabilistic neural network uses a flower pollination algorithm to optimize the smoothing factor of the probabilistic neural network, and the switching probability in the flower pollination algorithm is a nonlinear function that decreases with the number of iterations.

[0010] Furthermore, after acquiring the transformer partial discharge signal, the method further includes: performing noise reduction processing on the partial discharge signal using a complete set empirical mode decomposition method.

[0011] Furthermore, after extracting the discharge statistical features from the spectrum, the method further includes: using principal component analysis to perform dimensionality reduction processing on the discharge statistical features.

[0012] Furthermore, the switching probability in the flower pollination algorithm adopts a nonlinear function that decreases with the number of iterations, including:

[0013] nonlinear functions ,in and These are the upper and lower bounds of the switching probability, respectively. The maximum number of iterations, The nonlinear adjustment coefficient is... This represents the number of iterations.

[0014] Each iteration generates a random value rand, and then compares rand with... The size allows for switching between two pollination modes.

[0015] Furthermore, in the flower pollination algorithm, the globally optimal pollen location during the iteration process is subjected to Cauchy mutation.

[0016] Furthermore, the smoothing factor is a vector related to the dimension n of the discharge statistical feature. .

[0017] Furthermore, the two-dimensional spectrum includes a spectrum describing the relationship between discharge quantity and phase, a spectrum describing the relationship between maximum discharge quantity and phase, a spectrum describing the relationship between average discharge quantity and phase, and a spectrum describing the relationship between discharge phase and discharge number.

[0018] The present invention discloses a transformer partial discharge pattern recognition system based on an optimized probabilistic neural network, comprising:

[0019] The discharge signal acquisition unit is used to acquire the partial discharge signal of the transformer;

[0020] The discharge statistical feature extraction unit is used to establish a two-dimensional spectrum of the partial discharge phase distribution mode based on the partial discharge signal, and extract discharge statistical features from the spectrum.

[0021] A discharge pattern recognition unit is used to input the discharge statistical features into an optimized probabilistic neural network for pattern recognition to obtain partial discharge patterns.

[0022] The optimized probabilistic neural network uses a flower pollination algorithm to optimize the smoothing factor of the probabilistic neural network, and the switching probability in the flower pollination algorithm is a nonlinear function that decreases with the number of iterations.

[0023] The computer-readable storage medium of the present invention stores a computer program, which, when executed by a processor, implements the transformer partial discharge pattern recognition method based on an optimized probabilistic neural network.

[0024] The computer program product of the present invention includes a computer program that, when executed by a processor, implements the transformer partial discharge pattern recognition method based on an optimized probabilistic neural network.

[0025] Beneficial effects: Compared with the prior art, the advantages of the present invention are: (1) The present invention uses the flower pollination algorithm to optimize the smoothing factor of the PNN network and adaptively searches for the optimal smoothing factor, which can significantly improve the generalization ability and classification accuracy of the probabilistic neural network. (2) The present invention proposes an "adaptive switching probability mechanism": First, the original fixed switching probability p is modified into a function that dynamically changes with the number of iterations t. The algorithm achieves an intelligent balance between global exploration (cross-pollination) and local development (self-pollination) to expand the search range and improve the search accuracy of the smoothing factor vector. The algorithm can intelligently balance the global exploration and local development capabilities according to the iteration process. Compared with the fixed p in the original scheme, the adaptive strategy enables the algorithm to quickly locate the potential optimal region in the early stage of smoothing factor vector optimization and finely adjust the vector parameters in the later stage, thereby significantly improving the convergence speed while ensuring global convergence. (3) In response to the "premature" convergence problem that occurs in the smoothing factor optimization process, the Cauchy mutation operator is introduced in the population update process. When the population fitness variance has not changed significantly for several generations, the current optimal smoothing factor vector is adjusted. Perform Cauchy mutation. It can generate a large jump distance, helping the algorithm escape local constraints and find a better combination of smoothing factor vectors. Experimental verification shows that this hybrid strategy has a better final convergence error than the traditional FPA algorithm when dealing with the parameter optimization problem of PNN network after dimensionality reduction of high-dimensional features. (4) The present invention has a smoothing parameter The extraction process underwent secondary optimization, and a "smoothing factor adaptive mechanism" was proposed. The core improvement lies in using a single smoothing factor... Expanding to a smoothing factor vector related to the feature dimension n allows features of different dimensions to have independent weight adjustment mechanisms in classification decisions. This improves computational efficiency while suppressing noise in the original data. It can adaptively adjust to the sparsity of different feature dimensions. For feature dimensions with small inter-class distances, the algorithm automatically selects the smaller one. Sharpen the classification boundaries; for feature dimensions with high noise, select a larger one. This approach smooths out interference. Compared to a "compromise" strategy using a single smoothing factor, it significantly improves the pattern resolution of probabilistic neural networks. Attached Figure Description

[0026] Figure 1 This is a flowchart of the transformer partial discharge mode recognition method according to an embodiment of the present invention.

[0027] Figure 2 This is a diagram of the PNN architecture according to an embodiment of the present invention.

[0028] Figure 3 The flowchart shows the optimization process of the smoothing factor of the PNN network in the flower pollination algorithm of this invention.

[0029] Figure 4 This is a comparison chart of the smoothing factor optimization results in the embodiments of the present invention. Detailed Implementation

[0030] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0031] like Figure 1 As shown, the transformer partial discharge pattern recognition method based on optimized probabilistic neural network includes the following steps.

[0032] Step 1: Use the pulse current method to collect the discharge signal of partial discharge of the transformer.

[0033] Step 2: Use the Complete Empirical Mode Decomposition (CEEMD) method to denoise the signal and remove white noise interference caused by thermal noise from the transformer windings.

[0034] Complete Ensemble Empirical Mode Decomposition (CEEMD) is a method that thoroughly decomposes modes by adding different Gaussian white noise to the remaining components at each scale and then averaging them, thereby reducing aliasing and spurious components.

[0035] The specific process is as follows:

[0036] Step 2.1, assuming the noisy signal is Gaussian white noise was added to it multiple times. To obtain a noisy signal

[0037]

[0038] in The coefficients are used to adjust the signal-to-noise ratio of the original signal.

[0039] Step 2.2, for noisy signals conduct The EMD decomposition is performed, the overall mean is calculated and defined as follows: intrinsic mode function , can be represented as:

[0040]

[0041] Step 2.3, calculate the remaining stochasticity. :

[0042]

[0043] Step 2.4, Decomposition , Its first intrinsic mode component is taken as the intrinsic mode function of CEEMD. , can be represented as:

[0044]

[0045] Step 2.5, and so on. Remaining residual :

[0046]

[0047] Step 2.6, Decomposition , Its first intrinsic mode component is taken as the intrinsic mode function of CEEMD. , can be represented as:

[0048]

[0049] Step 2.7: If there is at most one extreme point in the remaining residual, the decomposition ends; otherwise, return to step 5 for iteration until the requirements are met, and finally obtain the remaining residual. Its expression is:

[0050]

[0051] in: This represents the total number of decomposed modes. In summary:

[0052]

[0053] Step 3: Construct a two-dimensional spectrum of the denoised data under the partial discharge phase distribution mode (PRPD) and extract eigenvalues. The specific process is as follows:

[0054] Four two-dimensional spectra were constructed under PRPD to describe the discharge quantity. With phase Spectrum of the relationship between them Maximum discharge amount With phase Spectrum of the relationship between them Average discharge quantity With phase Spectrum of the relationship between them and discharge phase With the number of discharges Spectrum of the relationship between them Extract the following feature parameters: skewness Steepness Number of peaks Cross-relationships Discharge factor Phase asymmetry Corrected cross-correlation factor These statistical characteristic parameters can accurately characterize the distribution characteristics of partial discharge signals from different dimensions. The extracted feature values ​​are normalized using the following formula:

[0055]

[0056] in, Represents the maximum value in a given dataset. This represents the minimum value in a given dataset.

[0057] Step 4: Perform dimensionality reduction on the extracted eigenvalues ​​using Principal Component Analysis (PCA). The specific operation method is as follows:

[0058] Step 4.1: Select initial analysis variables based on the research question: Select initial variables with strong correlation for analysis, and use Bartlett's test of sphericity for variable selection.

[0059] Step 4.2, calculate the covariance matrix. Assumption yes The original feature space matrix, where It is the number of samples in the data. This refers to the number of characteristic parameters. The formula for calculating the covariance matrix is ​​as follows:

[0060]

[0061] in, yes The average value, The resulting covariance matrix is A square formation.

[0062] Step 4.3, calculate the eigenvalues ​​and eigenvectors of the covariance matrix: First, calculate the covariance matrix. eigenvalues With the corresponding feature vector Then, the eigenvalues ​​are sorted by size, and the columns of the eigenvector matrix are rearranged in this order to obtain a new transformation matrix. Finally, the original feature space matrix is... Right multiplication Then you will get A new feature parameter Matrix operations are represented as follows:

[0063]

[0064] Step 4.4, determine the number of principal components. Calculate separately The contribution rate of a certain column of elements, i.e., the single contribution rate. and Center front The total contribution rate of the new elements, i.e., the cumulative contribution rate. The calculation formulas for both are as follows:

[0065]

[0066]

[0067] The selection of principal components is determined based on the two contribution rates mentioned above. Generally, when the cumulative contribution rate reaches 85% or higher, it is considered that the principal components are considered to be principal components. One element can replace the original. These are the characteristic parameters, also known as principal components.

[0068] Step 4.5, Extraction of principal components: Extraction The former Column elements.

[0069] Step 5: Input the dimensionality-reduced feature parameters into a probabilistic neural network (PNN) optimized with a smoothing factor based on a genetic algorithm for pattern recognition.

[0070] Probabilistic Neural Networks (PNNs) are a type of feedforward neural network derived from radial basis function networks. Their theoretical basis is the Bayesian minimum risk criterion (Bayesian decision theory), such as... Figure 2As shown, the PNN neural network consists of an input layer, hidden layers, a summing layer, and an output layer. The first layer is the input layer, which receives values ​​from the training samples and passes the data to the hidden layers. The number of neurons in this layer is equal to the length of the input vector. The second hidden layer is the radial layer, where each neuron has a center. This layer receives the sample input from the input layer, calculates the distance between the input vector and the center, and returns the last scalar value. The number of neurons in this layer is the same as the number of training samples. The third summing layer performs a weighted average of the outputs of hidden neurons belonging to the same class in the hidden layers. The number of neurons in the summing layer is the same as the total number of classes in the training samples. The fourth output layer consists of competing neurons, with the same number of neurons as the summing layer. It receives the output of the summing layer and performs simple thresholding.

[0071] Traditional PNNs apply a uniform smoothing parameter to the input feature dimension to obtain the probability density, i.e. Since the input feature vectors have different physical meanings and distribution characteristics in each dimension, forcibly applying the same smoothing factor to feature dimensions with different degrees of sparsity will lead to blurred classification boundaries and limit further improvement in recognition accuracy. Although the above formula optimizes the smoothing factor, it is essentially still an optimization of a single scalar and fails to solve the problem of heterogeneity between feature dimensions.

[0072] This invention addresses the issue of smoothing parameters. The extraction process underwent secondary optimization, proposing an adaptive smoothing factor mechanism. Furthermore, the Flower Pollination Algorithm (FPA) was employed to optimize the smoothing factor of the PNN network, adaptively searching for the optimal smoothing factor, referencing... Figure 3 The specific steps are as follows.

[0073] Step 5.1: PNN uses Parzen window probability density estimation, with a Gaussian kernel function, to obtain the probability density function. :

[0074]

[0075] in, Given a training sample set, The number of training sample categories, For category The number of samples, The feature dimension after PCA dimensionality reduction. The smoothing factor corresponding to the k-th feature dimension. For the first Class 1 Training samples.

[0076] Step 5.2, perform normalization on the sample values, the calculation is as follows:

[0077]

[0078] in, Let be the dimension of each feature vector sample. For the first Class 1 In the training samples, the th Each feature value.

[0079] Step 5.3: Input the normalized samples into the hidden layer as neurons in that layer. The selected activation function is the radial basis function, and each training sample corresponds to one neuron in the hidden layer. Finally, calculate the Euclidean distance between the sample to be identified and the training samples.

[0080]

[0081] in, For the normalized first The first sample to be identified 1 eigenvalue, For the training sample in step 5.2 Class 1 Each feature value.

[0082] Step 5.4: Substitute the Euclidean distance into the hidden layer activation function to obtain the initial probability matrix. The hidden layer activation function is as follows:

[0083]

[0084] Step 5.5 involves summing and averaging the initial probabilities of the hidden layer outputs belonging to the same type, as shown in the following formula:

[0085]

[0086] Step 5.6: Obtain the expected category of the sample and calculate the first... The sample to be identified is Probability of class sample And find the maximum value among them, that is, the category that belongs to the maximum value.

[0087]

[0088] Furthermore, this invention employs a flower pollination algorithm to optimize the smoothing factor of the PNN network. However, standard FPA has two drawbacks: First, it uses a fixed switching probability p to control the switching between the two pollination modes. This fixed switching strategy cannot dynamically adjust the search range according to the actual progress of the optimization process. In the early stages of iteration, the population diversity is high, requiring a larger global search probability to avoid getting trapped in local optima; while in the later stages of iteration, the solution space tends to converge, requiring a stronger local search to improve accuracy. A fixed parameter p makes it difficult to balance convergence speed and optimization accuracy, potentially leading to slow algorithm convergence or getting trapped in local optima. Second, it is prone to getting trapped in local optima during the optimization process. This invention proposes an adaptive switching probability mechanism, modifying the original fixed switching probability p into a function that dynamically changes with the number of iterations t. Furthermore, the Cauchy mutation operator is introduced during the population update process, as detailed below.

[0089] Step 5.7: Randomly generate a set of smoothing factors for the PNN network, which will serve as the initial population for the flower pollination algorithm. The PNN uses Parzen window probability density estimation, where the kernel function is a Gaussian function. The probability density formula is as follows:

[0090]

[0091] in For corresponding The set of smoothing factors for each category, arbitrarily The value of is any floating-point number between 0 and 1, and different values ​​are randomly selected. The initial population consists of 50 individuals.

[0092] Step 5.8: Construct the error value function and fitness function for iterative calculation, as shown below:

[0093]

[0094]

[0095] in, The number of samples in the training sample set. The number of classes in the training samples. and The first The corresponding sample of the th sample Predicted and actual values ​​for each category. It is a positive minimum value, to avoid the denominator having a value of 0.

[0096] Step 5.9: Based on the global and local pollination strategies, update the smoothing factor to form a new population and calculate the fitness value. Single-dimensional pollination corresponds to the cross-pollination process of flowers and is directly related to the algorithm's global search capability. The formula for updating offspring individuals is as follows:

[0097] (t)

[0098] In the formula, The first The second iteration and the first The th iteration in the Flowers; This represents the globally optimal pollen location in the current N-dimensional iteration of the population; This describes the characteristics of the pollen dispersal process; The step size adjustment factor for Levy flight; represents the conversion probability between the two pollination modes; rand is the random value generated in the current iteration. .

[0099] For the present Execute Cauchy mutation:

[0100]

[0101] in, These are standard Cauchy distribution random numbers. It is a variable-asynchronous length factor. Utilizing the long wing characteristic of the Cauchy distribution, a large random perturbation is generated, forcing the algorithm to escape the current local extremum region.

[0102] Levy's flight process can be represented as:

[0103]

[0104] in It is the standard Functions, based on empirical values This is the optimal value. s is the flight step length. Minimum step size threshold, constraint conditions The step size s of Levy's flight is limited to be significantly greater than a positive number close to zero. This ensures the long jump characteristic of Levy flight, prevents the algorithm from degenerating into a local search due to excessively small step size, and avoids the mathematical singularity of zero denominator.

[0105] Based on the above, the multi-dimensional global pollination calculation formula for its offspring individuals is as follows:

[0106]

[0107] in, The optimal pollen location in the current population across N dimensions.

[0108] Local pollination corresponds to the self-pollination process of flowers and is directly related to the algorithm's local search capability. The formula for calculating offspring individuals is as follows:

[0109]

[0110] In the formula: Let be a random number that follows a uniform distribution within this interval. , These are random individuals from different flowers of the same type.

[0111] The nonlinear decreasing function is:

[0112]

[0113] In the formula, and These are the upper and lower bounds of the switching probability, respectively. The maximum number of iterations, This is a nonlinear adjustment coefficient. When k is less than 1, the function exhibits a concave shape, prematurely weakening its global search capability and hindering its ability to escape local optima. Therefore, in this embodiment... The value ranges from 1.5 to 2. In the initial stage of the iteration, Maintaining a high level, the algorithm tends to perform global pollination operations. To broaden the search scope; as the number of iterations increases, As the number of pollination operations gradually decreases, the algorithm shifts its focus to local pollination operations. This is to improve the search accuracy of the smoothing factor vector.

[0114] The population of pollinated flowers is updated according to the above calculation strategy, and the adaptive function value and error value of each individual are calculated. The individual with the smallest error value is determined as the optimal factor in this iteration. If the error corresponding to this factor is greater than the error of the optimal factor in the previous iteration, the optimal factor value of this iteration is set as the optimal factor value of the previous iteration.

[0115] Step 5.10: Repeat step 5.9, iterating until the error reaches the expected accuracy. In this invention, the expected error value is set to 0.01. When the error calculated according to the error value function in step 5.8 is less than this value, the expected accuracy is considered to have been reached, and the iteration terminates.

[0116] Step 5.11, Algorithm Termination Conditions This is the optimal smoothing factor for the PNN network. The obtained smoothing factor is then fed into the probabilistic neural network, along with training samples, to determine the probabilistic neural network model. Finally, the sample to be identified is input, and the classification result is obtained.

[0117] To verify the effectiveness of the method proposed in this invention, a transformer partial discharge simulation experimental platform was constructed in a laboratory environment to simulate three typical partial discharge modes: corona discharge, surface discharge, and internal air gap discharge. 100 samples were collected for each mode, with 60 sets used as training samples and 40 sets as test samples. Specific comparative experimental data and analysis are as follows.

[0118] (1) Comparison of signal denoising performance

[0119] The noisy signal was processed using both the traditional EMD method and the CEEMD method used in this invention. Signal-to-noise ratio (SNR) and mean square error (MSE) were used as evaluation metrics.

[0120] Table 1. Experimental Results of Signal Denoising Performance Comparison

[0121]

[0122] Experimental results show that the CEEMD method improves the signal-to-noise ratio and significantly reduces the mean square error. This demonstrates that the CEEMD method effectively solves the mode mixing problem by adding white noise and averaging, and its denoising effect is significantly better than traditional methods, providing a high-quality signal source for subsequent feature extraction.

[0123] (2) Analysis of the effectiveness of feature dimensionality reduction

[0124] After constructing a two-dimensional spectrum under the partial discharge phase distribution mode (PRPD) and extracting eigenvalues, the original feature parameters in this simulation experiment had a dimension of 33. Principal component analysis (PCA) was used for dimensionality reduction, and the contribution rate of the principal components was analyzed.

[0125] Table 2 Results of Feature Dimensionality Reduction Effectiveness Analysis

[0126]

[0127] Experimental results show that the cumulative contribution rate of the first three principal components has reached 92.51%, exceeding the conventional standard of 85%. This demonstrates that using the PCA algorithm to reduce the dimensionality of feature values ​​can compress the feature dimension from 33 dimensions to 3 dimensions while retaining most of the original information, greatly reducing the complexity of the PNN network and improving computational efficiency.

[0128] (3) Performance comparison of pattern recognition algorithms

[0129] To demonstrate the advantages of the FPA-based optimized PNN method, three sets of comparative experiments were conducted: a traditional BP neural network, an unoptimized PNN network, and the hybrid strategy FPA-PNN network proposed in this invention. The comparison metrics included recognition accuracy, number of convergence iterations, and runtime. The simulation test results are shown in the table below:

[0130] Table 3 Performance Comparison Results of Pattern Recognition Algorithms

[0131]

[0132] Experimental results show that the FPA-PNN method achieves an average recognition accuracy of 98.6% on the test set, which is 7.4% higher than that of the unoptimized PNN and 11.1% higher than that of the traditional BP network. In terms of convergence speed, FPA-PNN only requires 42 iterations to reach the expected accuracy with an error value set at 0.01, which verifies the advantages of the flower pollination algorithm in terms of strong global search capability and fast convergence speed. In terms of running time, due to the adoption of PCA dimensionality reduction and efficient PNN structure, the running time of the method of this invention is basically the same as that of the unoptimized PNN and far superior to the BP neural network, meeting the real-time requirements of transformer partial discharge pattern recognition.

[0133] from Figure 4 It can be clearly seen that the hybrid strategy FPA-PNN performs better than the fixed smoothing factor, which in turn performs better than the neural network.

Claims

1. A transformer partial discharge pattern recognition method based on an optimized probabilistic neural network, characterized in that, Includes the following steps: Acquire the partial discharge signal of the transformer; A two-dimensional spectrum of the partial discharge phase distribution mode is established based on the partial discharge signal, and discharge statistical features are extracted from the spectrum. The discharge statistical features are input into an optimized probabilistic neural network for pattern recognition to obtain partial discharge patterns. The optimized probabilistic neural network uses a flower pollination algorithm to optimize the smoothing factor of the probabilistic neural network, and the switching probability in the flower pollination algorithm is a nonlinear function that decreases with the number of iterations.

2. The transformer partial discharge pattern recognition method based on an optimized probabilistic neural network according to claim 1, characterized in that, After acquiring the transformer partial discharge signal, the method further includes: performing noise reduction processing on the partial discharge signal using a complete set empirical mode decomposition method.

3. The transformer partial discharge pattern recognition method based on an optimized probabilistic neural network according to claim 1, characterized in that, After extracting the discharge statistical features from the spectrum, the method further includes: using principal component analysis to reduce the dimensionality of the discharge statistical features.

4. The transformer partial discharge pattern recognition method based on an optimized probabilistic neural network according to claim 1, characterized in that, The switching probability in the flower pollination algorithm adopts a nonlinear function that decreases with the number of iterations, including: nonlinear functions ,in and These are the upper and lower bounds of the switching probability, respectively. The maximum number of iterations, The nonlinear adjustment coefficient is... This represents the number of iterations. Each iteration generates a random value rand, and then compares rand with... The size allows for switching between two pollination modes.

5. The transformer partial discharge pattern recognition method based on an optimized probabilistic neural network according to claim 1, characterized in that, In the flower pollination algorithm, Cauchy mutation is performed on the globally optimal pollen location during the iteration process.

6. The transformer partial discharge pattern recognition method based on an optimized probabilistic neural network according to claim 1, characterized in that, The smoothing factor is a vector related to the dimension n of the discharge statistical characteristics. .

7. The transformer partial discharge pattern recognition method based on an optimized probabilistic neural network according to claim 1, characterized in that, The two-dimensional spectrum includes a spectrum describing the relationship between discharge quantity and phase, a spectrum describing the relationship between maximum discharge quantity and phase, a spectrum describing the relationship between average discharge quantity and phase, and a spectrum describing the relationship between discharge phase and discharge number.

8. A transformer partial discharge pattern recognition system based on an optimized probabilistic neural network, characterized in that, include: The discharge signal acquisition unit is used to acquire the partial discharge signal of the transformer; The discharge statistical feature extraction unit is used to establish a two-dimensional spectrum of the partial discharge phase distribution mode based on the partial discharge signal, and extract discharge statistical features from the spectrum. A discharge pattern recognition unit is used to input the discharge statistical features into an optimized probabilistic neural network for pattern recognition to obtain partial discharge patterns. The optimized probabilistic neural network uses a flower pollination algorithm to optimize the smoothing factor of the probabilistic neural network, and the switching probability in the flower pollination algorithm is a nonlinear function that decreases with the number of iterations.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the transformer partial discharge pattern recognition method based on an optimized probabilistic neural network according to any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the transformer partial discharge pattern recognition method based on an optimized probabilistic neural network according to any one of claims 1-7.

Images