Partial discharge feature selection method based on meta-heuristic swarm intelligence optimization algorithm

Abstract

The invention discloses a partial discharge feature selection method based on a meta-heuristic swarm intelligence optimization algorithm, and the method comprises the steps: extracting a time domain feature, a frequency domain feature and a time-frequency domain feature based on a partial discharge signal sample, and constructing a feature vector and an original feature set; randomly dividing the original feature set into a training set and a verification set; selecting a classifier for partial discharge type classification and identification, training the selected classifier by using the training set to obtain a classifier model, and setting a fitness function as the classification accuracy of the currently trained classifier model on the verification set; a feature subset with the optimal classification performance is obtained by adopting a meta-heuristic swarm intelligence optimization algorithm and through multiple rounds of iterative operation; and re-training the classifier by using the selected optimal feature subset to obtain a classification model, and evaluating the classification model to ensure good performance of the model on test data. According to the method, irrelevant or interference features are effectively removed, and the accuracy and reliability of partial discharge type classification and identification are improved.

Description

Technical Field

[0001] The present invention relates to the field of partial discharge pattern recognition, and in particular to a partial discharge feature selection method based on a meta-heuristic swarm intelligence optimization algorithm. Background Art

[0002] Partial discharge is the main cause of insulation degradation of electrical equipment, and is also an important sign and manifestation of insulation degradation. It is closely related to the degradation of insulating materials and the breakdown process of insulators, and can effectively reflect the latent defects and faults of insulation inside electrical equipment. According to the electric field distribution of insulation defects in electrical equipment, the defect types of partial discharge can usually be divided into the following basic types: metal tip discharge in gas or liquid insulation, air gap discharge when there is an air gap inside solid insulation, surface discharge along the surface of solid insulation, and suspended potential discharge caused by poor contact of metal bodies such as loose screws in electrical equipment, and free particle discharge caused by the presence of metal particles. The discharge type is closely related to the degree of harm of the discharge. Pattern recognition of the discharge type based on the detection signal of partial discharge is an important basis for analyzing the insulation defects of electrical equipment. However, the existing partial discharge pattern recognition method often directly sends all the extracted time domain or statistical features directly into the classifier for classification and recognition, which will cause some irrelevant and interference features to have an adverse effect on the discharge type recognition, thereby affecting the accuracy and reliability of partial discharge type recognition. Summary of the invention

[0003] The purpose of the present invention is to provide a partial discharge feature selection method based on a meta-heuristic swarm intelligence optimization algorithm, which removes redundant and interfering features through feature optimization technology, and selects a small number of features that are strongly associated with the discharge type to improve the accuracy and reliability of partial discharge type identification.

[0004] The technical solution adopted by the present invention is as follows: a partial discharge feature selection method based on a meta-heuristic swarm intelligence optimization algorithm, comprising the following steps:

[0005] Step S1, feature set construction: based on the partial discharge signal sample, extract the time domain features: mean, variance, peak, skewness, kurtosis; frequency domain features: power spectrum density, spectrum width, harmonic features; time-frequency domain features: wavelet transform coefficients, Hilbert transform features; combine all the extracted features to form a feature vector X = [x1, x2, ..., x n ], where n is the number of features extracted; the feature vectors corresponding to all partial discharge signal samples are combined to construct the original feature set Q = {X1, X2, ..., X m}, where m is the number of all samples;

[0006] Step S2, feature set division: randomly divide the original feature set Q into a training set and a validation set, where the training set is used to train the classifier model and the validation set is used to verify the classification accuracy of the classifier model;

[0007] Step S3, fitness function construction: select a classifier for partial discharge type classification and identification; use the training set to train the selected classifier to obtain a classifier model; the goal of feature selection is to select an optimal feature subset X b To maximize the performance of the classifier, the fitness function is set to the classification accuracy of the currently trained classifier model on the validation set, which is defined as follows: Among them, TC represents the number of correct classifications, and FC represents the number of incorrect classifications;

[0008] Step S4, feature selection: obtaining the feature subset with the best classification performance by adopting a meta-heuristic swarm intelligence optimization algorithm and performing multiple rounds of iterative operations;

[0009] Step S5, classification and evaluation: retrain the classifier using the selected optimal feature subset to obtain a classification model, and evaluate the classification model to ensure good performance of the model on the test data.

[0010] Further, step S4 includes:

[0011] Step S4-1, population initialization: set the number of individuals Np in the population, and for each individual, randomly select several eigenvalues ​​from the eigenvector in step S1 to initialize the individual position, i.e. the corresponding feature subset X;

[0012] Step S4-2, individual evaluation: use the feature subset X corresponding to each individual to perform classifier training and classifier verification, and calculate the fitness function value f(X);

[0013] Step S4-3, obtain the best individual and the worst individual: the best individual X b is the individual with the largest fitness function value, and the worst individual X w is the individual with the smallest fitness function value;

[0014] Step S4-4, iterative update: set the maximum number of iterations T and the current number of iterations t=1, set the adaptive parameters v and DF, and when t is less than T, iteratively update the individual state according to the following steps:

[0015] Step S4-4-1, individual position update: Randomly select parameters r1 and r2 in the range [1, Np], and for each individual, use the following formula to update the individual state:

[0016]

[0017] Among them, X t is the current individual position, X t+1 is the updated individual position, a and b are random numbers between (0,1), randn represents a normally distributed random number with a mean of 0 and a variance of 1, and X b is the optimal individual position,

[0018] Δx=rand(1,n)×|X b -X t |,

[0019]

[0020] In order to avoid falling into the local optimum, a mechanism to escape from the local optimum is set. When rand is less than the set DF, the individual state is updated according to the following formula:

[0021]

[0022] Among them, γ1 and γ2 are random numbers between (-1,1) and (-0.5,0.5), and are random numbers, generated by the following formulas:

[0023]

[0024] Among them, β represents a binary number, that is, 1 or 0, and rand represents a uniform random number between (0,1);

[0025] Step S4-4-2, individual evaluation: use the feature subset X corresponding to each individual t+1 , perform classifier training and classifier verification, and calculate the fitness function value f(X t+1 );

[0026] Step S4-4-3, update the best and worst individuals: sort the individuals in the population according to the fitness function value, and update the individual with the smallest fitness function value as the worst individual X w , update the individual with the largest fitness function value to the optimal individual X b ;

[0027] Step S4-4-4, add 1 to the current number of iterations t, and determine whether the previous number of iterations has reached the maximum number of iterations. If not, repeat steps S4-4-2 to S4-4-4. If reached, output the feature subset corresponding to the optimal individual as the selected optimal feature subset. This feature subset is the feature set with the best classification performance.

[0028] Furthermore, the ratio of the training set to the validation set in step S2 is 7:3.

[0029] Furthermore, in step S4-1, the number of individuals Np in the population is an integer between 20 and 50.

[0030] The beneficial effects of the present invention are as follows: the method of the present invention can remove those features that are irrelevant or interfere with the discharge type identification through feature selection, thereby reducing the influence of noise and redundant information on the classifier and improving the accuracy of partial discharge type identification; since only the features that contribute most to the classification are selected, the constructed model has stronger robustness to slight changes or noise in the input data and can maintain stable performance under different conditions; the meta-heuristic swarm intelligence optimization algorithm is used for feature selection, which realizes the automation of feature selection, reduces manual intervention and improves work efficiency. The present invention provides an effective technical means for insulation status monitoring and fault diagnosis of electrical equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 The present invention is a flowchart of a method for selecting features for identifying partial discharge types according to an embodiment of the present invention.

[0032] Figure 2 The present invention is a flowchart of performing feature selection through a meta-heuristic swarm intelligence optimization algorithm according to an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The technical solutions in the embodiments of the present invention are described clearly and completely below. 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 creative work are within the scope of protection of the present invention.

[0034] Figure 1 A flow chart of a method for selecting local discharge type identification features according to an embodiment of the present invention is shown. As shown in the figure, a method for selecting local discharge features based on a meta-heuristic swarm intelligence optimization algorithm comprises:

[0035] Collect a certain number of various types of partial discharge signal samples inside the switch cabinet using ultra-high frequency, transient or ultrasonic partial discharge detection devices. Preferably, the number of partial discharge signal samples of each type is not less than 50;

[0036] Step S1, feature set construction: based on the partial discharge signal sample, extract the time domain features: mean, variance, peak, skewness, kurtosis; frequency domain features: power spectrum density, spectrum width, harmonic features; time-frequency domain features: wavelet transform coefficients, Hilbert transform features; combine all the extracted features to form a feature vector X = [x1, x2, ..., x n], where n is the number of features extracted; the feature vectors corresponding to all partial discharge signal samples are combined to construct the original feature set Q = {X1, X2, ..., X m}, where m is the number of all samples;

[0037] Step S2, feature set division: randomly divide the original feature set Q into a training set and a validation set according to a certain ratio, wherein the training set is used to train the classifier model, and the validation set is used to verify the classification accuracy of the classifier model. Preferably, the ratio of the training set to the validation set is 7:3;

[0038] Step S3, fitness function construction: select a classifier for partial discharge type classification and identification, such as support vector machine, linear discriminant, KNN, etc.; use the training set to train the selected classifier to obtain the classifier model; the goal of feature selection is to select an optimal feature subset X b To maximize the performance of the classifier, the fitness function is set to the classification accuracy of the currently trained classifier model on the validation set, which is defined as follows: Among them, TC represents the number of correct classifications, and FC represents the number of incorrect classifications;

[0039] Step S4, feature selection: a meta-heuristic swarm intelligence optimization algorithm is used to obtain the feature subset with the best classification performance through multiple rounds of iterative operations. Figure 2 The flowchart of the embodiment of the present invention for performing feature selection by using a meta-heuristic swarm intelligence optimization algorithm is shown, including:

[0040] Step S4-1, population initialization: Set the number of individuals Np in the population. For each individual, randomly select several eigenvalues ​​from the eigenvector in step S1 to initialize the position of the individual, that is, the corresponding feature subset X. Preferably, Np is generally set to an integer between 20 and 50;

[0041] Step S4-2, individual evaluation: use the feature subset X corresponding to each individual to perform classifier training and classifier verification, and calculate the fitness function value f(X);

[0042] Step S4-3, obtain the best individual and the worst individual: the best individual X b is the individual with the largest fitness function value, and the worst individual X w is the individual with the smallest fitness function value;

[0043] Step S4-4, iterative update: set the maximum number of iterations T and the current number of iterations t=1, set the adaptive parameters v and DF, preferably, the maximum number of iterations T is generally set to 100, v is set to 0.8, and DF is set to 0.6. When t is less than T, iteratively update the individual state according to the following steps:

[0044] Step S4-4-1, individual position update: Randomly select parameters r1 and r2 in the range [1, Np], and for each individual, use the following formula to update the individual state:

[0045]

[0046] Among them, X t is the current individual position, X t+1 is the updated individual position, a and b are random numbers between (0,1), randn represents a normally distributed random number with a mean of 0 and a variance of 1, and X b is the optimal individual position,

[0047] Δx=rand(1,n)×|X b -X t |,

[0048]

[0049] In order to avoid falling into the local optimum, a mechanism to escape from the local optimum is set. When rand is less than the set DF, the individual state is updated according to the following formula:

[0050]

[0051] Among them, γ1 and γ2 are random numbers between (-1,1) and (-0.5,0.5), and are random numbers, generated by the following formulas:

[0052]

[0053] Among them, β represents a binary number, that is, 1 or 0, and rand represents a uniform random number between (0,1);

[0054] Step S4-4-2, individual evaluation: use the feature subset X corresponding to each individual t+1 , perform classifier training and classifier verification, and calculate the fitness function value f(X t+1 );

[0055] Step S4-4-3, update the best and worst individuals: sort the individuals in the population according to the fitness function value, and update the individual with the smallest fitness function value as the worst individual X w , update the individual with the largest fitness function value to the optimal individual X b ;

[0056] Step S4-4-4, add 1 to the current iteration number t, and determine whether the previous iteration number has reached the maximum iteration number. If not, repeat steps S4-4-2 to S4-4-4. If it has reached the maximum iteration number, output the feature subset X corresponding to the optimal individual. b As the optimal feature subset selected, this feature subset is usually the feature set with the best classification performance;

[0057] Step S5, classification and evaluation: using the selected optimal feature subset X b Retrain the classifier to obtain the classification model; use cross-validation and other methods to evaluate the classification model, calculate evaluation indicators such as accuracy, recall, F1-score, etc., to ensure the good performance of the model on the test data.

[0058] The above contents are merely examples and explanations of the structure of the present invention. The technicians in this technical field may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the structure of the present invention or exceed the scope defined by the claims, they should all fall within the protection scope of the present invention.

Claims

1. A partial discharge feature selection method based on a meta-heuristic swarm intelligence optimization algorithm, characterized in that: The steps include: Step S1, feature set construction: based on the partial discharge signal samples, extract time domain features: mean, variance, peak, skewness, and kurtosis; Frequency domain features: power spectrum density, spectrum width, harmonic features; time-frequency domain features: wavelet transform coefficients, Hilbert transform features; all extracted features are combined to form a feature vector X = [x1, x2, ..., x n ], where n is the number of features extracted; the feature vectors corresponding to all partial discharge signal samples are combined to construct the original feature set Q = {X1, X2, ..., X m }, where m is the number of all samples; Step S2, feature set division: randomly divide the original feature set Q into a training set and a validation set, where the training set is used to train the classifier model and the validation set is used to verify the classification accuracy of the classifier model; Step S3, fitness function construction: select a classifier for partial discharge type classification and identification; use the training set to train the selected classifier to obtain a classifier model; the goal of feature selection is to select an optimal feature subset X b To maximize the performance of the classifier, the fitness function is set to the classification accuracy of the currently trained classifier model on the validation set, which is defined as follows: Among them, TC represents the number of correct classifications, and FC represents the number of incorrect classifications; Step S4, feature selection: obtaining the feature subset with the best classification performance by adopting a meta-heuristic swarm intelligence optimization algorithm and performing multiple rounds of iterative operations; Step S5, classification and evaluation: retrain the classifier using the selected optimal feature subset to obtain a classification model, and evaluate the classification model to ensure good performance of the model on the test data.

2. A method for selecting local discharge features based on a meta-heuristic swarm intelligence optimization algorithm as claimed in claim 1, characterized in that: Step S4 includes: Step S4-1, population initialization: set the number of individuals Np in the population, and for each individual, randomly select several eigenvalues ​​from the eigenvector in step S1 to initialize the individual position, i.e. the corresponding feature subset X; Step S4-2, individual evaluation: use the feature subset X corresponding to each individual to perform classifier training and classifier verification, and calculate the fitness function value f(X); Step S4-3, obtain the best individual and the worst individual: the best individual X b is the individual with the largest fitness function value, and the worst individual X w is the individual with the smallest fitness function value; Step S4-4, iterative update: set the maximum number of iterations T and the current number of iterations t=1, set the adaptive parameters v and DF, and when t is less than T, iteratively update the individual state according to the following steps: Step S4-4-1, individual position update: Randomly select parameters r1 and r2 in the range [1, Np], and for each individual, use the following formula to update the individual state: Among them, X t is the current individual position, X t+1 is the updated individual position, a and b are random numbers between (0,1), randn represents a normally distributed random number with a mean of 0 and a variance of 1, and X b is the optimal individual position, Δx=rand(1,n)×|X b -X t |, In order to avoid falling into the local optimum, a mechanism to escape from the local optimum is set. When rand is less than the set DF, the individual state is updated according to the following formula: Among them, γ1 and γ2 are random numbers between (-1,1) and (-0.5,0.5), and are random numbers, generated by the following formulas: Among them, β represents a binary number, that is, 1 or 0, and rand represents a uniform random number between (0,1); Step S4-4-2, individual evaluation: use the feature subset X corresponding to each individual t+1 , perform classifier training and classifier verification, and calculate the fitness function value f(X t+1 ); Step S4-4-3, update the best and worst individuals: sort the individuals in the population according to the fitness function value, and update the individual with the smallest fitness function value as the worst individual X w , update the individual with the largest fitness function value to the optimal individual X b ; Step S4-4-4, add 1 to the current number of iterations t, and determine whether the previous number of iterations has reached the maximum number of iterations. If not, repeat steps S4-4-2 to S4-4-4. If reached, output the feature subset corresponding to the optimal individual as the selected optimal feature subset. This feature subset is the feature set with the best classification performance.

3. A method for selecting local discharge features based on a meta-heuristic swarm intelligence optimization algorithm as claimed in claim 1, characterized in that: The ratio of the training set to the validation set in step S2 is 7:

3.

4. The method for selecting local discharge features based on a meta-heuristic swarm intelligence optimization algorithm according to claim 1, characterized in that: In step S4-1, the number of individuals Np in the population is an integer between 20 and 50.

Images