Cable partial discharge identification method and device based on improved decision tree-rotating forest

Abstract

The invention discloses a cable partial discharge identification method and device based on an improved decision tree-rotating forest, and the method comprises the steps: carrying out the wavelet packet decomposition energy feature extraction of a TRPD ultrasonic signal, calculating the zero-crossing rate of the ultrasonic signal, and carrying out the classification of TRPD ultrasonic features through employing an improved decision tree algorithm based on adaptive weighted information adjustment gain. Multi-dimensional information such as occurrence time, phase and energy distribution of partial discharge can be comprehensively described, characteristic information of cable partial discharge can be more comprehensively acquired, and the accuracy of partial discharge detection and identification is improved; the Bayesian fusion algorithm based on multi-stage fusion is adopted, the results of different feature classifiers are fused, the defects of a single classification algorithm are overcome, and the reliability of the partial discharge type recognition result is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of cable partial discharge detection, and in particular to a cable partial discharge identification method and equipment based on an improved decision tree-rotation forest. Background Art

[0002] As power cables are widely used in urban power transmission and distribution systems, their safe operation has become crucial. Partial discharge is an early sign of cable insulation failure, so accurate detection and identification of it is of great significance. At present, although there are many partial discharge detection methods, there are still some problems when applying ultrasonic detection to cable partial discharge diagnosis, such as unclear ultrasonic detection indicators, imperfect quantitative detection methods, inaccurate partial discharge analysis model establishment, and the recognition rate of partial discharge types needs to be further improved.

[0003] Ultrasonic testing faces challenges in cable partial discharge diagnosis, including the problem of unclear ultrasonic testing indicators. This means that in practical applications, it may be difficult to determine which ultrasonic signals are directly related to partial discharge activity and how to accurately quantify these signals for effective diagnosis.

[0004] When dealing with strong random signals, the traditional method of partial discharge pattern recognition technology relies on statistical operators for feature extraction. This method has limitations in the classification of DC cable partial discharge signals with strong randomness. For example, the method of feature extraction using statistical operators such as skewness and mutual correlation coefficient may not accurately capture all the characteristics of partial discharge signals.

[0005] Although some studies have proposed pattern recognition methods based on deep learning, such as convolutional neural networks (CNNs), to improve the recognition rate, these methods may still face challenges in practical applications. For example, the partial discharge signals collected on site contain a lot of noise, which may affect the effectiveness of pattern recognition. In addition, the effectiveness of pattern recognition of partial discharge signals under different noise intensities needs further study.

[0006] In summary, the existing partial discharge detection methods cannot fully describe the information of partial discharge, and thus cannot accurately understand the partial discharge phenomenon, and cannot fully obtain the characteristic information of cable partial discharge, resulting in poor partial discharge type identification results. Summary of the invention

[0007] Based on the problems raised by the above background technology, the purpose of the present invention is to provide a cable partial discharge identification method based on an improved decision tree-rotation forest, which solves the problem that the existing partial discharge detection method cannot fully describe the information of partial discharge, and thus cannot have an accurate understanding of the partial discharge phenomenon, and cannot fully obtain the characteristic information of cable partial discharge, thereby resulting in poor partial discharge type identification results.

[0008] The present invention is achieved through the following technical solutions:

[0009] The first aspect of the present invention provides a cable partial discharge identification method based on an improved decision tree-rotation forest, comprising the following steps:

[0010] Step S1, obtaining a cable partial discharge monitoring signal, performing ultrasonic signal feature extraction on the cable partial discharge monitoring signal, and obtaining a TPRD ultrasonic feature, a PRPD ultrasonic feature, and an EPRD ultrasonic feature;

[0011] Step S2, using an improved decision tree algorithm based on adaptive weighted information to adjust gain to perform classification prediction on the TPRD ultrasonic features, and using a rotation forest algorithm to perform integrated prediction on the PRPD ultrasonic features and the EPRD ultrasonic features;

[0012] Step S3: A Bayesian fusion algorithm based on multi-stage fusion performs decision fusion on the results of classification prediction and the results of integrated prediction, and determines the type of cable partial discharge according to the result of decision fusion.

[0013] In the above technical scheme, the TRPD ultrasonic signal is decomposed by wavelet packets to extract energy features and calculate the zero-crossing rate of the ultrasonic signal, and the TRPD ultrasonic features are classified by adopting an improved decision tree algorithm based on adaptive weighted information to adjust the gain. The PRPD ultrasonic signal is feature extracted by phase feature extraction based on high-order statistics, and the Cohen-class time-frequency distribution feature is extracted for the ERPD ultrasonic feature. The PRPD ultrasonic features and ERPD ultrasonic features are classified by the rotation forest algorithm. TRPD focuses on the characteristics of the discharge pulse on the time axis, PRPD focuses on the relationship between discharge and phase, and ERPD analyzes partial discharge from the energy perspective. The combination of the three technologies can comprehensively describe the multi-dimensional information such as the occurrence time, phase and energy distribution of partial discharge, so that researchers can have a more complete understanding of the partial discharge phenomenon. The above feature extraction method and classification method can more comprehensively obtain the characteristic information of cable partial discharge, and improve the accuracy of partial discharge detection and identification.

[0014] The Bayesian fusion algorithm based on multi-stage fusion is used to fuse the results of different feature classifiers, which makes up for the shortcomings of a single classification algorithm and improves the reliability of partial discharge type identification results. It solves the problem that the existing partial discharge detection method cannot fully describe the information of partial discharge, thus cannot accurately understand the partial discharge phenomenon, cannot fully obtain the characteristic information of cable partial discharge, and thus causes poor partial discharge type identification results.

[0015] In an optional embodiment, the TPRD ultrasound feature is classified and predicted using an improved decision tree algorithm based on adaptive weighted information to adjust gain, comprising the following steps:

[0016] Establishing a training data set using the TPRD ultrasound features;

[0017] Calculating the weighted information entropy of the training data set, and constructing a decision tree with the weighted information entropy;

[0018] Starting from the leaf nodes of the decision tree, pruning evaluation is performed on each non-leaf node in turn upward, and the decision tree is pruned according to the results of the pruning evaluation;

[0019] A test data set is established with the TPRD ultrasound features, and a pruned decision tree is used to perform classification prediction on the test data set to obtain a predicted category.

[0020] In an optional embodiment, starting from the leaf node of the decision tree, pruning evaluation is performed on each non-leaf node upward in sequence, including the following steps:

[0021] For a non-leaf node N, calculate the original subtree T corresponding to the non-leaf node N N The weighted classification error Error on the training dataset w (T N );

[0022] Mark the non-leaf node N as a leaf node N, and calculate the weighted classification error Error of the subtree corresponding to the leaf node N w (N);

[0023] If Error w (N)≤Error w (T N ), then the non-leaf node N is pruned.

[0024] In an optional embodiment, the integrated prediction of the PRPD ultrasonic feature and the EPRD ultrasonic feature is performed using a rotating forest algorithm, comprising the following steps:

[0025] Establishing a data set with the PRPD ultrasound features and the EPRD ultrasound features;

[0026] Establishing K decision trees, determining parameters of the decision trees based on the data set, and training the decision trees based on the parameters;

[0027] The trained decision tree is used to predict the PRPD ultrasonic features and the EPRD ultrasonic features to obtain K category results, and a final predicted category is determined from the K category results.

[0028] In an optional embodiment, determining the parameters of the decision tree based on the data set, and training the decision tree based on the parameters comprises the following steps:

[0029] The data set includes n samples and m features;

[0030] For each decision tree t, k features are randomly selected from the m features to form a feature subset F t ;

[0031] For the feature subset F t Perform feature rotation and use the feature subset F after feature rotation t Train the decision tree.

[0032] In an optional embodiment, a Bayesian fusion algorithm based on multi-stage fusion performs decision fusion on the classification prediction results and the integrated prediction results, including the following steps:

[0033] Determine the prior probabilities of the TPRD ultrasonic feature, the PRPD ultrasonic feature, and the EPRD ultrasonic feature, and calculate the conditional probabilities of the classifier in the improved decision tree algorithm and the classifier in the rotation forest algorithm respectively;

[0034] Perform preliminary fusion based on the prior probability to obtain a preliminary fusion probability distribution;

[0035] The predicted probability of the classifier for the sample is calculated according to the conditional probability, and the preliminary fusion probability distribution is secondary fused using the Bayesian fusion algorithm based on the predicted probability to obtain the final fusion probability distribution.

[0036] In an optional embodiment, performing ultrasonic signal feature extraction on the cable partial discharge monitoring signal to obtain the PRPD ultrasonic feature includes the following steps:

[0037] Extracting a PRPD ultrasonic signal from the cable partial discharge monitoring signal;

[0038] Dividing one cycle of the PRPD ultrasonic signal into a plurality of phase intervals;

[0039] Calculating the third-order cumulant and the fourth-order cumulant of the PRPD ultrasonic signal in each phase interval respectively;

[0040] Based on the numerical values, increasing and decreasing trends and fluctuations of the third-order cumulants and the fourth-order cumulants in different phase intervals, PRPD ultrasonic features are extracted from the PRPD ultrasonic signal.

[0041] In an optional embodiment, performing ultrasonic signal feature extraction on the cable partial discharge monitoring signal to obtain EPRD ultrasonic features includes the following steps:

[0042] Extracting EPRD ultrasonic signal from the cable partial discharge monitoring signal;

[0043] The fuzzy function of the EPRD ultrasonic signal is calculated by using a numerical integration method;

[0044] Constructing a rectangular kernel function, bringing the fuzzy function and the rectangular kernel function into the Cohen-like time-frequency distribution formula for integral calculation, and obtaining the Cohen-like time-frequency distribution;

[0045] Signal classification is performed on the Cohen-like time-frequency distribution to obtain EPRD ultrasound features.

[0046] A second aspect of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a cable partial discharge identification method based on an improved decision tree-rotation forest is implemented.

[0047] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a cable partial discharge identification method based on an improved decision tree-rotation forest.

[0048] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0049] By extracting energy features from TRPD ultrasonic signals through wavelet packet decomposition and calculating the zero-crossing rate of ultrasonic signals, and adopting an improved decision tree algorithm based on adaptive weighted information to adjust gain, the TRPD ultrasonic features can be classified, which can comprehensively describe the multi-dimensional information such as the occurrence time, phase and energy distribution of partial discharge, and can more comprehensively obtain the characteristic information of cable partial discharge, thus improving the accuracy of partial discharge detection and identification; the Bayesian fusion algorithm based on multi-stage fusion is adopted to fuse the results of different feature classifiers, which makes up for the shortcomings of a single classification algorithm and improves the reliability of partial discharge type identification results. It solves the problem that the existing partial discharge detection method cannot comprehensively describe the information of partial discharge, thus cannot accurately understand the partial discharge phenomenon, cannot comprehensively obtain the characteristic information of cable partial discharge, and thus leads to poor partial discharge type identification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without creative work. In the drawings:

[0051] Figure 1A schematic flow chart of a cable partial discharge identification method based on an improved decision tree-rotation forest provided in Example 1 of the present invention;

[0052] Figure 2 A schematic diagram of the structure of an electronic device provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0053] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0054] Example 1

[0055] Figure 1 A schematic diagram of a flow chart of a cable partial discharge identification method based on an improved decision tree-rotation forest provided in Example 1 of the present invention, such as Figure 1 As shown in FIG. 1 , the cable partial discharge identification method based on the improved decision tree-rotation forest includes the following steps:

[0056] Step S1, obtaining a cable partial discharge monitoring signal, performing ultrasonic signal feature extraction on the cable partial discharge monitoring signal, and obtaining a TPRD ultrasonic feature, a PRPD ultrasonic feature, and an EPRD ultrasonic feature;

[0057] Step S2, using an improved decision tree algorithm based on adaptive weighted information to adjust gain to perform classification prediction on the TPRD ultrasonic features, and using a rotation forest algorithm to perform integrated prediction on the PRPD ultrasonic features and the EPRD ultrasonic features;

[0058] Step S3: A Bayesian fusion algorithm based on multi-stage fusion performs decision fusion on the results of classification prediction and the results of integrated prediction, and determines the type of cable partial discharge according to the result of decision fusion.

[0059] It should be noted that the TRPD ultrasonic signal is decomposed by wavelet packets to extract energy features and calculate the zero-crossing rate of the ultrasonic signal, and the TRPD ultrasonic features are classified by adopting an improved decision tree algorithm based on adaptive weighted information to adjust the gain. The PRPD ultrasonic signal is extracted by phase feature extraction based on high-order statistics, and the Cohen-class time-frequency distribution feature is extracted for the ERPD ultrasonic feature. The PRPD ultrasonic features and ERPD ultrasonic features are classified by the rotation forest algorithm. TRPD focuses on the characteristics of the discharge pulse on the time axis, PRPD focuses on the relationship between discharge and phase, and ERPD analyzes partial discharge from the energy perspective. The combination of the three technologies can comprehensively describe the multi-dimensional information such as the occurrence time, phase and energy distribution of partial discharge, so that researchers can have a more complete understanding of the partial discharge phenomenon. The above feature extraction method and classification method can more comprehensively obtain the characteristic information of cable partial discharge and improve the accuracy of partial discharge detection and identification.

[0060] The Bayesian fusion algorithm based on multi-stage fusion is used to fuse the results of different feature classifiers, which makes up for the shortcomings of a single classification algorithm and improves the reliability of partial discharge type identification results. It solves the problem that the existing partial discharge detection method cannot fully describe the information of partial discharge, thus cannot accurately understand the partial discharge phenomenon, cannot fully obtain the characteristic information of cable partial discharge, and thus causes poor partial discharge type identification results.

[0061] In an optional embodiment, performing ultrasonic signal feature extraction on the cable partial discharge monitoring signal to obtain the TPRD ultrasonic feature includes the following steps:

[0062] Extracting a TPRD ultrasonic signal from the cable partial discharge monitoring signal;

[0063] Decomposing the TPRD ultrasonic signal into different frequency bands by using wavelet basis functions, and calculating the energy in each frequency band for the wavelet packet coefficients in each frequency band;

[0064] Calculating the average zero-crossing rate of the TPRD ultrasonic signal within a certain time window;

[0065] A TPRD ultrasound feature is extracted from the TPRD ultrasound signal based on the energy and the average zero-crossing rate in each frequency band.

[0066] It should be noted that in the TRPD (partial discharge) mode of the partial discharge signal, it is necessary to extract the TPRD ultrasonic features. First, the TPRD ultrasonic signal is extracted from the cable partial discharge monitoring signal, and the TPRD features of the TPRD ultrasonic signal are extracted from the frequency domain characteristics and the time domain characteristics respectively.

[0067] As for the frequency domain characteristics of TRPD, this embodiment uses wavelet packet decomposition features for feature extraction, wherein the Daubechies wavelet basis function is selected, the number of decomposition layers is determined to be 4, and the pulse waveform parameters and amplitude features in the TPRD ultrasonic signal are decomposed into different frequency bands.

[0068] For each frequency band, the energy is obtained by summing the squares of the wavelet packet coefficients in the frequency band. Specifically, the calculation formula is as follows:

[0069]

[0070] Among them, E j represents the energy of the jth frequency band after decomposition, c jk represents the kth wavelet packet coefficient in the jth frequency band.

[0071] In terms of TRPD time domain characteristics, this embodiment extracts features by calculating the zero-crossing rate of the TRPD ultrasonic signal, where the zero-crossing rate refers to the ratio of the number of times the signal changes from positive to negative or from negative to positive to the total number of samples within a certain time window. For a discrete ultrasonic signal sequence x(n), the calculation formula for the zero-crossing rate ZCR is:

[0072]

[0073] where N is the length of the signal sequence and sgn(·) is the sign function.

[0074] The first few frequency bands with large energy are selected from the frequency bands as TRPD ultrasonic features. At the same time, the average zero-crossing rate within a certain time window is counted and used as the TRPD ultrasonic feature.

[0075] There are two methods for analyzing TRPD signals from different perspectives. Wavelet packet decomposition feature extraction focuses on the frequency domain characteristics of the signal, while zero-crossing rate focuses on the time domain characteristics of the signal. Combining these two features can more comprehensively describe and identify TRPD signals.

[0076] In an optional embodiment, performing ultrasonic signal feature extraction on the cable partial discharge monitoring signal to obtain the PRPD ultrasonic feature includes the following steps:

[0077] Extracting a PRPD ultrasonic signal from the cable partial discharge monitoring signal;

[0078] Dividing one cycle of the PRPD ultrasonic signal into a plurality of phase intervals;

[0079] Calculating the third-order cumulant and the fourth-order cumulant of the PRPD ultrasonic signal in each phase interval respectively;

[0080] Based on the numerical values, increasing and decreasing trends and fluctuations of the third-order cumulants and the fourth-order cumulants in different phase intervals, PRPD ultrasonic features are extracted from the PRPD ultrasonic signal.

[0081] It should be noted that the PRPD ultrasonic signal (Phase Resolved Partial Discharge) is a technology for detecting partial discharge phenomena, which identifies and locates the discharge source by analyzing the relationship between the discharge signal and the voltage phase. This embodiment calculates the high-order statistics of the PRPD ultrasonic signal and extracts the phase feature based on the high-order statistics. Specifically, a cycle (0-360°) of the PRPD ultrasonic signal is divided into 36 phase intervals, each interval is 10°. This helps to divide the periodic characteristics of the signal into finer phase intervals for more detailed analysis. For a specific phase value, determine the phase interval to which it belongs. This step classifies the local discharge events in the signal into specific phase intervals in preparation for subsequent statistical analysis. For each phase interval, the third-order cumulative amount and the fourth-order cumulative amount of the PRPD signal in the interval are calculated. These cumulative amounts are important indicators in high-order statistics, and they can provide non-Gaussian characteristic information of the signal.

[0082] Among them, for the discrete PRPD ultrasonic signal sequence, the third-order cumulant C 3x (m,n) and the fourth-order cumulant C 4x The calculation formulas for (m,n,p) are:

[0083]

[0084] The above formula extracts phase-related features from high-order statistics, where k represents the discrete time point index during the partial discharge signal acquisition process. It represents the time points at which the cable partial discharge signal is sampled at different times. x(k) represents the cable partial discharge signal value collected at the time point. m, n, and p are specific time delay parameters used to introduce different time intervals in the time series of the cable partial discharge signal. m represents the first time delay relative to a reference time point, n is the second time delay, and p is the third time delay. These features include the numerical value, increase or decrease trend, and fluctuation of the cumulative amount in different phase intervals, which help to reveal the nonlinear and non-Gaussian characteristics of the signal and are of great significance for the classification and identification of the signal. The extraction of PRPD ultrasonic features from the PRPD ultrasonic signal is completed by observing the numerical value, increase or decrease trend, and fluctuation of the cumulative amount in different phase intervals.

[0085] In an optional embodiment, performing ultrasonic signal feature extraction on the cable partial discharge monitoring signal to obtain EPRD ultrasonic features includes the following steps:

[0086] Extracting EPRD ultrasonic signal from the cable partial discharge monitoring signal;

[0087] The fuzzy function of the EPRD ultrasonic signal is calculated by using a numerical integration method;

[0088] Constructing a rectangular kernel function, bringing the fuzzy function and the rectangular kernel function into the Cohen-like time-frequency distribution formula for integral calculation, and obtaining the Cohen-like time-frequency distribution;

[0089] Signal classification is performed on the Cohen-like time-frequency distribution to obtain EPRD ultrasound features.

[0090] It should be noted that different fault sources have different energy characteristics when generating partial discharge. Through ERPD feature extraction, the energy distribution pattern of partial discharge can be analyzed to identify the characteristics of the fault source.

[0091] First, the numerical integration method is used to approximate the fuzzy function of the EPRD ultrasonic signal. The fuzzy function is the basis of the time-frequency analysis of the signal. It provides the representation of the signal in the time-frequency domain. By calculating the fuzzy function, the energy distribution of the signal can be obtained. Specifically, the calculation process is as follows:

[0092]

[0093] Among them, A x (τ,ν) is the blur function of the signal x(t).

[0094] Construct a rectangular kernel function Φ(τ,ν), bring the fuzzy function and the rectangular kernel function into the Cohen-like time-frequency distribution formula for integral calculation, and obtain the Cohen-like time-frequency distribution. Specifically, the calculation process is as follows:

[0095]

[0096] Among them, C x (t,f) represents the Cohen-type time-frequency distribution of the signal x(t), τ is the time delay parameter. It represents the relative displacement of the two signals in time. ν represents the frequency variable. e -j2πνt It is a frequency-dependent exponential function that converts the signal in the time domain to the frequency domain. By analyzing different frequencies ν, we can understand the characteristics of the signal in the frequency domain. "j" usually represents the imaginary unit, that is, In this formula, e -j2πνt It is a complex exponential function, where "j" makes this function have real and imaginary parts, which plays an important role in frequency domain analysis. Through this complex exponential function, the signal in the time domain can be converted to the frequency domain for analysis.

[0097] ERPD feature extraction is performed using Cohen-like time-frequency distribution, which can combine the time domain and frequency domain information of the signal, and more comprehensively display the frequency characteristics of the signal changing over time and the distribution of different frequency components in time, which is particularly important for the analysis of non-stationary signals. It can highlight the energy distribution of the signal within the local time and frequency range, and help discover local features such as transient components, mutation points, and rapid changes in frequency in the ERPD signal.

[0098] Finally, the Cohen-like time-frequency distribution is monitored and its signals are classified to identify different fault source characteristics.

[0099] In an optional embodiment, the TPRD ultrasound feature is classified and predicted using an improved decision tree algorithm based on adaptive weighted information to adjust gain, comprising the following steps:

[0100] Establishing a training data set using the TPRD ultrasound features;

[0101] Calculating the weighted information entropy of the training data set, and constructing a decision tree with the weighted information entropy;

[0102] Starting from the leaf nodes of the decision tree, pruning evaluation is performed on each non-leaf node in turn upward, and the decision tree is pruned according to the results of the pruning evaluation;

[0103] A test data set is established with the TPRD ultrasound features, and a pruned decision tree is used to perform classification prediction on the test data set to obtain a predicted category.

[0104] It should be noted that the training data set D is established based on the TPRD ultrasound feature. The sample x in the training data set D i Corresponding to a eigenvector F i and the class label y i , and weight ω i .

[0105] First, for the root node of the decision tree, calculate the weighted information entropy of the training data set D w (D) and use it as the initial information entropy of the root node of the decision tree. At the same time, the weighted information gain IG of each attribute in the training data set is calculated w (D, A), select the attribute with the largest weighted information gain as the partition attribute A of the root node max .

[0106] For each non-leaf node N of the decision tree, the non-leaf node N corresponds to the data set D N , category label set Y N. Check whether the termination conditions are met (samples all belong to the same category, the number of samples is less than the threshold, the information gain is less than the threshold, and the weight variance is less than the threshold). If not, calculate the weighted information gain of each attribute and select the attribute with the maximum gain as the partition attribute. Divide the data into subsets based on the attribute values ​​and create child nodes for each subset.

[0107] Specifically, if D N The samples in all belong to the same category, namely D N The sample is less than the node sample quantity threshold |D N |<β, the number of samples is less than the information gain threshold IG w (D N ,A)<α, or the variance of the sample weight in the node is less than the weight variance threshold Var w (D N )<∈, then mark the node as a leaf node, and its category is D N The category with the largest number of samples.

[0108] For each sample, the entropy value of its predicted probability distribution can be used to measure the uncertainty of the weight, where the entropy value H is calculated as follows:

[0109]

[0110] Among them, p i is the predicted probability that the sample belongs to the i-th category.

[0111] Each time a new sample is processed, its weight is adaptively adjusted according to the impact of the new sample on the model.

[0112] Otherwise, calculate D N The weighted information gain IG of each attribute A in w (D N ,A), select the attribute with the maximum weighted information gain as the partition attribute of the node.

[0113] According to A max The value of D N Divide into multiple subsets D Nv (v is A max ), create a child node for each subset and assign the corresponding samples to the corresponding child nodes.

[0114] Then, each non-leaf node is evaluated to determine whether it needs to be pruned. Specifically, starting from the leaf node, each non-leaf node is traversed upwards, and each leaf node is calculated to determine whether it needs to be pruned.

[0115] For each sample x in the test dataset T j , starting from the root node of the decision tree, according to the sample Fj The feature vectors of are decided in the decision tree in turn until they reach the leaf node, and the category of the leaf node is the sample x j The predicted category.

[0116] In an optional embodiment, starting from the leaf node of the decision tree, pruning evaluation is performed on each non-leaf node upward in sequence, including the following steps:

[0117] For a non-leaf node N, calculate the original subtree T corresponding to the non-leaf node N N The weighted classification error Error on the training dataset w (T N );

[0118] Mark the non-leaf node N as a leaf node N, and calculate the weighted classification error Error of the subtree corresponding to the leaf node N w (N);

[0119] If Error w (N)≤Error w (T N ), then the non-leaf node N is pruned.

[0120] It should be noted that for each non-leaf node N, let its original subtree be T N , calculate T N Weighted classification error on the training dataset Error w (T N ).

[0121] Calculated by cross-validation and other methods, for each sample x i , whose contribution to the classification error is in is the sample x i Actual category y i and predicted category The error cost function.

[0122] Mark node N as a leaf node (i.e. cut off all its child nodes) and calculate the weighted classification error Error at this time w (N).

[0123] If Error w (N)≤Error w (T N ), then prune the non-leaf node N, that is, prune all the child nodes of the non-leaf node N and turn it into a leaf node, category D N is the category with the largest number of samples; otherwise, retain node N and its subtree.

[0124] In an optional embodiment, the integrated prediction of the PRPD ultrasonic feature and the EPRD ultrasonic feature is performed using a rotating forest algorithm, comprising the following steps:

[0125] Establishing a data set with the PRPD ultrasound features and the EPRD ultrasound features;

[0126] Establishing K decision trees, determining parameters of the decision trees based on the data set, and training the decision trees based on the parameters;

[0127] The trained decision tree is used to predict the PRPD ultrasonic features and the EPRD ultrasonic features to obtain K category results, and a final predicted category is determined from the K category results.

[0128] It should be noted that the datasets established with PRPD ultrasonic features and EPRD ultrasonic features provide training data for the rotation forest algorithm, where the dataset includes n samples and m features, and each sample includes a feature vector x i and the class label y i .

[0129] Establish K decision trees and the size of the feature subset used by each decision tree. In this embodiment, the feature subset size is set to k (k <m)。

[0130] The parameters of the decision tree are determined based on the data set, including the selection of feature subsets and feature rotation. The decision tree is trained using the parameters after feature rotation. Feature rotation can randomly change the direction of the feature space, which helps the algorithm to be more robust to noise and outliers. Through rotation, the distribution of noise in the feature space becomes more random, reducing their impact on the decision tree training process. By training multiple decision trees in different feature subspaces, the rotation forest algorithm can capture the characteristics of the data in different directions, thereby improving the generalization ability of the model on unseen data. Data after feature rotation may reveal new structures in the data that may not be easily detected in the original feature space. By training the decision tree in the rotated feature space, these structures can be captured more accurately, thereby improving the accuracy of classification.

[0131] New sample x new Each decision tree predicts K categories, and the category with the most votes is voted as the final prediction category. If there are multiple categories with the same number of votes, one of them can be randomly selected, or other strategies (such as category probability averaging in random forest) can be used to determine the final category.

[0132] In an optional embodiment, determining the parameters of the decision tree based on the data set, and training the decision tree based on the parameters comprises the following steps:

[0133] The data set includes n samples and m features;

[0134] For each decision tree t, k features are randomly selected from the m features to form a feature subset F t ;

[0135] For the feature subset F t Perform feature rotation and use the feature subset F after feature rotation t Train the decision tree.

[0136] It should be noted that for each decision tree t (t = 1, 2, ..., K), k features are randomly selected from the m features to form a feature subset F t ; Establish relevant sample subset D from the data of decision tree t t , extract F t And the relevant sample subset D t and standardized.

[0137] Calculate the relevant sample subset D t The covariance matrix C t , decomposed into the eigenvector matrix V t and the eigenvalue diagonal matrix Λ t , use V t Rotate sample x i,t Get x i ' ,t , complete the feature rotation, and use the rotated data to train the decision tree T t .

[0138] In an optional embodiment, a Bayesian fusion algorithm based on multi-stage fusion performs decision fusion on the classification prediction results and the integrated prediction results, including the following steps:

[0139] Determine the prior probabilities of the TPRD ultrasonic feature, the PRPD ultrasonic feature, and the EPRD ultrasonic feature, and calculate the conditional probabilities of the classifier in the improved decision tree algorithm and the classifier in the rotation forest algorithm respectively;

[0140] Perform preliminary fusion based on the prior probability to obtain a preliminary fusion probability distribution;

[0141] The predicted probability of the classifier for the sample is calculated according to the conditional probability, and the preliminary fusion probability distribution is secondary fused using the Bayesian fusion algorithm based on the predicted probability to obtain the final fusion probability distribution.

[0142] It should be noted that the prior probability P(C i ), where C i represents the i-th category.

[0143] For a classifier of the improved decision tree algorithm and a specific input sample x, assume that its prediction result is category C j , according to the performance of the classifier on the training set P(x|C j ), calculate the conditional probability, which can be obtained by counting the probability that the classifier belongs to category C in the training set j The sample prediction is approximately estimated as the frequency of the current feature value.

[0144] The same operation is performed on a classifier of the rotation forest, and its prediction for sample x is category C j The conditional probability P(x|C j ).

[0145] In this embodiment, preliminary fusion and secondary fusion are performed. In the preliminary fusion, the prediction results of multiple classifiers are combined to improve the accuracy of prediction; in the secondary fusion, the preliminary fusion results are further combined with the results of other classifiers to improve the overall prediction performance.

[0146] Among them, in the preliminary fusion stage, the classifiers with higher accuracy in the training set are selected for preliminary fusion. Using the weighted average method, the probability distribution after preliminary fusion is calculated according to the weight and predicted probability distribution of each classifier. Specifically, the TPRD ultrasound features, PPRD ultrasound features and EPRD ultrasound features with higher accuracy in the training set are selected for preliminary fusion. Each ultrasound feature corresponds to a classifier. Three classifiers C1, C2, and C3 are arbitrarily selected. The predicted probability distribution of each classifier for the sample is P1(x), P2(x), and P3(x). Using the weighted average method, first determine the weights w1, w2, and w3 of each classifier. The probability distribution after preliminary fusion is In this embodiment, the weight of each classifier can be determined by those skilled in the art according to the performance index of the classifier.

[0147] In the secondary fusion stage, for a new input sample x i , and perform fusion calculations according to Bayes’ theorem. This involves combining the probability distribution after the initial fusion with the predicted probability distribution of the classifiers that did not participate in the initial fusion, and calculating the posterior probability after fusion.

[0148] For each category C i , the probability distribution after preliminary fusion is P initial (x), and there is a classifier C n+1 ,C n-2 ,…,C n+m , whose predicted probability distribution is P 3+1 (x),P 3+2 (x),…,P 3+m(x), the posterior probability after fusion is:

[0149]

[0150] In the above formula, P(C i |x) means that under the condition that the data is x, the category is C i This is the posterior probability that the partial discharge belongs to category C after observing the specific partial discharge signal feature x. i The probability estimate of P(x|C i ) indicates that in the known category C i The probability of data x appearing under the condition of . In cable partial discharge monitoring, it can represent the probability of observing a specific partial discharge signal feature when it is known that the partial discharge belongs to a specific type.

[0151] Furthermore, P(x|C i ) can be combined with the improved decision tree algorithm and the rotation forest algorithm to predict the sample x as category C. i The conditional probability is calculated as follows:

[0152]

[0153] Among them, w j and u k They represent the weights of each classifier in the improved decision tree algorithm and the rotating forest algorithm respectively; α and β represent the algorithm weight coefficients of the improved decision tree algorithm and the rotating forest algorithm.

[0154] In this embodiment, the weight of each classifier can be determined according to the performance of the classifier on the validation set; α+β=1.

[0155] According to the Bayesian fusion formula, the final fusion probability distribution is:

[0156] Furthermore, at different fusion stages, the weights can be dynamically adjusted according to the performance of the classifiers. In the initial fusion stage, if the accuracy of a classifier decreases, its weight can be reduced; in the secondary fusion stage, if a classifier is highly consistent with the initial fusion result, its weight can be increased.

[0157] Compare the posterior probabilities of each category after fusion, and select the category with the largest posterior probability as the final fusion result.

[0158] Example 2

[0159] Figure 2 A schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present invention is shown in FIG. Figure 2As shown, the electronic device includes a processor 21, a memory 22, an input device 23 and an output device 24; the number of processors 21 in the computer device can be one or more. Figure 2 A processor 21 is taken as an example; the processor 21, the memory 22, the input device 23 and the output device 24 in the electronic device can be connected by a bus or other means. Figure 2 The example of connecting through bus is taken in the following.

[0160] The memory 22 is a computer-readable storage medium that can be used to store software programs, computer executable programs and modules. The processor 21 executes various functional applications and data processing of the electronic device by running the software programs, instructions and modules stored in the memory 22, that is, the cable partial discharge identification method based on the improved decision tree-rotation forest of Example 1 is implemented.

[0161] The memory 22 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system and at least one application required for a function; the data storage area may store data created according to the use of the terminal, etc. In addition, the memory 22 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some instances, the memory 22 may further include a memory remotely arranged relative to the processor 21, and these remote memories may be connected to the electronic device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0162] The input device 23 can be used to receive the ID and password input by the user, etc. The output device 24 is used to output the network configuration page.

[0163] Example 3

[0164] Embodiment 3 of the present invention further provides a computer-readable storage medium, wherein the computer-executable instructions, when executed by a computer processor, are used to implement the cable partial discharge identification method based on the improved decision tree-rotation forest as provided in Embodiment 1.

[0165] An embodiment of the present invention provides a storage medium containing computer executable instructions, and its computer executable instructions are not limited to the method operations provided in Example 1, and can also execute related operations in the cable partial discharge identification method based on improved decision tree-rotation forest provided in any embodiment of the present invention.

[0166] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A cable partial discharge identification method based on improved decision tree-rotation forest, characterized in that: The steps include: Step S1, obtaining a cable partial discharge monitoring signal, performing ultrasonic signal feature extraction on the cable partial discharge monitoring signal, and obtaining a TPRD ultrasonic feature, a PRPD ultrasonic feature, and an EPRD ultrasonic feature; Step S2, using an improved decision tree algorithm based on adaptive weighted information to adjust gain to perform classification prediction on the TPRD ultrasonic features, and using a rotation forest algorithm to perform integrated prediction on the PRPD ultrasonic features and the EPRD ultrasonic features; Step S3: A Bayesian fusion algorithm based on multi-stage fusion performs decision fusion on the results of classification prediction and the results of integrated prediction, and determines the type of cable partial discharge according to the result of decision fusion.

2. The cable partial discharge identification method based on improved decision tree-rotation forest according to claim 1 is characterized in that: The improved decision tree algorithm based on adaptive weighted information gain adjustment is used to classify and predict the TPRD ultrasound features, including the following steps: Establishing a training data set using the TPRD ultrasound features; Calculating the weighted information entropy of the training data set, and constructing a decision tree with the weighted information entropy; Starting from the leaf nodes of the decision tree, pruning evaluation is performed on each non-leaf node in turn upward, and the decision tree is pruned according to the results of the pruning evaluation; A test data set is established with the TPRD ultrasound features, and the pruned decision tree is used to perform classification prediction on the test data set to obtain a predicted category.

3. The cable partial discharge identification method based on improved decision tree-rotation forest according to claim 2 is characterized in that: Starting from the leaf node of the decision tree, pruning evaluation is performed on each non-leaf node upward in sequence, including the following steps: For a non-leaf node N, calculate the original subtree T corresponding to the non-leaf node N N The weighted classification error Error on the training dataset w (T N ); Mark the non-leaf node N as a leaf node N, and calculate the weighted classification error Error of the subtree corresponding to the leaf node N w (N); If Error w (N)≤Error w (T N ), then the non-leaf node N is pruned.

4. The cable partial discharge identification method based on improved decision tree-rotation forest according to claim 1 is characterized in that: The integrated prediction of the PRPD ultrasonic features and the EPRD ultrasonic features is performed using a rotation forest algorithm, comprising the following steps: Establishing a data set with the PRPD ultrasound features and the EPRD ultrasound features; Establishing K decision trees, determining parameters of the decision trees based on the data set, and training the decision trees based on the parameters; The trained decision tree is used to predict the PRPD ultrasonic features and the EPRD ultrasonic features to obtain K category results, and a final prediction category is determined from the K category results.

5. The cable partial discharge identification method based on improved decision tree-rotation forest according to claim 4 is characterized in that: Determining the parameters of the decision tree based on the data set, and training the decision tree based on the parameters, comprises the following steps: The data set includes n samples and m features; For each decision tree t, k features are randomly selected from the m features to form a feature subset F t ; For the feature subset F t Perform feature rotation and use the feature subset F after feature rotation t Train the decision tree.

6. The cable partial discharge identification method based on improved decision tree-rotation forest according to claim 1 is characterized in that: The Bayesian fusion algorithm based on multi-stage fusion makes decision fusion on the results of classification prediction and ensemble prediction, including the following steps: Determine the prior probabilities of the TPRD ultrasonic feature, the PRPD ultrasonic feature, and the EPRD ultrasonic feature, and calculate the conditional probabilities of the classifier in the improved decision tree algorithm and the classifier in the rotation forest algorithm respectively; Perform preliminary fusion based on the prior probability to obtain a preliminary fusion probability distribution; The predicted probability of the classifier for the sample is calculated according to the conditional probability, and the preliminary fusion probability distribution is secondary fused using the Bayesian fusion algorithm based on the predicted probability to obtain the final fusion probability distribution.

7. The cable partial discharge identification method based on improved decision tree-rotation forest according to claim 1 is characterized in that: Extracting ultrasonic signal features from the cable partial discharge monitoring signal to obtain PRPD ultrasonic features includes the following steps: Extracting a PRPD ultrasonic signal from the cable partial discharge monitoring signal; Dividing one cycle of the PRPD ultrasonic signal into a plurality of phase intervals; Calculating the third-order cumulant and the fourth-order cumulant of the PRPD ultrasonic signal in each phase interval respectively; Based on the numerical values, increasing and decreasing trends and fluctuations of the third-order cumulants and the fourth-order cumulants in different phase intervals, PRPD ultrasonic features are extracted from the PRPD ultrasonic signal.

8. The cable partial discharge identification method based on improved decision tree-rotation forest according to claim 1 is characterized in that: Extracting ultrasonic signal features from the cable partial discharge monitoring signal to obtain EPRD ultrasonic features comprises the following steps: Extracting EPRD ultrasonic signal from the cable partial discharge monitoring signal; The fuzzy function of the EPRD ultrasonic signal is calculated by using a numerical integration method; Constructing a rectangular kernel function, bringing the fuzzy function and the rectangular kernel function into the Cohen-like time-frequency distribution formula for integral calculation, and obtaining the Cohen-like time-frequency distribution; Signal classification is performed on the Cohen-like time-frequency distribution to obtain EPRD ultrasound features.

9. An electronic device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the cable partial discharge identification method based on the improved decision tree-rotation forest as claimed in any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the cable partial discharge identification method based on the improved decision tree-rotation forest as described in any one of claims 1 to 8 is implemented.

Images