Method and system for identifying internal defects of suspension type porcelain insulator and pillar type porcelain sleeve

Abstract

The invention relates to the technical field of power electronics, and discloses a suspension type porcelain insulator and pillar type porcelain bushing internal defect identification method and system, and the method comprises the steps: collecting an ultrasonic detection signal of a suspension type porcelain insulator or pillar type porcelain bushing through an ultrasonic probe, and generating a signal sample set; carrying out filtering processing on the ultrasonic detection signals in the sample set by adopting a wavelet transform denoising algorithm; eMD empirical mode decomposition is carried out on the preprocessed ultrasonic detection signals, and feature parameters in the signals are extracted to form a feature matrix; constructing an integrated learning model through a weighted combination multilayer perceptron MLP and a random forest RF algorithm, and inputting the feature matrix into the integrated learning model for analysis; and identifying the internal defect type of the suspension type porcelain insulator or the pillar type porcelain sleeve according to an analysis result output by the integrated learning model. Effective identification of ultrasonic detection signals of internal defects of porcelain insulators or pillar type porcelain sleeves is realized through empirical mode decomposition and ensemble learning.

Description

Technical Field

[0001] The present invention relates to the technical field of power electronics, and in particular to a method and system for identifying internal defects of a suspension porcelain insulator and a pillar-type porcelain bushing. Background Art

[0002] Suspension porcelain insulators and support-type porcelain bushings are widely used in power systems due to their excellent mechanical strength, high-temperature resistance, and good electrical insulation properties, playing an important role in insulation and support. However, as they age, some suspension porcelain insulators and support-type porcelain bushings may develop internal defects. If these defects are not discovered in a timely manner, they may lead to serious power failures and safety accidents. Therefore, during operation and maintenance, it is urgent to adopt advanced non-destructive testing technologies to effectively detect internal defects in suspension porcelain insulators and support-type porcelain bushings, thereby preventing power system accidents and ensuring the safe operation of power facilities. Ultrasonic testing technology is a convenient and efficient non-destructive testing technology that can not only detect the internal structural information of the workpiece under test, but also has a fast response time, safe use, and a long service life. It also has excellent electrical insulation and anti-electromagnetic interference capabilities, making it particularly suitable for fault detection in power system equipment.

[0003] While ultrasonic testing technology has been studied for internal defect detection in suspension porcelain insulators and post-type porcelain bushings, it has only successfully demonstrated the feasibility of ultrasonic testing for internal defects in composite insulators. Most studies rely solely on subjective judgment of ultrasonic testing signals from defective insulators, resulting in misjudgments and difficulty distinguishing. However, methods for intelligently identifying defect signals and processing ultrasonic testing signals for suspension porcelain insulators and post-type porcelain bushings are lacking. Therefore, ultrasonic testing for internal defects in suspension porcelain insulators and post-type porcelain bushings, along with intelligent identification of defect signals, are pressing technical challenges in this field. Summary of the Invention

[0004] The purpose of the present invention is to address the shortcomings of the above-mentioned prior art and to provide a method and system for identifying internal defects of suspension porcelain insulators and pillar-type porcelain bushings. The method adopts empirical mode decomposition and an integrated learning method based on multi-layer perceptron (MLP) and random forest algorithm (RF) to effectively identify the ultrasonic detection signals of internal defects of porcelain insulators or pillar-type porcelain bushings (no defects, pores, cracks).

[0005] On the one hand, a method for identifying internal defects of suspension porcelain insulators and pillar-type porcelain bushings is provided, comprising the following steps: S1: The ultrasonic detection signal of the suspension porcelain insulator or the pillar porcelain bushing is collected by an ultrasonic probe to generate a signal sample set, and the ultrasonic detection signal in the sample set is filtered using a wavelet transform denoising algorithm; S2: Perform EMD empirical mode decomposition on the pre-processed ultrasonic detection signal, and then extract the characteristic parameters in the signal to form a feature matrix; S3: inputting the feature matrix into an ensemble learning model constructed by weighted combination of multilayer perceptron (MLP) and random forest (RF) algorithm for analysis; S4: Identify the internal defect type of the suspension porcelain insulator or the pillar-type porcelain bushing according to the analysis result output by the integrated learning model.

[0006] Furthermore, in step S1, the wavelet transform denoising algorithm includes: The signal is decomposed into high-frequency noise components and low-frequency effective components through wavelet decomposition. The high-frequency noise components are suppressed based on a preset threshold. The denoised signal is then reconstructed by wavelet to obtain a filtered ultrasonic detection signal. The formula of the wavelet transform denoising algorithm is as follows: in, , represents the basic wavelet function, is the translation of the wavelet function, is the scale expansion of the wavelet function, and the basic wavelet function satisfies .

[0007] Furthermore, in step S2, performing EMD empirical mode decomposition on the preprocessed ultrasonic detection signal includes: S21: First, the preprocessed original signal is used as the signal to be decomposed , set the number of intrinsic mode function IMF decompositions , ; S22: Calculating the signal to be decomposed The maximum and minimum values ​​of the spectrum are fitted into a smooth curve by the cubic spline interpolation method to generate the upper and lower envelope spectra, and the mean of the upper and lower envelope spectra is calculated. ; S23: Calculate the difference signal between the signal to be decomposed and the mean value , determine the difference signal Whether the criterion of the intrinsic mode function IMF is met, If satisfied, extract the IMF component , , and the signal to be decomposed With the IMF component The difference between ; Otherwise, the difference signal As the new signal to be decomposed, ; S24: Repeat steps S22 to S23 until all IMF components are extracted and the residual term remains Approaches 0, then the signal to be decomposed It is expressed as follows: The decomposed IMF components are used as characteristic parameters to construct the characteristic matrix.

[0008] Furthermore, in step S3, the ensemble learning model constructed by weighted combination of multilayer perceptron MLP and random forest RF algorithm includes: S31: Split the feature matrix extracted by EMD into a training set, a validation set, and a test set in proportion, and input the training set into training a multi-layer perceptron MLP model and a random forest RF model respectively; S32: Dynamically assign weights of the multilayer perceptron MLP and the random forest RF based on the performance of each model applied to the validation set. If the difference in accuracy between the two on the validation set is less than a preset threshold, the ensemble learning model adopts weighted combination output; If the difference in accuracy between the two on the validation set is greater than the preset threshold, the single model with better performance is selected as the output of the ensemble learning model; S33: Input the test set into the ensemble learning model, analyze the output classification results, evaluate the performance indicators of the model in the defect recognition task, and appropriately adjust the hyperparameters of the multilayer perceptron MLP and random forest RF according to the indicators to further improve the model recognition effect.

[0009] Preferably, in step S31, the training of the multi-layer perceptron (MLP) model further comprises: S311: Normalize the input training set data to ensure that the value ranges of different features are similar to speed up the model convergence; S312: Initialize weights and biases, calculate predicted values ​​through forward propagation, and perform backpropagation optimization based on the cross-entropy loss function, using stochastic gradient descent to update weights and biases; S313: Continuously repeat the forward propagation and backpropagation processes until the loss function converges or the set number of iterations is reached, input new data for forward propagation, and output the predicted defect results.

[0010] More preferably, in step S31, the training of the random forest RF model further includes: S314: Generate multiple training subsets by sampling with replacement from the training set using the Bagging method; S315: Build a decision tree on each subset and randomly select feature subspaces for node partitioning; S316: Multiple decision trees are integrated through a voting mechanism to obtain the classification decision of the random forest.

[0011] Furthermore, step S32 further includes: If the performance difference between the multilayer perceptron MLP model and the random forest RF model is less than a preset threshold, the weights of the multilayer perceptron MLP and the random forest RF in the ensemble learning model are determined in a weighted combination manner. The weight calculation formula is as follows: in, and are the weights of the multilayer perceptron MLP model and the random forest RF model, and These are the accuracies of the multi-layer perceptron MLP model and the random forest RF model on the validation set; If the performance difference between the multilayer perceptron (MLP) model and the random forest (RF) model is greater than a preset threshold, the model with higher accuracy on the validation set is selected as the final ensemble learning model.

[0012] Furthermore, step S33 further includes: The test set is input into the ensemble learning model, and the calculated weights are used. and For each sample, the ensemble learning model makes a weighted decision based on the output of the multi-layer perceptron MLP and random forest RF to obtain the final defect classification label. If a weighted combination is selected, the final prediction output is .

[0013] On the other hand, a system for identifying internal defects of suspension porcelain insulators and support porcelain bushings is provided, comprising: The signal acquisition and preprocessing module is used to collect ultrasonic detection signals of suspension porcelain insulators or pillar-type porcelain bushings through ultrasonic probes to generate signal sample sets, and filter the ultrasonic detection signals in the sample sets using a wavelet transform denoising algorithm; The feature extraction module is used to perform EMD empirical mode decomposition on the pre-processed ultrasonic detection signal and then extract the characteristic parameters in the signal to form a feature matrix; An integrated learning module is used to construct an integrated learning model by weighted combination of multi-layer perceptron MLP and random forest RF algorithm, and input the feature matrix into the integrated learning model for analysis; A classification and identification module is used to identify the internal defect type of the suspension porcelain insulator or the pillar-type porcelain bushing according to the analysis result output by the integrated learning model.

[0014] At the same time, a computer-readable storage medium is provided, on which a computer program is stored, characterized in that the program includes a boot program and an application program, and when executed by a processor, implements the internal defect identification method of the suspension porcelain insulator and the support porcelain bushing described in any one of the above items.

[0015] In addition, an electronic device is provided, comprising: one or more processors; a storage device for storing one or more programs, which, when executed by the one or more processors, enables the one or more processors to implement the method for identifying internal defects of suspension porcelain insulators and support porcelain bushings as described in any one of the above items.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This paper solves the problems of poor generalization ability and high misclassification rate of traditional single models through integrated learning of multi-layer perceptron (MLP) and random forest (RF), combining the nonlinear feature extraction capability of MLP and the robust classification performance of RF. The present invention uses EMD to decompose non-stationary and nonlinear ultrasonic signals, extract multi-scale intrinsic mode functions (IMFs), accurately characterize defect characteristics, overcome the limitations of traditional time-frequency analysis methods that are poorly adaptable to complex signals, and provide highly discriminative input data for the model. The present invention suppresses high-frequency noise and reconstructs effective signals through wavelet decomposition, significantly reducing the impact of environmental noise and equipment interference on detection results, improving signal quality, reducing the risk of false detection and missed detection, and ensuring the reliability of subsequent analysis; The present invention dynamically adjusts weights based on the performance differences between MLP and RF on the validation set. This mechanism improves the system's adaptability to different data distributions and avoids performance fluctuations caused by fixed weights. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 This is a flow chart of a method for identifying internal defects of a suspension porcelain insulator and a pillar-type porcelain bushing according to the present invention; Figure 2 This is an example diagram of the wavelet transform denoising effect of a defect-free ultrasonic detection signal of the present invention; Figure 3 This is a flow chart of an empirical mode decomposition (EMD) method of the present invention; Figure 4This is a schematic diagram of the basic framework of a multi-layer perceptron MLP of the present invention; Figure 5 A schematic diagram of the flow of a random forest RF algorithm of the present invention; Figure 6 A schematic diagram of the defect recognition effect of an empirical mode decomposition-multi-layer perceptron MLP-random forest RF method of the present invention; Figure 7 This is a schematic diagram comparing the defect recognition effects of various methods of the present invention; Figure 8 This is a structural block diagram of an internal defect identification system for suspension porcelain insulators and support porcelain bushings of the present invention. DETAILED DESCRIPTION

[0018] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0019] This paper proposes a method for identifying internal defects in composite insulators based on empirical mode decomposition (EMD), a multilayer perceptron (MLP), and a random forest algorithm. First, an ultrasonic probe suitable for inspecting composite insulator materials is selected. A detection area is selected based on the composite insulator's specifications and the locations prone to internal defects. An ultrasonic detector is used to acquire a sample set of ultrasonic inspection signals. The ultrasonic inspection signals collected by the hardware system are then filtered using a wavelet transform denoising algorithm. EMD decomposition is then performed on the preprocessed ultrasonic inspection signals to extract their characteristic parameters. An ensemble learning method combining EMD with a multilayer perceptron (MLP) and a random forest algorithm (RF) is used to validate the ultrasonic inspection signals for internal defects (including absence of defects, pores, and cracks) in porcelain insulators or post-type porcelain bushings. Finally, a MATLAB GUI module is used for model training and visualization of the defect identification results, enabling intelligent identification of internal defects in composite insulators.

[0020] The specific implementation of the present invention is described below with reference to the accompanying drawings and embodiments.

[0021] Example 1 See also Figure 1 The embodiment provides a method for identifying internal defects of suspension porcelain insulators and support porcelain bushings. The technical solution includes the following steps: S1: The ultrasonic detection signal of the suspension porcelain insulator or the pillar porcelain bushing is collected by an ultrasonic probe to generate a signal sample set, and the ultrasonic detection signal in the sample set is filtered using a wavelet transform denoising algorithm; S2: Perform EMD empirical mode decomposition on the pre-processed ultrasonic detection signal, and then extract the characteristic parameters in the signal to form a feature matrix; S3: inputting the feature matrix into an ensemble learning model constructed by weighted combination of multilayer perceptron (MLP) and random forest (RF) algorithm for analysis; S4: Identify the internal defect type of the suspension porcelain insulator or the pillar-type porcelain bushing according to the analysis result output by the integrated learning model.

[0022] Among them, the ultrasonic probe selects a small-diameter single-crystal straight probe that can transmit and receive ultrasonic waves independently, which is suitable for detecting composite insulator materials. At the same time, we select the detection area based on the specifications of the composite insulator and the location where internal defects are likely to occur. Then, we apply coupling agent to the surface of the ultrasonic probe, and make close contact with the outer surface of the porcelain insulator or the pillar-type porcelain bushing, so that the ultrasonic wave can be transmitted to the inside of the insulator without loss. On this basis, we use an ultrasonic detector to obtain a set of ultrasonic detection signal samples. In this embodiment, field tests were conducted on simulated ultrasonic detection data of suspended porcelain bushings and actual suspended insulators. Before testing, the parameters of the ultrasonic detection instrument were calibrated. A 2.5MHz-5MHz ultrasonic independent transmitter and receiver, a small-diameter single-crystal straight probe was selected, and a gel-type coupling agent was fully applied to eliminate air interference. This system is used to identify ultrasonic detection signals of internal defects of suspended porcelain insulators and pillar-type porcelain bushings.

[0023] The ultrasonic detection signal will be affected by the background noise generated by the material center and the electrical noise in the detection instrument, which will affect the recognition results. Therefore, we use the wavelet transform denoising algorithm described in step S1 to filter the ultrasonic detection signal and verify the denoising effect, including: The signal is decomposed into high-frequency noise components and low-frequency effective components through wavelet decomposition, the high-frequency noise components are suppressed based on a preset threshold, and then the denoised signal is reconstructed by wavelet to obtain a filtered ultrasonic detection signal.

[0024] After wavelet transform, the ultrasonic detection signal will be decomposed into a superposition of a series of basic wavelet functions. The formula of the wavelet transform denoising algorithm is as follows: in, , represents the basic wavelet function, is the translation of the wavelet function, is the scale expansion of the wavelet function, and the basic wavelet function satisfies .

[0025] Based on the wavelet transform denoising algorithm, the ultrasonic detection signal filtering processing effect of internal defects of suspension porcelain insulators and pillar porcelain bushings is as follows: Figure 2 shown.

[0026] Furthermore, the ultrasonic detection signal is a non-stationary, nonlinear signal that can be considered to be composed of a finite number of intrinsic mode functions (IMFs). The idea behind the empirical mode decomposition (EMD) method is to decompose any signal into several IMFs and then perform envelope processing, i.e., Hilbert transform, on each IMF component. Each IMF component after decomposition satisfies two criteria: 1) At any time, the mean of the upper and lower envelope spectra composed of the maximum and minimum values ​​of the IMF is 0; 2) In the entire time series, the difference between the number of maximum and minimum points does not exceed 1.

[0027] Specifically, the pre-processed ultrasonic detection signal is subjected to EMD empirical mode decomposition as described in step S2, such as Figure 3 As shown, further comprising: S21: First, the preprocessed original signal is used as the signal to be decomposed , set the number of intrinsic mode function IMF decompositions , ; S22: Calculating the signal to be decomposed The maximum and minimum values ​​are fitted into a smooth curve by the cubic spline interpolation method to generate the upper and lower envelope spectra and , and calculate the mean of the upper and lower envelope spectra ; S23: Calculate the difference signal between the signal to be decomposed and the mean value , determine the difference signal Whether the criterion of the intrinsic mode function IMF is met, If satisfied, extract the IMF component , , and the signal to be decomposed With the IMF component The difference between ; Otherwise, the difference signal As the new signal to be decomposed, ; S24: Repeat steps S22 to S23 until all IMF components are extracted and the residual term remains Approaches 0, then the signal to be decomposed It is expressed as follows: The decomposed IMF components are used as characteristic parameters to construct the characteristic matrix.

[0028] In step S3, an ensemble learning model is constructed by weightedly combining a multilayer perceptron (MLP) and a random forest (RF) algorithm. Ensemble learning combines multiple weak learners, using sample weighting and learner weighting to achieve a strong learner with significantly superior generalization performance compared to a single learner. Ensemble learning requires that each learner possess a certain level of accuracy and variability. The basic idea is to construct different base learners by applying different sampling or preprocessing methods to the training data and then combine them into an ensemble model to improve overall prediction performance and generalization ability.

[0029] This method aims to efficiently and accurately identify defects in ultrasonic signals through an integrated learning model of MLP (Multi-Layer Perceptron) and RF (Random Forest). This method combines the feature extraction capabilities of MLP and the classification performance of RF, and achieves the optimal combination of the two models through a weighted strategy, thereby improving the accuracy and robustness of ultrasonic signal defect recognition. Specifically, it includes: S31: Split the feature matrix extracted by EMD into a training set, a validation set, and a test set in proportion, and input the training set into training a multi-layer perceptron MLP model and a random forest RF model respectively; S32: Dynamically assign weights of the multilayer perceptron MLP and the random forest RF based on the performance of each model applied to the validation set. If the difference in accuracy between the two on the validation set is less than a preset threshold, the ensemble learning model adopts weighted combination output; If the difference in accuracy between the two on the validation set is greater than the preset threshold, the single model with better performance is selected as the output of the ensemble learning model; S33: Input the test set into the ensemble learning model, analyze the output classification results, evaluate the performance indicators of the model in the defect recognition task, and appropriately adjust the hyperparameters of the multilayer perceptron MLP and random forest RF according to the indicators to further improve the model recognition effect.

[0030] This paper aims to efficiently and accurately identify defects in ultrasonic signals through an integrated learning model combining an MLP (Multi-Layer Perceptron) and a RF (Random Forest). This method combines the feature extraction capabilities of the MLP with the classification performance of the RF, achieving an optimal combination of the two models through a weighted strategy, thereby improving the accuracy and robustness of ultrasonic signal defect recognition.

[0031] Among them, training the multi-layer perceptron MLP model further includes: S311: Normalize the input training set data to ensure that the value ranges of different features are similar to speed up the model convergence; S312: Initialize weights and biases, calculate predicted values ​​through forward propagation, and perform backpropagation optimization based on the cross-entropy loss function, using stochastic gradient descent to update weights and biases; S313: Continuously repeat the forward propagation and backpropagation processes until the loss function converges or the set number of iterations is reached, input new data for forward propagation, and output the predicted defect results.

[0032] Specifically, MLP can also be called a multi-layer perceptron, and its basic framework is as follows Figure 4 shown.

[0033] Multilayer Perceptron is a multilayer neural network, which has neurons. Assume there are L layers of network. 、 To connect The output of the layer network, f is the mapping relationship between input and output, The output value predicted by the neural network is the predicted defect type, is the real defect type, The feature vector extracted from the ultrasonic detection echo signal.

[0034] The forward propagation formula is: The objective function is: Back propagation uses the gradient descent method, assuming The objective function is The gradient, The objective function is gradient. The objective function is The derivative of the layer network activation function is expressed as follows: Update the gradient using gradient descent: In this embodiment, the process of MLP training and prediction is as follows: Step 1: Data Preparation Prepare the training dataset, including the input feature matrix and the corresponding target output ,Then the input features are normalized to ensure that the value ranges of different ,features are similar to speed up the convergence of the model.

[0035] Step 2: Initialize weights and biases Randomly initialize the weights and biases of each connection in the network. Usually the weights are initialized to small random values ​​to avoid gradient explosion or gradient vanishing.

[0036] Step 3: Forward Propagation The features extracted from the ultrasonic signal are input into the input layer of the network as training samples; Each hidden layer performs weighted summation based on the output of the previous layer and performs nonlinear transformation through the activation function; The output of the hidden layer passes through the last layer to obtain the predicted value of the network.

[0037] Step 4: Calculate the loss function The network output is compared with the target output y and the loss function is calculated to measure the prediction error.

[0038] Step 5: Backpropagation Use the chain rule to calculate the gradient of the loss function with respect to each weight; Update the weights and biases in the network using gradient descent or its variants.

[0039] Step 6: Iterative training The forward propagation and back propagation processes are repeated continuously, and multiple rounds of iterative training are carried out until the loss function converges or the set number of iterations is reached.

[0040] Step 7: Prediction and Application After training is completed, new data is input for forward propagation and the predicted defect results are output.

[0041] Training the random forest RF model includes: S314: Generate multiple training subsets by sampling with replacement from the training set using the Bagging method; S315: Build a decision tree on each subset and randomly select feature subspaces for node partitioning; S316: Multiple decision trees are integrated through a voting mechanism to obtain the classification decision of the random forest.

[0042] Specifically, random forest is defined as a tree with the same probability distribution of random vectors and the input vector Constructed decision tree The random forest uses the Bagging method to extract multiple sample data subsets from the training set of the data set with replacement, and then obtains the random subspace of the feature parameter group space as the parameter set to establish multiple decision trees. The classification results of the random forest are obtained through the construction, identification, and voting of the decision trees. The process is as follows: Figure 5 shown.

[0043] For a training set (in and are the kth training set data and its corresponding known category), we can give The training set consists of samples randomly selected with replacement , where each training set The number of samples is Due to different are independent of each other, Each sample in is also independent of each other, so a sample Maybe in the training set appears repeatedly in , or not at all. For a given classification algorithm , using any training set A classifier can be trained , so we can get the combined classifier .

[0044] For random forest, although the bagging combination classifier has defects, random forest can still use the bagging algorithm to resample the training data set without replacement to obtain multiple training subsets. .

[0045] Based on the trained Multi-Layer Perceptron (MLP) model and Random Forest (RF) model described above, we perform a weighted combination of the two models through ensemble learning. Ensemble learning combines multiple weak learners, weighting both samples and learners, to achieve a strong learner with significantly superior generalization performance compared to a single learner. Ensemble learning requires that each learner possess a certain level of accuracy and variability. The basic idea is to construct different base learners by applying different sampling or preprocessing methods to the training data, and then combine them into an ensemble model to improve overall prediction performance and generalization capabilities.

[0046] In this embodiment, the specific steps of the integrated learning model of MLP (Multi-layer Perceptron) and RF (Random Forest) are as follows: Step 1: Use the features extracted by EMD decomposition as the input of the recognition model; Step 2: Model training and hyperparameter selection (1) MLP model training: a) Input the extracted features into the MLP training model and train it by adjusting the hyperparameters (number of hidden layers, number of neurons, learning rate, etc.); b) MLP uses its internal multi-layer structure to extract deep features from the input data and learn complex nonlinear relationships; c) During the training process, the model parameters are continuously updated to minimize the error and gradually fit the data characteristics.

[0047] (2) RF model training: a) Input the same features into the RF model and train it with different hyperparameter combinations (such as the number of trees, maximum depth, etc.); b) RF classifies the data by integrating multiple decision trees and uses a voting mechanism to form the final classification decision.

[0048] Step 3: Model validation and optimal parameter selection a) Apply the trained MLP and RF models to the validation set and calculate the accuracy of each model on the validation set and ; b) If the current hyperparameter combination fails to achieve the best results, return to step 2 and reselect different hyperparameter combinations until the best performing MLP model on the validation set is found. and RF model .

[0049] Step 4: Model performance comparison and weight calculation (1) Performance difference judgment Set a threshold To measure the performance difference between MLP and RF models, that is Is it less than the threshold?

[0050] (2) Weight distribution strategy: a) If (That is, the performance difference between the MLP and RF models is within an acceptable range), the weights of MLP and RF are determined by weighted combination. The weight calculation formula is: in, and are the weights of the multilayer perceptron MLP model and the random forest RF model respectively.

[0051] b) If (i.e., the performance difference between the MLP and RF models is large), the model with better performance is selected as the final model classification model: Step 5: Test set prediction and defect identification (1) Combination model prediction: input the test set into the combination model, and calculate the weights and Make predictions (2) Classification decision: a) For each sample, the combined model makes a weighted decision based on the output of the MLP and RF to obtain the final defect classification label; b) If weighted combination is selected, the final prediction output is .

[0052] Step 6: Model Performance Evaluation (1) Analyze the classification results of the test set and evaluate the model's classification accuracy, confusion matrix and other indicators to determine the performance of the model in the defect recognition task; According to the test results, if the model performance does not meet the expected goals, you can appropriately adjust the hyperparameters of MLP and RF, or try to improve the feature extraction method to further improve the model recognition effect.

[0053] The internal defect recognition effect of suspension porcelain insulators and pillar porcelain bushings based on empirical mode decomposition method and multi-layer perception-random forest algorithm proposed in this invention is as follows: Figure 6 As shown, it can be found that the present invention can effectively identify ultrasonic detection signals of types such as umbrella sleeve air holes, cracks, interface debonding and no defects in suspension porcelain insulators and pillar porcelain bushings, with a total recognition rate of 95.3%.

[0054] In the prior art, there are some defect identification methods, including mathematical statistics analysis method, support vector machine method and decision tree method, among which, The advantage of the statistical analysis method is that the method is simple and the workload is small. The characteristics of the detection signal are closely related to the defect type, and the ultrasonic detection signal echo is used to identify defects based on these properties. Because it is difficult for computers to identify the physical and geometric characteristics of the signal, easy-to-identify mathematical and statistical features are used. In general, the mathematical statistics that describe the signal include: mean, standard deviation, correlation coefficient, covariance, etc. Therefore, when identifying the signal through a computer, the signal is first analyzed in the time domain or frequency domain to find the relevant digital features that characterize the essence of the signal. However, its reliability is poor, and it has high requirements for data integrity and accuracy. In addition, improper feature selection will seriously affect the scientific nature of the results. Therefore, this method has a large subjective influence and does not meet the needs of intelligent identification.

[0055] The basic idea of ​​the support vector machine method is to find an optimal classification hyperplane that maximizes the classification interval while distinguishing samples. Although it has a large feature space, it does not rely on the entire dataset and is free from local minima. SVMs are generally only used for binary classification problems. In practical applications, ultrasonic testing signals can produce many types of defects, and indirect methods are often used to identify multiple defect types. Indirect methods treat multi-classification problems as multiple binary classification problems, offering advantages in training speed and classification accuracy. However, their processing efficiency is insufficient when dealing with a large number of data samples, and there is no universal solution for nonlinear problems. Sometimes, it is difficult to find a suitable kernel function, which introduces new problems.

[0056] Decision tree classification is a classification algorithm based on a tree structure. A decision tree consists of a root node, internal nodes, and leaf nodes. In a decision tree, each internal node partitions the dataset according to a specific feature, with different branches corresponding to different values ​​of that feature. The branching relationships between layers represent different feature selection paths. Data samples are classified sequentially along the tree's branches from top to bottom, ultimately reaching the leaf nodes to obtain classification or prediction results. Decision trees are characterized by their simple structure and high interpretability. Each branch represents a specific feature selection criterion. Nodes on the same layer are not connected, and the tree structure is unidirectional, with the output of each node propagating only to its child nodes. When a decision tree partitions the training data, highly specific branches may form at the leaf nodes. This can cause the model to memorize too much training data, resulting in poor generalization and reduced predictive performance on new data. Overfitting is particularly severe when the data volume is small or noisy. Furthermore, decision trees are susceptible to noisy data during construction, affecting model accuracy and stability.

[0057] like Figure 7 As shown in the figure, the multi-layer perception-random forest algorithm (MLP-RF) ensemble learning method used in the present invention performs better than both support vector machines (SVMs) and decision trees (DTs) for defect recognition using the same feature dataset. The SVM achieves an overall recognition accuracy of 88%, while the DT achieves an overall recognition accuracy of 90%. This demonstrates the highest recognition accuracy of the proposed method, demonstrating its practicality and application value, achieving excellent results in detecting and identifying internal defects in composite insulators.

[0058] This embodiment also provides a system for identifying internal defects of suspension porcelain insulators and support porcelain bushings. Figure 8 Shown, including: The signal acquisition and preprocessing module is used to collect ultrasonic detection signals of suspension porcelain insulators or pillar-type porcelain bushings through ultrasonic probes to generate signal sample sets, and filter the ultrasonic detection signals in the sample sets using a wavelet transform denoising algorithm; The feature extraction module is used to perform EMD empirical mode decomposition on the pre-processed ultrasonic detection signal and then extract the characteristic parameters in the signal to form a feature matrix; An integrated learning module is used to construct an integrated learning model by weighted combination of multi-layer perceptron MLP and random forest RF algorithm, and input the feature matrix into the integrated learning model for analysis; A classification and identification module is used to identify the internal defect type of the suspension porcelain insulator or the pillar-type porcelain bushing according to the analysis result output by the integrated learning model.

[0059] Specifically, the present invention develops an intelligent ultrasonic detection and diagnosis system for internal defects in composite insulators based on the MATLAB software platform and GUI module. The system consists of a signal acquisition and preprocessing module, a feature extraction module, an MLP-RF integrated learning module, a classification and recognition module, and a comprehensive visualization display module.

[0060] It should be noted that the steps in the method for identifying internal defects of suspension porcelain insulators and support-type porcelain bushings provided in this embodiment can be implemented using corresponding modules in the internal defect identification system for suspension porcelain insulators and support-type porcelain bushings. Those skilled in the art can refer to the technical solution of the system to implement the step flow of the method, that is, the embodiments in the system can be understood as preferred examples for implementing the method, which will not be elaborated here.

[0061] In addition to implementing the system and its various devices provided by the present invention in purely computer-readable program code, it is entirely possible to implement the same functions of the system and its various devices provided by the present invention in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, embedded microcontrollers, etc. by logically programming the method steps. Therefore, the system and its various devices provided by the present invention can be considered a hardware component, and the devices included therein for implementing the various functions can also be considered as structures within the hardware component; the devices for implementing the various functions can also be considered as both software modules implementing the method and structures within the hardware component.

[0062] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention, and the scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the principles of the present invention are within the scope of protection of the present invention. It should be noted that improvements and modifications that do not depart from the principles of the present invention, which are apparent to those skilled in the art, should also be considered within the scope of protection of the present invention.

[0063] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for identifying internal defects of suspension porcelain insulators and support porcelain bushings, characterized in that: The steps include: S1: The ultrasonic detection signal of the suspension porcelain insulator or the pillar porcelain bushing is collected by an ultrasonic probe to generate a signal sample set, and the ultrasonic detection signal in the sample set is filtered using a wavelet transform denoising algorithm; S2: Perform EMD empirical mode decomposition on the pre-processed ultrasonic detection signal, and then extract the characteristic parameters in the signal to form a feature matrix; S3: inputting the feature matrix into an ensemble learning model constructed by weighted combination of multilayer perceptron (MLP) and random forest (RF) algorithm for analysis; S4: Identify the internal defect type of the suspension porcelain insulator or the pillar-type porcelain bushing according to the analysis result output by the integrated learning model.

2. The method for identifying internal defects of suspension porcelain insulators and support porcelain bushings according to claim 1, characterized in that: In step S1, the wavelet transform denoising algorithm includes: The signal is decomposed into high-frequency noise components and low-frequency effective components through wavelet decomposition. The high-frequency noise components are suppressed based on a preset threshold. The denoised signal is then reconstructed by wavelet to obtain a filtered ultrasonic detection signal. The formula of the wavelet transform denoising algorithm is as follows: in, , represents the basic wavelet function, is the translation of the wavelet function, is the scale expansion of the wavelet function, and the basic wavelet function satisfies .

3. The method for identifying internal defects of suspension porcelain insulators and support porcelain bushings according to claim 1, characterized in that: In step S2, performing EMD empirical mode decomposition on the preprocessed ultrasonic detection signal further includes: S21: First, the preprocessed original signal is used as the signal to be decomposed , set the number of intrinsic mode function IMF decompositions , ; S22: Calculating the signal to be decomposed The maximum and minimum values ​​of the spectrum are fitted into a smooth curve by the cubic spline interpolation method to generate the upper and lower envelope spectra, and the mean of the upper and lower envelope spectra is calculated. ; S23: Calculate the difference signal between the signal to be decomposed and the mean value , determine the difference signal Whether the criterion of the intrinsic mode function IMF is met, If satisfied, extract the IMF component , , and the signal to be decomposed With the IMF component The difference between ; Otherwise, the difference signal As the new signal to be decomposed, ; S24: Repeat steps S22 to S23 until all IMF components are extracted and the residual term remains Approaches 0, then the signal to be decomposed It is expressed as follows: The decomposed IMF components are used as characteristic parameters to construct the characteristic matrix.

4. The method for identifying internal defects of suspension porcelain insulators and post-type porcelain bushings according to claim 1, characterized in that: In step S3, the ensemble learning model constructed by weighted combination of multilayer perceptron (MLP) and random forest (RF) algorithm further includes: S31: Split the feature matrix extracted by EMD into a training set, a validation set, and a test set in proportion, and input the training set into training a multi-layer perceptron MLP model and a random forest RF model respectively; S32: Dynamically assign weights of the multilayer perceptron MLP and the random forest RF based on the performance of each model applied to the validation set. If the difference in accuracy between the two on the validation set is less than a preset threshold, the ensemble learning model adopts weighted combination output; If the difference in accuracy between the two on the validation set is greater than the preset threshold, the single model with better performance is selected as the output of the ensemble learning model; S33: Input the test set into the ensemble learning model, analyze the output classification results, evaluate the performance indicators of the model in the defect recognition task, and appropriately adjust the hyperparameters of the multilayer perceptron MLP and random forest RF according to the indicators to further improve the model recognition effect.

5. The method for identifying internal defects of suspension porcelain insulators and support porcelain bushings according to claim 4, characterized in that: In step S31, the training of the multi-layer perceptron (MLP) model further includes: S311: Normalize the input training set data to ensure that the value ranges of different features are similar to speed up the model convergence; S312: Initialize weights and biases, calculate predicted values ​​through forward propagation, and perform backpropagation optimization based on the cross-entropy loss function, using stochastic gradient descent to update weights and biases; S313: Continuously repeat the forward propagation and backpropagation processes until the loss function converges or the set number of iterations is reached, input new data for forward propagation, and output the predicted defect results.

6. The method for identifying internal defects of suspension porcelain insulators and post-type porcelain bushings according to claim 4, characterized in that: In step S31, the training of the random forest RF model further includes: S314: Generate multiple training subsets by sampling with replacement from the training set using the Bagging method; S315: Build a decision tree on each subset and randomly select feature subspaces for node partitioning; S316: Multiple decision trees are integrated through a voting mechanism to obtain the classification decision of the random forest.

7. The method for identifying internal defects of suspension porcelain insulators and post-type porcelain bushings according to claim 4, characterized in that: Step S32 further includes: If the performance difference between the multilayer perceptron MLP model and the random forest RF model is less than a preset threshold, the weights of the multilayer perceptron MLP and the random forest RF in the ensemble learning model are determined in a weighted combination manner. The weight calculation formula is as follows: in, and are the weights of the multilayer perceptron MLP model and the random forest RF model, and These are the accuracies of the multi-layer perceptron MLP model and the random forest RF model on the validation set; If the performance difference between the multilayer perceptron (MLP) model and the random forest (RF) model is greater than a preset threshold, the model with higher accuracy on the validation set is selected as the final ensemble learning model.

8. The method for identifying internal defects of suspension porcelain insulators and support porcelain bushings according to claim 6, characterized in that: Step S33 further includes: The test set is input into the ensemble learning model, and the calculated weights are used. and For each sample, the ensemble learning model makes a weighted decision based on the output of the multi-layer perceptron MLP and random forest RF to obtain the final defect classification label. If a weighted combination is selected, the final prediction output is .

9. A system for identifying internal defects of suspension porcelain insulators and support porcelain bushings, characterized in that: include: The signal acquisition and preprocessing module is used to collect ultrasonic detection signals of suspension porcelain insulators or pillar-type porcelain bushings through ultrasonic probes to generate signal sample sets, and filter the ultrasonic detection signals in the sample sets using a wavelet transform denoising algorithm; The feature extraction module is used to perform EMD empirical mode decomposition on the pre-processed ultrasonic detection signal and then extract the characteristic parameters in the signal to form a feature matrix; An integrated learning module is used to construct an integrated learning model by weighted combination of multi-layer perceptron MLP and random forest RF algorithm, and input the feature matrix into the integrated learning model for analysis; A classification and identification module is used to identify the internal defect type of the suspension porcelain insulator or the pillar-type porcelain bushing according to the analysis result output by the integrated learning model.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method for identifying internal defects of a suspension porcelain insulator and a post-type porcelain bushing is implemented according to any one of claims 1 to 8.

11. An electronic device, characterized in that: include: one or more processors; A storage device for storing one or more programs, which, when executed by the one or more processors, enables the one or more processors to implement the method for identifying internal defects of suspension porcelain insulators and post-type porcelain bushings as described in any one of claims 1 to 8.

Images