Power transmission line fault diagnosis method and system based on ada-stacking ensemble learning

Abstract

The application discloses a power transmission line fault diagnosis method and system based on Ada-Stacking integrated learning, and the power transmission line fault diagnosis method comprises the following steps: firstly, constructing an Ada-Stacking integrated deep learning model; then, optimizing model parameters by using a parrot optimization algorithm; and finally, substituting the optimal model parameters into the Ada-Stacking integrated deep learning model to perform power transmission line fault diagnosis. In application, the Ada-Stacking integrated deep learning model is obtained by combining AdaBoost homogeneous integrated learning on the basis of Stacking heterogeneous integrated learning, so that the classification precision of the model can be improved, and the power transmission line fault diagnosis precision can be finally improved.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology for power transmission line engineering, specifically relating to a fault diagnosis method and system for power transmission lines based on Ada-Stacking ensemble learning. Background Technology

[0002] With the development of the social economy and the growth of electricity load, the power grid is constantly expanding, and the structure of the transmission network is becoming increasingly complex. Power transmission lines bear the important functions of transmitting, distributing, and sending electricity, and their normal operation is a strong guarantee for the safe and stable operation of the power system. Therefore, timely and effective prevention and handling of short-circuit faults in transmission lines is not only necessary for the safe operation of the power system, but also an important measure for environmental protection and sustainable development. Currently, some transmission line fault diagnosis methods based on machine learning and artificial intelligence have deeply explored the spatial and temporal relationships in multi-dimensional features, achieving end-to-end fault diagnosis of transmission lines. However, in complex dynamic environments, they inevitably suffer from low diagnostic accuracy and weak robust generalization ability. Therefore, how to select and improve fault diagnosis methods to adapt to efficient fault diagnosis in complex dynamic transmission line environments is a topic worthy of in-depth research.

[0003] A review of existing technical literature revealed that the paper "Design and Research of Intelligent Monitoring Method for Multi-Objective Transmission Lines Based on Machine Learning" uses an adaptive boosting (AdaBoost) algorithm and a radial basis function (RBF) neural network model to diagnose fault types and anomalies in transmission line sensors and systems, achieving automatic fault location and classification. The paper "Research on Power Line Fault Diagnosis Based on SVM-RF" combines the advantages of Support Vector Machine (SVM) and Random Forest (RF), compensating for SVM's insensitivity to missing information and mitigating RF's bias towards larger attribute values ​​when constructing decision trees, ensuring high accuracy in fault type identification even with missing fault information. These papers demonstrate that machine learning algorithms can learn fault characteristics and patterns from data, achieving effective fault diagnosis. However, traditional machine learning algorithms often have limitations in diagnostic accuracy and generalization ability.

[0004] Current literature integrates multiple basic learners to improve diagnostic performance. The paper "Fault Diagnosis of HVDC Transmission Systems Based on Ensemble Learning" designs a fault diagnosis method based on ensemble learning, significantly improving the speed, accuracy, and robustness of fault diagnosis, achieving fault identification and rapid diagnosis of HVDC systems. The paper "Lightning Fault Identification Method for HVDC Transmission Lines Based on Ensemble Learning" combines the Whale Optimization Algorithm (WOA) and ensemble learning (Adaboost.M2) to achieve effective and accurate identification of lightning faults in transmission lines under different fault distances and lightning current amplitudes. However, the performance of ensemble learning models largely depends on the selection of hyperparameters, and existing intelligent algorithms for parameter optimization suffer from slow convergence speed and poor global search capabilities, leading to further improvements in the diagnostic accuracy and efficiency of the methods proposed in the aforementioned papers. Summary of the Invention

[0005] The purpose of this invention is to address the aforementioned problems in the existing technology by providing a transmission line fault diagnosis method and system based on Ada-Stacking ensemble learning that can improve diagnostic accuracy.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] In a first aspect, the present invention provides a method for diagnosing transmission line faults based on Ada-Stacking ensemble learning, the method comprising:

[0008] S1. Construct an Ada-Stacking ensemble deep learning model, which is obtained by integrating a meta-learner with multiple strong-base learners. The meta-learner is trained using the output of the strong-base learner for classification and recognition. The strong-base learner is obtained by weighting multiple identical weak-base learners, and the weak-base learners in different strong-base learners have different categories.

[0009] S2. Use the parrot optimization algorithm to optimize the model parameters of the Ada-Stacking ensemble deep learning model and obtain the optimal model parameters;

[0010] S3. Substitute the optimal model parameters into the Ada-Stacking integrated deep learning model for transmission line fault diagnosis.

[0011] S2 includes:

[0012] S21. Initialize the positions of all individuals in the population;

[0013] S22. Calculate the fitness value of all individuals in the population according to the following steps;

[0014] A1. Randomly divide the collected transmission line fault sample data into an initial training set and an initial test set;

[0015] A2. Substitute the model parameters corresponding to the individual positions to be calculated into the Ada-Stacking ensemble deep learning model. Utilize the initial training set to perform k-fold cross-validation on each strong-base learner in the model, obtaining the training results of each strong-base learner on the initial training set. For each strong-base learner obtained after each round of training in the k-fold cross-validation, test the results using the initial test set. Use a voting method to obtain the optimal test result of each strong-base learner on the initial test set from the k rounds of test results. The model parameters include the parameters of the meta-learner, the number of weak-base learners in the strong-base learner, and the number of folds (k) in the cross-validation.

[0016] A3. First, perform a correlation analysis on each strong foundation learner based on the training results of each strong foundation learner, retain the strong foundation learners with low correlation, then use the training results of the retained strong foundation learners as the training set of the meta-learner, use the optimal test results of the retained strong foundation learners as the test set of the meta-learner, and calculate the classification accuracy of the meta-learner. The classification accuracy is the fitness value.

[0017] S23. Find the optimal individual position in the population based on the fitness value, and update the position of all individuals in the population according to the optimal individual position. The optimal individual is the individual with the largest fitness value in the population.

[0018] S24. Repeat S22-S23 iteratively until the maximum number of iterations is reached. At this point, the model parameters corresponding to the optimal individual position in the population are the optimal model parameters.

[0019] In S21, the initial positions of all individuals in the population include:

[0020] C1. Initialize the population using chaotic mapping according to the following formula to obtain the initial population:

[0021]

[0022]

[0023] In the above formula, X 0 This represents the initial population, with an initial population size of N; The initial population X 0 The positions of the i-th and (i+1)-th individuals are defined; ub and lb are the upper and lower bounds of the search space, respectively; rand(0, 1) is a random number in the range [0, 1]; a and b are control parameters; mod is the modulo function; sin is the sine function.

[0024] C2. Generate the inverse solution for each individual in the initial population using the following formula to obtain the inverse population:

[0025]

[0026] In the above formula, θ represents a reverse population; θ represents a 1×R random vector that follows a normal distribution, and R represents the dimension of the population individuals;

[0027] C3. Merge the reverse population with the initial population, calculate the fitness value of all individuals in the merged population and sort them in descending order, and select the N individuals with the highest fitness values ​​as the final population.

[0028] S23 includes:

[0029] D1. Select the action to update the individual position based on the following piecewise function:

[0030]

[0031]

[0032]

[0033] In the above formula, S t Let represent the behavior chosen by the individual in the t-th iteration; rand(0, 1) is a random number in the range [0, 1]; C1(t) and C2(t) represent the cosine adaptive weights for choosing foraging behavior and choosing communication behavior, respectively; c max c min These represent the upper and lower bounds of the adaptive weights, respectively; Max iter Indicates the total number of iterations;

[0034] D2. Smooth the updated individual positions according to the following formula:

[0035]

[0036] In the above formula, This represents the position of the i-th individual in the (t+1)-th iteration after smoothing. λ represents the positions of the i-th, i-1-th, and i+1-th individuals in the (t+1)-th iteration, respectively; λ is the weight coefficient; sign(*) represents the sign function.

[0037] In A3, the k-fold cross-validation specifically involves: dividing the initial training set into k subsets; for each subset, using the remaining k-1 subsets as the training set to train each strong-base learner once; then, using this subset as the validation set to input into each trained strong-base learner to obtain the classification result. The classification result of each strong-base learner for all subsets is the training result of each strong-base learner for the initial training set; wherein, the step of using the remaining k-1 subsets as the training set to train the strong-base learner specifically involves:

[0038] B1. Initialize the weights of all sample data in the training set, which consists of the remaining k-1 subsets;

[0039] B2. Input the training set into the m-th weak-base learner, obtain the diagnostic error of the weak-base learner, and then calculate the weight of the m-th weak-base learner according to the following formula:

[0040]

[0041] In the above formula, α m ε m Let represent the weights and diagnostic error of the m-th weak-base learner, respectively;

[0042] B3. Update the weights of all samples in the training set according to the following formula:

[0043]

[0044] In the above formula, D m+1 (g) represents the weight of the g-th sample data in the training set during the (m+1)-th iteration; D m (g) represents the weight of the g-th sample data in the training set during the m-th iteration; y g h represents the true label of the g-th sample data. m (g) represents the classification result of the m-th weak basis learner on the g-th sample data; G is the total number of sample data in the training set.

[0045] B4. Repeat steps B2-B3 until the weights of all weak-based learners are calculated. The strong-based learners are then obtained by weighted summation according to the following formula:

[0046]

[0047] In the above formula, H(x) represents the classification result of the strong-base learner for sample data x; M represents the total number of weak-base learners in the strong-base learner; h m (x) represents the classification result of the m-th weak basis learner on the sample data x.

[0048] Secondly, the present invention provides a transmission line fault diagnosis system based on Ada-Stacking ensemble learning, the transmission line fault diagnosis system including a model building module, a parameter optimization module, and a fault diagnosis module;

[0049] The model building module is used to build an Ada-Stacking ensemble deep learning model, which is obtained by integrating a meta-learner with multiple strong-base learners. The meta-learner is trained using the output of the strong-base learners for classification and recognition. The strong-base learner is obtained by weighting multiple identical weak-base learners, and the weak-base learners in different strong-base learners have different categories.

[0050] The parameter optimization module is used to optimize the model parameters of the Ada-Stacking ensemble deep learning model using the parrot optimization algorithm to obtain the optimal model parameters.

[0051] The fault diagnosis module is used to input the optimal model parameters into the Ada-Stacking integrated deep learning model to perform fault diagnosis on the transmission line to be diagnosed.

[0052] The parameter optimization module includes a population location initialization module, a fitness value calculation module, a population location update module, and an optimal model parameter output module.

[0053] The population location initialization module is used to initialize the location of all individuals in the population;

[0054] The fitness value calculation module is used to calculate the individual fitness value according to the following steps:

[0055] A1. Randomly divide the collected transmission line fault sample data into an initial training set and an initial test set;

[0056] A2. Substitute the model parameters corresponding to the individual positions to be calculated into the Ada-Stacking ensemble deep learning model. Use the initial training set to perform k-fold cross-validation on each strong base learner in the model to obtain the training results of each strong base learner on the initial training set. After each round of training in the k-fold cross-validation, each strong base learner is tested using the initial test set to obtain the test results. A voting method is used to obtain the optimal test result of each strong base learner on the initial test set from the k round test results of each strong base learner on the initial test set. The model parameters include the parameters of the meta-learner, the number of weak base learners in the strong base learner, and the number of folds k in the cross-validation.

[0057] A3. First, perform a correlation analysis on each strong foundation learner based on the training results of each strong foundation learner, retain the strong foundation learners with low correlation, then use the training results of the retained strong foundation learners as the training set of the meta-learner, use the optimal test results of the retained strong foundation learners as the test set of the meta-learner, and calculate the classification accuracy of the meta-learner. The classification accuracy is the fitness value.

[0058] The population location update module is used to find the optimal individual location in the population based on the fitness value, and update the location of all individuals in the population according to the optimal individual location. The optimal individual is the individual with the largest fitness value.

[0059] The optimal model parameter output module is used to output the model parameters corresponding to the position of the best individual in the population when the maximum number of iterations is reached. These model parameters are the optimal model parameters.

[0060] The population location initialization module initializes the location of all individuals in the population according to the following steps:

[0061] C1. Initialize the population using chaotic mapping according to the following formula to obtain the initial population:

[0062]

[0063]

[0064] In the above formula, X 0 This represents the initial population, with an initial population size of N; The initial population X 0 The positions of the i-th and (i+1)-th individuals are defined; ub and lb are the upper and lower bounds of the search space, respectively; rand(0, 1) is a random number in the range [0, 1]; a and b are control parameters; mod is the modulo function; sin is the sine function.

[0065] C2. Generate the inverse solution for each individual in the initial population using the following formula to obtain the inverse population:

[0066]

[0067] In the above formula, θ represents a reverse population; θ represents a 1×R random vector that follows a normal distribution, and R represents the dimension of the population individuals;

[0068] C3. Merge the reverse population with the initial population, calculate the fitness value of all individuals in the merged population and sort them in descending order, and select the N individuals with the highest fitness values ​​as the final population.

[0069] The population location update module updates the location of all individuals in the population according to the following steps:

[0070] D1. Select the action to update the individual position based on the following piecewise function:

[0071]

[0072]

[0073]

[0074] In the above formula, S t Let represent the behavior chosen by the individual in the t-th iteration; rand(0, 1) is a random number in the range [0, 1]; C1(t) and C2(t) represent the cosine adaptive weights for choosing foraging behavior and choosing communication behavior, respectively; c max c min These represent the upper and lower bounds of the adaptive weights, respectively; Max iter Indicates the total number of iterations;

[0075] D2. Smooth the updated individual positions according to the following formula:

[0076]

[0077] In the above formula, This represents the position of the i-th individual in the (t+1)-th iteration after smoothing. λ represents the positions of the i-th, i-1-th, and i+1-th individuals in the (t+1)-th iteration, respectively; λ is the weight coefficient; sign(*) represents the sign function.

[0078] The k-fold cross-validation specifically involves: dividing the initial training set into k subsets; for each subset, first using the remaining k-1 subsets as the training set to train each strong-base learner once; then using this subset as the validation set to input into each trained strong-base learner to obtain the classification result. The classification result of each strong-base learner for all subsets is the training result of each strong-base learner for the initial training set; wherein, the step of using the remaining k-1 subsets as the training set to train the strong-base learner is as follows:

[0079] B1. Initialize the weights of all sample data in the training set, which consists of the remaining k-1 subsets;

[0080] B2. Input the training set into the m-th weak-base learner, obtain the diagnostic error of the weak-base learner, and then calculate the weight of the m-th weak-base learner according to the following formula:

[0081]

[0082] In the above formula, α m ε m Let represent the weights and diagnostic error of the m-th weak-base learner, respectively;

[0083] B3. Update the weights of all samples in the training set according to the following formula:

[0084]

[0085] In the above formula, D m+1 (g) represents the weight of the g-th sample data in the training set during the (m+1)-th iteration; D m (g) represents the weight of the g-th sample data in the training set during the m-th iteration; y g h represents the true label of the g-th sample data. m (g) represents the classification result of the m-th weak basis learner on the g-th sample data; G is the total number of sample data in the training set.

[0086] B4. Repeat steps B2-B3 until the weights of all weak-based learners are calculated. The strong-based learners are then obtained by weighted summation according to the following formula:

[0087]

[0088] In the above formula, H(x) represents the classification result of the strong-base learner for sample data x; M represents the total number of weak-base learners in the strong-base learner; h m (x) represents the classification result of the m-th weak basis learner on the sample data x.

[0089] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0090] 1. This invention discloses a transmission line fault diagnosis method based on Ada-Stacking ensemble learning. It constructs an Ada-Stacking ensemble deep learning model, optimizes the model parameters using the Parrot Optimization algorithm, and then substitutes the optimal model parameters into the Ada-Stacking ensemble deep learning model for transmission line fault diagnosis. In this design, the Ada-Stacking ensemble deep learning model is obtained by combining AdaBoost homogeneous ensemble learning with Stacking heterogeneous ensemble learning to improve the model's classification accuracy, ultimately improving the accuracy of transmission line fault diagnosis. Therefore, this invention can improve the accuracy of transmission line fault diagnosis.

[0091] 2. This invention discloses a transmission line fault diagnosis method based on Ada-Stacking ensemble learning. On one hand, k-fold cross-validation is performed on each strong-base learner in the model using an initial training set to obtain the training results of each strong-base learner on the initial training set. Since there are k subsets, the training is performed in k rounds. Each subsequent round of training is based on the strong-base learners obtained in the previous round. After each round of training, the strong-base learners are tested using the initial test set to obtain test results. A voting method is used to obtain the optimal test result of each strong-base learner on the initial test set from the k rounds of test results. The training results and the optimal test results are used as the training set and test set of the meta-learner, respectively. This avoids random errors caused by data partitioning, reduces the randomness and chance of model classification results, and improves its generalization ability. On the other hand, since the greater the difference between strong-base learners, the higher the classification accuracy of Stacking heterogeneous ensemble learning, correlation analysis is performed on the training results of each strong-base learner. Only strong-base learners with low correlation are retained to improve the classification accuracy of the Ada-Stacking ensemble deep learning model. Therefore, this invention can not only avoid random errors caused by data partitioning and improve the generalization ability of the model, but also further improve the classification accuracy of the model.

[0092] 3. This invention discloses a transmission line fault diagnosis method based on Ada-Stacking ensemble learning. When generating the Parrot Optimization algorithm population, an initial population is first obtained through chaotic mapping. This is because chaotic mapping can generate complex, random, nonlinear dynamics with characteristics such as randomness, uniformity, and orderliness, increasing the randomness and diversity of the initial population, enabling better exploration of the solution space and improving the algorithm's search efficiency. On the other hand, the inverse solution for each individual in the generated initial population is obtained through reverse learning, resulting in a reverse population, which further expands the population. This is because generating the opposite solution has a higher probability of finding the global optimum than generating a random solution. Finally, the reverse population is merged with the initial population, and the fitness values ​​of all individuals in the merged population are calculated and sorted in descending order. The top N individuals with the highest fitness values ​​are selected as the final population used by the Parrot Optimization algorithm. Therefore, this invention can improve the search efficiency and global search capability of the Parrot Optimization algorithm.

[0093] 4. This invention discloses a transmission line fault diagnosis method based on Ada-Stacking ensemble learning. In each iteration of the Parrot Optimization Algorithm, a piecewise function is used to randomly select one behavior from four options—foraging, staying, communicating, and fear of strangers—to update the individual positions in the population. This piecewise function design increases the probability of selecting foraging behavior in the initial exploration phase, which helps to better explore the search solution space and ensures that the algorithm can discover potential excellent solutions as soon as possible. As the number of iterations increases, the probability of selecting foraging behavior gradually decreases, while the probability of selecting communicating behavior gradually increases, to promote the transmission and sharing of information within the population. This helps to better utilize existing excellent solutions to guide the search, ultimately reducing the risk of getting trapped in local optima and accelerating the model's convergence. To reduce fluctuations and corners in the optimal solution, a smoothing process is embedded to smooth the solution and reduce the amplitude of jumps. Therefore, this invention not only reduces the risk of getting trapped in local optima and accelerates the model's convergence, but also smooths the fluctuations in the optimal solution. Attached Figure Description

[0094] Figure 1 This is a flowchart of the method described in Example 1.

[0095] Figure 2 This is a structural diagram of the Ada-Stacking integrated deep learning model in the method described in Example 1.

[0096] Figure 3 This is a flowchart of the process of optimizing model parameters using the parrot optimization algorithm in the method described in Example 1.

[0097] Figure 4 This is a flowchart of training the Ada-Stacking ensemble deep learning model in the method described in Example 1.

[0098] Figure 5 This is a schematic diagram of the system described in Example 2.

[0099] Figure 6 This is a schematic diagram of the device described in Example 3. Detailed Implementation

[0100] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings.

[0101] Example 1:

[0102] See Figure 1 A method for diagnosing transmission line faults based on Ada-Stacking ensemble learning is proposed, which proceeds in the following steps:

[0103] Step 1: Construct the Ada-Stacking ensemble deep learning model; see [link / reference] Figure 2The Ada-Stacking ensemble deep learning model is obtained by integrating a meta-learner with w strong-based learners. The meta-learner is trained using the outputs of the strong-based learners for classification and recognition. The weak-based learners within different strong-based learners belong to different categories. Each strong-based learner is obtained by weighting M identical weak-based learners. By sequentially training a series of identical weak-based learners and weighting their outputs, a strong-based learner is constructed. This allows subsequent learners to focus on samples misdiagnosed by earlier learners, gradually improving the overall performance of the strong-based learner. When w=7, Support Vector Machine (SVM), Ridge Regression (RR), Random Forest (RF), Backpropagation (BP), K-Nearest Neighbors (KNN), Decision Tree (DT), and Lightweight Gradient Boosting Machine (Light GBM) can be selected as weak-based learners to analyze and mine data features from multiple perspectives. The corresponding strong-based learners generated by these weak-based learners are AB-SVM, AB-RR, AB-RF, AB-BP, AB-KNN, AB-DT, and AB-Light, respectively. GBM; Considering the generalization performance of the model, the Extreme Gradient Boosting Tree (XGBoost) model with strong generalization performance is selected as the meta-learner to prevent overfitting.

[0104] Step 2: Optimize the model parameters of the Ada-Stacking ensemble deep learning model using the Parrot Optimization Algorithm, and output the optimal model parameters; for example... Figure 3 As shown, parameter optimization using the Parrot algorithm is performed according to the following steps:

[0105] Step 2.1: Initialize the positions of all individuals in the population according to the following steps:

[0106] Step 2.11: Initialize the population using chaotic mapping according to the following formula to obtain the initial population:

[0107]

[0108]

[0109] In the above formula, X 0 This represents the initial population, with an initial population size of N; The initial population X 0 The positions of the i-th and (i+1)-th individuals are defined; ub and lb are the upper and lower bounds of the search space, respectively; rand(0, 1) is a random number in the range [0, 1]; a and b are control parameters; mod is the modulo function; sin is the sine function.

[0110] Step 2.12: Generate the inverse solution for each individual in the initial population according to the following formula to obtain the inverse population:

[0111]

[0112] In the above formula, θ represents a reverse population; θ represents a 1×R random vector that follows a normal distribution, and R represents the dimension of the population individuals;

[0113] Step 2.13: Merge the reverse population with the initial population, calculate the fitness value of all individuals in the merged population and sort them in descending order. Select the N individuals with the highest fitness values ​​as the final population. The fitness value is the classification accuracy of the Ada-Stacking ensemble deep learning model. Figure 4 As shown, the steps for calculating the fitness value are as follows;

[0114] Step A1: Randomly divide the collected transmission line fault sample data into an initial training set and an initial test set;

[0115] Step A2: Substitute the model parameters corresponding to the individual positions to be calculated into the Ada-Stacking ensemble deep learning model. Use the initial training set to perform k-fold cross-validation on each strong base learner in the model to obtain the training results of each strong base learner on the initial training set. After each round of training in the k-fold cross-validation, each strong base learner is tested using the initial test set to obtain the test results. A voting method is used to obtain the optimal test result of each strong base learner on the initial test set from the k round test results of each strong base learner on the initial test set. The model parameters include the parameters of the meta-learner, the number M of weak base learners in the strong base learner, and the number of folds k in the cross-validation. The parameters of the meta-learner include the learning rate, the maximum depth of the tree, and the regularization terms L1 and L2.

[0116] Specifically, the k-fold cross-validation involves dividing the initial training set into k subsets. Since there are k subsets, the training iterations are k rounds. For each subset, the remaining k-1 subsets are used as the training set to train each strong-base learner once. Then, this subset is used as the validation set and input into the trained strong-base learner to obtain the classification result. The classification result of each strong-base learner for all subsets is the training result of each strong-base learner for the initial training set. Each round of training is as follows:

[0117] Step A21: Initialize the weights of all sample data in the training set, which consists of the remaining k-1 subsets;

[0118] Step A22: Input the training set into the m-th weak-based learner in the strong-based learner, obtain the diagnostic error of the weak-based learner, and then calculate the weight of the m-th weak-based learner according to the following formula:

[0119]

[0120] In the above formula, α m ε m Let represent the weights and diagnostic error of the m-th weak-base learner, respectively;

[0121] Step A23: Update the weights of all sample data in the training set according to the following formula:

[0122]

[0123] In the above formula, D m+1 (g) represents the weight of the g-th sample data in the training set during the (m+1)-th iteration; D m (g) represents the weight of the g-th sample data in the training set during the m-th iteration; y g h is the true label of the g-th sample data, with a value of +1 or -1; m (g) represents the classification result of the m-th weak-base learner on the g-th sample data, with a value of +1 or -1; G is the total number of sample data in the training set; M is the number of weak-base learners in the strong-base learner;

[0124] Step A24: Repeat steps A22-A23 until the weights of all weak-based learners in the strong-based learner are calculated. Then, sum the outputs of all weak-based learners using the following formula to obtain the output of the strong-based learner:

[0125]

[0126] In the above formula, H(x) represents the classification result of the strong-base learner for sample data x; M represents the total number of weak-base learners in the strong-base learner; h m (x) represents the classification result of the m-th weak basis learner on the sample data x;

[0127] Step A3: Perform correlation analysis on the training results of each strong basis learner, retain the three strong basis learners with low correlation, then use the training results of these three strong basis learners as the training set of the meta-learner, use the best test results of these three strong basis learners as the test set of the meta-learner, and calculate the classification accuracy ε of the meta-learner.

[0128] Step A4: Determine whether the classification accuracy ε is greater than the best accuracy Best_accuracy; if yes, output the classification accuracy obtained at this time as the fitness value; if not, return to step A1 to continue iterative calculation.

[0129] Step 2.2: Calculate the fitness value of all individuals in the population. See steps A1-A4 for the calculation steps.

[0130] Step 2.3: Select the individual with the highest fitness value in the population as the optimal individual of the population, and update the positions of all individuals in the population according to the position of the optimal individual; wherein, the update step is specifically as follows;

[0131] Step 2.31: Based on the following piecewise function, randomly select one behavior from the four behaviors of foraging, dwelling, communicating, and fear of strangers to update the individual's location:

[0132]

[0133]

[0134]

[0135] In the above formula, S t Let represent the behavior chosen by the individual in the t-th iteration; rand(0, 1) is a random number in the range [0, 1]; C1(t) and C2(t) represent the cosine adaptive weights for choosing foraging behavior and choosing communication behavior, respectively; c max c min These represent the upper and lower bounds of the adaptive weights, respectively; Max iter Indicates the total number of iterations;

[0136] The location update formula for foraging behavior is:

[0137]

[0138]

[0139] In the above formula, This represents the position of the i-th individual in the t-th and t+1-th iterations; X represents the average position of the population at the t-th iteration; best This represents the position from initialization to the current best position found, and also indicates the owner's position; This refers to the movement of an individual relative to its owner's position. This indicates the individual's observation of the entire population's location to further determine the direction of food; Levy(dim) represents the Levy distribution, used to describe the parrot's flight patterns, and can be obtained according to the rules in the following formula;

[0140]

[0141] In the above formula, Γ(·) is the gamma function; N(0, dim) is a normal distribution from 0 to dimension dim; μ, σ, and v are intermediate quantities; γ takes the value of 1.5;

[0142] The position update formula for dwell behavior is:

[0143]

[0144] In the above formula, ones(1, dim) represents a variable with all 1s in dimension dim; X best Levy(dim) represents the process of an individual flying to its host; rand(0,1) and ones(1,dim) represent the process of randomly stopping at a certain part of the host's body.

[0145] The location update formula for communication behavior is:

[0146]

[0147] In the above formula, This refers to the process by which an individual joins a group of parrots to communicate. This represents the process of an individual flying out immediately after communication; P is a randomly generated value within the range [0,1].

[0148] The location update formula for fear of strangers is:

[0149]

[0150] In the above formula, This indicates the process of an individual redirecting its flight back to its owner; This represents the process by which an individual distances themselves from strangers;

[0151] Step 2.32: To reduce fluctuations and corner points in the optimal solution, the updated individual positions are smoothed according to the following formula:

[0152]

[0153] In the above formula, This represents the position of the i-th individual in the (t+1)-th iteration after smoothing. λ represents the positions of the i-th, (i-1)-th, and (i+1)-th individuals at the (t+1)-th iteration, respectively; λ is the weight coefficient; sign(*) represents the sign function;

[0154] Step 2.4: Repeat steps 2.2-2.3 iteratively until the maximum number of iterations is reached. At this point, the model parameters corresponding to the optimal individual position in the population are the optimal model parameters.

[0155] Step 3: Substitute the optimal model parameters into the Ada-Stacking integrated deep learning model for transmission line fault diagnosis.

[0156] Performance verification:

[0157] To verify the effectiveness of the transmission line fault diagnosis method described in this invention, a three-phase transmission line simulation system with a voltage source level of 400kV at both ends, an operating frequency of 50Hz, and a line length of 300km was built using the MATLAB / Simulink simulation platform. Ten types of faults, as shown in Table 1, were simulated experimentally. 1000 samples were collected for each type of fault, and 900 samples from each type of fault were selected to form the transmission line fault sample data. The Ada-Stacking ensemble deep learning model in the proposed method was trained using this transmission line fault sample data. The optimized model parameters of the trained Ada-Stacking ensemble deep learning model are shown in Table 2. The trained Ada-Stacking ensemble deep learning model was then used to diagnose the remaining samples.

[0158] Table 1. 10 Types of Transmission Line Faults

[0159]

[0160]

[0161] Table 2. Model parameter optimization results

[0162]

[0163] 1. The fault diagnosis performance of the method proposed in this invention (Method 1) is compared with other methods, including Stacking ensemble deep learning model (Method 2), Ada-Stacking ensemble deep learning model without intelligent algorithm optimization (Method 3), Ada-Stacking ensemble deep learning model using the Grey Wolf Optimization Algorithm for parameter optimization (Method 4), Ada-Stacking ensemble deep learning model using the traditional Parrot Algorithm for parameter optimization (Method 5), and Support Vector Machine (SVM) (Method 6). The comparison results are shown in Table 3.

[0164] Table 3 Comparison of Fault Diagnosis Performance

[0165]

[0166] As can be seen from Table 3, the method proposed in this invention has the highest accuracy, recall, precision, and F1-score, reaching 92%, 89%, 94%, and 90%, respectively. This indicates that it performs best in terms of overall prediction accuracy, is most effective in detecting real positive samples, has the highest accuracy when predicting positive samples, and is the best when considering both prediction accuracy and positive sample detection capability.

[0167] 2. The fault diagnosis performance of the method proposed in this invention was compared with other methods under different noise levels. The comparison results are shown in Table 4:

[0168] Table 4 Comparison of fault diagnosis performance under different noise levels

[0169]

[0170] As shown in Table 4, with the increase of noise impact, the fault diagnosis performance indicators of all methods generally decline. This is because noise interference affects the model's ability to learn and generalize data, resulting in a decrease in the model's accuracy and stability. However, compared with other methods, the method proposed in this invention still maintains a high level in all fault diagnosis performance indicators, indicating that it has strong anti-noise interference ability as well as optimal robustness and generalization ability.

[0171] Example 2:

[0172] See Figure 5 A transmission line fault diagnosis system based on Ada-Stacking ensemble learning includes a model building module, a parameter optimization module, and a fault diagnosis module.

[0173] The model building module is used to build an Ada-Stacking ensemble deep learning model. The Ada-Stacking ensemble deep learning model is obtained by integrating a meta-learner with multiple strong-base learners. The meta-learner is trained using the output of the strong-base learners for classification and recognition. The strong-base learners are obtained by weighting multiple identical weak-base learners. The weak-base learners in different strong-base learners have different categories. The model building module is specifically used to implement step 1 in embodiment 1.

[0174] The parameter optimization module is used to optimize the model parameters of the Ada-Stacking ensemble deep learning model using the parrot optimization algorithm to obtain the optimal model parameters; the parameter optimization module is specifically used to implement step 2 in embodiment 1.

[0175] Furthermore, the parameter optimization module includes a population position initialization module, a fitness value calculation module, a population position update module, and an optimal model parameter output module; the population position initialization module is used to initialize the position of all individuals in the population according to the following steps:

[0176] Step 2.11: Initialize the population using chaotic mapping according to the following formula to obtain the initial population:

[0177]

[0178]

[0179] In the above formula, X 0 This represents the initial population, with an initial population size of N; The initial population X 0 The positions of the i-th and (i+1)-th individuals are defined; ub and lb are the upper and lower bounds of the search space, respectively; rand(0, 1) is a random number in the range [0, 1]; a and b are control parameters; mod is the modulo function; sin is the sine function.

[0180] Step 2.12: Generate the inverse solution for each individual in the initial population according to the following formula to obtain the inverse population:

[0181]

[0182] In the above formula, θ represents a reverse population; θ represents a 1×R random vector that follows a normal distribution, and R represents the dimension of the population individuals;

[0183] Step 2.13: Merge the reverse population with the initial population, calculate the fitness value of all individuals in the merged population and sort them in descending order, and select the N individuals with the highest fitness values ​​as the final population.

[0184] The fitness value calculation module is used to calculate the fitness value according to the following steps:

[0185] Step A1: Randomly divide the collected transmission line fault sample data into an initial training set and an initial test set;

[0186] Step A2: Substitute the model parameters corresponding to the individual positions to be calculated into the Ada-Stacking ensemble deep learning model. Use the initial training set to perform k-fold cross-validation on each strong base learner in the model to obtain the training results of each strong base learner on the initial training set. After each round of training in the k-fold cross-validation, each strong base learner is tested using the initial test set to obtain the test results. A voting method is used to obtain the optimal test result of each strong base learner on the initial test set from the k round test results of each strong base learner on the initial test set. The model parameters include the parameters of the meta-learner, the number M of weak base learners in the strong base learner, and the number of folds k in the cross-validation. The parameters of the meta-learner include the learning rate, the maximum depth of the tree, and the regularization terms L1 and L2.

[0187] Specifically, the k-fold cross-validation involves dividing the initial training set into k subsets. Since there are k subsets, the training iterations are k rounds. For each subset, the remaining k-1 subsets are used as the training set to train each strong-base learner once. Then, this subset is used as the validation set and input into the trained strong-base learners to obtain the classification results. The classification results of each strong-base learner for all subsets are the training results of each strong-base learner for the initial training set. Each round of training is as follows:

[0188] Step A21: Initialize the weights of all sample data in the training set, which consists of the remaining k-1 subsets;

[0189] Step A22: Input the training set into the m-th weak-based learner in the strong-based learner, obtain the diagnostic error of the weak-based learner, and then calculate the weight of the m-th weak-based learner according to the following formula:

[0190]

[0191] In the above formula, α m ε m Let represent the weights and diagnostic error of the m-th weak-base learner, respectively;

[0192] Step A23: Update the weights of all sample data in the training set according to the following formula:

[0193]

[0194] In the above formula, D m+1 (g) represents the weight of the g-th sample data in the training set during the (m+1)-th iteration; D m (g) represents the weight of the g-th sample data in the training set during the m-th iteration; y g h is the true label of the g-th sample data, with a value of +1 or -1; m (g) represents the classification result of the m-th weak-base learner on the g-th sample data, with a value of +1 or -1; G is the total number of sample data in the training set; M is the number of weak-base learners in the strong-base learner;

[0195] Step A24: Repeat steps A22-A23 until the weights of all weak-based learners in the strong-based learner are calculated. Then, sum the outputs of all weak-based learners using the following formula to obtain the output of the strong-based learner:

[0196]

[0197] In the above formula, H(x) represents the classification result of the strong-base learner for sample data x; M represents the total number of weak-base learners in the strong-base learner; h m(x) represents the classification result of the m-th weak basis learner on the sample data x;

[0198] Step A3: Perform correlation analysis on the training results of each strong basis learner, retain the three strong basis learners with low correlation, then use the training results of these three strong basis learners as the training set of the meta-learner, use the best test results of these three strong basis learners as the test set of the meta-learner, and calculate the classification accuracy ε of the meta-learner.

[0199] Step A4: Determine whether the classification accuracy ε is greater than the preset best accuracy Best_accuracy; if yes, output the classification accuracy obtained at this time as the fitness value; if not, return to step A1 to continue iterative calculation.

[0200] The population position update module is used to find the individual with the highest fitness value in the population and take it as the optimal individual, and update the position of all individuals in the population according to the position of the optimal individual;

[0201] Furthermore, the population location update module updates the location of all individuals in the population according to the following steps:

[0202] Step 2.31: Select the action to update the individual position based on the following piecewise function:

[0203]

[0204]

[0205]

[0206] In the above formula, S t Let represent the behavior chosen by the individual in the t-th iteration; rand(0, 1) is a random number in the range [0, 1]; C1(t) and C2(t) represent the cosine adaptive weights for choosing foraging behavior and choosing communication behavior, respectively; c max c min These represent the upper and lower bounds of the adaptive weights, respectively; Max iter Indicates the total number of iterations;

[0207] Step 2.32: Smooth the updated individual positions according to the following formula:

[0208]

[0209] In the above formula, This represents the position of the i-th individual in the (t+1)-th iteration after smoothing. λ represents the positions of the i-th, (i-1)-th, and (i+1)-th individuals at the (t+1)-th iteration, respectively; λ is the weight coefficient; sign(*) represents the sign function;

[0210] The optimal model parameter output module is used to output the model parameters corresponding to the position of the best individual in the population when the maximum number of iterations is reached. These model parameters are the optimal model parameters.

[0211] The fault diagnosis module is used to substitute the optimal model parameters into the Ada-Stacking integrated deep learning model to perform fault diagnosis on the transmission line to be diagnosed; the fault diagnosis module is specifically used to implement step 3 in embodiment 1.

[0212] Example 3:

[0213] See Figure 6 A transmission line fault diagnosis device based on Ada-Stacking ensemble learning includes a memory and a processor; the memory is used to store computer program code and transmit the computer program code to the processor; the processor is used to execute the method described in Example 1 according to the instructions in the computer program code.

[0214] Example 4:

[0215] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in Example 1.

[0216] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0217] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0218] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0219] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0220] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A transmission line fault diagnosis method based on Ada-Stacking ensemble learning, characterized in that: The transmission line fault diagnosis method includes: S1. Construct an Ada-Stacking ensemble deep learning model, which is obtained by integrating a meta-learner with multiple strong-base learners. The meta-learner is trained using the output of the strong-base learner for classification and recognition. The strong-base learner is obtained by weighting multiple identical weak-base learners, and the weak-base learners in different strong-base learners have different categories. S2. Use the parrot optimization algorithm to optimize the model parameters of the Ada-Stacking ensemble deep learning model and obtain the optimal model parameters; S3. Substitute the optimal model parameters into the Ada-Stacking integrated deep learning model for transmission line fault diagnosis. S2 includes: S21. Initialize the positions of all individuals in the population; S22. Calculate the fitness value of all individuals in the population according to the following steps; A1. Randomly divide the collected transmission line fault sample data into an initial training set and an initial test set; A2. Substitute the model parameters corresponding to the individual positions to be calculated into the Ada-Stacking ensemble deep learning model. Use the initial training set to perform k-fold cross-validation on each strong base learner in the model to obtain the training results of each strong base learner on the initial training set. After each round of training in the k-fold cross-validation, each strong base learner is tested using the initial test set to obtain the test results. A voting method is used to obtain the optimal test result of each strong base learner on the initial test set from the k round test results of each strong base learner on the initial test set. The model parameters include the parameters of the meta-learner, the number of weak base learners in the strong base learner, and the number of folds k in the cross-validation. A3. First, perform a correlation analysis on each strong foundation learner based on the training results of each strong foundation learner, retain the strong foundation learners with low correlation, then use the training results of the retained strong foundation learners as the training set of the meta-learner, use the optimal test results of the retained strong foundation learners as the test set of the meta-learner, and calculate the classification accuracy of the meta-learner. The classification accuracy is the fitness value. S23. Find the optimal individual position in the population based on the fitness value, and update the position of all individuals in the population according to the optimal individual position. The optimal individual is the individual with the largest fitness value in the population. S24. Repeat S22-S23 iteratively until the maximum number of iterations is reached. At this point, the model parameters corresponding to the optimal individual position in the population are the optimal model parameters.

2. The transmission line fault diagnosis method based on Ada-Stacking ensemble learning according to claim 1, characterized in that: In S21, the initial positions of all individuals in the population include: C1. Initialize the population using chaotic mapping according to the following formula to obtain the initial population: ； ； In the above formula, This represents the initial population, with an initial population size of N; , The initial population The Middle The, the The individual's location; , These are the upper and lower bounds of the search space, respectively. A random number in the range [0,1]; , All are control parameters; For the remainder function; It is a sine function; C2. Generate the inverse solution for each individual in the initial population using the following formula to obtain the inverse population: ； In the above formula, Indicates a reverse population; Let R represent a 1×R random vector that follows a normal distribution, where R represents the dimension of the population. C3. Merge the reverse population with the initial population, calculate the fitness value of all individuals in the merged population and sort them in descending order, and select the N individuals with the highest fitness values ​​as the final population.

3. The transmission line fault diagnosis method based on Ada-Stacking ensemble learning according to claim 1 or 2, characterized in that: S23 includes: D1. Select the action to update the individual position based on the following piecewise function: ； ； ； In the above formula, Indicates the first The behavior chosen by the individual in the next iteration; A random number in the range [0,1]; , These represent the cosine adaptive weights for choosing foraging behavior and choosing communication behavior, respectively. , These represent the upper and lower bounds of the adaptive weights, respectively. Indicates the total number of iterations; D2. Smooth the updated individual positions according to the following formula: ； In the above formula, Indicates the smoothed-out first... During the nth iteration The location of each individual; , , They represent the first During the nth iteration The, the The, the The location of each individual; These are the weighting coefficients; Represents a symbolic function.

4. The transmission line fault diagnosis method based on Ada-Stacking ensemble learning according to claim 1 or 2, characterized in that: In A3, the k-fold cross-validation specifically involves: dividing the initial training set into k subsets; for each subset, using the remaining k-1 subsets as the training set to train each strong-base learner once; then, using this subset as the validation set to input into each trained strong-base learner to obtain the classification result. The classification result of each strong-base learner for all subsets is the training result of each strong-base learner for the initial training set; wherein, the step of using the remaining k-1 subsets as the training set to train the strong-base learner specifically involves: B1. Initialize the weights of all sample data in the training set, which consists of the remaining k-1 subsets; B2. Input the training set into the m-th weak-base learner, obtain the diagnostic error of the weak-base learner, and then calculate the weight of the m-th weak-base learner according to the following formula: ； In the above formula, , Let represent the weights and diagnostic error of the m-th weak-base learner, respectively; B3. Update the weights of all samples in the training set according to the following formula: ； In the above formula, For the training set in the (m+1)th iteration The weights of each sample data; For the training set in the m-th iteration The weights of each sample data; For the first The true label of each sample data; For the m-th weak basis learner, the first... Classification results of individual sample data; The total number of sample data in the training set; B4. Repeat steps B2-B3 until the weights of all weak-based learners are calculated. The strong-based learners are then obtained by weighted summation according to the following formula: ； In the above formula, This indicates that the strong-base learner is effective for sample data. The classification results; This represents the total number of weak-base learners in a strong-base learner set. This indicates that the m-th weak basis learner is used with respect to the sample data. The classification results.

5. A transmission line fault diagnosis system based on Ada-Stacking ensemble learning, characterized in that: The transmission line fault diagnosis system includes a model building module, a parameter optimization module, and a fault diagnosis module. The model building module is used to build an Ada-Stacking ensemble deep learning model, which is obtained by integrating a meta-learner with multiple strong-base learners. The meta-learner is trained using the output of the strong-base learners for classification and recognition. The strong-base learner is obtained by weighting multiple identical weak-base learners, and the weak-base learners in different strong-base learners have different categories. The parameter optimization module is used to optimize the model parameters of the Ada-Stacking ensemble deep learning model using the parrot optimization algorithm to obtain the optimal model parameters. The fault diagnosis module is used to input the optimal model parameters into the Ada-Stacking integrated deep learning model to perform fault diagnosis on the transmission line to be diagnosed. The parameter optimization module includes a population location initialization module, a fitness value calculation module, a population location update module, and an optimal model parameter output module. The population location initialization module is used to initialize the location of all individuals in the population; The fitness value calculation module is used to calculate the individual fitness value according to the following steps: A1. Randomly divide the collected transmission line fault sample data into an initial training set and an initial test set; A2. Substitute the model parameters corresponding to the individual positions to be calculated into the Ada-Stacking ensemble deep learning model. Use the initial training set to perform k-fold cross-validation on each strong base learner in the model to obtain the training results of each strong base learner on the initial training set. After each round of training in the k-fold cross-validation, each strong base learner is tested using the initial test set to obtain the test results. A voting method is used to obtain the optimal test result of each strong base learner on the initial test set from the k round test results of each strong base learner on the initial test set. The model parameters include the parameters of the meta-learner, the number of weak base learners in the strong base learner, and the number of folds k in the cross-validation. A3. First, perform a correlation analysis on each strong foundation learner based on the training results of each strong foundation learner, retain the strong foundation learners with low correlation, then use the training results of the retained strong foundation learners as the training set of the meta-learner, use the optimal test results of the retained strong foundation learners as the test set of the meta-learner, and calculate the classification accuracy of the meta-learner. The classification accuracy is the fitness value. The population location update module is used to find the optimal individual location in the population based on the fitness value, and update the location of all individuals in the population according to the optimal individual location. The optimal individual is the individual with the largest fitness value. The optimal model parameter output module is used to output the model parameters corresponding to the position of the best individual in the population when the maximum number of iterations is reached. These model parameters are the optimal model parameters.

6. The transmission line fault diagnosis system based on Ada-Stacking ensemble learning according to claim 5, characterized in that: The population location initialization module initializes the location of all individuals in the population according to the following steps: C1. Initialize the population using chaotic mapping according to the following formula to obtain the initial population: ； ； In the above formula, This represents the initial population, with an initial population size of N; , The initial population The Middle The, the The individual's location; , These are the upper and lower bounds of the search space, respectively. A random number in the range [0,1]; , All are control parameters; For the remainder function; It is a sine function; C2. Generate the inverse solution for each individual in the initial population using the following formula to obtain the inverse population: ； In the above formula, Indicates a reverse population; Let R represent a 1×R random vector that follows a normal distribution, where R represents the dimension of the population. C3. Merge the reverse population with the initial population, calculate the fitness value of all individuals in the merged population and sort them in descending order, and select the N individuals with the highest fitness values ​​as the final population.

7. The transmission line fault diagnosis system based on Ada-Stacking ensemble learning according to claim 5 or 6, characterized in that: The population location update module updates the location of all individuals in the population according to the following steps: D1. Select the action to update the individual position based on the following piecewise function: ； ； ； In the above formula, Indicates the first The behavior chosen by the individual in the next iteration; A random number in the range [0,1]; , These represent the cosine adaptive weights for choosing foraging behavior and choosing communication behavior, respectively. , These represent the upper and lower bounds of the adaptive weights, respectively. Indicates the total number of iterations; D2. Smooth the updated individual positions according to the following formula: ； In the above formula, Indicates the smoothed-out first... During the nth iteration The location of each individual; , , They represent the first During the nth iteration The, the The, the The location of each individual; These are the weighting coefficients; Represents a symbolic function.

8. The transmission line fault diagnosis system based on Ada-Stacking ensemble learning according to claim 5 or 6, characterized in that: The k-fold cross-validation specifically involves: dividing the initial training set into k subsets; for each subset, first using the remaining k-1 subsets as the training set to train each strong-base learner once; then using this subset as the validation set to input into each trained strong-base learner to obtain the classification result. The classification result of each strong-base learner for all subsets is the training result of each strong-base learner for the initial training set; wherein, the step of using the remaining k-1 subsets as the training set to train the strong-base learner is as follows: B1. Initialize the weights of all sample data in the training set, which consists of the remaining k-1 subsets; B2. Input the training set obtained in B1 into the m-th weak-base learner to obtain the diagnostic error of the weak-base learner, and then calculate the weight of the m-th weak-base learner according to the following formula: ； In the above formula, , Let represent the weights and diagnostic error of the m-th weak-base learner, respectively; B3. Update the weights of all samples in the training set according to the following formula: ； In the above formula, For the training set in the (m+1)th iteration The weights of each sample data; For the training set in the m-th iteration The weights of each sample data; For the first The true label of each sample data; For the m-th weak basis learner, the first... Classification results of individual sample data; The total number of sample data in the training set; B4. Repeat steps B2-B3 until the weights of all weak-based learners are calculated. The strong-based learners are then obtained by weighted summation according to the following formula: ； In the above formula, This indicates that the strong-base learner is effective for sample data. The classification results; This represents the total number of weak-base learners in a strong-base learner set. This indicates that the m-th weak basis learner is used with respect to the sample data. The classification results.

Images