A method for identifying parameters of a classifier based on extreme learning machine

Abstract

This invention discloses a method for classifier parameter identification based on Extreme Learning Machine (ELM). Step 1: Divide the classification dataset into training and test datasets. Step 2: Construct an ELM model and use a modified whale optimization algorithm to obtain the optimal initialization parameters for the multi-novel ELM. Step 3: Train the ELM model online using the training dataset, evaluate the model using the structural risk loss function, identify and update the output weights, and complete the classification training of the obtained dataset to refine the ELM model parameters. Step 4: Input the test dataset to be classified into the multi-novel ELM model trained online in Step 3, and identify the category of the test dataset online. If new data is input, repeat Step 3 to classify the newly input classification dataset. This invention improves the classification accuracy of the ELM model.

Description

Technical Field

[0001] This invention relates to the field of classifier parameter identification technology, and in particular to a method for classifier parameter identification of extreme learning machines. Background Technology

[0002] As engineering applications demand increasingly higher standards, methods for preprocessing and correctly classifying objects such as images and text are constantly evolving. Machine learning is one of the mainstream methods for classification. Most mathematical models rely on system identification, and multiple innovation theory, a branch of system identification, aims to extend innovation length and fully utilize useful information from data. In other words, multiple innovation theory introduces the information learned in moving data window modeling, including not only current information but also some past information. In machine learning, the Extreme Learning Machine (ELM), with its simple single-hidden-layer feedforward neural network structure and the characteristic of using randomly assigned input weights, eliminates the need for repeated training. Compared to other traditional machine learning algorithms, ELM offers advantages in training speed while maintaining accuracy. Its kernel version, ELMwithkernel (KELM), changes the original ELM mapping method to a kernel function implicit mapping, reducing uncertainty and improving performance while retaining the original advantages. Therefore, KELM is also widely used in the field of classifier parameter identification.

[0003] In practical engineering applications, there will be different new data inputs. If only past data is used as the basis, it will be difficult to maintain the performance of the classifier. In complex real-world environments, the collected input data is extremely noisy, and some input data may even be detrimental to the identification of classifier parameters. Therefore, the ability to minimize or even eliminate the influence of adverse data on parameter identification becomes very important in classifier parameter identification methods. Summary of the Invention

[0004] This invention provides a method for classifier parameter identification based on Extreme Learning Machine (ELM), which solves the problem that ELM has limited ability to process unfavorable data with poor correlation, and improves the classification accuracy of ELM model.

[0005] To achieve the above effects, the technical solution of the present invention is as follows:

[0006] A method for classifier parameter identification based on extreme learning machine includes the following steps:

[0007] Step 1: Obtain the classification dataset and divide it into training and testing datasets. In the training dataset, each element is the input (x, y) for each online training session, containing m data points. x is an n-dimensional column vector, where n represents the number of features in each input sample; y is a t-dimensional output column vector, where t represents the number of categories. Normalize the data in both the training and testing datasets.

[0008] Step 2: Construct the Extreme Learning Machine model and use the improved whale optimization algorithm to obtain the optimal initialization parameters of the multi-innovation Extreme Learning Machine: regularization factor C and kernel function parameter γ; the improvement process of the whale optimization algorithm is as follows: introduce an adaptive weight strategy and a Levy flight search perturbation mechanism to update the position of the whale individual in the whale optimization algorithm;

[0009] Step 3: Train the Extreme Learning Machine model online using the training dataset, evaluate the Extreme Learning Machine model using the structural risk loss function, identify and update the output weights, and complete the classification training of the obtained dataset.

[0010] Step 4: Input the test dataset to be classified into the multi-innovation extreme learning machine model trained online in Step 3, and identify the category of the test dataset online. If there is new data input, repeat Step 3 to classify the new input classification dataset.

[0011] In the above scheme, the Extreme Learning Machine (ELM) is a multi-novel kernel ELM model. This invention combines multi-novel theory with ELM, establishing a multi-data moving window during online training. This improves data utilization, strengthens the connection between the model and the current data, and provides an effective approach for optimizing the accuracy of the ELM algorithm. Multi-novel theory introduces the idea that the information learned through moving data window modeling includes not only current information but also some past information; ELM, as an optimization algorithm, transforms explicit mappings into implicit mappings.

[0012] Furthermore, the basic model of the Extreme Learning Machine in step 2 is shown below:

[0013] Hβ=Y (1)

[0014] In the formula, Y is the Extreme Learning Machine model, i.e., a linear matrix equation, and β=[β1,β2,...,β i ] T β represents the output weights of the extreme learning machine. i =[β i1 ,β i2 ,...,β it The dimension of the output weights of the Extreme Learning Machine depends on the number of hidden layer nodes i and the number of output label classes t; H is the feature mapping matrix, expressed as:

[0015]

[0016] In the formula, g() is the activation function, and (a,b) are the input weights between the input layer and the hidden layer of the extreme learning machine. The input weights are randomly assigned.

[0017] By solving the linear matrix equation of the extreme learning machine model (1), the least squares solution β is obtained. The least squares solution β is the output weight of the extreme learning machine. The output weight of the extreme learning machine is transformed into the following form.

[0018] β=H T (I / C+HH T ) -1 Y (3)

[0019] In the formula, C represents the regularization factor, which is a constant; I is the identity matrix of the corresponding dimension.

[0020] Furthermore, step 2 obtains the optimal initialization parameters for the Multiple Innovation Extreme Learning Machine, specifically including the following steps:

[0021] Step 2.1: Set the initialization parameters for the whale optimization algorithm. The initialization parameters include the whale population size N, the number of dimensions D, the maximum number of iterations T_max, and the upper and lower bounds U of the whale population in each dimension. d ;

[0022] Step 2.2: Use a grid search strategy to search for whale populations X = {X1, X2, ..., X} within the required area. N}, where X i (i = 1, 2, ..., N) represents the location information of the i-th individual whale, X i ={X i1 ,X i2 ,...,X iD}, X ij (j = 1, 2, ..., D) represents the position information of the j-th dimension of the i-th individual whale;

[0023] Step 2.3: Calculate the fitness function value F for each individual whale in the whale population using the F-score criterion, and select and retain the optimal fitness value F. best The optimal fitness value F best The corresponding optimal individual whale X best As the current globally optimal solution;

[0024] Step 2.4: Introduce an adaptive weighting strategy and the Lévy flight search perturbation mechanism to update the position of individual whales in the whale optimization algorithm, thus improving the algorithm; that is, introduce an adaptive weighting strategy to apply to the optimal whale individual X.best Adjust the optimal individual whale X best The weight ratio is adjusted to fully leverage the role of individual whales in location updates; the Levy flight search perturbation mechanism is introduced to enhance the whale optimization algorithm's ability to escape local optima;

[0025] Step 2.5: Iteratively update the whale optimization algorithm using a mutation improvement mechanism, that is, perform Gaussian mutation based on manually set conditional probabilities, and use a survival-of-the-fittest mechanism. If the updated or mutated whale population X... t Its fitness value is better than the best whale individual X in the previous iteration. best Then the current global optimal solution is replaced by X. t If the whale population X is updated or mutated t The fitness value is worse than the best whale individual X in the previous iteration. best Then the current global optimal solution is the best individual whale X from the previous iteration. best The update continues until the number of iterations t reaches the maximum number of iterations T_max.

[0026] Furthermore, the fitness function value F in step 2.3 is:

[0027]

[0028] In the formula, Classes represents the number of categories; Recall i For recall rate, Precision i For precision, it is expressed as:

[0029]

[0030] Among them, TP i FP i TN i and FN i These represent the number of true positive, false positive, true negative, and false negative results in the classification, respectively.

[0031] Furthermore, the position update of the individual whale in step 2.4 is shown in the following formula:

[0032]

[0033] In the formula, t is the number of iterations, q is a probability randomly generated between (0,1); A=a·(2r1-1) is the coefficient vector, and a is the convergence factor that linearly decreases from 2 to 0. bl is a random number between [0,1], b is a logarithmic spiral constant, usually b=1, and l represents a random number between [-1,1]. ω represents the adaptive weighting strategy.

[0034] Furthermore, in the Levi flight search perturbation mechanism in step 2.4, α is the random step size, and s is the perturbation step size. The formula for calculating s is as follows:

[0035]

[0036] In the formula, u and v are random numbers distributed according to the standard normal distribution; λ is a random number between [0,2], and Γ() is the gamma function.

[0037] Furthermore, the mutation and improvement mechanism in step 2.5 is shown in the following equation:

[0038] X new =X(t+1)(1+Gaussion(τ)), (8)

[0039]

[0040] In the formula, X new The whale population is updated to X(t+1), where Gaussian function is used and τ is the Gaussian kernel parameter; X(t+1) is the whale population for the next iteration.

[0041] Furthermore, the structural risk loss function in step 3 is shown in the following equation:

[0042]

[0043] In the formula, p is the innovation length, and β = [β1, β2, ..., β i ] T ,β i =[β i1 ,β i2 ,…,β it Y(p,j) and Φ(p,j) are the output sliding window and feature mapping matrix based on the innovation length p, respectively; the output sliding window Y(p,j) and feature mapping matrix Φ(p,j) are defined as follows:

[0044]

[0045] In the formula, h(j) is the explicit mapping vector;

[0046] Solving the structural risk loss function equation (10), we obtain the output weight β:

[0047] β=(I / C+Φ(p,m)Φ T (p,m)) -1 Y(p,m) (12).

[0048] Furthermore, the identification and update of the output weights in step 3 specifically includes:

[0049] Step 3.1: Obtain a training dataset containing m samples, and denote the training dataset as (x... i ,y i ), i = 1, 2, ..., m, where x i =[x i1 ,x i2 ,...,x in ] T y i =[y i1 ,y i2 ,…,y it ] T We take the classification dataset of d = m / 10 as the test dataset and the remaining 9m / 10 of the classification dataset as the training dataset for real-time training updates.

[0050] Step 3.2: Transform the explicit feature mapping method into an implicit mapping method; the training dataset is the input data, and the kernel function k(x) is used. i ,x j Construct a kernel matrix Ω = HH with the input data. T Through formula derivation, combined with the kernel matrix Ω under multiple innovation theory, m Choose the kernel matrix Ω d The initial output weight β1 = (I) is obtained using formula (13). p / C+Ω d ) -1 Y d ;

[0051] β=(I / C+Ω) -1 Y (13)

[0052]

[0053] Step 3.3: Based on the kernel matrix Ω d And the initial output weight β1 calculation error E1=Y d -Ω d •β1, The error term E is calculated after each update of the kernel matrix Ω and output weight β. m Then, calculate the new kernel matrix Ω according to formula (14). m Through derivation based on the kernel matrix Ω m Error term E m and β m-1 For the output weight β m Perform online updates.

[0054] Furthermore, in step 3.3, the output weight β m Represented as:

[0055]

[0056] In the formula, n m =β m-1 Ω m (:,1:p(m-1)),r n =I p / C+Ω m (:,p(m-1)+1:pm).

[0057] In the above scheme, this invention proposes a classifier parameter identification method based on Extreme Learning Machine (ELM). It optimizes the ELM classifier by combining multiple innovation theory, improves the utilization of past data by establishing a sliding data window, and increases the connection between current data and previously constructed models. After preprocessing the collected classification dataset, the optimization method is used for online training to complete the real-time model construction, thereby completing the classification of the data to be classified and improving the accuracy of classifier parameter identification. Simultaneously, the whale optimization algorithm is used to optimize model parameter initialization to achieve optimal model performance.

[0058] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0059] The training and updating process of the Extreme Learning Machine (ELM) model in this invention reduces unnecessary repetitive training, saves time, solves the problem of the limited ability of the ELM to handle unfavorable data with poor correlation, improves the classification accuracy of the ELM model, and uses a mutation improvement mechanism for iterative updates to enable the ELM model of this invention to achieve the optimal classification effect. Attached Figure Description

[0060] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the invention. To better illustrate this embodiment, some components in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0061] Figure 1 This is a schematic diagram of the classifier parameter identification method based on extreme learning machine of the present invention;

[0062] Figure 2 This is a schematic diagram illustrating the classification of the Ecoli dataset under different p-values ​​according to the present invention;

[0063] Figure 3 This is a schematic diagram comparing the optimization performance of the improved whale algorithm of this invention with other group-based optimization algorithms on the Musk (Version1) dataset. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] Example 1

[0066] For easier understanding, please refer to Figure 1 A method for classifier parameter identification based on extreme learning machine includes the following steps:

[0067] Step 1: Obtain the classification dataset and divide it into a training dataset (X, Y) and a test dataset (TX, TY). In the training dataset, each element is the input (x, y) for each online training session, containing m data points. x is an n-dimensional column vector, where n represents the number of features in each input sample; y is a t-dimensional output column vector, where t represents the number of classes. Generally, y represents a label vector. Normalize the data in both the training and test datasets.

[0068] The classification dataset used is the UCI (University of California Irvine) dataset.

[0069] Step 2: Construct the Extreme Learning Machine model and use the improved whale optimization algorithm to obtain the optimal initialization parameters of the multi-innovation Extreme Learning Machine: regularization factor C and kernel function parameter γ; the improvement process of the whale optimization algorithm is as follows: introduce an adaptive weight strategy and a Levy flight search perturbation mechanism to update the position of the whale individual in the whale optimization algorithm;

[0070] The Extreme Learning Machine (ELM) model is shown below:

[0071] Hβ=Y (1)

[0072] In the formula, Y is the extreme learning machine model, i.e., a linear matrix equation, and β = [β1, β2, ..., β]. i ] T β represents the output weights of the extreme learning machine. i =[β i1 ,β i2 ,…,β it The dimension of the output weights of the Extreme Learning Machine depends on the number of hidden layer nodes i and the number of output label classes t; H is the feature mapping matrix, expressed as:

[0073]

[0074] In the formula, g() is the activation function, and (a,b) are the input weights between the input layer and the hidden layer of the extreme learning machine. The input weights are randomly assigned.

[0075] By solving the linear matrix equation of the extreme learning machine model (1), the least squares solution β is obtained. The least squares solution β is the output weight of the extreme learning machine. The output weight of the extreme learning machine is transformed into the following form.

[0076] β=H T (I / C+HH T ) -1 Y (3)

[0077] In the formula, C represents the regularization factor, which is a constant; I is the identity matrix of the corresponding dimension.

[0078] In the specific implementation process, obtaining the optimal initialization parameters for the Multiple Innovation Extreme Learning Machine includes the following steps: Step 2.1: Set the initialization parameters for the whale optimization algorithm. The initialization parameters include the whale population size N, the number of dimensions D, the maximum number of iterations T_max, and the upper and lower bounds U of the whale population in each dimension. d ;

[0079] Step 2.2: Use a grid search strategy to search for whale populations X = {X1, X2, ..., X} within the required area. N}, where X i (i = 1, 2, ..., N) represents the location information of the i-th individual whale, X i ={X i1 ,X i2 ,...,X iD}, X ij (j = 1, 2, ..., D) represents the position information of the j-th dimension of the i-th individual whale;

[0080] Step 2.3: Calculate the fitness function value F for each individual whale in the whale population using the F-score criterion, and select and retain the optimal fitness value F. best The optimal fitness value F best The corresponding optimal individual whale X best As the current globally optimal solution;

[0081] In the specific implementation process, the fitness function value F is:

[0082]

[0083] In the formula, Classes represents the number of categories; Recall i For recall rate, Precision iFor precision, it is expressed as:

[0084]

[0085] Among them, TP i FP i TN i and FN i These represent the number of true positive, false positive, true negative, and false negative results in the classification, respectively.

[0086] Step 2.4: Introduce an adaptive weighting strategy and the Lévy flight search perturbation mechanism to update the position of individual whales in the whale optimization algorithm, thus improving the algorithm; that is, introduce an adaptive weighting strategy to apply to the optimal whale individual X. best Adjust the optimal individual whale X best The weight ratio is adjusted to fully leverage the role of individual whales in location updates; the Levy flight search perturbation mechanism is introduced to enhance the whale optimization algorithm's ability to escape local optima;

[0087] In the specific implementation process, the position update of an individual whale is shown in the following formula:

[0088]

[0089] In the formula, t is the number of iterations, q is a probability randomly generated between (0,1); A=a·(2r1-1) is the coefficient vector, and a is the convergence factor that linearly decreases from 2 to 0. bl is a random number between [0,1], b is a logarithmic spiral constant, usually b=1, and l represents a random number between [-1,1]. ω represents the adaptive weighting strategy.

[0090] In the specific implementation process, in the Levy flight search perturbation mechanism in step 2.4, α is the random step size, s is the perturbation step size, and the formula for calculating s is as follows:

[0091]

[0092] In the formula, u and v are random numbers distributed according to the standard normal distribution; λ is a random number between [0,2], and Γ() is the gamma function.

[0093] Step 2.5: Iteratively update the whale optimization algorithm using a mutation improvement mechanism, that is, perform Gaussian mutation based on manually set conditional probabilities, and use a survival-of-the-fittest mechanism. If the updated or mutated whale population X... t Its fitness value is better than the best whale individual X in the previous iteration. best Then the current global optimal solution is replaced by X. t If the whale population X is updated or mutated tThe fitness value is worse than the best whale individual X in the previous iteration. best Then the current global optimal solution is the best individual whale X from the previous iteration. best The update continues until the number of iterations t reaches the maximum number of iterations T_max.

[0094] In practical implementation, the mutation and improvement mechanism is as follows:

[0095] X new =X(t+1)(1+Gaussion(τ)), (8)

[0096]

[0097] In the formula, X new The whale population is updated to X(t+1), where Gaussian function is used and τ is the Gaussian kernel parameter; X(t+1) is the whale population for the next iteration.

[0098] It should be noted that this invention improves the whale optimization algorithm by introducing adaptive weights and the Levy flight mechanism in the position update stage, and a Gaussian mutation is performed with probability (human-set) after the position update.

[0099] Step 3: Train the Extreme Learning Machine model online using the training dataset, evaluate the Extreme Learning Machine model using the structural risk loss function, identify and update the output weights, complete the classification training of the obtained dataset, and improve the parameters of the Extreme Learning Machine model.

[0100] The structural risk loss function is shown in the following equation:

[0101]

[0102] In the formula, p is the innovation length, and β = [β1, β2, ..., β i ] T ,β i =[β i1 ,β i2 ,…,β it Y(p,j) and Φ(p,j) are the output sliding window and feature mapping matrix based on the innovation length p, respectively; the output sliding window Y(p,j) and feature mapping matrix Φ(p,j) are defined as follows:

[0103]

[0104] In the formula, h(j) is the explicit mapping vector;

[0105] Solving the structural risk loss function equation (10), we obtain the output weight β:

[0106] β=(I / C+Φ(p,m)Φ T (p,m)) -1 Y(p,m) (12).

[0107] It should be noted that in equation 11, j is a number from 1 to m. When j = m, √(p,j) and √(p,m) are the same.

[0108] Specifically, the identification and update of output weights in step 3 includes:

[0109] Step 3.1: Obtain a training dataset containing m samples, and denote the training dataset as (x... i ,y i ), i = 1, 2, ..., m, where x i =[x i1 ,x i2 ,...,x in ] T y i =[y i1 ,y i2 ,…,y it ] T We take the classification dataset of d = m / 10 as the test dataset and the remaining 9m / 10 of the classification dataset as the training dataset for real-time training updates.

[0110] Step 3.2: Transform the explicit feature mapping method into an implicit mapping method; the training dataset is the input data, and the kernel function k(x) is used. i ,x j Construct a kernel matrix Ω = HH with the input data (γ). T Through formula derivation, combined with the kernel matrix Ω under multiple innovation theory, m Choose the kernel matrix Ω d The initial output weight β1 = (I) is obtained using formula (13). p / C+Ω d ) -1 Y d ;

[0111] β=(I / C+Ω) -1 Y (13)

[0112]

[0113] Step 3.3: Based on the kernel matrix Ω d And the initial output weight β1 calculation error E1=Y d -Ω d •β1, The error term E is calculated after each update of the kernel matrix Ω and output weight β. mThen, calculate the new kernel matrix Ω according to formula (14). m Through derivation based on the kernel matrix Ω m Error term E m and β m-1 For the output weight β m Perform online updates.

[0114] In the specific implementation process, the output weight β m Represented as:

[0115]

[0116] In the formula, n m =β m-1 Ω m (:,1:p(m-1)),r n =I p / C+Ω m (:,p(m-1)+1:pm);

[0117] It should be noted that Ip is a p-dimensional identity matrix, and Ω... m (:,1:p(m-1)) represents the matrix Ω m The matrix consisting of all rows and all elements from the first column to the p(m-1)th column, followed by Ω. m The explanation for (:,p(m-1)+1:pm) is similar. The kernel function k(x) i ,x j Construct a kernel matrix Ω = HH with the input data (γ). T It is a transformation from extreme learning machine to extreme learning machine, from explicit mapping to implicit mapping, hence the existence of multi-innovation (core) extreme learning machines.

[0118] If a new similar dataset is introduced that requires retraining, the parameter identification method of the present invention can be used to continue updating the model parameters in step 3.3 on the basis of the already trained Extreme Learning Machine model.

[0119] Step 4: Input the test dataset to be classified into the multi-innovation extreme learning machine model trained online in Step 3, and identify the category of the test dataset online. If there is new data input, repeat Step 3 to classify the new input classification dataset.

[0120] The training and updating process of the Extreme Learning Machine (ELM) model in this invention reduces unnecessary repetitive training, saves time, solves the problem of the limited ability of the ELM to process unfavorable data with poor correlation, and improves the classification accuracy of the ELM model.

[0121] Example 2

[0122] Specifically, based on Example 1, the solution will be described in conjunction with specific embodiments to further demonstrate its technical effects. Specifically:

[0123] In another embodiment of the present invention, Figure 2 This diagram illustrates the classification of samples in the Ecoli dataset under different innovation lengths (p). It shows that with p=1, the number of classification errors is relatively high (samples not coinciding with AC in the diagram are considered classification errors; EC represents the predicted results). Compared to p=1, the parameter identification performance improves with increasing p value (p=7 is optimal, corresponding to the Ecoli dataset accuracy in Table 2), proving the effectiveness of the algorithm presented in this invention.

[0124] Figure 3 This is a graphical representation of the performance of the improved whale algorithm on the Musk (Version 1) dataset, comparing it with other swarm-based optimization algorithms. Figure 3 The suffix MIKOSELM in the legend refers to the online extreme learning machine with multiple novel kernels for classifier parameter identification proposed in this invention; the prefixes (DE, PSO, GA, MWOA, WOA) represent the swarm intelligence optimization algorithms combined, namely differential evolution algorithm, particle swarm optimization algorithm, genetic algorithm, whale optimization algorithm of this invention, and whale optimization algorithm, respectively.

[0125] To verify the effectiveness of the present invention, simulation test cases were conducted. The first objective of the test cases was to verify that the Extreme Learning Machine (ELM) algorithm improves the classification accuracy relative to the Extreme Learning Machine (i.e., p=1) under different innovation lengths p. The dataset for the first test case was selected from some classification datasets in the UCI database. The detailed contents of the dataset are shown in the table below:

[0126] Table 1 UCI Dataset Information

[0127] UCI dataset Training dataset test set category property Ecoli 235 101 8 7 Parkinsons 136 59 2 23 WDBC 398 171 2 30 Musk (Version 1) 333 143 2 166

[0128] In Test Case 1, the same initialization values ​​were selected for the test cases. The kernel function k() was set to the RBF kernel function, and the penalty parameter and kernel parameter were set to 2 and 8 respectively. The test case results are shown in the table below:

[0129] Table 2 Test Set Accuracy

[0130]

[0131] The test results show that for the classification dataset used, the accuracy of the Extreme Learning Machine algorithm (p>1) is improved to a certain extent (1.34%~7.92%). Different datasets achieve an optimal improvement with different p values, which proves the effectiveness and flexibility of the Extreme Learning Machine algorithm of this invention and is also a new idea for improving the accuracy of machine learning algorithms.

[0132] The second objective of the test case is to verify the effectiveness of the whale optimization algorithm. Test case two uses the LowResolutionSpectromete dataset from the UCI database to compare the optimization performance with other swarm-based optimization algorithms under the same parameters. The initialization parameters for the whale optimization algorithm are: whale population size set to 20, maximum number of iterations set to 25; for the particle swarm optimization (PSO) algorithm, both the self-learning coefficient and the global learning coefficient are set to 2, the inertia coefficient is set to 1, and the maximum initial velocity is limited to 1 / 10 of the value range length; the crossover rate and mutation rate in the differential evolution algorithm (DE) are set to 0.3 and 0.5, respectively; in the genetic algorithm (GA), the crossover rate is set to 0.8, and the mutation rate is set to 0.05. The resulting curves for test case two are shown below. Figure 3 ;

[0133] Test Example 2 shows that the Whale Optimization Algorithm (MWOA) used in this invention has a larger output weight and a faster convergence speed compared to the Differential Evolution Algorithm (DE); it also has a more significant and better optimization effect compared to the Particle Swarm Optimization Algorithm (PSO) and the Genetic Algorithm (GA). Therefore, the method of this invention can effectively optimize model parameters.

[0134] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for classifier parameter identification based on Extreme Learning Machine, characterized in that, Includes the following steps: Step 1: Obtain the classification dataset and divide it into training and testing datasets. In the training dataset, each element is the input (x, y) for each online training session, containing m data points. x is an n-dimensional column vector, where n represents the number of features in each input sample; y is a k-dimensional output column vector, where k represents the number of categories. Normalize the data in both the training and testing datasets. Step 2: Construct the Extreme Learning Machine model and use the improved whale optimization algorithm to obtain the optimal initialization parameters of the multi-innovation Extreme Learning Machine: regularization factor C and kernel function parameter γ; the improvement process of the whale optimization algorithm is as follows: introduce an adaptive weight strategy and a Levy flight search perturbation mechanism to update the position of the whale individual in the whale optimization algorithm; Step 3: Train the Extreme Learning Machine model online using the training dataset, evaluate the Extreme Learning Machine model using the structural risk loss function, identify and update the output weights, and complete the classification training of the obtained dataset. The structural risk loss function in step 3 is shown in the following formula: （10） In the formula, p For the length of the new message, ,Y ( p,j ) and Φ( p,j These are the output sliding window and the feature mapping matrix, respectively, based on the innovation length p; the output sliding window Y ( p,j ) and the feature mapping matrix Φ( p,j The definition is as follows: （11） In the formula, h ( j ) is an explicit mapping vector; Solve the structural risk loss function equation (10) to obtain the output weights. β : （12）； In the formula, m represents the dimension of the input vector. y(j) Indicates the output vector. Represents the feature data mapping window, This represents the output data window; Step 4: Input the test dataset to be classified into the multi-innovation extreme learning machine model trained online in Step 3, and identify the category of the test dataset online. If there is new data input, repeat Step 3 to classify the new input classification dataset.

2. The method for classifier parameter identification based on Extreme Learning Machine according to claim 1, characterized in that, The basic model of the Extreme Learning Machine in step 2 is shown below: （1） In the formula, Y represents the Extreme Learning Machine model, i.e., a linear matrix equation. The output weights of the extreme learning machine. The dimension of the output weights of an Extreme Learning Machine depends on the number of hidden layer nodes. i and the number of output tag classes k； H is the feature mapping matrix, represented as: （2） In the formula, g () is the activation function; By solving the linear matrix equation containing the Limit Learning Machine model (1), the least squares solution is obtained. β Least squares solution β This refers to the output weights of the Extreme Learning Machine, which are transformed into the following form; （3） In the formula, C The regularization factor is a constant; I is the identity matrix of the corresponding dimension.

3. The method for classifier parameter identification based on extreme learning machine according to claim 2, characterized in that, Step 2 obtains the optimal initialization parameters for the Multiple Innovations Extreme Learning Machine, which includes the following steps: Step 2.1: Set the initialization parameters for the whale optimization algorithm, including the whale population size. N Dimensions D Maximum number of iterations T_max The upper and lower limits of whale population values ​​in various dimensions U d ; Step 2.2: Use a grid search strategy to search for whale populations within the required area. ,in, X i Indicates the first i The location information of individual whales, where i = 1, 2, ..., N. , X ij Indicates the first i The first individual whale j The positional information of the dimension, where j = 1, 2, ..., D; Step 2.3: Use F-score The evaluation criteria involve calculating the fitness function value F of individual whales in each whale population and selecting and retaining the optimal fitness value. F best The optimal fitness value F best The corresponding optimal individual whale X best As the current globally optimal solution; Step 2.4: Introduce an adaptive weight strategy and a Lévy flight search perturbation mechanism to update the position of individual whales in the whale optimization algorithm, thus improving the algorithm; that is, introduce an adaptive weight strategy to apply to the optimal whale individual. X best Adjusting to the optimal individual whale X best The weight ratio is adjusted to fully leverage the role of individual whales in location updates; the Levy flight search perturbation mechanism is introduced to enhance the whale optimization algorithm's ability to escape local optima; Step 2.5: Iteratively update the whale optimization algorithm using a mutation improvement mechanism, that is, perform Gaussian mutation based on manually set conditional probabilities, and use a survival-of-the-fittest mechanism. If the updated or mutated whale population... X t The fitness value is better than the best whale individual in the previous iteration. X best Then the current global optimal solution is replaced by X t If the whale population is updated or mutated X t The fitness value is worse than the best whale in the previous iteration. X best Then the current global optimal solution is the best whale individual from the previous iteration. X best until the number of iterations... t Reaching the maximum number of iterations T_max Update ends at this time.

4. The method for classifier parameter identification based on Extreme Learning Machine according to claim 3, characterized in that, The fitness function value F in step 2.3 is: （4） In the formula, Classes represents the number of categories; Recall i For recall rate, Precision i For precision, it is expressed as: （5） Among them, TP i FP i TN i and FN i These represent the number of true positive, false positive, true negative, and false negative results in the classification, respectively.

5. The method for classifier parameter identification based on Extreme Learning Machine according to claim 4, characterized in that, The position update of the individual whales in step 2.4 is shown in the following formula: （6） In the formula, t Let q be the number of iterations, and q be a probability randomly generated between (0,1). Let be the coefficient vector, and 'a' be the linear convergence factor decreasing from 2 to 0. ; bl A random number between [0,1] b It is a logarithmic spiral constant. b =1, l Represents a random number between [-1, 1]; ω represents the adaptive weighting strategy.

6. The method for classifier parameter identification based on extreme learning machine according to claim 5, characterized in that, In step 2.4, in the Levy flight search perturbation mechanism, α is the random step size. s For the perturbation step size, s The calculation formula is as follows: （7） In the formula, u and v These are random numbers distributed according to a standard normal distribution. λ A random number between [0, 2]. Γ () represents the gamma function. This represents the scale parameter.

7. The method for classifier parameter identification based on extreme learning machine according to claim 6, characterized in that, The mutation and improvement mechanism in step 2.5 is shown in the following formula: （8） （9） In the formula, X new Update the whale population for X(t+1). Gaussion () represents the Gaussian function, and τ represents the Gaussian kernel parameter; X ( t +1) represents the whale population for the next iteration, and F() represents the fitness function.

Images