Excitation waveform estimation method based on dimension reduction data driving

Abstract

The invention provides an excitation waveform estimation method based on dimension reduction data driving, and belongs to the technical field of magnetic electrical elements. The technical problem that a traditional model method is insufficient in accuracy and anti-interference capacity is solved. According to the technical scheme, the method comprises the following steps of 1) performing dimension reduction preprocessing on original data to obtain a feature sample; 2) determining each layer structure of the MLP neural network; 3) improving a star-graffiti optimization algorithm by using a mating strategy; 4) training network parameters of the MLP neural network in combination with an improved star-graffiti optimization algorithm to obtain a preliminarily updated MLP neural network; and 5) predicting the excitation waveform of the magnetic core by using an FNOA-MLP neural network algorithm. The method has the beneficial effect that the data in low-rank approximation has stronger model expression ability.

Description

Technical Field

[0001] The present invention relates to the technical field of magnetic electrical components, and in particular to an excitation waveform estimation method based on dimension reduction data driving. Background Art

[0002] As global energy demand continues to rise, especially in the industrial and electronic fields, improving energy efficiency has become an urgent need. In this context, the importance of magnetic components, including transformers, inductors, magnetic cores, etc., in the efficiency of power conversion and transmission has become increasingly prominent. The rapid development of new energy technologies has a high dependence on efficient power conversion and storage systems, and magnetic components play a fundamental role in these systems. Magnetic components are key components to ensure the safe and stable operation of electronic equipment. Based on Faraday's law of electromagnetic induction, they are composed of magnetic cores, wires, brackets, etc., and are responsible for realizing the conversion of electrical energy and magnetic energy. According to their uses, magnetic components are mainly divided into electronic transformers and inductors. Electronic transformers use the principle of electromagnetic induction to realize the conversion or transmission of electrical energy, converting high voltage into low voltage; while inductors, as a kind of energy storage element, are mainly used for signal filtering, noise suppression, current stabilization and electromagnetic interference suppression.

[0003] In AC electromagnetic equipment, the excitation process is to generate a magnetic field by applying AC current to the winding. This process is essential for the normal operation of equipment such as transformers and motors. The excitation waveform can be any of a sine wave, square wave, triangle wave, sawtooth wave or complex waveform, among which the sine wave is the most common due to its compatibility with the ideal AC power supply. The shape of the excitation waveform has a significant impact on the efficiency of the equipment. Sinusoidal excitation can provide higher efficiency, while the appearance of irregular waveforms may cause additional heat loss, reduce the reliability of the equipment, and may cause increased vibration and noise. In addition, the excitation waveform also affects the power factor of the equipment. A good waveform helps maintain a high power factor, thereby reducing power loss. Analysis of the excitation waveform is a key step in evaluating equipment performance, implementing fault diagnosis, and optimizing operating parameters. Therefore, accurate estimation of the excitation waveform is essential to ensure equipment performance.

[0004] How to solve the above technical problems is the subject faced by the present invention. Summary of the invention

[0005] The present invention aims to solve the technical problems of insufficient anti-interference ability in the excitation waveform estimation method and excessive calculation amount of the traditional data-driven method. The present invention provides an excitation waveform estimation method based on dimension reduction data drive. The method introduces low-rank approximation technology based on rows and columns, and constructs an efficient data processing framework, aiming to reduce the complexity of large-scale data set processing and improve processing efficiency. By processing only a part of the original data to construct an approximate data representation while retaining the intrinsic structure of the data, this method is particularly suitable for dimension reduction processing of excitation waveform data, and can extract key feature parameters related to excitation waveform estimation. In addition, the present invention also improves the nutcracker algorithm, and uses the improved nutcracker algorithm to optimize the initial weights of each layer of the neural network to improve the training efficiency and generalization ability of the network; experimental results show that this neural network driven by dimension reduction data and intelligently optimized has higher accuracy in excitation waveform estimation, and its prediction error is significantly lower than that of the traditional MLP neural network and the MLP neural network optimized by the traditional swarm intelligence algorithm.

[0006] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is specifically: an excitation waveform estimation method based on dimension reduction data drive, comprising the following steps:

[0007] Step 1) Acquire experimental data measured from four different magnetic core materials, including magnetic flux density, magnetic waveform type, etc. Perform dimensionality reduction preprocessing on the original data to obtain feature samples;

[0008] Step 2) Designing the input layer of the MLP neural network based on the feature samples and determining the structure of each layer of the MLP neural network;

[0009] Step 3) Using mating strategy to improve the nutcracker optimization algorithm;

[0010] Step 4) The network parameters of the MLP neural network are trained in combination with the improved Nudibranch optimization algorithm to obtain a preliminarily updated MLP neural network. The feature samples are input into the preliminarily updated MLP neural network for training, and back propagation is performed according to a preset loss function to further update the MLP neural network;

[0011] Step 5) Use the FNOA-MLP neural network algorithm to predict the core excitation waveform.

[0012] The data structure of the reaction excitation waveform is sorted out, and the core data is extracted and classified according to the factors related to the excitation waveform to obtain the classified data. Excitation waveform types: sine wave, triangle wave and trapezoidal wave; magnetic flux density (B ti ): Magnetic flux density refers to the magnetic flux per unit area perpendicular to the magnetic field lines, with a total of 1024 sampling points (sampling at the same interval within one cycle); peak value (Bpeak ): refers to the maximum amplitude value reached by the waveform in a specific cycle; average value (B): reflects the central trend or DC component of the waveform in a complete cycle; fluctuation amplitude (B rms ): How much a waveform deviates from its mean value.

[0013] B peak =Max(B ti ) (1)

[0014]

[0015] Feature samples are extracted from the classified data and the data of the magnetic core itself.

[0016] The data indicators that can reflect the excitation waveform are mined by row- and column-based low-rank approximation decomposition. In the low-rank approximation, it is first necessary to select a portion of representative rows and columns from the original matrix A. The selection of these rows and columns can be determined by statistical leverage scores or other methods to ensure that they can represent the information of the entire matrix to the greatest extent. The original matrix A is approximated by selecting some columns (C), rows (R) and a diagonal matrix (U) of the original matrix A. A set of rows and columns is found whose combination can effectively capture the important characteristics of the matrix, such as compressing and simplifying the data set while retaining its core information. The decomposition mainly includes the following three steps: (1) Column selection: select a portion of columns from the matrix A to form the matrix C; (2) Row selection: select a portion of rows from the matrix A to form the matrix R; (3) Diagonal matrix construction: create a diagonal matrix U, whose non-diagonal elements are usually set to 0 or a small value to further compress information. Combining these three steps, we can approximate A as A≈CUR, where the multiplication of C, R and U can obtain a good approximation of the original matrix. The calculation steps are:

[0017] (1) Given an input matrix A∈K m*n , is a matrix with m rows and n columns. The importance of each column / row is estimated using the statistical leverage score.

[0018]

[0019] π j is the ratio of the 2-norm of the column to the 2-norm of all columns, is the jth coordinate vector in the i-th column. j As an indicator of its statistical influence, select the matrix C and matrix R that best represent the original data.

[0020] (2) Define matrix W as the matrix consisting of the intersection elements of matrix C and matrix R, that is, the intersection points of rows and columns that are selected simultaneously. Calculate the SVD decomposition of matrix W.

[0021] W=XΣY T (5)

[0022] Σ is a diagonal matrix containing the singular values ​​of the matrix W. X and Y are matrices consisting of left and right singular vectors, respectively. (3) Calculate the generalized inverse matrix. Calculate the generalized inverse matrix Σ based on the matrix Σ. + This is done by replacing the non-zero elements on the diagonal with their reciprocals.

[0023] (4) Combining the SVD decomposition results and the generalized inverse matrix Σ + , calculate the matrix

[0024] U=YΣX T (6)

[0025] U is a dense matrix that maps C and R to an approximate original matrix A.

[0026] (5) Use C, U and R matrices to reconstruct the approximation of the original matrix A, that is,

[0027] A=CUR=CYΣ + X T R (7)

[0028] Three matrices with low-rank approximation of rows and columns are obtained. The combination of three matrices C, U and R is used to approximate the original matrix A. The Spearman correlation coefficient is selected to analyze the correlation of various indicators, and the parameters that best represent the excitation waveform are selected, which are divided into magnetic flux density, peak value, average value and fluctuation amplitude after data dimensionality reduction.

[0029] As an excitation waveform estimation method based on dimension reduction data driving of the present invention, the specific steps of step 2) are as follows:

[0030] Based on the magnetic flux density, peak value, average value and fluctuation amplitude, the number of hidden layer nodes of the MLP neural network is determined according to the number of feature samples and the number of nodes in the output layer.

[0031]

[0032] Among them, N is the number of hidden layer nodes; n is the number of output layer nodes; m is the number of input layer nodes.

[0033] As an excitation waveform estimation method based on dimension reduction data driving of the present invention, the specific steps of step 3) are as follows:

[0034] Intelligent algorithms can optimize the initial weights of each layer of the neural network to improve model performance and training efficiency. Swarm intelligence algorithms are an optimization method inspired by the behavior of biological groups in nature. In neural networks, these algorithms are often used to find the optimal weights and bias parameters to achieve higher prediction accuracy and faster training speed. Each particle represents a set of weights in the network, and the position of the particle corresponds to the value of the weight. By continuously iteratively updating the position and speed of the particles, the weight configuration that minimizes the network error is eventually found. Neural networks combined with intelligent optimization algorithms can overcome the problem that traditional back propagation is prone to falling into local optimal solutions.

[0035] Nutcracker Optimization Algorithm (NOA) is a population-based heuristic algorithm that simulates the behavior of nutcrackers in finding, storing, and recovering food in different seasons. In the NOA algorithm, each particle represents a set of weights in the network, and the position of the particle corresponds to the value of the weight. NOA mainly consists of two main stages: the first stage foraging and storage strategy, and the second stage cache and retrieval strategy.

[0036] In the first stage of foraging, the individual's position is initialized.

[0037]

[0038] represents the average position of the population in the tth iteration; A, B, and C are random individuals that are not equal to each other and are not equal to i; γ is a random number generated by the Levy flight strategy; μ and r are random numbers from 0 to 1; U is the upper limit of the problem dimension; L is the lower limit of the problem dimension. Randomly initialized populations can enhance global search capabilities.

[0039] The first stage is the storage phase, which involves storing the food obtained through foraging. If a seed is found, the nutcracker will transport it to the storage area and bury it:

[0040]

[0041] represents the new position of the storage area of ​​individual i in the tth iteration; represents the optimal position of the individual in the tth iteration; τ1, τ2, τ3 are random numbers between 0 and 1; l represents a factor that decreases linearly from 1 to 0.

[0042] In the second stage of caching, Nutcracker memorizes by updating two reference points:

[0043]

[0044] represents the first reference point of individual i in the tth iteration; represents the second reference point of individual i in the tth iteration; RP represents the random position; θ is a random number from 0 to Π.

[0045]

[0046] Using α to ensure the convergence of the NOA algorithm and optimize the reference point for the next iteration. Using linear reduction iterations improves the convergence speed of the algorithm to avoid falling into local minima.

[0047] As an excitation waveform estimation method based on dimension reduction data driving of the present invention, the specific steps of step 3) are as follows:

[0048] In the second stage of retrieval, the ordinary NOA algorithm determines which position is better by comparing the fitness values ​​of the reference point and the current position, and assigns the best position point to the next generation, so that the nutcracker individuals can always search in a better area, but this will waste the reference point and the excellent parts of other individuals outside the optimal individual, making the convergence speed slower. This paper gives a new strategy for the retrieval stage of the nutcracker, the fertility strategy (Nutcracker Optimization Algorithm with Fertility Strategy, FNOA). In each generation of individuals and reference points, the two individuals with the best fitness are selected for mating, and their optimal parts are combined and assigned to the next generation. The new individuals are as follows:

[0049]

[0050] β and β1 are two numbers randomly generated according to normal distribution, e l is the exponential constant, V1 and V2 are generated by:

[0051]

[0052]

[0053] The reproductive strategy reorganizes the best substances to retrieve the next generation with the best characteristics, which can greatly speed up the convergence of the algorithm.

[0054] The fitness value is calculated as follows:

[0055]

[0056] Among them, Y is the real data, is the output of the neural network after the weight represented by the current particle is brought into the neural network, and n is the number of data samples. The smaller the fitness, the higher the quality of the current solution. Thus, the optimal particle located at the optimal position is selected. The above process continues until the preset total number of iterations is met or the fitness function reaches the convergence condition, where the convergence condition is fitness = le -4 . Thus, the optimal particle located at the optimal position is selected.

[0057] As an excitation waveform estimation method based on dimension reduction data driving of the present invention, the specific steps of step 4) are as follows:

[0058] Based on the MLP neural network, the number of neuron nodes is used as the input for initializing the population dimension. The position of each particle corresponds to the weight of the neural network node. The particle position is gradually updated using the Nutcracker algorithm principle, and the optimal position of the particle and the population is updated based on the fitness function evaluation results. This process continues until the preset total number of iterations is met or the fitness function reaches the convergence condition, thereby obtaining a preliminarily optimized MLP neural network model.

[0059] The feature samples after dimension reduction preprocessing are input into this model for training. Back propagation is performed according to the preset loss function until the total number of iterations or the loss function convergence condition is met, and finally the updated MLP neural network prediction model is obtained.

[0060] As the excitation waveform estimation method based on dimension reduction data driving of the present invention, the specific steps of step 5) are as follows:

[0061] The feature parameters extracted in step 1) are used as the training set data for model training, and the excitation waveform type is used as the target value for model training. When the model is repeatedly trained on the same data sequence, it may overfit to these specific data samples and ignore other potential samples that are equally important for generalization. The training set data is shuffled each time during training, which enhances its generalization ability and reduces the risk of overfitting.

[0062] The predicted categories and actual categories are used to construct the error cross entropy loss function to measure the difference between the probability distributions:

[0063]

[0064] Among them, m is the total number of samples, p(x i ) is the type distribution of the i-th real data, q(x i ) is the category distribution predicted by the ith model.

[0065] Compared with the prior art, the present invention has the following beneficial effects:

[0066] (1) When processing 1024 or more high-dimensional magnetic flux density data, the traditional data-driven method usually requires a large amount of labeled data to train the model. The low-rank approximation decomposition data index based on rows and columns is introduced. It is only necessary to extract some representative rows and columns from the original data to construct a low-rank approximation matrix, and extract feature samples, thereby realizing efficient compression and analysis of the original data. This sampling method not only reduces the amount of required data, but also maintains the structure and internal relationship of the data, so that the decomposed feature parameters are accurate and stable.

[0067] (2) The present invention introduces an improvement strategy into the intelligent optimization algorithm, which can avoid the waste of high-quality solutions during the algorithm iteration process, accelerate the algorithm convergence speed, and make the algorithm more stable.

[0068] (3) The present invention introduces an intelligent optimization algorithm in the neural network process to optimize the initial weights of each layer of the neural network, which not only improves the model performance and training efficiency, but also ensures the superiority of the algorithm in terms of accuracy and convergence speed. The final optimization is the FNOA-MLP neural network algorithm, which is used to complete waveform estimation with high recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0070] Figure 1 It is a diagram showing the effect of the improved strategy provided by the present invention on different test functions.

[0071] Figure 2 It is a flow chart of the FNOA-MLP algorithm provided by the present invention.

[0072] Figure 3 Comparison diagram of the FNOA-MLP algorithm estimation and the actual waveform provided by the present invention.

[0073] Figure 4 A comparison chart of the error between the FNOA-MLP algorithm estimate and the true waveform provided by the present invention. DETAILED DESCRIPTION

[0074] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. Of course, the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0075] Example 1

[0076] The object of study in this embodiment is a magnetic electrical component. In AC electromagnetic equipment, the excitation process is to generate a magnetic field by applying an AC current to the winding. This process is essential for the normal operation of equipment such as transformers and motors. The excitation waveform can be any of a sine wave, square wave, triangle wave, sawtooth wave or complex waveform.

[0077] See also Figures 1 to 4 This embodiment provides an excitation waveform estimation method based on dimension reduction data driving, comprising the following steps:

[0078] Step 1) Acquire experimental data measured from four different magnetic core materials, including magnetic flux density, magnetic waveform type, etc. Perform dimensionality reduction preprocessing on the original data to obtain feature samples;

[0079] Step 2) Designing the input layer of the MLP neural network based on the feature samples and determining the structure of each layer of the MLP neural network;

[0080] Step 3) Using mating strategy to improve the nutcracker optimization algorithm;

[0081] Step 4) The network parameters of the MLP neural network are trained in combination with the improved Nudibranch optimization algorithm to obtain a preliminarily updated MLP neural network. The feature samples are input into the preliminarily updated MLP neural network for training, and back propagation is performed according to a preset loss function to further update the MLP neural network;

[0082] Step 5) Use the FNOA-MLP neural network algorithm to predict the core excitation waveform.

[0083] Specifically, step 1) includes the following steps:

[0084] The magnetic core to be tested is in a circular shape, with the excitation winding and induction winding wound evenly on the core. The signal generator generates a sine wave or other waveform of a given frequency, which is applied to the excitation winding as an excitation source through a high-frequency power amplifier. According to Ampere's loop law, the magnetic field intensity acts on the magnetic core of the magnetic material to generate a magnetic flux density (B ti ), the magnetic flux per unit area perpendicular to the magnetic field direction, in one cycle, the same spacing is sampled, a total of 1024 sampling points. The magnetic core is made of four materials, each material is tested at different temperatures (25, 50, 70, 90), a total of 12400 sets of data. Peak (B peak ): refers to the maximum amplitude value reached by the waveform in a specific cycle; average value (B): reflects the central trend or DC component of the waveform in a complete cycle; fluctuation amplitude (B rms ): How much a waveform deviates from its mean value.

[0085] B peak =Max(Bti ) (1)

[0086]

[0087] Feature samples are extracted from the classified data and the data of the magnetic core itself.

[0088] Define A as A∈K 1024*12400 A matrix containing all the experimentally measured data. The importance of each column / row is estimated using a statistical leverage score.

[0089]

[0090] π j is the ratio of the 2-norm of the column to the 2-norm of all columns, is the jth coordinate vector in the i-th column. j As an indicator of its statistical influence, select the matrix C and matrix R that best represent the original data.

[0091] Define matrix W as the intersection of matrix C and matrix R, that is, the intersection of rows and columns that are selected at the same time. Compute the SVD decomposition of matrix W.

[0092] W=XΣY T (5)

[0093] Σ is a diagonal matrix containing the singular values ​​of the matrix W, and X and Y are matrices consisting of the left and right singular vectors respectively.

[0094] Calculate the generalized inverse matrix. Calculate the generalized inverse matrix Σ based on the matrix Σ. + This is done by replacing the non-zero elements on the diagonal with their reciprocals.

[0095] Combining the SVD decomposition result and the generalized inverse matrix Σ + , calculate the matrix

[0096] U=YΣX T (6)

[0097] U is a dense matrix that maps C and R to an approximate original matrix A.

[0098] Use C, U and R matrices to reconstruct the approximation of the original matrix A, that is

[0099] A=CUR=CYΣ + X T R (7)

[0100] Three matrices with low-rank approximation of rows and columns are obtained. The combination of three matrices C, U and R is used to approximate the original matrix A. The Spearman correlation coefficient is selected to analyze the correlation of various indicators, and the parameters that best represent the excitation waveform are selected, which are divided into magnetic flux density, peak value, average value and fluctuation amplitude after data dimensionality reduction.

[0101] Taking the magnetic flux density data after dimensionality reduction as an example, the specific steps of extracting voltage characteristic value indicators by dimensionality reduction are as follows: First, extract all data, Z j (j=1, 2, ..., n) is the magnetic flux density data of the jth test, and n is the total number of tests. j Decompose the magnetic flux density data Z for each test according to the principle of row and column low-rank approximation. j A singular value υ can be obtained to represent its characteristic information i Finally, the data obtained from all the tests are arranged in order to form a matrix. The Spearman correlation coefficient is selected to analyze the correlation of each index, and the parameters that best represent the excitation waveform are selected, which are the magnetic flux density, peak value, average value and fluctuation amplitude after dimensionality reduction.

[0102] Specifically, step 2) includes the following steps:

[0103] Based on the magnetic flux density, peak value, average value and fluctuation amplitude, the number of hidden layer nodes of the MLP neural network is determined according to the number of feature samples and the number of nodes in the output layer.

[0104]

[0105] Among them, N is the number of hidden layer nodes; n is the number of output layer nodes; m is the number of input layer nodes.

[0106] Specifically, step 3) includes the following steps:

[0107] Nutcracker Optimization Algorithm (NOA) is a population-based heuristic algorithm that simulates the behavior of nutcrackers in finding, storing, and recovering food in different seasons. In the NOA algorithm, each particle represents a set of weights in the network, and the position of the particle corresponds to the value of the weight. NOA mainly consists of two main stages: the first stage foraging and storage strategy, and the second stage cache and retrieval strategy.

[0108] In the first stage of foraging, the individual's position is initialized.

[0109]

[0110] represents the average position of the population in the tth iteration; A, B, and C are random individuals that are not equal to each other and are not equal to i; γ is a random number generated by the Levy flight strategy; μ and r are random numbers from 0 to 1; U is the upper limit of the problem dimension; L is the lower limit of the problem dimension. Randomly initialized populations can enhance global search capabilities.

[0111] The first stage is the storage phase, which involves storing the food obtained through foraging. If a seed is found, the nutcracker will transport it to the storage area and bury it:

[0112]

[0113] represents the new position of the storage area of ​​individual i in the tth iteration; represents the optimal position of the individual in the tth iteration; τ1, τ2, τ3 are random numbers between 0 and 1; l represents a factor that decreases linearly from 1 to 0.

[0114] In the second stage of caching, Nutcracker memorizes by updating two reference points:

[0115]

[0116] represents the first reference point of individual i in the tth iteration; represents the second reference point of individual i in the tth iteration; RP represents the random position; θ is a random number from 0 to Π.

[0117]

[0118] Using α to ensure the convergence of the NOA algorithm and optimize the reference point for the next iteration. Using linear reduction iterations improves the convergence speed of the algorithm to avoid falling into local minima.

[0119] In the second stage of retrieval, the ordinary NOA algorithm determines which position is better by comparing the fitness values ​​of the reference point and the current position, and assigns the best position point to the next generation, so that the nutcracker individuals can always search in a better area, but this will waste the reference point and the excellent parts of other individuals outside the optimal individual, making the convergence speed slower. This paper gives a new strategy for the retrieval stage of the nutcracker, the fertility strategy (Nutcracker Optimization Algorithm with Fertility Strategy, FNOA). In each generation of individuals and reference points, the two individuals with the best fitness are selected for mating, and their optimal parts are combined and assigned to the next generation. The new individuals are as follows:

[0120]

[0121] β and β1 are two numbers randomly generated according to normal distribution, e l is the exponential constant, V1 and V2 are generated by:

[0122]

[0123] The reproductive strategy reorganizes the best substances to retrieve the next generation with the best characteristics, which can greatly speed up the convergence of the algorithm.

[0124] The fitness value is calculated as follows:

[0125]

[0126] Among them, Y is the real data, is the output of the neural network after the weight represented by the current particle is brought into the neural network, and n is the number of data samples. The smaller the fitness, the higher the quality of the current solution. Thus, the optimal particle located at the optimal position is selected. The above process continues until the preset total number of iterations is met or the fitness function reaches the convergence condition, where the convergence condition is fitness = le -4 . Thus, the optimal particle located at the optimal position is selected.

[0127] Specifically, step 4) includes the following steps:

[0128] Based on the MLP neural network, the number of neuron nodes is used as the input for initializing the population dimension. The position of each particle corresponds to the weight of the neural network node. The particle position is gradually updated using the Nutcracker algorithm principle, and the optimal position of the particle and the population is updated based on the fitness function evaluation results. This process continues until the preset total number of iterations is met or the fitness function reaches the convergence condition, thereby obtaining a preliminarily optimized MLP neural network model.

[0129] The feature samples after dimension reduction preprocessing are input into this model for training. Back propagation is performed according to the preset loss function until the total number of iterations or the loss function convergence condition is met, and finally the updated MLP neural network prediction model is obtained.

[0130] Specifically, step 5) includes the following steps:

[0131] The feature parameters extracted in step 1) are used as the training set data for model training, and the excitation waveform type is used as the target value for model training. When the model is repeatedly trained on the same data sequence, it may overfit to these specific data samples and ignore other potential samples that are equally important for generalization. The training set data is shuffled each time during training, which enhances its generalization ability and reduces the risk of overfitting.

[0132] The predicted categories and actual categories are used to construct the error cross entropy loss function to measure the difference between the probability distributions:

[0133]

[0134] Among them, m is the total number of samples, p(x i ) is the type distribution of the i-th real data, q(x i ) is the category distribution predicted by the ith model.

[0135] In summary, it can be seen that the excitation waveform estimation method based on dimensionality reduction data driven has better results, the model accuracy is higher than the traditional neural network model, and it has certain engineering value.

[0136] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for estimating excitation waveform based on dimension reduction data drive, characterized in that: The following steps are involved: Step 1) obtaining experimental data, magnetic flux density, and magnetic waveform type measured from four different magnetic core materials, and performing dimensionality reduction preprocessing on the original data to obtain feature samples; Step 2) Designing the input layer of the MLP neural network based on the feature samples and determining the structure of each layer of the MLP neural network; Step 3) Using mating strategy to improve the nutcracker optimization algorithm; Step 4) training the network parameters of the MLP neural network in combination with the improved Nudibranch optimization algorithm to obtain a preliminarily updated MLP neural network, inputting the feature samples into the preliminarily updated MLP neural network for training, and performing back propagation according to a preset loss function to further update the MLP neural network; Step 5) Use the FNOA-MLP neural network algorithm to predict the core excitation waveform.

2. The excitation waveform estimation method based on dimension reduction data driving according to claim 1 is characterized in that: The step 1) comprises the following steps: 1-1) Sort out the data structure of the reaction excitation waveform, extract and classify the core data according to the factors related to the excitation waveform, and obtain the classified data; Excitation waveform types: sine wave, triangle wave and trapezoidal wave; magnetic flux density (B ti ): Magnetic flux density refers to the magnetic flux per unit area perpendicular to the magnetic field lines. There are 1024 sampling points in total. The same sampling interval is used in one cycle. Peak value B peak : refers to the maximum amplitude value reached by the waveform in a specific period; average value It reflects the central trend or DC component of the waveform in a complete cycle; the fluctuation amplitude B rms : the degree to which the waveform deviates from its mean value; B peak =Max(B ti ) (1) Extract feature samples from the classified data and the data of the magnetic core itself; 1-2) Use row- and column-based low-rank approximation decomposition to mine data indicators of the reaction excitation waveform. In the low-rank approximation, first select a portion of representative rows and columns from the original matrix A. The selection of these rows and columns is determined by statistical leverage scores or other methods to ensure that they represent the information of the entire matrix to the greatest extent. By selecting some columns C, rows R and a diagonal matrix U of the original matrix A to represent the original matrix A, find a set of rows and columns whose combination captures the important characteristics of the matrix. The decomposition includes the following three steps: (1) Column selection: select a part of the columns from matrix A to form matrix C; (2) Row selection: select a portion of rows from matrix A to form matrix R; (3) Diagonal matrix construction: Create a diagonal matrix U with non-diagonal elements set to 0 or small values ​​to further compress information. Combining these three steps, A is expressed as A≈CUR. The calculation steps are: (a) Given an input matrix A∈K m*n , is a matrix with m rows and n columns, and the importance of each column or row is estimated using the statistical leverage score; π j is the ratio of the 2-norm of the column to the 2-norm of all columns, is the jth coordinate vector in the i-th column, where π j As a measure of their statistical influence, the matrices C and R representing the original data are selected; (b) Define matrix W as the matrix consisting of the intersection elements of matrix C and matrix R, that is, the intersection points of rows and columns that are selected at the same time, and calculate the SVD decomposition of matrix W; W=XΣY T (5) Σ is a diagonal matrix containing the singular values ​​of the matrix W, and X and Y are matrices consisting of left and right singular vectors respectively; (c) Calculate the generalized inverse matrix. Calculate its generalized inverse matrix Σ based on the matrix Σ. + This is done by replacing the non-zero elements on the diagonal with their reciprocals; (d) Combining the SVD decomposition result and the generalized inverse matrix Σ + , calculate the matrix U=YΣ + X T (6) U is a dense matrix that maps C and R to the approximate original matrix A; (e) Use C, U and R matrices to reconstruct the approximation of the original matrix A, that is, A=CUR=CYΣ + X T R (7) A low-rank approximation decomposition based on rows and columns was obtained to mine data indicators that can reflect the excitation waveform. Among the three matrices in the low-rank approximation, the combination of three matrices C, U and R was used to approximate the original matrix A. The Spearman correlation coefficient was selected to analyze the correlation of various indicators, and the parameters that best represent the excitation waveform were selected, which are divided into flux density, peak value, average value and fluctuation amplitude after data dimensionality reduction.

3. The excitation waveform estimation method based on dimension reduction data driving according to claim 1 is characterized in that: The step 3) comprises the following: The Nutcracker Optimization Algorithm is a population-based heuristic algorithm that simulates the behavior of Nutcrackers in finding, storing, and recovering food in different seasons. In the NOA algorithm, each particle represents a set of weights in the network, and the position of the particle corresponds to the value of the weight. NOA consists of two main stages: the first stage foraging and storage strategy, and the second stage cache and retrieval strategy; In the first stage of foraging, the individual's position is initialized; represents the average position of the group in the tth iteration; A, B, and C are random individuals that are not equal to each other and are not equal to i; γ is a random number generated by the Levy flight strategy; μ and r are random numbers between 0 and 1; U is the upper limit of the problem dimension; L is the lower limit of the problem dimension, and the randomly initialized population enhances the global search capability; The first stage is the storage phase, which involves storing the food obtained through foraging. If seeds are found, the nutcrackers will transport them to the storage area and bury them: represents the new position of the storage area of ​​individual i in the tth iteration; represents the optimal position of the individual in the tth iteration; τ1, τ2, τ3 are random numbers between 0 and 1; l represents a factor that decreases linearly from 1 to 0; In the second stage of caching, Nutcracker memorizes by updating two reference points: represents the first reference point of individual i in the tth iteration; represents the second reference point of individual i in the tth iteration; RP represents the random position; θ is a random number from 0 to Π; Use α to ensure the convergence of the NOA algorithm, optimize the reference point of the next iteration, and use linear reduction iteration; In the second stage of retrieval, the common NOA algorithm determines which position is better by comparing the fitness values ​​of the reference point and the current position, and assigns the best position point to the next generation, so that the nutcracker individuals can always search in a better area. This gives the nutcrackers a new strategy for the retrieval stage. The reproductive strategy selects the two individuals with the best fitness in each generation and the reference point for mating, combines their best parts, and assigns them to the next generation. The new individuals are as follows: β and β1 are two numbers randomly generated according to normal distribution, e l is the exponential constant, V1 and V2 are generated by: The reproductive strategy reorganizes the best materials to retrieve the next generation with the best characteristics; The fitness value is calculated as follows: Among them, Y is the real data, is the output of the neural network after the weight represented by the current particle is brought into the neural network, n is the number of data samples, the smaller the fitness is, the higher the quality of the current solution is, so the optimal particle located at the optimal position is selected; the above process continues until the preset total number of iterations is met or the fitness function reaches the convergence condition, where the convergence condition is fitness = le -4 , thereby selecting the optimal particle located at the optimal position.

4. The excitation waveform estimation method based on dimension reduction data driving according to claim 1 is characterized in that: The step 4) comprises the following steps: Based on the multi-layer perceptron neural network, the number of neuron nodes is used as the input for initializing the population dimension. The position of each particle corresponds to the weight of the neural network node. The particle position is gradually updated using the Nutcracker algorithm principle, and the optimal position of the particle and the group is updated according to the fitness function evaluation result. This process continues until the preset total number of iterations is met or the fitness function reaches the convergence condition, thereby obtaining a preliminarily optimized MLP neural network model. The feature samples after dimensionality reduction preprocessing are input into this model for training, and back propagation is performed according to the preset loss function until the total number of iterations or the loss function convergence condition is met, and finally the updated MLP neural network prediction model is obtained.

5. The excitation waveform estimation method based on dimension reduction data driving according to claim 1 is characterized in that: The step 5) comprises the following: The feature parameters extracted in step 1) are used as the training set data for model training, and the excitation waveform type is used as the target value for model training. When the model is repeatedly trained on the same data sequence, it will over-fit to these specific data samples and ignore other potential samples that are equally important for generalization. The data of the training set is disrupted each time during training, which enhances its generalization ability and reduces the risk of overfitting. The predicted categories and actual categories are used to construct the error cross entropy loss function to measure the difference between the probability distributions: Among them, m is the total number of samples, p(x i ) is the type distribution of the i-th real data, q(x i ) is the category distribution predicted by the ith model.

Images