CN-122017700-A - J-A hysteresis model parameter identification method and system based on physical information neural network

CN122017700ACN 122017700 ACN122017700 ACN 122017700ACN-122017700-A

Abstract

The invention discloses a J-A hysteresis model parameter identification method and a J-A hysteresis model parameter identification system based on a physical information neural network, which belong to the technical field of data processing and aim at solving the problem that the calculation efficiency and the physical rationality are difficult to consider in the existing ferromagnetic material parameter identification technology; the method realizes the deep fusion of data driving and physical mechanism, accelerates the solving process by using a neural network, greatly improves the computing efficiency, ensures the physical interpretability of the parameters by physical equation constraint, effectively avoids the generation of non-physical solutions, and provides reliable technical support for high-precision hysteresis modeling of electromagnetic equipment.

Inventors

DONG JIANYANG
XU XIAOWEN
YANG SHIYOU

Assignees

浙江大学

Dates

Publication Date: 20260512
Application Date: 20260410

Claims (10)

1. The J-A hysteresis model parameter identification method based on the physical information neural network is characterized by comprising the following steps of: The method comprises the steps of obtaining hysteresis loop actual measurement data of a ferromagnetic material to be identified, wherein the hysteresis loop actual measurement data comprise magnetic field intensity and corresponding magnetization state thereof, taking the magnetic field intensity as input, taking the magnetization state as output, and embedding a differential equation of a J-A hysteresis model as a physical constraint loss term into a loss function to construct a physical information neural network; Taking the measured data of the hysteresis loop, a differential equation of a J-A hysteresis model and a priori physical range of the J-A hysteresis model parameter as parallel supervision information to construct an objective function containing network parameters and parameters to be identified of the J-A hysteresis model; Performing iterative training on the physical information neural network by minimizing the objective function, and synchronously updating network parameters of the physical information neural network and parameters to be identified of the J-A hysteresis model in the back propagation process of each iteration; And extracting a target optimized value of the parameter to be identified from the physical information neural network meeting the convergence condition as a parameter identification result of the ferromagnetic material to be identified.
2. The method for identifying parameters of a J-a hysteresis model based on a physical information neural network according to claim 1, wherein the step of constructing the physical information neural network by taking the magnetic field strength as an input, taking the magnetization state as an output, and embedding the differential equation of the J-a hysteresis model as a physical constraint loss term into a loss function comprises: constructing a neural network comprising an input layer, a plurality of fully connected hidden layers and an output layer; taking the normalized magnetic field intensity as the input of the neural network, taking the normalized magnetization state as the output of the neural network, and taking the parameters to be identified of the J-A hysteresis model as trainable variables of the neural network; Calculating a first derivative of the magnetization state of the output end to the magnetic field intensity of the input end through automatic differentiation, and simultaneously calculating a theoretical derivative of the magnetization state of the output end to the magnetic field intensity of the input end, which is determined by the J-A hysteresis model differential equation, according to the parameter to be identified of the current iteration; and constructing a physical constraint loss term by using a residual error between the first derivative and the theoretical derivative, and embedding the physical constraint loss term into an overall loss function of the neural network to construct the physical information neural network.
3. The method for identifying parameters of a J-a hysteresis model based on a physical information neural network according to claim 1, wherein the step of using the measured hysteresis loop data, differential equation of the J-a hysteresis model, and a priori physical range of the J-a hysteresis model parameters as parallel supervision information is as follows: constructing a data fitting loss term as first supervision information according to the deviation between the predicted magnetization state and the normalized measured magnetization state output by the physical information neural network; taking a physical constraint loss term embedded in the physical information neural network loss function as second supervision information; constructing a parameter boundary constraint loss term as third supervision information according to the deviation degree between the current value of the parameter to be identified and the prior physical range; And carrying out weighted combination on the first supervision information, the second supervision information and the third supervision information to obtain an objective function taking the network parameter and the parameter to be identified as optimization variables.
4. A method for identifying parameters of a J-a hysteresis model based on a physical information neural network according to claim 3, wherein said step of constructing a parameter boundary constraint loss term according to the degree of deviation between the current value of the parameter to be identified and the prior physical range is as follows: presetting a physical value interval for each parameter to be identified, wherein the interval comprises a lower limit value and an upper limit value; Judging whether the current value of each parameter to be identified falls into a corresponding physical value interval or not respectively; Calculating a lower boundary penalty amount based on a difference value between the lower limit value and the current value for parameters with the current value lower than the lower limit value, and calculating an upper boundary penalty amount based on a difference value between the current value and the upper limit value for parameters with the current value higher than the upper limit value; And respectively giving corresponding weight coefficients to the lower boundary penalty and the upper boundary penalty of each parameter, and then summing to generate the parameter boundary constraint loss term, wherein the weight coefficients are positively correlated with the lower boundary penalty or the upper boundary penalty.
5. The method for identifying parameters of a J-a hysteresis model based on a physical information neural network according to claim 1, wherein the step of iteratively training the physical information neural network with the objective function is as follows: configuring an adaptive moment estimation optimizer, and setting an initial learning rate and a learning rate attenuation strategy; In each iteration, inputting the normalized magnetic field intensity into the physical information neural network, calculating a predicted magnetization state through forward propagation, and calculating a total loss value of the current iteration according to the objective function; respectively calculating gradients of the total loss value to the network parameter and the parameter to be identified through a back propagation algorithm; based on the gradient of the parameter to be identified, synchronously updating the network parameter and the parameter to be identified by using the self-adaptive moment estimation optimizer; According to the learning rate attenuation strategy, the current learning rate is adjusted, and the weight coefficient of each loss item in the objective function is dynamically adjusted; Repeating the iterative process until the preset convergence condition is met.
6. The method for identifying parameters of a J-a hysteresis model based on a physical information neural network according to claim 5, wherein the steps of adjusting a current learning rate according to the learning rate decay strategy and dynamically adjusting weight coefficients of each loss term in the objective function are as follows: Based on a preset attenuation period and an attenuation factor, carrying out stepwise attenuation on the initial learning rate according to the current iteration times to obtain an updated current learning rate; Acquiring gradient norms of each loss item to network parameters in the current iteration; according to the gradient norm proportion of each loss term, recalculating the weight coefficient of each loss term in the objective function according to a gradient normalization method; And (3) acting the updated current learning rate and the recalculated weight coefficient on a parameter updating process of the subsequent iteration.
7. The method for identifying parameters of a J-A hysteresis model based on a physical information neural network according to claim 1, wherein the convergence condition comprises at least one of an objective function value lower than a first preset threshold, a physical constraint loss term value lower than a second preset threshold, a variation of the objective function value within a continuous preset iteration number lower than a third preset threshold, and the iteration number reaching a preset maximum iteration number.
8. The J-a hysteresis model parameter identification system based on a physical information neural network, which is applicable to the J-a hysteresis model parameter identification method based on the physical information neural network as claimed in any one of claims 1 to 7, and is characterized by comprising the following steps: the data acquisition module is used for acquiring the hysteresis loop actual measurement data of the ferromagnetic material to be identified; the network construction module is used for constructing a physical information neural network by taking the magnetic field intensity as input, taking the magnetization state as output and taking a differential equation of the J-A hysteresis model as a physical constraint loss term to embed a loss function; The objective function construction module is used for constructing an objective function containing network parameters and parameters to be identified of the J-A hysteresis model by taking the measured data of the hysteresis loop, a differential equation of the J-A hysteresis model and the prior physical range of the J-A hysteresis model parameters as parallel supervision information; the training optimization module is used for carrying out iterative training on the physical information neural network by minimizing the objective function, and synchronously updating network parameters of the physical information neural network and parameters to be identified of the J-A hysteresis model in the back propagation process of each iteration; and the parameter extraction module is used for extracting the target optimization value of the parameter to be identified from the physical information neural network meeting the convergence condition as the parameter identification result of the ferromagnetic material to be identified.
9. An electronic device comprising a processor, a memory and a computer program stored on the memory and executable on the processor, the computer program implementing the method of any one of claims 1-7 when executed by the processor.
10. A computer readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the method according to any one of claims 1-7.

Description

J-A hysteresis model parameter identification method and system based on physical information neural network Technical Field The invention relates to the technical field of data processing, in particular to a J-A hysteresis model parameter identification method and system based on a physical information neural network. Background In engineering applications such as power transformers, motors, electromagnetic devices and the like, the accurate representation of the hysteresis characteristics of ferromagnetic materials is a core foundation for equipment performance analysis and optimization design, and a Jiles-Athereon hysteresis model (J-A hysteresis model for short) is a mainstream hysteresis modeling scheme due to the fact that the physical meaning of parameters is clear and the application range is wide, and the parameter identification precision and efficiency directly determine modeling effects. The existing mainstream intelligent optimization identification methods such as genetic algorithm, particle swarm optimization and simulated annealing all convert parameter identification into pure numerical value optimization, and have inherent technical barriers that firstly identification relies on a large number of iterations and hysteresis curve forward calculation, the calculation efficiency is low, engineering real-time requirements are difficult to meet, secondly the parameter initial value and algorithm super-parameter are highly sensitive, the stability of identification results is poor, thirdly, only curve fitting errors are used as optimization targets, physical constraint of a J-A model is not integrated, unreasonable solution that mathematical fitting reaches the standard but physical meaning fails is easy to generate, and model prediction is extremely easy to fail under multiple conditions. The core concept of the prior art is to realize parameter optimization by minimizing hysteresis loop fitting errors through an intelligent optimization algorithm, and the inherent defect is that the physical constraint of a differential equation of a J-A model is not embedded into the whole optimizing process, so that the identification efficiency, the stability of a result and the physical rationality cannot be simultaneously considered, and the landing application of hysteresis modeling in the scenes of real-time monitoring, fault diagnosis and the like is severely restricted. Disclosure of Invention The invention aims to solve the problem that the existing ferromagnetic material parameter identification technology is difficult to consider the calculation efficiency and the physical rationality, and provides a J-A hysteresis model parameter identification method and a J-A hysteresis model parameter identification system based on a physical information neural network, which construct a joint loss function by taking hysteresis loop actual measurement data, a J-A model differential equation and a parameter priori range as parallel supervision information, and network parameters and J-A parameters to be identified are synchronously updated in the back propagation, so that the identification result is ensured to strictly meet the magnetization physical rule of the ferromagnetic material while the parameter identification time is greatly shortened, and the modeling precision and engineering application reliability are obviously improved. In a first aspect, the technical scheme provided in the embodiment of the invention is that the J-A hysteresis model parameter identification method based on a physical information neural network comprises the following steps: The method comprises the steps of obtaining hysteresis loop actual measurement data of a ferromagnetic material to be identified, wherein the hysteresis loop actual measurement data comprise magnetic field intensity and corresponding magnetization state thereof, taking the magnetic field intensity as input, taking the magnetization state as output, and embedding a differential equation of a J-A hysteresis model as a physical constraint loss term into a loss function to construct a physical information neural network; Taking the measured data of the hysteresis loop, a differential equation of a J-A hysteresis model and a priori physical range of the J-A hysteresis model parameter as parallel supervision information to construct an objective function containing network parameters and parameters to be identified of the J-A hysteresis model; Performing iterative training on the physical information neural network by minimizing the objective function, and synchronously updating network parameters of the physical information neural network and parameters to be identified of the J-A hysteresis model in the back propagation process of each iteration; And extracting a target optimized value of the parameter to be identified from the physical information neural network meeting the convergence condition as a parameter identification result of the ferromagnetic material