CN-121980957-A - Switch hand feeling multi-scale modeling method based on physical information neural network

CN121980957ACN 121980957 ACN121980957 ACN 121980957ACN-121980957-A

Abstract

The invention discloses a switch hand feeling multi-scale modeling method based on a physical information neural network, firstly, a one-dimensional nonlinear elastic damping partial differential equation PDE and a normal differential equation ODE for describing the overall deformation behavior of a key are established to form a hybrid PDE-ODE dynamic model. And secondly, designing a Bayesian physical information neural network B-PINN with uncertainty quantization capability according to a dynamics model. And then dynamically increasing the data sampling density in the abrupt region, and adopting an event-driven strategy to adjust the training weight distribution so as to strengthen the self-adaptive training capacity of the neural network B-PINN. And finally, generating an elastic modulus and damping coefficient probability distribution map by using the trained network model, and mapping the subjective hand feeling scores of the users into a mechanical objective function and inverting parameters. The invention overcomes the defects of poor generalization and insufficient physical consistency of the pure data driving method and improves the reliability of the design parameters of the switch.

Inventors

PAN JIAHAO
Guan Yacun
YANG JIAXIN
PAN LIYAN
YU HONGBO

Assignees

杭州电子科技大学

Dates

Publication Date: 20260505
Application Date: 20260212

Claims (11)

1. The switch hand feeling multi-scale modeling method based on the physical information neural network is characterized by comprising the following steps of: S1, establishing a one-dimensional nonlinear elastic damping partial differential equation PDE and a normal differential equation ODE for describing the overall deformation behavior of a key, and forming a hybrid PDE-ODE dynamics model through nonlinear boundary condition coupling; S2, designing a Bayesian physical information neural network B-PINN with uncertainty quantization capability according to the constructed hybrid PDE-ODE dynamic model; S3, dynamically increasing the data sampling density near the mutation area, adjusting the training weight distribution by adopting an event-driven strategy, and strengthening the self-adaptive training capacity of the neural network B-PINN; S4, generating a probability distribution map of the elastic modulus and the damping coefficient by utilizing the B-PINN network model which is subjected to event-driven strategy optimization training, mapping the subjective hand feeling scores of the user into quantifiable mechanical objective functions, inverting parameters, and realizing closed-loop optimization of the product.
2. The method for modeling switch hand feeling multi-scale based on physical information neural network according to claim 1, wherein the establishment process of the one-dimensional nonlinear elastic damping partial differential equation PDE is as follows: For a mechanical key switch, a one-dimensional simplifying assumption is adopted along the key pressing direction, and a displacement field is simplified into a scalar function along the pressing direction , Is defined as a one-dimensional calculation domain in the key pressing direction The upper part of the upper part is provided with a plurality of grooves, Is the effective stroke length of the key; Is a time variable; The nonlinear elastic damping partial differential equation PDE describing the elastic response, damping effect and contact collision behavior is established as follows: Wherein the method comprises the steps of Is the density of the material; ; Is one-dimensional stress, determined by nonlinear constitutive relation, i.e , Nonlinear constitutive function Is that , Is a linear elastic modulus of the material, Is a second-order nonlinear coefficient which is a first-order nonlinear coefficient, Is a third-order nonlinear coefficient; is a viscous damping coefficient; is the collision dissipation factor; is a microcontact closed state variable and is used for realizing macro-microcosmic coupling; Is a Heaviside step function; is critical displacement; for a regularized approximation of the Dirac function, Is a regularization parameter; Is the critical speed of contact closure; as a Dirac function.
3. The method for modeling switch hand feeling multi-scale based on physical information neural network according to claim 2, wherein the ordinary differential equation ODE is established as follows: microcontact closure state variables The value range is , Indicating the separation of the contacts and, Indicating that the contacts are fully closed; Is described by a normal differential equation ODE: Wherein, the Characterizing the time response characteristics of contact closure for a contact kinetic rate constant; For the current contact force, determining from the response of the macroscopic stress field at the contact location; the characteristic contact force is a critical force threshold that causes the contacts to begin to close; indicating that the value is taken when the value in brackets is greater than 0, otherwise 0.
4. The method for modeling switch feel multiscale based on physical information neural network according to claim 3, wherein the specific process of forming the hybrid PDE-ODE dynamics model by nonlinear boundary condition coupling is as follows: In one dimension, the coupling of macroscopic PDE to microscopic strain ODE is achieved by macroscopic equivalent strain obtained by length averaging Realizing; when the microstructure is equivalent to a system of discrete elastic elements, the calculation of the macroscopic equivalent strain is converted from an integral form to a discrete summation form as follows: wherein Is the first The length of the individual microcolumns; is the microscopic unit number; Is the first The lumped parameter model equivalent the microcomponent into discrete mass-spring-damping unit, microcosmic local strain Description of the first embodiment Local deformation of the individual elastic elements: wherein Is the first The deformation of the individual microcosmic units; By macroscopic equivalent strain Calculating one-dimensional effective elastic modulus in combination with mechanical properties of the series elastic elements in a time domain On the spatial domain The effective elastic modulus is defined as: ; ; Setting a switch base as a fixed boundary, setting the surface of a key cap as a load applying boundary, setting the displacement and the speed at the initial moment to be zero, and forming a PDE-ODE dynamic model, wherein the boundary conditions and the initial conditions are as follows: fixed boundary: the base of the switch is rigidly fixed: ; load application boundary-user applies pressing force on the key cap surface: , wherein, Is that A one-dimensional stress at the location; the negative sign on the right side of the equation indicates that the pressing produces a compressive stress state with the direction pointing to the inside of the switch; if given Total pressing force at moment of time Then: wherein The effective sectional area of the key cap; initial displacement condition, the switch is in a static state when not pressed, and the initial displacement is zero: ; Initial speed condition, the switch does not move at the initial moment, and the speed is zero: 。
5. the method for modeling switch hand feeling multi-scale based on physical information neural network according to claim 4, wherein the specific implementation process of step S2 is as follows: s2.1, constructing a physical constraint prior based on PDE residual errors of a hybrid PDE-ODE dynamic model; S2.2, constructing a deep neural network structure comprising an input layer, a hidden layer and an output layer based on physical constraint priori, and designing a composite loss function; S2.3, obtaining a final B-PINN mathematical model through random gradient descent update, and finally outputting a prediction result comprising a mean value and a standard deviation through a variational distribution approximation posterior on the basis of prior distribution.
6. The method for modeling the switch hand feeling multi-scale based on the physical information neural network according to claim 5, wherein the specific implementation process of the step S2.1 is as follows: designing a B-PINN network with uncertainty quantization Wherein Training parameters for the B-PINN network, optimally solving the parameters in the mixed PDE-ODE dynamic model through the B-PINN network, For the standard deviation of variation distribution, solving B-PINN by adopting a variation inference method, wherein the loss function is as follows , wherein, Representation is directed to variation distribution Is a mathematical expectation of (a); Fitting a term to the data; Is a PDE residual term; Is a boundary condition item; is an initial condition item; weighting coefficients for each loss term; KL divergence; Data fitting term Wherein Is the total number of data points used for training; For the B-PINN neural network at the first Each space-time coordinate point A predicted displacement value output by the position; For the true displacement observation value corresponding to the coordinate point, Calculating for norms; PDE residual terms : Wherein, the Training sampling points for PDE residuals; Representation pair Squaring the residual value calculated by the point; Boundary condition item , wherein, Training sampling points for boundary conditions; for the predicted stress value of the network, Is the first The moments of the sampling points; Initial condition item , wherein, Training sample points for initial conditions.
7. The method for modeling switch hand feeling multi-scale based on physical information neural network according to claim 6, wherein the specific implementation process of step S2.2 is as follows: PDE residual construction physical constraint prior based on hybrid PDE-ODE dynamic model, definition of prior distribution in B-PINN, prior distribution Combines physical constraint and material characteristic priori knowledge, and the specific form is as follows: , wherein, Is a physical constraint item; Is a material property term; Physical constraints Is based on the prior of PDE residual error, and presumes parameters The gaussian distribution is satisfied: material property term Define parameter sets to be solved for the whole Is instantiated for the elastic modulus when building a priori in particular Using means and variances of experimental or molecular dynamics simulations in combination with known ranges of material properties Specific prior of elastic modulus Set to truncated gaussian: Wherein the method comprises the steps of Expressed in terms of Mean value(s), A gaussian distribution probability density function that is a variance; assuming a distribution of variation For a diagonal gaussian distribution: , wherein, Is the mean vector; Is a vector of the standard deviation, Representation pair Performing element-by-element squaring; The analytical solution is as follows: Wherein Is the network of The mean value of the individual parameters in the variation distribution; Is the network of Standard deviation of the individual parameters in the variation distribution; Representation of Relative to KL divergence of (c).
8. The method for modeling the switch hand feeling multi-scale based on the physical information neural network according to claim 7, wherein the specific implementation process of the step S2.1 is as follows: Introducing a re-parameterization technique separates the non-guided random sampling operation from the gradient back propagation path: Updating by random gradient descent And ; The final B-PINN mathematical model is based on prior distribution, approximates the posterior through variational distribution, and finally outputs a prediction result comprising a mean value and a standard deviation: Wherein, the Expressed in the parameter value of When the network pairs the space-time points A predicted value of displacement; To distribute in variation A lower variance operator.
9. The method for modeling the switch hand feeling multi-scale based on the physical information neural network according to claim 8, wherein the specific implementation process of the step S3 is as follows: s3.1, in the training process of the network B-PINN, monitoring the contact state in real time according to the current predicted displacement field, and when the predicted displacement is close to the critical threshold value of contact closure When the method is used, a mutation area is automatically identified, the density of sampling points is dynamically increased in the area, and a training data set with non-uniform distribution is generated: Wherein the method comprises the steps of For the density of the sampling points, Controlling the sampling enhancement degree of the critical area; controlling the width of the enhancement region; S3.2, in combination with the adjustment of sampling density, calculating the total loss function When the data points of the critical area are assigned a higher weight: When (when) Weight at time The value is greater than When in use; s3.3, dynamically adjusting the weight of PDE residual terms in total loss according to real-time residual distribution calculated by PDE equation in the training process Wherein For PDE weights before adjustment; Is the weight decay coefficient.
10. The method for modeling switch hand feeling multi-scale based on physical information neural network according to claim 9, wherein the specific implementation process of step S4 is as follows: S4.1, carrying out forward reasoning on the trained B-PINN network for multiple times through Monte Carlo Dropout technology to generate elastic modulus And damping coefficient By counting the mean and variance of multiple reasoning results, constructing a probability density function of parameters, and screening out a physical parameter solution set with confidence coefficient larger than a confidence coefficient threshold based on the distribution map The confidence is the cumulative probability of the parameter value in the probability density function; S4.2, carrying out non-parameterized probability density estimation on the screened parameters, calculating covariance matrixes among the parameters, and constructing a joint probability density function based on the covariance matrixes: Wherein the method comprises the steps of To screen out the total number of parameters greater than the confidence threshold; To de-centralize the first The modulus of elasticity and damping coefficient of the data points; Is covariance matrix; S4.3, utilizing Dropout mechanism as approximate implementation of Bayesian variational inference, and using probability in inference stage Randomly shielding neurons and simulating the effect of model integration; S4.4 based on The result of the second Monte Carlo forward propagation sampling is used for calculating the statistical mean and variance of displacement and force response, and the uncertainty of a prediction stage is quantified: Wherein, the Is shown in the first The value of the parameter obtained by forward reasoning of the second Monte Carlo is At the time of time, time space point A predicted value of displacement; S4.5, under the frame of B-PINN, simultaneously carrying out the process of macroscopic parameters ) And microcosmic parameters [ ] ) Performing joint inversion updating; s4.6, converting subjective handfeel into an optimizable engineering index, firstly establishing a corresponding relation between user scores and touch characteristics, then establishing experience mapping between the touch characteristics and mechanical parameters based on experimental data regression analysis, finally constructing an objective function containing subjective score weights, optimizing in a screened high-confidence parameter space, and outputting an optimal material parameter combination meeting the objective handfeel.
11. The method for modeling switch hand feeling multi-scale based on physical information neural network according to claim 10, wherein the specific implementation process of step S4.6 is as follows: s4.6.1, establishing subjective scores Dividing the scores of the users into different intervals, and defining specific touch descriptors for each interval; S4.6.2 based on pre-acquisition Group score-force curve pairing data, mapping is established by regression analysis: , Wherein the regression coefficient 、、、、 Fitting by a least square method to obtain; Representing the target effective elastic modulus; is a target viscous damping coefficient; S4.6.3 constructing a global optimization objective function according to the mapping relation, wherein the objective function constrains the deviation of inversion parameters and empirical regression parameters so as to minimize the objective function Searching in a high-confidence parameter space subjected to uncertainty screening as a criterion, and finally solving the optimal material parameter combination capable of accurately reproducing the target hand feeling in an inverse way: wherein 、 Is a weight coefficient.

Description

Switch hand feeling multi-scale modeling method based on physical information neural network Technical Field The invention belongs to the technical field of mechanical design and touch simulation, and particularly relates to a switch hand feeling multi-scale modeling method based on a Physical Information Neural Network (PINN). Background With the rapid development of industrial automation and intelligent manufacturing, mechanical key switches are used as important interfaces for information interaction, and touch sense of the mechanical key switches directly influences user experience and equipment performance. Modern electronics, medical equipment and industrial control systems place increasing demands on the speed of response, durability and feel feedback of the switches. The traditional modeling method is often based on simplified physical assumption or pure data driving, and complex dynamic behaviors of a switch under actual working conditions are difficult to comprehensively reflect, so that a novel modeling method which takes into consideration multi-scale, multi-physical field coupling and discontinuous event description is needed. At present, modeling of switch hand feeling mainly depends on two major technologies, namely, one is based on a physical model of a traditional Partial Differential Equation (PDE) and a normal differential equation (ODE), macroscopic displacement and strain are described by constructing a nonlinear elastic damping system, and microscopic transient events such as contact closing are described by using the ODE, and the other is a pure data driving method, such as a time sequence prediction model based on a long short term memory network (LSTM) and control strategy optimization based on reinforcement learning. These methods can fit complex responses in large data environments, but often ignore physical constraints. In recent years, physical Information Neural Networks (PINN) have become an important means to integrate data-driven and physical modeling advantages due to their ability to embed physical knowledge into the neural network training process. Despite the advantages of the above methods, there are significant limitations in that (1) the conventional PDE-ODE method is difficult to accurately capture in terms of dealing with the discontinuity and local abrupt changes when the contacts are closed, such as transient collisions and insufficient description of the energy dissipation effects, (2) the pure data-driven model, while having improved prediction accuracy, lacks physical constraints, results in insufficient generalization capability, and in particular, is difficult to perform high-precision parametric inversion in the data sparse region, and (3) the current method also fails to form a system-effective solution in terms of multi-scale coupling and uncertainty quantification, and the insufficient data sampling and uncertainty assessment of the critical region further limit the reliability and popularization thereof in engineering practice. Disclosure of Invention Aiming at the problems, the invention provides a switch hand feeling multi-scale modeling method based on a physical information neural network. According to the method, a mixed PDE-ODE model is adopted to finely describe a macroscopic continuous medium and a microscopic discontinuous event, parameter uncertainty quantification is achieved through a Bayesian PINN framework, an event-driven sampling strategy is introduced to enhance critical area data density, and finally a training model is optimized through multi-scale inversion. In order to achieve the above purpose, the method of the invention is specifically implemented as follows: S1, establishing a one-dimensional nonlinear elastic damping dynamics partial differential equation describing the overall deformation behavior of a key, introducing a contact closure state variable and establishing a corresponding ordinary differential equation, wherein the two equations are coupled through nonlinear boundary conditions at the tail end of a structure to form a hybrid PDE-ODE dynamics model; s1.1, aiming at the characteristic that the mechanical response of the mechanical key switch in the pressing process is concentrated in the pressing direction, adopting a one-dimensional simplifying assumption along the pressing direction of the key, and simplifying the displacement field into a scalar function along the pressing direction , wherein,Is defined as a one-dimensional calculation domain in the key pressing directionApplying; Is a time variable; is the effective stroke length of the key; The nonlinear elastic damping partial differential equation PDE describing the elastic response, damping effect and contact collision behavior is established as follows: Wherein the method comprises the steps of Is the density of the material;; Is one-dimensional stress, determined by nonlinear constitutive relation, i.e ,Nonlinear constitutive functionIn particular to,Is a l