A physical information neural network modeling method of a magneto-rheological damper

Abstract

The application discloses a physical information neural network modeling method of a magneto-rheological damper, and comprises the following steps: S1, acquiring data and establishing a multi-dimensional input feature space of the magneto-rheological damper containing time sequence historical states; S2, establishing a solid-liquid double-structure model, specifically, establishing an explicit calculation model of mechanical responses of solid phase / liquid phase of the magneto-rheological damper; S3, constructing a neural network with a feature vector as input and damping force as output, specifically, constructing a multi-objective composite loss function which integrates data constraints, physical residuals and energy dissipation consistency; S4, adopting a progressive dynamic training strategy to optimize parameters of the physical information neural network, and outputting a high-precision damping force prediction model of the magneto-rheological damper. According to the application, accurate prediction of the damping force of the magneto-rheological damper is realized, and the consistency and rationality of key physical characteristics such as hysteresis and energy dissipation are considered, thereby providing reliable model support for dynamic performance evaluation, simulation analysis and control application of the magneto-rheological damper.

Description

Technical Field

[0001] This invention relates to the technical field of dampers, and in particular to a method for modeling the physical information of magnetorheological dampers using a neural network. Background Technology

[0002] There are some difficulties and shortcomings in the existing modeling methods for magnetorheological dampers in engineering applications. For example: (1) It is difficult to express the complex behaviors of magnetorheological dampers, such as the difference before and after yielding caused by changes in current, strong nonlinear hysteresis characteristics and memory effect, in a unified and stable manner. (2) The model is highly dependent on the completeness of parameters under computational conditions. In practical applications, there are often problems such as complex modeling process and high implementation cost, making it difficult to balance modeling efficiency and application requirements. (3) Under conditions of limited data and changing operating conditions, the existing methods have problems of weak generalization ability and insufficient physical interpretation. In particular, it is difficult to ensure the consistency and credibility of key data such as dissipated energy. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide a neural network modeling method for the physical information of magnetorheological dampers. To achieve the above-mentioned objective and other advantages of the present invention, a neural network modeling method for the physical information of magnetorheological dampers is provided, comprising: S1. Acquire data and establish a multidimensional input feature space for the magnetorheological damper that includes the temporal history state; S2. Establish a solid-liquid dual-structure model, specifically an explicit calculation model of the solid / liquid mechanical response of the magnetorheological damper. S3. Construct a neural network with feature vectors as input and damping force as output, specifically by constructing a multi-objective composite loss function that integrates data constraints, physical residuals, and energy dissipation consistency. S4. A progressive dynamic training strategy is adopted to optimize the parameters of the physical information neural network and output a high-precision prediction model of the damping force of the magnetorheological damper.

[0004] Preferably, the construction of the neural network in step S2, which takes the feature vector as input and the damping force as output, includes: The yield shear stress is established using a polynomial function. With damping viscosity coefficient Mathematical mapping relationship between excitation current and its variation; The solid mechanical behavior of the damping fluid before yielding and the fluid mechanical behavior after yielding are defined. We introduce a microscopic relative deformation based on a displacement threshold and design a smooth step function to continuously and smoothly characterize the phase ratio of the transition from solid to liquid.

[0005] Preferably, the data includes piston displacement, velocity, current, and historical status data.

[0006] Preferably, step S3 specifically includes establishing a multi-objective composite loss function based on PINN. In the design of the neural network loss function, two physical dimension constraints are introduced, which include: (1) Physical mechanism loss: the mean square error residual between the network-predicted damping force and the mechanism damping force calculated in step S2; (2) Energy dissipation loss: using the integral of damping force and velocity in the time domain Construct the energy equation.

[0007] Preferably, step S4 is a progressive network training strategy based on adaptive dynamic parameter adjustment. To address the gradient conflict and local optima problems common in multi-objective optimization, dynamic weight scheduling is implemented, specifically including a staged weight adjustment mechanism and a phase sharpness dynamic adjustment mechanism.

[0008] Compared with the prior art, the advantages and positive effects of the present invention are: By incorporating the solid-liquid conversion mechanism within the magnetorheological damper into the neural network training process, and combining data constraints with physical constraints during training, the model can maintain an effective representation of the damping force response even with limited data. Simultaneously, the introduction of energy consistency constraints enhances the model's reliability in representing dissipation characteristics, thereby reducing the risk of physical inconsistencies arising from relying solely on data fitting, making it more suitable for engineering applications.

[0009] In situations with limited data, noise, or changing operating conditions, this method aims to accurately predict the damping force of magnetorheological dampers while ensuring consistency and rationality in key physical characteristics such as hysteresis and energy dissipation. This provides reliable model support for the dynamic performance evaluation, simulation analysis, and control applications of magnetorheological dampers. Attached Figure Description

[0010] Figure 1 A flowchart of a neural network modeling method for the physical information of a magnetorheological damper according to the present invention; Figure 2 A comparison of the PINN prediction model and experimental results under different operating conditions using the physical information neural network modeling method for the magnetorheological damper according to the present invention; Figure 3 A comparison and verification diagram of the model prediction results for the extrapolation working conditions of the physical information neural network modeling method for the magnetorheological damper according to the present invention. Detailed implementation method. 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, and 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.

[0011] Reference Figure 1 A neural network modeling method for the physical information of a magnetorheological damper, comprising: Step S1: Construct a multidimensional input feature space for the magnetorheological damper that includes the temporal history state.

[0012] Magnetorheological dampers (MRDs) exhibit significant hysteresis nonlinearity and memory effects; their damping force at the current moment depends not only on the current state but also on the historical motion trajectory. This invention first obtains piston displacement sequences under different operating conditions. velocity sequence With excitation current sequence .

[0013] First-order historical state extraction is performed on time series data, taking the previous time step as an example. displacement With speed As an additional feature, with the current time The states are concatenated to construct a five-dimensional time series input tensor. Its expression is as follows: ; This five-dimensional input matrix serves as the data input to the physical information neural network, providing the network with the state information required to characterize the historical memory effect in a strongly nonlinear system of a damper.

[0014] Step S2: Establish a solid-liquid constitutive, continuously differentiable physical mechanism model for the magnetorheological damper. To meet the requirements of gradient backpropagation in the physical information neural network, the non-differentiable operators in the constitutive model need to be modified. This invention establishes a smooth and continuously differentiable physical constraint model, with the specific changes as follows: (1) Extracting magnetorheological dampers at different currents The response characteristics under the damping fluid yield shear stress With viscosity coefficient It is parameterized as a physical function that varies with current.

[0015] (2) Differentiable substitution is applied to velocity-dependent non-smooth operators to eliminate gradient jumps near zero-point velocities. This is achieved by... Replace hard sign function ,use Alternate velocity absolute value Based on this, the dominant forces of the system in the unyielding solid phase are defined respectively. The dominant force in the post-yield liquid phase : ; ; in, This is the equivalent stiffness coefficient; To reflect the displacement deviation of the micro-particle chain deformation, where For microscopic yielding center; It is a nonlinear velocity index; and These are numerical parameters that control the direction of speed and the width of smoothing, respectively.

[0016] (3) Design a continuous phase weighting factor based on a displacement threshold. This replaces the traditional hard-condition judgment. When the absolute value of the displacement deviation... Exceeding the ultimate yield displacement threshold At this point, a phase transition occurs in the system. Phase weights are constructed using a sigmoid-type function: ; In the formula, This is a phase sharpness parameter. When in the solid phase... When in the liquid phase, .

[0017] (4) Combining the above phase weighting factors, the weighted superposition method is used to output the overall physical mechanism prediction power of the system. And at each time step, the micro yield center is updated smoothly and recursively. : ; This differentiable model not only retains the physical meaning of the damping fluid from solidification to flow, but also ensures the continuity of the gradient throughout the entire operating domain, directly embedding a neural network as a physical prior constraint.

[0018] Step S3: Construct a multi-objective composite loss function; the training objective of the PINN model is to simultaneously satisfy the constraints of experimental data fitting and physical and mechanical laws. This invention defines the total loss function as: ; In the formula, and These are adaptive weights for physical loss and energy loss, respectively. Their specific composition is as follows: (5) Loss of experimental data items : The predicted damping force used to ensure the network output Compared with the actual damping force collected in the experiment Maintain consistency in both the main path and local shape: ; in, This represents the number of data samples.

[0019] (6) Physical consistency loss Damping force predicted by constrained neural network Without deviating from the solid-liquid two-phase continuous differentiable dynamic model established in step S2 This gives the black box model physical interpretability: ; in, For the number of points, These are the parameters to be identified in the physical model.

[0020] (7) Consistency loss of energy dissipation The core engineering aspect of magnetorheological dampers is energy dissipation. The energy dissipated by the damper over a given time period... It can be expressed as a discrete integral of force and velocity. This invention innovatively constrains the cumulative energy dissipation predicted by the network. Compared with the actual energy dissipation in the experiment Error between: ; The energy loss term imposes a constraint on the area of ​​the loop in the force-displacement hysteresis curve, which greatly improves the physical rationality and robustness of the model when extrapolating across operating conditions and with varying currents.

[0021] Step S4: Implement a progressive network training strategy based on adaptive dynamic parameter adjustment; implement dynamic weight scheduling to address the gradient conflict and local optima problems common in multi-objective optimization.

[0022] (8) Staged weight adjustment mechanism: In the early stage of training: pure data fitting is the main focus, and the weights are set... This allows the model to quickly capture the main features of the data. Mid-training: Linearly and smoothly increasing the weights of the physical mechanism loss. To reach the target value, physical constraints are gradually introduced into the network. In the later stages of training: based on the initial convergence of the model, energy dissipation loss weights are gradually introduced and linearly increased. This ensures the conservation of macroscopic energy.

[0023] (9) Dynamic adjustment mechanism for phase sharpness: Let the phase sharpness parameter in step S2 be set as follows: With training rounds Linear increase. In the early stages of training, smaller sharpness provides broad, smooth gradient guidance; in the later stages of training, larger sharpness accurately approximates the real non-smooth physical yield hard features, effectively accelerating model convergence and improving the edge accuracy of hysteresis loops.

[0024] The Physically Constrained Neural Network (PINN) constructed in this invention is used to fit and verify the damping force-velocity hysteresis relationship of a magnetorheological damper under multiple operating conditions. Figure 2 The experimental results and PINN prediction results under six typical working conditions are compared, including hysteresis curves under different displacement amplitudes, excitation frequencies and excitation current combinations.

[0025] from Figure 2 The PINN model can accurately reconstruct the main shape characteristics of the damping force-velocity hysteresis loop under various operating conditions, and the experimental curves and predicted curves maintain a high degree of consistency in most speed ranges. The coefficients of determination (R²) for the six operating conditions are shown in the figures. 2 The overall range is 0.9901–0.9984, with an average value of approximately 0.9947. This result indicates that PINN exhibits good fitting accuracy and stability under different amplitude, frequency, and current conditions.

[0026] To further verify the predictive ability of the constructed PINN model outside the training distribution, this paper selects the 4mm–1Hz operating condition as an independent external test set, and compares and analyzes the model prediction results with experimental data under three excitation current conditions of 0.25A, 0.5A, and 0.75A. The results are as follows: Figure 3 As shown.

[0027] Depend on Figure 3 The results show that the PINN model can reconstruct the overall profile of the damping force-velocity hysteresis curve well under all three extrapolation conditions, and the corresponding coefficient of determination R0 is [value missing]. 2 The values ​​were 0.9583, 0.9658, and 0.9752, respectively. Although the fitting accuracy decreased compared to the observed operating conditions, it remained at a high level overall, indicating that the model still has good predictive ability in the unobserved operating condition of low frequency and large amplitude. Especially in the plateau section of the high-speed region and at the outer contour of the hysteresis loop, the predicted results maintained good consistency with the experimental curves, indicating that the proposed method can stably capture the dominant energy dissipation characteristics of the magnetorheological damper. In summary, this invention addresses the shortcomings of existing methods by proposing a physical information neural network modeling method for magnetorheological dampers. Based on the integration of the solid-liquid two-phase constitutive mechanism and the PINN method, a high-fidelity physical information model of the magnetorheological damper is constructed, realizing an accurate description and continuous characterization of the mechanical behavior before and after yielding. This not only makes up for the shortcomings of existing methods but also provides reliable technical support for the dynamic performance analysis, control design, and engineering application of magnetorheological dampers.

[0028] The number of devices and processing scale described herein are for simplification purposes. Applications, modifications, and variations of this invention will be readily apparent to those skilled in the art. Although embodiments of the invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for this invention, and further modifications can be readily implemented by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, this invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A method for modeling the physical information of a magnetorheological damper using a neural network, characterized in that, Includes the following steps: S1. Acquire data and establish a multidimensional input feature space for the magnetorheological damper that includes the temporal history state; S2. Establish a solid-liquid dual-structure model, specifically an explicit calculation model of the solid / liquid mechanical response of the magnetorheological damper. S3. Construct a neural network with feature vectors as input and damping force as output, specifically by constructing a multi-objective composite loss function that integrates data constraints, physical residuals, and energy dissipation consistency. S4. A progressive dynamic training strategy is adopted to optimize the parameters of the physical information neural network and output a high-precision prediction model of the damping force of the magnetorheological damper.

2. The physical information neural network modeling method for a magnetorheological damper as described in claim 1, characterized in that, The construction of the neural network in step S2, which takes the feature vector as input and the damping force as output, includes: The yield shear stress is established using a polynomial function. With damping viscosity coefficient Mathematical mapping relationship between excitation current and its variation; The solid mechanical behavior of the damping fluid before yielding and the fluid mechanical behavior after yielding are defined. We introduce a microscopic relative deformation based on a displacement threshold and design a smooth step function to continuously and smoothly characterize the phase ratio of the transition from solid to liquid.

3. The physical information neural network modeling method for a magnetorheological damper as described in claim 1, characterized in that, The data includes piston displacement, speed, current, and historical status data.

4. The physical information neural network modeling method for a magnetorheological damper as described in claim 1, characterized in that, Step S3 specifically includes establishing a multi-objective composite loss function based on PINN. In the design of the neural network loss function, two physical dimension constraints are introduced, which include: (1) Physical mechanism loss: the mean square error residual between the network-predicted damping force and the mechanism damping force calculated in step S2; (2) Energy dissipation loss: using the integral of damping force and velocity in the time domain Construct the energy equation.

5. The physical information neural network modeling method for a magnetorheological damper as described in claim 1, characterized in that, Step S4 is specifically a progressive network training strategy based on adaptive dynamic parameter adjustment. To address the gradient conflict and local optima problems common in multi-objective optimization, dynamic weight scheduling is implemented, specifically including a staged weight adjustment mechanism and a phase sharpness dynamic adjustment mechanism.

Images