Control Current Prediction For Magnetorheological Dampers Based On Stacking Fusion Model

With the development of the automotive industry and the times,people have higher requirements for the comfort and durability of vehicles during driving.When encountering uneven roads,vehicles rely on the dampers and elastic devices in the vehicle suspension to ensure smooth driving of the vehicle body.Magnetorheological(MR)dampers have good application prospects due to their material properties,but due to their highly nonlinear and hysteretic characteristics under current excitation,building parametric models may bring a lot of problems such as parameter identification and difficulty in solving inverse models.Establishing accurate control current prediction models is the key to achieving automatic control of MR damper systems.Based on previous research methods,this paper considers using machine learning,deep learning models,and model fusion methods to study the current prediction problem of magnetorheological damper control.Firstly,this article uses the proposed Bouc＿Wen model with high accuracy builds a mechanical model of a magnetorheological damper for simulation modeling.Simulation modeling is performed under sinusoidal displacement excitation to observe the characteristics of magnetorheological damper data-high nonlinearity and hysteresis.Simulation modeling is performed under random displacement excitation to simulate different road conditions and vehicle driving conditions on actual roads and obtain data sources.Considering that there are fewer feature dimensions in the original data set,which may make it difficult to achieve the best effect of the model by under fitting.In this paper,feature derivation is used to expand the variables into 16 variables after feature expansion.Seven variables are removed using correlation coefficients,VIF values,mutual information values,and recursive feature deletion.The effectiveness of feature engineering is verified by the model effects before and after feature engineering.This paper selects KNN,SVR,decision tree,XGBoost,Light BGM models that can describe nonlinearity in machine learning to construct MRD control current prediction models,and utilizes the strong fitting ability of neural networks to construct FNN,DFNN,and LSTM models.Using grid search to find the optimal parameters of each model,it is verified from the model effect that the SVR model does not have applicability,and Light GBM has better prediction accuracy.Finally,in order to further enhance the generalization ability and prediction accuracy of the inverse model,a Stacking fusion model is constructed,which has a higher improvement in all evaluation indicators compared to a single model,indicating that the Stacking model has a more accurate control current prediction effect.