NARX Neural Network Modelling of Hydraulic Suspension Dampers for Steady-state and Variable Temperature Operation

Summary Two NARX-type neural networks are developed for modelling nonlinear dynamic characteristics of passive twin-tube hydraulic dampers used in vehicle suspension systems. Quasi-isothermal and variable temperature NARX models are rigorously tested and compared with a state-of-the-art physical model proposed by Duym and Reybrouck (1998) and Duym (2000). Measured damper data, generated under isothermal and temperature varying conditions, is used for NARX training, physical model calibration, and predictive comparisons. Test kinematics include high amplitude sinusoidal displacements up to 14 Hz, and realistic random road profiles. The NARX models are trained via ‘teacher forcing’ and the feedforward backpropagation algorithm using both ‘Early Stopping’ and Bayesian Regularisation. Stable network design is also examined using the minimum posterior prediction error as the criterion for selecting a good network from a small number of tests. Calibration of the physical model proves highly complicated owing to considerable nonlinearity-in-the-parameters, requiring use of Sequential Quadratic Programming with an implicitly nonlinear constraint. The paper shows that NARX neural network modelling is vastly superior in terms of calibration efficiency, and prediction times, whilst offering roughly similar, if not better, model accuracy.