• EPS systems
• hardware in the loop
• machine learning
• neural networks
• steering system models
• steering systems
• time-series forecasting

In the development of vehicles, creating an accurate model of the steering system plays a key role to obtain sufficient performances. To achieve a reliable model, it is essential to accurately approximate the non-measurable phenomena. In literature, there are numerous mathematical models to approximate friction forces, joint clearances and kinematic non-linearities, but having a model that includes all of them is challenging and it requires high computational resources. This study aims to develop a method to identify the non-measurable forces of the steering systems, through a data-driven approach. An HiL simulator is used to collect all the data, including an EPS steering system in the simulation loop. Different neural network architectures are tested, with the aim to predict the aforementioned forces using only data from the EPS-ECU unit, in view of future developments on real vehicles. Techniques of time series forecasting and RNN recurrent neural networks are adopted, and a Bayesian optimization is performed for the hyperparameters. The analysis shows that predictions strongly depends on the network architecture used and its hyperparameters. The results reveal that the trained network is capable of performing the desired prediction on the test maneuvers in most cases.