Prediction of follower jumps in cam-follower mechanisms: The benefit of using physics-inspired features in recurrent neural networks