The LSTM Models with Gated Recurrent Unit Applied for Shaft Torque Estimation of Two-Mass System

The paper deals with the application of neural networks designed to estimate torsional torque. The estimator is used in a control structure that utilizes a state controller for a two-mass system. Two topologies of a deep neural network are tested. Typical Deep Neural Network (DNN) is combined with a parallel branch of an LSTM or GRU layer. Neural models do not require direct identification of the system parameters. To train the network, easily attained variables such as motor speed and electromagnetic torque are used. The result shows a high precision of the estimated signal in comparison to the signal measured from the experimental bench. The networks are evaluated for different number of training epochs, the reference trajectories differ from the signals applied for training.