A Physics Guided Reversible Residual Neural Network Model Applied to Build Forward and Inverse Model

To satisfy the application requirements of the forward/inverse model for various control algorithms used in servo systems, a physics-guided reversible residual neural network (PGRRNN) model is proposed, which is applied to build forward/inverse model for turntable servo systems. With the special design of the network structure, both forward and inverse models can be obtained through one-way training of the model, so that the modeling efficiency is improved. Combining the gated recurrent unit and the fully connected layer as the nonlinear term of the reversible residual unit, the PGRRNN model can capture the relevant information between time series data in the training process, so that the prediction accuracy can be effectively improved. Moreover, by selecting the linear sweeping frequency signals with different amplitudes in series as the input signal, the system can be fully excited and the generalization ability can be effectively improved. Experimental results show that the forward/inverse model built by the proposed method has high precision and good generalization ability. The prediction accuracy of the forward/inverse model is better than that of the models built by the traditional modeling method and our previous work. In addition, the accuracy and real-time performance can be guaranteed simultaneously by adjusting the length of the input sequence and the number of residual units, so it has a certain practical value.