Applying neural networks for plant model simulation in embedded control system

In debugging and testing the control software used in distributed embedded systems such as automotive systems, simulations are conducted if the plants to be controlled are unavailable. In the simulation, models artificially based on physical principles are used. However, in some cases, models cannot be acquired partly because of confidentiality issues of the manufacturers, and only the input and output data of the models can be obtained. This paper presents a plant simulation for testing control software based on input and output data using a neural network-recurrent neural network (RNN) with long short-term memory-suitable for dealing with time series data, such as simulation data. In simulations, for comparison, we also evaluate the method of system identification. In experiments, a model with numerous nonlinear blocks is used. The results show that RNN yields much better results than system identification.