Hardware Implementation of Low-complexity Deep Learning-based Model Predictive Controller

Model predictive control (MPC) is an advanced control strategy that predicts the future behavior of the plant while considering the system dynamics and constraints. This optimization-based control algorithm needs huge amount of computational resources as it solves the optimization problem at each sampling time. This computational load demands powerful hardware and light-weight algorithms for implementing MPC on an embedded systems with limited computational resources. Deep neural network (DNN) is a attractive alternative for MPC as it provide the close approximation for various linear/nonlinear functions. In this paper, a feed forward neural network (FFNN) and recurrent neural network (RNN)-based linear MPC is developed. These neural network algorithms which are trained offline can efficiently approximate MPC control law which can be easily executed on low-level embedded hardware. The closed-loop performance is verified on a hardware-in-the-loop (HIL) co-simulation on ARM microcontroller. The performance of proposed DNN-MPC is demonstrated with a case study of a 2 degree-of-freedom (DOF) helicopter. Hardware result shows that the DNN-MPC is faster and consumes less memory as compared to MPC while retaining most of the performance indices.