Reservoir computing for a MEMS mirror-based laser beam control on FPGA

Abstract

In this paper, a small-world network-based reservoir computing (SWN-RC) is introduced to a micro-electromechanical system (MEMS) mirror-based laser scanner to achieve high-accuracy and low-delay laser trajectory control. The benefits of SWN-RC are confirmed through a comprehensive simulation, comparing it with reservoir computing (RC) based on regular and random networks. Subsequently, the RC control module is designed and implemented on a cost-optimized field-programmable gate array (FPGA). To balance the resource consumption and the processing delay, we use a half-parallel architecture for the large-scale matrix multiplications. In addition, the weight matrices of the RC are expressed by the 12-bit fixed-point data, which sufficiently suppresses the quantization noise. Furthermore, we simplify the activation function as a piecewise linear function and store the values in the read-only memory (ROM), resulting in a 76.6% reduction in ROM utilization. Finally, the SWN-RC, regular-RC, and random-RC control modules are implemented on the FPGA board and experimentally tested in the MEMS mirror-based laser scanner system. To the authors’ knowledge, it is the first reported RC-based MEMS mirror control system implemented on FPGA. In addition, the PID control is also tested as a baseline experiment. The results indicate that the RC control greatly outperforms the PID control with a 57.18% reduction in delay and over a 58.83% reduction in root mean square error (RMSE). Among the RC controls, the SWN-RC exhibits the best performance than the others. The SWN-RC achieves a further 14.03% and 12.42% reduction in RMSE compared to regular-RC and random-RC, respectively.