Energy-efficient pipelined DTW architecture on hybrid embedded platforms

It is predicted that fifty billion sensor-based devices are to be connected to the Internet by 2020 with the fast development of Internet of Things (IoT). Stream data mining on these tremendous sensor-based devices has become an urgent task. Dynamic time warping (DTW) is a popular similarity measure, which is the foundation of stream data mining. In the last decade, DTW has been well accelerated with software and reconfigurable hardware optimizations. However, energy-efficiency has not been considered, which is critical for data mining on these devices. In this paper, we propose an energy-efficient DTW acceleration architecture for stream data mining on sensor-based devices, which is based on a hybrid embedded platform of ARM and field programmable gate array (FPGA). Software optimizations for DTW are implemented on ARM, and pipelined DTW is implemented on FPGA for further accelerations. A pilot study is performed with three widely adopted stream data mining tasks: similarity search, classification, and anomaly detection. The results show that the performance improvements vary for different configurations, and the achieved average speedup and energy efficiency improvement are 7.52× and 4.23×, respectively.