A Memristor-Based Infrared Reservoir Computing System for Dynamic Target Perception With Spatial–Temporal Features

Abstract

The rapid expansion of the Internet of Things (IoT) necessitates unprecedented resource efficiency in sensor-edge hardware to implement artificial intelligence (AI). A memristor-based infrared reservoir computing (MIRC) system was proposed in this work, incorporating a physical mask, a pyroelectric infrared (PIR) sensor based on LiTaO3 (LT), and a memristor based on LiNiO3 (LN). The PIR sensor with the physical mask can encode spatiotemporal features into electrical signals with high data fidelity. Subsequently, the analog sensor signals can be directly fed to the delay-based memristor reservoir. It is worth mentioning that the memristor shows a large dynamic space with rich reservoir states and a nonlinear short-term memory effect, so that the spatiotemporal features of PIR signals can be mapped more finely to the conductance of the memristor. These advancements in memristor performance enhance the system’s ability to process complex spatiotemporal data, allowing for the perception of targets with varying sizes, velocities, and distances, with results indicating a high distinction for each target. Compared to conventional methods, this approach significantly saves hardware resources, minimizes data transmission requirements, and simplifies backend algorithmic processing, providing a promising framework for advancing IoT technologies with integrated AI capabilities at the sensor edge.