Optical-Cavity-Based Reservoir Computing for Low-Power Time-Series Analysis in Industrial Internet of Things Applications

Abstract

Industrial time-series data analysis poses significant challenges due to multichannel sensor inputs, complex temporal dependencies, and strong dynamic variability. While large generative models (LGMs) can process temporal information effectively, their reliance on large training datasets and substantial computational power renders them unsuitable for power-constrained, distributed applications, including the industrial Internet of Things (IIoT), intelligent manufacturing, and autonomous driving, where fast response and long-term stable operation are often required. To address this limitation, we propose a physics-based reservoir computing system that uses an optical cavity for time-series prediction. The physical cavity intrinsically mixes input data with various time delays, thereby generating high-dimensional, randomly delayed responses and effective nonlinear features. These features are then combined through a linear readout to produce the final prediction. In a proof-of-concept evaluation, we show that by extracting a compact representation from only five hidden nodes in the cavity, the output can perform classification task with an accuracy of 87.5%. By increasing the number of hidden-layer (reservoir) outputs to moderate values of approximately 50 to 100, the system performs benchmark tasks such as nonperiodic Mackey–Glass (MG) time-series prediction with a normalized root-mean-square error (NRMSE) of 0.0843, and anomaly detection on the Skoltech anomaly benchmark (SKAB) dataset with an $F1$ score of 0.784. These findings demonstrate the potential of optical cavities for low-power, real-time time-series analysis in industrial applications. They also highlight the potential of physics-based hardware as an efficient accelerator for edge artificial intelligence.