Cascaded Reservoir Computing for Temporal Sensor Data: Integrating Physical Dynamics with Echo State Networks

Md Razuan Hossain1, Imran Fahad2, Braylen Robinson2, Sai Swaminathan2, Hritom Das3
1Utah Valley University, 2The University of Tennessee, Knoxville, 3Oklahoma State University


Abstract

Reservoir Computing (RC) offers an energy-efficient framework for processing temporal data. This paper presents a two-stage reservoir computing architecture for efficient and accurate classification of temporal sensory data, particularly vibrational signals captured from the Hall sensors. The proposed architecture leverages a physical reservoir consisting of a mechanical system to perform intrinsic dynamic encoding of input signals, followed by a software-based Echo State Network (ESN) that acts as a second-stage reservoir for nonlinear classification. Experiments conducted in two distinct environments - a terrain surface dataset and a machine shop dataset- demonstrate the effectiveness of this cascaded design. Our proposed method achieves a reduction in memory overhead, lower computational complexity, and maintains a competitive accuracy. Principal Component Analysis and confusion matrix evaluations confirm enhanced separability and improved class-specific feature alignment provided by the second-stage ESN, validating its utility for real-time classification tasks such as resource-constrained edge, embedded devices, etc. In contrast to deep learning baselines such as LSTM and 1D-CNN, our cascaded ESN achieves comparable accuracy with two orders of magnitude fewer trainable parameters and drastically reduced runtime. This compactness and efficiency make the proposed architecture particularly well-suited for deployment in low-power edge and embedded systems, where conventional deep learning models are impractical.