Year: 2026
Type: preprint
Abstract: Joint angle prediction is relevant to wearable motion monitoring, human–machine interaction, exoskeleton control, and rehabilitation assessment. Flexible piezoelectric sensors offer mechanical compliance, low cost, and fast dynamic response, but their voltage outputs are affected by random noise, sensor–skin contact, and local deformation fluctuations. These factors lead to low signal-to-noise ratio, nonstationarity, and nonlinear signal–angle relationships. This study proposes an NAE-Transformer (Noise-Aware Efficient Transformer) for joint angle prediction using a single flexible piezoelectric sensor. The model forms a dual-representation input by combining a Kalman-smoothed trend signal with a sliding-window noise-envelope feature, and uses a hybrid sparse attention encoder to capture temporal dependencies. Experiments were performed on a custom-built f lexible piezoelectric sensor dataset, with a passive mechanical arm angle sensor used as the reference standard. Compared with LSTM (long short-term memory), Transformer, Informer, and their Kalman-filter-enhanced variants, NAE-Transformer obtained lower prediction error and higher goodness of fit. The results indicate that joint use of trend smoothing, local fluctuation representation, and hybrid sparse attention can improve piezoelectric-signal-based joint angle estimation under noisy and nonstationary sensing conditions.
Source: Research Square
Institution: Zhengzhou University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7233852392
