I-KAN: Reconstructing Over-Range Inertial Signals

Inertial sensors are widely used in modern society, covering automotive, aerospace, robotics, and consumer electronics. However, these widely used sensors suffer from range limitations, causing signal saturation and information loss under high-dynamic motion conditions. To address this problem, we introduce the Kolmogorov-Arnold Networks (KAN) into inertial sensor signal processing for the first time and propose the I-KAN model. Unlike traditional deep learning models that treat signals as discrete point sequences, KAN treats the signal as a continuous curve through embedded spline functions, which allows it to infer the lost signal segments by exploring the continuity and fluctuation patterns of signal curves. Moreover, considering deep learning models suffer from the hallucination problem that generates chaotic and spurious outputs, we propose the Generation Hallucination Entropy (GHE), which quantifies and reduces hallucinations by enhancing the consistency of outputs for similar inputs, thus improving the stability and reliability of the model. Given the absence of research and dataset for reconstructing inertial over-range signals, we release the first inertial over-range signal recovery dataset (IOSRD) in Github, which consists of inertial sensor data from 10 smartphones. The results demonstrate that I-KAN effectively handles varying degrees of signal saturation and sets a benchmark in inertial over-range signal reconstruction.