Handcrafted Time-Domain Features
Time-domain statistical features extracted manually from time-series signals such as sEMG, including mean absolute value, waveform length, and zero-crossing rate. These features can significantly improve classification accuracy.
Timeline (last 90 days)
Handcrafted time-domain features improve per-task accuracy by up to about 40% compared with raw signals.
Under class-incremental drift, only experience replay preserves the accuracy gain brought by handcrafted features; EWC and naive fine-tuning cannot maintain this advantage.
Handcrafted time-domain features typically include mean absolute value (MAV), waveform length (WL), zero crossings (ZC), slope sign changes (SSC), and root mean square (RMS), computed on fixed-length sliding windows to compress high-dimensional raw sEMG sequences into low-dimensional descriptors.
Design principles for a cross-user sEMG gesture recognition system include: adopting experience replay as the core continual learning strategy; using stateless training with shuffled windows; and incorporating handcrafted time-domain features to improve baseline performance.
All Facts (4)
Handcrafted time-domain features improve per-task accuracy by up to about 40% compared with raw signals.
50%UnverifiedUnder class-incremental drift, only experience replay preserves the accuracy gain brought by handcrafted features; EWC and naive fine-tuning cannot maintain this advantage.
50%UnverifiedHandcrafted time-domain features typically include mean absolute value (MAV), waveform length (WL), zero crossings (ZC), slope sign changes (SSC), and root mean square (RMS), computed on fixed-length sliding windows to compress high-dimensional raw sEMG sequences into low-dimensional descriptors.
50%UnverifiedDesign principles for a cross-user sEMG gesture recognition system include: adopting experience replay as the core continual learning strategy; using stateless training with shuffled windows; and incorporating handcrafted time-domain features to improve baseline performance.
50%