Automatic segment filtering procedure for processing non-stationary signals.

Area of Science:

• Biomechanics
• Signal Processing
• Data Analysis

Background:

• Computing time derivatives is crucial in biomechanical data processing.
• Signal noise amplification during differentiation reduces accuracy.
• Traditional low-pass Butterworth filters use a single cut-off frequency, which is suboptimal for non-stationary signals with varying frequency components.

Purpose of the Study:

• To introduce a novel method, the Automatic Segment Filtering Procedure (ASFP), for processing biomechanical signals.
• To address the limitations of fixed cut-off frequencies in low-pass filters for non-stationary signals.
• To improve the accuracy of signal derivatives in biomechanical analysis.

Main Methods:

• The Automatic Segment Filtering Procedure (ASFP) was developed, utilizing different Butterworth filter cut-off frequencies for distinct signal segments.
• Signal segments with varying energy content were identified using the Teager-Kaiser Energy Operator.
• The Autocorrelation-Based Procedure (ABP) was employed to determine segment-specific filter cut-off frequencies.

Main Results:

• The ASFP achieved a root mean square error (RMSE) of 16.4 rad s⁻² (26.6%) in estimating acceleration from the Dowling (1985) dataset.
• A standard approach using a single ABP-determined cut-off frequency resulted in a higher RMSE of 25.5 rad s⁻² (41.4%).
• A Generalized Cross-Validated Quintic Spline filter yielded an RMSE of 23.6 rad s⁻² (38.4%), also less accurate than ASFP.

Conclusions:

• The ASFP offers a significant improvement in estimating acceleration from biomechanical data compared to conventional methods.
• This adaptive filtering approach effectively preserves high-frequency content in non-stationary signals.
• The ASFP is advantageous for accurate biomechanical data processing, particularly for signals with time-varying characteristics.