Data-driven Cut-off Frequency Optimization for Biomechanical Sensor Data Pre-Processing

• Published in: Data Science – Analytics and Applications pp. 20 - 25
• Main Authors: Bernhart, Severin, Venek, Verena, Kranzinger, Christina, Kremser, Wolfgang, Martínez, Aaron
• Format: Book Chapter
• Language: English
• Published: Wiesbaden Springer Fachmedien Wiesbaden 2022
• Subjects: automated filtering techniques; clustering algorithm; cut-off frequency optimization; human motion data analytics
• ISBN: 3658362944; 9783658362942
• DOI: 10.1007/978-3-658-36295-9_3

The pre-processing of biomechanical sensor data often involves signal filters for noise removal in order to improve the performance of segmentation and machine learning algorithms. However, finding an optimal value for the filter’s cut-off frequency is time consuming, as researchers have to rely on heuristics and experience. Therefore, we introduce a method called FcOpt for automatically estimating an optimal cut-off frequency for noise filtering in one-dimensional biomechanical data. The method resamples the input data and applies three automated cut-off frequency determination methods, pools their individually suggested cut-off frequencies with a k-means cluster algorithm and provides an optimal cut-off frequency for filtering one-dimensional data streams. We demonstrate FcOpt in the context of a ski turn segmentation algorithm. This methodology counteracts the susceptibility for incongruously identifying cutoff frequencies by automated methods caused by high sampling rates. FcOpt suggests a cut-off frequency of 2.63 Hz instead of the originally proposed 3 Hz. Filtering with the suggested cut-off frequency on average deviates from the original temporal accuracy of the ski turn segmentation by 1.0 ms, which corresponds to only 0.08% in relation to the mean turn duration. Although FcOpt cannot entirely replace heuristics for cut-off frequency determination yet, it is an easy tool for researchers who want to improve the signal pre-processing for their segmentation algorithms. It lays the groundwork for future developments in the area of data-driven filter design.