Time-Series Data to Refined Insights: A Feature Engineering-Driven Approach to Gym Exercise Recognition

Abstract

Machine learning-based sports activity recognition has captured a lot of interest in recent years. Automatic activity recognition not only reduces cost and time but is very helpful in analyzing health-sensitive data acquired using smart wearable technology. Gym activity recognition by incorporating smart wearable technology comes within the scope of this topic. This paper present a system for classifying gym activities using feature engineering techniques applied to time series data. The collected time series data consists of an athlete’s body movement using an internal 3-axis accelerometer build into the zephyr bio-harness 3 device. The data were gathered by implementing a six-week fitness routine trying to target six muscle groups, preceded by one day of rest and recovery each week. The raw time-series data of the accelerometer is transformed to extract new features from it for identifying gym activities. The feature engineering techniques applied in this research are not limited to gym activity recognition but can be extended to any domain involving time-series data. The collected data was just three features, which are the reading of the tri-axial accelerometer signal as vertical, lateral, and sagittal axes. In order to formulate new features, basic concepts of statistics and mathematics were applied to the data. furthermore, we trained six GridSearchCV-based classifiers on the extracted features and tested their performance in four different types of experiments.

Keywords

• Internet-of-Things (IoT)
• feature engineering
• gym activity recognition
• machine learning
• smart workout