Evaluation of a Combined Conductive Fabric-Based Suspender System and Machine Learning Approach for Human Activity Recognition

Abstract

Accelerometer-based human activity recognition (HAR) wearable systems are location-centric and noisy, needing multiple sensors with complex signal processing and filtering mechanisms. A recently reported alternative approach using a wearable suspender integrated with strain sensors and machine learning presented a viable option for nonlocalized measurement with less noise and better recognition capabilities. The washability and wearability of the strain sensor instrumented suspenders due to the physical wires are limited, and the power consumption is higher, which needs to be minimized to extend the battery life of the wearable device. This article proposes an improved body-worn suspender-based HAR system built using a conductive knit jersey fabric material that overcomes the existing strain sensor-based wearable device’s limitations and at the same time provides improved sensitivity. The proposed suspender system recognizes 14 human activities using machine learning and deep learning algorithms with the best accuracy of 98.11%. A performance comparison of machine learning models based on two dimensionality reduction techniques using kernel and linear discriminatory analysis was conducted. The kernel-based method outperformed the linear one in recognizing human activities across all classifiers. The durability of the wearable is tested by washing the sensor, and the recognition capabilities were consistent before and after the wash.