Towards Model Development for Sensor-Based Activity Recognition at the Construction Site

Digital solutions for the construction industry are promising to improve energy consumption, tool life cycle, tool design, productivity, safety, health, and risk management. In this study we assess the feasibility of using accelerometer data obtained from a low-power Micro-Electro-Mechanical Systems (MEMS) sensor directly placed on the tool, to identify screwdriver tool usage types. We focus on the performance evaluation of several distinct features and machine learning (ML) techniques regarding their accuracy and model size. To establish a comprehensive data set, we first collect data, identify fitfor-purpose usage classes and, subsequently, apply a variety of feature engineering and ML techniques to the established problem. As two distinct usage class groups, we identify, runtime classes ("Drilling", "Screwing" and "Unscrewing") and the nonruntime classes ("Preparing the Tool", "Carrying the Tool", "Transportation of the Tool", and "No Movement"). The paper proposes two tree-based models Decision Tree Classifier (DTC) and Gradient Boosting Machine (GBM), for which we assess various techniques of automated and handcrafted feature extraction. We design an iterative feature selection method to identify the most important ones from more than 4000 features. Further, we evaluated the neural networks Long Short-Term Memory (LSTM) and Temporal Convolutional Network (TCN), which process time-series data, and the Minimally Random Convolutional Kernel Transform (MINIROCKET). The experimental evaluation focuses on accuracy and model size. The MINIROCKET is the best-suited model with a balanced accuracy of 94.1% and a model size of 377.5 kB, enabling realtime processing in small micro-controller or even in Bluetooth low energy modules.