Learning from acceleration data to differentiate the posture, dynamic and static work of the back: am experimental setup

Abstract

Information on body posture, postural change, and dynamic and static work is essential in understanding biomechanical exposure and has many applications in ergonomics and healthcare. This study aimed at evaluating the possibility of using triaxial acceleration data to classify postures and to differentiate between dynamic and static work of the back in an experimental setup, based on a machine learning (ML) approach. A movement protocol was designed to cover the essential degrees of freedom of the back, and a subject wearing a triaxial accelerometer implemented this protocol. Impulses and oscillations from the signals were removed by median filtering, then the filtered dataset was fed into two ML algorithms, namely a multilayer perceptron with back propagation (MLPBNN) and a random forest (RF), with the aim of inferring the most suitable algorithm and architecture for detecting dynamic and static work, as well as for correctly classifying the postures of the back. Then, training and testing subsets were delimitated and used to evaluate the learning and generalization ability of the ML algorithms for the same classification problems. The results indicate that ML has a lot of potential in differentiating between dynamic and static work, depending on the type of algorithm and its architecture, and the data quantity and quality. In particular, MLPBNN can be used to better differentiate between dynamic and static work when tuned properly. In addition, static work and the associated postures were better learned and generalized by the MLPBNN, a fact that could provide the basis for cheap real-world offline applications with the aim of getting time-scaled postural profiling data by accounting for the static postures. Although it wasn't the case in this study, on bigger datasets, the use of MLPBPNN may come at the expense of high computational costs in the training phase. The study also discusses the factors that may improve the classification performance in the testing phase and sets new directions of research.