Comparing Supervised Machine-Learning Algorithms to Measure the Number of Technical Actions in Industrial Environments

The counting of technical actions during work cycle is fundamental to identify biomechanical risks. One of the most widespread indices to assess the risk of upper limb musculoskeletal disorders is the Occupational Repetitive Action (OCRA), where the counting requires the video analysis of workstation performed by a skilled operator. This approach is questionable due to the lack of objectivity, leading to low level of repeatability and reproducibility. Thus, the automatization of such approach is still an untapped challenge. In this context, the paper aims at evaluating the feasibility to use machine-learning algorithms fed with inertial data to automatically count the technical actions in industrial environments. From this perspective, we performed a comparative analysis of seven machine-learning algorithms fed with features extracted from joint angles related to shoulder, elbow and wrist. Data was gathered from one worker in six different real industrial scenarios. The worker was sensorized with 17 inertial sensors. An RGB camera was also used during the experimental protocol in order to create the reference labelled actions. The performance of the classifiers was computed considering the confusion matrix and the overall accuracy. In addition. Two approaches were tested, where the first one consisted in a multiclass classification (M), whereas the second one in a binary one (B). Specifically, M consisted in the recognition of each specific technical action before the counting; whereas B in the recognition of technical actions with respect to no-actions phase, regardless the type of the action. Support vector machine with a quadratic kernel achieved the best accuracy, which was greater than 77.7% for all the tested scenarios when considering a binary approach. The findings of this study could represent a step forward in the automatization of technical action counting for the computation of the OCRA index.