Physical exertions recognition using surface electromyography and inertial measurements for occupational ergonomics

Automation of ergonomic risk level assessments using predictive models is limited to the recognition of certain activities with the intention of calculating some ergonomic risk factors, this means that the predictive models are only useful to recognize the activities that were taught. In this thesis, a framework was developed to automatically recognize the physical exertions done by an operator during a manual work task towards the automation of Job Strain Index (JSI) assessment. The framework includes the use of a wearable device that captures surface electromyography (sEMG) signals and inertial measurements called Mindrove armband, and provides the data treatments that maximized the training accuracy of a Cubic Support Vector Machine (CSVM) model, which was responsible for predicting the exertions depending on the behavior of the data. To determine the best data treatments, full factorial experiments were designed and analyses of variance (ANOVA) were performed. Thus, the best data treatments to obtain a maximum average training accuracy of 93.29% and testing accuracy of 94.31% of the CSVM were the filtering of raw signals with a 4th order Butterworth filter, the rectification of sEMG signals, the outliers' removal in data via Hampel identifier, the computation of the RMS envelope and normalization of sEMG, the zero calibration for inertial measurements, and the extraction of 126 statistical features. Additionally, the Visual Signal Analyzer App (SIANA) was developed for data processing, which works under the proposed framework. Automatic recognition of any physical exertion means that an automated JSI can be applied to any manual work task, thereby identifying more quickly the risk level of a work task in order to modify it to avoid occupational diseases and accidents.