Abstract

Rotating shifts have become prevalent in many industries, leading to a growing concern about the impact of fatigue on workers performance and safety. Thus, it is useful to develop a method to predict the fatigue of workers with rotating shifts. This thesis aims at contributing to the development of such method by building data-driven models to predict level of fatigue.

We use random forest classifier and random forest regressor to build two fatigue prediction models. A third model is built by a combination of random forest classifier and regressor. Two imbalanced datasets from different groups of workers in the same industry are used. We explore two strategies to deal with imbalanced datasets: random over-sampling and class weights.
We select features with feature importance of random forest and discover that a set of 19 features, selected from 38 original features, gives best performance.

We obtain good prediction accuracy on both datasets. The combined model reaches mean absolute error of 0.93 and 0.83 on two datasets, on a 9-level scale of fatigue. In the area of high level of fatigue, which in real work is of particular interest, our model can predict with average 85\% confidence that the true level falls into +-1 range of prediction.

We conclude that fatigue can be predicted with high confidence, based on a dataset of sleep patterns, work schedules and demographic data. Future work will focus on model generalization to datasets from different industries or geographical areas; and the discovery of other sets of features that give better prediction.