Author: Mr Pedro Melo STSM Period: 2021-07-26 – 2021-08-09
Hosting institution: Ss. Cyril and Methodius University
PURPOSE OF THE STMS
With commercial development has come an increase in work. Despite being advantageous for different economies and workers, this work makes people more stressed and, above all, tired. Nowadays, feelings of overwork and exhaustion are common. In addition to the sensations described above, fatigue is also one of the main consequences experienced.
Fatigue can be technically defined as tiredness (physical or mental) resulting from effort or different activities.
Therefore, this STSM action is based on the search for more effective methods for detecting fatigue.
Moreover, these methods can be less intrusive and easier to apply in everyday life, making their application in the real world much more enjoyable.
When applied to the context of this STSM action, new techniques used in the software engineering business can prove to be genuinely compelling and exciting. These allow the study of how applicable these technologies are in real contexts and will enable the knowledge of their limits and consequent adjustment of these to the expectations of their use. An example of these technologies is the use of
machine learning to aid fatigue detection.
The motivation for this topic has to do with the interconnection of the study of new and innovative tools, technologies, and areas of study, with the fact that they intrinsically help the human being; in the limit, they can reduce derived accidents, increase the quality of life and make the world safer.
For the development of this project and other projects typically involving machine learning, one of the main factors for the success of the chosen method is the quality of the assets (datasets) used to train the models. These assets can be obtained in different ways, but given the short duration of this STSM, we decided to use datasets publicly available through the analysis of existing publications.
DESCRIPTION OF WORK CARRIED OUT DURING THE STSMS
During my visit, I started by going through a list of publications with topics similar in some way to the topic of fatigue. The goal was to analyze the number of datasets that could be used for machine learning processes. During this process, it was possible to conclude that the number of freely and openly available datasets is considerably low.
After this phase, we contacted an author who made the dataset he had created upon formal request. After approval and help from Dr. Eftim Zdravevski, the author was contacted, and he made the dataset available. Thus, it was possible to download it and make it available internally to the work team.
Later on, sensors easily used in mobile situations, such as ECG and temperature, were also analyzed.
During this STSM action, we developed data processing methods and we’re aiming to publish a paper that will contain the results of this processing as well as a critical view of the same results. After visualization and interpretation of the results, the results and trends obtained were discussed. Brainstorming and idea generation sessions for future improvements were also held.
DESCRIPTION OF THE MAIN RESULTS OBTAINED
The following chapter describes the results obtained during this STSM action and also introduces and details the indicators used to evaluate them.
From the analysis of the works, it was possible to conclude that only 80 out of the 296 checked to discriminate the datasets they have used or made available. Moreover, of these 80, only 20 also provided the corresponding source code.
Since we contacted the author of the publication referred to in the previous chapter, it was possible to obtain local access to a «semi-private» dataset for use.
The following table summarises the values obtained in the course of this task.
During the training and test phases of the model, the model categorizes each of the information samples accessible in the same stage. In addition, after training, the model is tested to search its capability to classify all samples correctly.
After this process finishes, the Loss is measured. Loss corresponds to the disagreement between the classification delivered by the sample and its correct class. Thus, Validation Loss queries the loss rate in each of the training epochs, and Loss is the loss value after the testing phase. The results obtained focus on the objectives stated for COST Action 16226 related to monitoring physical fatigue.
Since the duration of the STSM action does not allow the realization of the second proposal presented in the initial chapter, writing a publication will be resumed after the conclusion of the STSM action.