Classificação de perfis de produtividade usando redes neurais artificiais a partir de registros eletroencefalográficos: uma aplicação na Mineradora Vale S.A., Complexo S11D Eliezer Batista

Abstract

Results and methods from neuroscience can already be applied on a routine scale. Applied neuroscience is in use, for example, to measure and study brain activity under high demand using electroencephalography (EEG). In certain work environments, human productivity is a direct function of brain activity. The intense recruitment of cognitive functions such as sustained attention and working memory influence productivity directly. The labor demands experienced by workers in mining companies are an example. This work is a part of a bigger project called “Usando Treinamento Cognitivo para o Desenvolvimento de Operadores de Alto Desempenho” from the Vale S. A. mining company and, therefore, all of the workers that participated in this study worked for that company. We used EEG to measure productivity during a task without the constraints typically found in experiments conducted in the laboratory. To this aim, we created a machine learning algorithm to analyze the resting EEG recorded before and after a 4D simulation, where mining workers (37 ± 7 years old) specialized to operate giant high-capacity shovels. The simulated task consisted of the operator using the shovel in a digging and loading routine. The task was not planned for our research. It was part of the worker’s routine training. That is, we took advantage of the workers’ training to carry out the study. Recordings were preprocessed using a band-pass filter (0.5-100 Hz) followed by filtering using ICA (Independent Component Analysis). After each operator finish the simulation, their productivity was measured by VALE S.A technical staff. The main parameter for good productivity was the amount of ore excavated. The operators were divided into groups according to ranges of productivity. The productivity was used as a label for the learning of the algorithm that consisted of an artificial neural network of the type inception. The number of neurons and layers was optimized using Bayesian optimization. The features extracted by the inception were the input to 13 classifiers. The classifier chosen for the final algorithm was the best, that is the one providing the best accuracy in the productivity classification. The training set contained 80% of the data. A hold-out validation was used to test the accuracy of the final algorithm using 20% of the data. The resulting accuracy when the operators were divided into four groups of productivity reached 91.35%. When there were only two groups of productivity the accuracy peaked at 95.05%. Our results showed that even under no laboratory constraints – during the regular training of the operators and using resting EEG – our algorithm succeed and it is ready to be used in future field operations. We have a prototype that is patentable.