VASCONCELOS, B. R. A.; http://lattes.cnpq.br/7355135638886385; VASCONCELOS, Bruno Rafael Araújo.
Résumé:
The identification of failures in industrial plant engines has great value for companies, as it aims to avoid consequences, such as: the burning or damage of equipment, the death of workers and environmental catastrophes, among others. The motors used in thermoelectric plants are taken out of operation (turned off) whenever a serious failure occurs. For each engine malfunction, in the power generation process, an alarm is sent to a control center, which is kept in an event history. Each alarm is associated with a single fault and, therefore, a sequence of alarms can be seen as a sequence of failures. Identifying when shutdowns will occur can help operators avoid them, through corrections in advance, to reduce losses in the production process. Given this scenario, the present research aimed, from the history of alarms, to extract characteristics and structure them to train a prediction and prognosis model. The model, built, consists of a Machine Learning Model. The proposed approach showed satisfactory results. Of the total of 57 valid cases in which there were disconnections, the technique provided a rate for the context, of 48 hits. Of the 507 in which there were no disconnections, the model got 390 cases right,
considering a prediction threshold of 0.5 and using the cross-validation technique (k-fold, with k = 10). The approach to predicting shutdowns therefore made it possible to make prognoses on the engines, anticipating shutdowns in advance. The results presented here, the number of false negatives shows that the model can present significant results when trained with a greater number of examples of disconnection. Thus, the proposed approach proved to be viable for predicting engine shutdowns at a thermoelectric plant.