Estratégia para detecão de falhas em equipamentos industriais com baixa repetibilidade de eventos usando redes neurais

Abstract

Many of the techniques used in the area of predictive maintenance seek to teach the models at least two operating situations: one corresponding to normal periods, in which there is no evidence of failure, and another covering periods prior to a failure event. However, in several types of equipment, the reality is that the repeatability of failure events is low. In this work a modeling strategy that seeks to model the normal operating period is presented. The use of fault event data is only for validation and adjustment of decision thresholds. This strategy constitutes a framework, generic enough to be applied to different equipment, and has been used in predictive maintenance software for different equipment families: control valve, reciprocating compressors and pumps. The approach is built on top of an Artificial Neural Network architecture, which takes care of learning the normal operating pattern, and is subsequently used as a comparative baseline for diagnosing the equipment, effectively implementing a digital twin of its response. A case study is performed on real data, collected in the period 2015-2018, from sensors related to one of the control valves of a large Brazilian chemical industry. The whole step of preprocessing, cleaning, model building and decision making is presented and discussed. The strategy presented was able to identify degradation trends in 6 out of 10 fault records that, in conjunction with alarm generation based on fixed thresholds, would make it possible to generate alarms up to 20 days in advance if the model were operating in production (online).