Predictive maintenance in industrial plants: real application of Machine Learning models for prognostics

The fundamental role that maintenance plays in the operating costs and productivity of plants has led companies and researchers to shift their interest on this issue. The last frontier of innovation, made possible by the advent of the fourth industrial revolution, is predictive maintenance. It aims to obtain an accurate forecast of the useful life of plant components necessary for the interventions’ management. Consequently, the use of IoT technologies and the analysis of big data acquired increase the assets’ productivity. In this context, a methodology for the prognostics of industrial plant components has been developed using a Machine Learning model. Particularly, a Support Vector Machine (SVM) classification method has been used for the diagnosis of a specific failure mode. This model, which identifies the hypersurface that stems the function classes of the machinery, allows to monitor over time the distance of its operating point from the separation hypersurface. Using a time series analysis model, it is possible to predict the moment in which the machine undergoes a change of state, hence its remaining useful life (RUL). The application of this methodology to a real case study has given the possibility to validate the proposed diagnostic model, building an SVM model with an accuracy of 100%. In this study is also explained the idea for the implementation to the case study of the prognostic model that will be explored and validated in subsequent studies. The analysed machine is a multistage centrifugal compressor, which extracts gas from the condenser of a geothermal power plant. The investigated failure mode was the compressor surge. The developed method also lays the foundations for the implementation of an industrial control system simply setting the distance between classes provided by the SVM classification model to the desired target value