A Deep Learning Approach for Short Term Prediction of Industrial Plant Working Status

Abstract

Predictive Maintenance has gained more and more research and commercial interests, being a pivotal topic for improving the efficiency of many production industrial plants to minimize downtimes, as well as to reduce operational costs for interventions. Solutions reviewed in literature are increasingly based on machine learning and deep learning methods for prediction of fault proneness with respect to normal working conditions. Many state-of-the art solutions are not actually applied in real scenarios, and have restrictions to be executed in real-time in the production environment. In this paper, a framework for predictive maintenance is presented. It has been built upon a deep learning model based on Long-Short Term Memory Neural Networks, LSTM and Convolutional LSTM. The proposed model provides a one-hour prediction of the plant status and indications on the areas in which the intervention should be performed by using explainable LSTM technique. The solution has been validated against real data of ALTAIR chemical plant, demonstrating an high accuracy with the capability of being executed in real-time in a production operative scenario. The paper also introduced business intelligence tools on maintenance data and the architectural infrastructure for the integration of predictive maintenance approach.