Early detection of critical process failures in industrial systems using machine learning for predictive maintenance

Abstract

The progress of Industry 4.0 has led to significant transformations in the perception and functioning of industrial environments, placing particular emphasis on maintenance tactics. In this context, predictive maintenance (PdM), supported by machine learning (ML) methods, has established itself as an essential tool for anticipating operational errors, minimizing unexpected interruptions, and enhancing process efficiency. This analysis offers a systematic literature review with the aim of examining the application of machine learning algorithms to predict critical errors in industrial processes. To achieve this, the PRISMA methodology was used in combination with the PICOC approach. The first query in the Scopus database yielded 328 results. Twenty-seven articles were obtained that met requirements such as the use of algorithmic models, industrial environments, implementation of PdM and availability of evaluation tools. The main findings showed that the deep learning models with the best predictive performance included LSTM networks, hybrid CNN architectures and small IoT solutions. The most frequent failures were found to be mechanical, electrical and environmental, especially in industries such as manufacturing, energy and transportation. This report assesses the current state of the art and makes recommendations on possible directions for the development of intelligent industrial maintenance.