Aplicação de ferramenta de machine learning para predição de falhas em equipamentos em indústria de papel

Abstract

The evolution of technologies in recent years has been increasingly fast and, driven by the 4th industrial revolution, innovations have also become more accessible and applicable. This work meets the need to search for new tools to guarantee and increase the availability of equipment and machinery in a paper and pulp mill. In order to ensure the availability of assets, it is necessary to know the correct mode of operation of the equipment, its failure modes and vital variables for monitoring, thus seeking to predict when the assets have the first defect indicators and prevent them from reaching a functional failure (total stop). In search of this early prediction of equipment failures, we find in Machine Learning an opportunity to enhance the analysis and identification of defects, correlating several vital variables of the assets simultaneously (multi-variable analysis) and repeatedly in short intervals of time (every 5 min). Then we developed a pilot project to evaluate the machine learning tools on the market, covering 10 groups of assets in a paper industry. The assets selected for the pilot were carefully selected. The standard operational model is the result of training the artificial intelligence tool, so that the machine identifies the most subtle deviations in the vital variables of the assets and signals it with early warnings of potential failures. The pilot was implemented in 4 months, with the participation of the project team of a digital technology company together with the maintenance engineering team of the paper industry. As a result of the application of the failure prediction tool, it was possible to identify and treat 2 potentials of failure in the first months of use. This predictability allows the field maintenance teams to plan the activity with more time in order to maximize the operation of the machines, with safety and guaranteed productivity. We consider the pilot project extremely successful, due to the ease of application of the tool and also due to the accuracy of predicted and anticipated failures.