Explainable Prediction of Machine-Tool Breakdowns Based on Combination of Natural Language Processing and Classifiers

Prognostics and Health Management (PHM) process has been developed to enhance predictive maintenance (PM) policies and decision support (DS). One of the PHM modules is fault diagnostics, it allows identifying and predicting future faults. Among diagnostic techniques, one can find Natural Language Processing (NLP) that can be used to exploit textual monitoring data such as logging data for fault prediction. However, there exists some unstructured texts that reduce the data quality and provide an un-explainable prediction of faults. To remedy this situation, this paper proposes a NLP methodology for system breakdown prediction. This methodology starts by cleaning textual data. Then, cleaned data and their labels, which represent the breakdown origin, are injected into feature encoding models. These two previous steps address special and redundant characters and non-standard spelling terms. Thus, they allow classifier models to learn mapping input texts to their corresponding labels without confusion for the fault prediction, making these predictions explainable. The proposed methodology is applied to real logging data carried out from a machine tool of a French company SCODER. The machine tool generates six failure labels that classifiers learn to predict. The prediction accuracy obtained by the proposed methodology, compared to existing methods, is promising and can be useful for a failure prognostics.