A generic fault prognostics algorithm for manufacturing industries using unsupervised machine learning classifiers

This paper presents a generic methodology for fault forecasting or prognosis in industrial equipment. Particularly, this technique regards training some unsupervised machine learning model using high amount of historical process data from such equipment as input and stop data as reference. The goal is to correlate as strongly as possible the anomalies found by the model in the process data with upcoming faults, according to a forecasting horizon. In this way, the outlier detection model serves for fault forecasting. In this work the pre-processing and automatic feature selection phases, which are of significant importance, are also described. Such trained model would be useful for an industrial operator if executed in real time, based on online process data, since potential anomaly alerts raised by the model could enable predictive maintenance. This method has been applied to real industrial data related to aluminium and plastic production. The experimental results, with Matthews Correlation Coefficient up to 0.73 for the binary classification problem formulated to evaluate forecasting, provide strong evidence that machine learning models are capable of successfully forecasting upcoming faults before their occurrence, despite the general difficulty to find useful information in the process data for fault forecasting.