Data-driven Industrial Machine Failure Detection in Imbalanced Environments

Machine failure often leads to unplanned downtime in industrial manufacturing, which could result in a significant loss in the manufacturer’s revenue. Several machine learning-based approaches have been proposed to alleviate the problem by instantly detecting occurring failures or predicting any potential breakdowns. However, there still exist limitations and issues that require attention. These include the difficulty of collecting real-world industrial data, especially big data, the challenge of feature selections and the under-representation of machine failure events in the data. In this paper, we present the use of a small predictive maintenance dataset with basic supervised learning algorithms for industrial machine failure detection. Moreover, we show the need of handling the imbalanced class distribution in such data for more accurate detection. Several non-deep learning algorithms were used for the classification task, and data resampling methods were applied to improve the model performance. Results show that decision tree could provide promising classification results, and with an undersampling method, the detection accuracy of 91% could be achieved.