An unsupervised multi-class ensemble classifier for identifying equipment failure mechanisms from maintenance records

Unstructured data in equipment maintenance records contain valuable information regarding failures. The ability to classify failure incidents into contributing failure mechanisms can help in improving equipment design and maintenance plans to achieve higher equipment uptime. Supervised learning approaches for automated extraction of failure mechanisms from unstructured data are impractical due to the manual labeling effort. Further, due to the complexities inherent in unstructured data, it can be beneficial to utilize multiple base classifier algorithms for analyzing the maintenance records from different perspectives. In this paper, we propose a novel unsupervised multi-class ensemble classifier (UMEC) model to automatically extract failure mechanisms from unstructured maintenance records by leveraging continuous scores generated by multiple base classifiers. The model decomposes the unsupervised multi-class classification problem into multiple binary classifiers using error-correcting output codes (ECOC). We encode the multi-class classification problems to multiple binary classes by using maximum as an order statistic to reduce multi-class scores to binary-classes. We also address the issue of unbalanced datasets in unsupervised classification. We study the influence of different types of noise structure (including feature noise and class mislabeling noise) over the classifiers and demonstrate the effectiveness of our approach using simulated and real-world industrial data.