Mid Sweden University

The importance of preventing machine failures and reducing costly unplanned downtime has led to extensive research aiming to develop methods that predict maintenance needs. In this context, data-driven and particularly Deep Learning (DL) based methods for anomaly detection, fault diagnosis, and health prognosis have been studied extensively because of their ability to handle the complexity of the sensor data describing the health state of machines. However, many challenging factors exist before it is possible to utilize these methods in practice. These primarily include the lack of labelled failure events, the heterogeneous nature of the data, and the occurrence of multi-component fault scenarios. Currently, most studies ignore these aspects and focus on scenarios limited to a laboratory environment, which means there is a need to develop methods that can be deployed in practice. Therefore, this thesis suggests methods for these challenges and gives insight, aiming to reduce the gap between research-defined scenarios and scenarios found in industrial environments. To achieve this, different areas in the context of DL for Predictive Maintenance (PdM) are examined, including multivariate anomaly detection, fault diagnosis, and Remaining Useful Life (RUL) prediction methods.

One of the contributions is a threshold-setting procedure that optimizes anomaly detection models with the user's support and a novel separate scoring method, and outperforms state-of-the-art alternatives for deployments in industrial applications. A published dataset of bearing faults from an industrial environment is also described, which is beneficial when developing and evaluating methods. In addition, a novel DL method for fault diagnosis of bearings using vibration data constructed with knowledge enrichment, time-based contextual enrichment, and a transfer learning technique is suggested. This method can be deployed on any machine without historical faults and outperforms state-of-the-art methods. Lastly, the most significant contribution is a prognostic hybrid framework for multi-component fault scenarios in rotating machines using vibration data that utilizes advancements in methods for anomaly detection, fault diagnosis, and RUL prediction of machines.

In summary, this thesis suggests novel methods for PdM adapted for industrial applications that can be used on a general basis and provides insights that lower the gap between research-defined scenarios and scenarios found in industrial environments.