A Monitoring Framework for Multi-Fault Detection in Rotating Shafts Integrating Deep Learning and Signal Reconstruction

Abstract

Shafts and rotors are essential components in many engineering applications, often playing a critical role in system functionality. Unexpected failures can result in costly downtime or pose risks to human safety, with both situations to be prevented. To mitigate these risks, there is a growing interest in monitoring systems that can proactively detect potential faults and provide real-time diagnostic insights to operators. However, rotor damage often leads to imbalances that can adversely affect other components, such as supports, or induce secondary faults, like cracks caused by high forces associated with rotation. Detecting both the imbalance and the presence of additional faults enhances the assessment of operational risks, enabling early alerts for catastrophic failures and safety-critical concerns. This study presents a monitoring framework for detecting multiple sources of damage in a transmission shaft by analyzing the accelerations at the support bearings during rotation. The rotor is experimentally tested on a test rig, with imbalance masses placed at various locations to simulate the effects of bullet impacts. In addition, a second damage source, specifically a support fault, is also artificially introduced and tested.. The proposed methodology combines a CNN for imbalance diagnosis with a signal reconstruction model for anomaly detection, both using the same acceleration data. This dual approach enables accurate localization of imbalances and detection of unknown faults while minimizing data requirements. The framework is designed to encourage future implementations in real-world applications potentially subjected to multiple faults.