Application of artificial intelligence techniques to condition monitoring and predictive maintenance of mechanical components

In the era of Industry 4.0, the reliability of complex mechanical systems is paramount, yet unexpected equipment failures remain a critical challenge, leading to significant economic losses and operational disruptions. This thesis presents the development and validation of a systematic, multi-stage framework for early fault diagnosis by applying Artificial Intelligence (AI) techniques to Condition Monitoring (CM) data. The core of the proposed methodology is a hybrid approach that integrates physics-informed feature engineering with unsupervised anomaly detection models. A key contribution is the novel strategy of using one sensing modality, such as online oil debris analysis, to supervise the definition of a robust 'healthy' baseline for training the unsupervised models. The framework processes multi-sensor data—primarily vibration and ultrasound—to generate both a General Health Index (GHI) for an overall assessment of system health and component-level Specific Health Indices (SHIs) for targeted diagnostics. The effectiveness of this framework is validated across two distinct, large-scale industrial case studies: a run-to-failure durability test on a heavy-duty mechanical gearbox, and a series of developmental tests on automotive electric powertrains, including Accelerated Durability and End-of-Line testing. In the gearbox case study, the methodology successfully identified an incipient bearing fault, providing over 800 hours of lead time before catastrophic failure. In the electric powertrain applications, the framework demonstrated its adaptability in detecting a range of anomalies, from lubrication system degradation to the onset of severe mechanical faults, and its viability for real-time, onboard implementation was verified. Ultimately, this research contributes a validated, adaptable, and scalable framework for implementing data-driven predictive maintenance. It demonstrates a practical and effective methodology that bridges the gap between theoretical AI models and the demands of real-world industrial machinery, offering a clear pathway to enhanced operational reliability and efficiency.