The growing wind energy sector faces a continuous challenge in optimizing performance and mitigating operational risks. Among these concerns, the premature failure of wind turbine rolling element bearings poses a significant financial and logistical hurdle. Early detection of bearing degradation plays a crucial role in minimizing downtime, preventing cascading system failures, and optimizing maintenance expenditures. While traditionally, expert systems utilizing pre-defined vibration thresholds have served as the primary diagnostic tool, their limitations in scalability and adaptability necessitate the exploration of alternative approaches. This study investigates the efficacy of deep learning in bearing anomaly detection compared to conventional expert systems. It proposes a deep autoencoder architecture, trained on extensive vibration data from a healthy turbine. This model extracts salient features from complex signals, enabling the identification of subtle deviations from the established baseline, indicative of potential bearing anomalies. Notably, this approach eliminates the cumbersome and error-prone process of manually setting thresholds, potentially fostering greater scalability and adaptability across diverse operating conditions. To comprehensively evaluate the proposed deep learning model, a comparative analysis alongside a well-established expert system was conducted. Both methodologies were applied to a year-long dataset of vibration data encompassing both healthy and anomalous operating periods for a machine. The results demonstrate the deep learning model’s potential advantage in anomaly detection accuracy. While the expert system exhibited limitations, potentially leading to false positives and overlooking incipient anomalies, the deep learning model displayed promising sensitivity and specificity. It may be able to pinpoint anomalies sooner and minimize false alarms, offering potentially valuable opportunities for proactive maintenance interventions and avoiding catastrophic failures. Beyond its immediate implications for wind turbine health management, this study suggests potential for the transformative potential of deep learning in predictive maintenance across other industrial sectors. The inherent data-driven nature of deep learning algorithms has the potential to mitigate dependence on expert knowledge and manual threshold adjustments, potentially paving the way for automated and scalable anomaly detection systems. Such advancements could potentially enhance operational efficiency, improve resource allocation, and ultimately contribute to a more sustainable and resilient energy landscape. In conclusion, this research provides initial evidence for the potential superiority of deep learning in wind turbine bearing anomaly detection compared to traditional expert systems. The proposed deep autoencoder model, characterized by its promising accuracy, scalability, and potential for adaptability, represents a potentially transformative development in predictive maintenance with far-reaching implications for the wind energy industry and beyond. As technological advancements evolve, deep learning stands poised to potentially contribute to a future of efficient and sustainable energy production.
Deep Learning, Vibration, Wind Turbines, Rolling Bearings, Predictive Maintenance
