Presentations

Abstract

Introduction Condition monitoring of wind turbines (WTs) is essential for early fault detection, reducing downtimes, enabling predictive maintenance, and for improving the cost effectiveness of wind energy. SCADA-based normal behavior models (NBMs) are employed as a data-driven approach to characterize a wind turbine’s condition and for fault diagnosis. However, NBMs require a substantial amount of training data. When NBMs are trained on scarce and non-representative data, the resulting NBMs can exhibit significantly inferior fault diagnosis performance. To address this challenge, we propose a novel transfer learning and domain adaptation approach that yields more accurate fault diagnosis with only few weeks of training data from new wind farms by leveraging knowledge from existing wind farms. In our work, we present a domain mapping approach. Domain mapping involves ”mapping” data from a target domain, represented by a WT lacking representative data, to resemble data from a source domain, defined by a substantially different WT with abundant and representative training data. This process ultimately allows us to directly use the source WT's reliable NBM with limited data from our different target WT, enabling earlier and more reliable fault diagnosis. Methods Domain adaptation is critical to overcome the limitations of data scarcity, as WT-specific NBMs are tailored to the data they are trained on. Applying a model developed for one WT to data from a substantially different WT can lead to inaccurate predictions and unreliable fault diagnosis. This is due to a shift in e.g., operational behavior or geographical locations. Our work demonstrates how to artificially make the WTs identical by transforming their sensor measurements into one another. A critical component in this approach is that the mapping should preserve the sample’s content when translating it to another domain. For example, a SCADA measurement capturing a WT running idly should remain in an idle state following the domain mapping, and critically, anomalous behaviour should be mapped to anomalous behaviour across WT domains. Our proposed model is a domain mapping network that can translate SCADA samples from one WT (“source domain”) into resembling samples from another WT with only limited, non-representative training data available (“target domain”) and vice versa. The network is based on CycleGAN coupled with two physics-informed consistency losses to preserve the SCADA content during mapping. Once the domain mapping network is trained, the scarce target WT’s SCADA data is mapped to the source domain. The mapped SCADA data from the data-scarce target WT can then be used to obtain the expected normal behaviour of the target WT states through the pretrained source NBM for anomaly detection. Results Our findings show significantly improved fault diagnosis compared to NBMs trained on scarce data. Moreover, our domain mapping approach outperforms conventional fine-tuning across all considered varying data scarcity degrees from 1 week to 2 months. Our presented technique enables earlier and more reliable fault diagnosis for newly installed wind farms, demonstrating a novel and promising research direction to improve anomaly detection performance when faced with a lack of SCADA data.