Traversing blades and where to find them in visual-language latent landscapes: Exploring contextual computer-vision domains and model compression for tower-radar rotor monitoring
Christian Kexel, Sercan Alipek, Jochen Moll
TL;DR: This article focuses on the radar-only computer-vision task of classifying rotors without complementary costly instrumentation, and measured radargrams from field experiments are complemented with the novel synthetic dataset SiWiRoRa as well as further open imagery.
Tower-radar computer vision (TRCV) represents an emerging application-oriented field of study. Here, image-type measurements are acquired from radar transceivers bound to the mast of wind power turbines and these radargrams subsequently get analyzed using data-driven algorithms. TRCV shows promise for reducing downtime and increasing safety of such renewable energy plants, accordingly, the respective monitoring systems should long-term perform real-time recognition of moving objects like rotor blades and their condition, disregard maintenance workers, identify birds and bats or unauthorized aerial vehicles, in order to trigger appropriate multi-faceted responses. This article focuses on the radar-only computer-vision task of classifying rotors without complementary costly instrumentation [1]. Progress in TRCV for blade monitoring remains, however, hindered by the limited publicly available data [2]. For this specialized task, recently [3] general-purpose feature extractors have proven robust to environmental and operational conditions and as a valuable building block in TRCV processing pipelines. Such large models, like OpenCLIP, have been pre-trained on internet-scale amounts of text-image pairs. They however (i) commonly still require additional data for transfer to dedicated use cases, as well as (ii) exhibit power demands or latencies incompatible with real-time resource-constrained application scenarios, models hence need to be compressed. In this paper, measured radargrams from field experiments are therefore complemented with the novel synthetic dataset SiWiRoRa as well as further open imagery. Here, several specialized image-type datasets are carefully compiled to span a context around the focal measured radargrams. Second, the main geometrical directions of this context landscape are explained by classical image statistics and with human perceptions collected from annotators. Third, large models are distilled towards lightweight capacity where it is found that both the synthetic dataset but also seemingly unrelated imagery can improve performance in blade classification. Accordingly, routes for enhancing the SiWiRoRa dataset are suggested and moreover implemented so as to further advance TRCV. References [1] Alipek, Sercan, et al. "Potential and Limitations of Anomaly Detection via Tower-Radar Monitoring of Wind Turbine Blades in Regular Operation with Convolutional Networks." EWSHM (2024). [2] Mälzer, Moritz et al. "Radar-based structural monitoring of wind turbines blades: Field results from two operational wind turbines." IWSHM (2023). [3] Kexel, Christian et al. "Mast-Bound and Too Curious: Overcoming Drift in Wind-Tower Radar for Blade Monitoring Using Pre-Training, Augmentation and Weight Consolidation Due to Correlated Conditions." LATAM-SHM (2026)