In Pursuit of Open-Weights Models for Intrinsic Interpretability in Wind Turbine Blade Monitoring with Tower-Mounted Radar

Wind turbines are critical infrastructure whose economical deployment benefits from blade monitoring. Tower-radar remote sensing generates radargrams from a mast-bound active sensor for this and related purposes. Initial machine learning classifiers are publicly available, yet they have not been stress-tested. Suitable test imagery remains scarce. This paper addresses these gaps through interconnected investigations, supported by two data contributions: a synthetic surrogate benchmark (spanning eight image dimensions in a full-factorial design) and an enrichment of the empirical WiRoRa dataset augmented with human annotations and machine-generated ones where the latter can also serve as an on-the-fly labeling solution in the field. The studies report: multi-class anomaly filtering; a sensitivity analysis revealing that the end-to-end-trained (E2E) classifier is poorly calibrated, while its pretrained counterpart is substantially more stable; adversarial vulnerability evaluation showing that the E2E model is also more easily fooled; an analytical derivation of when/why bottleneck training on auxiliary imagery improves representations; a surrogate test confirming the bottleneck hypothesis; and a preliminary Mixture-of-Experts pilot for enhanced traceability as well as scalability that performs environmental/operational metaparameter regression as an archetypal example. Together, the results expose failure modes of existing classifiers and chart a path toward intrinsically interpretable systems for structural health monitoring and beyond.