Introduction Wind turbine main shaft crack detection is crucial for operational safety and maintenance planning. Conventional feature based diagnosis generalizes poorly to complex or unseen cracks, and deep learning is constrained by scarce and imbalanced defect data. This study proposes an acoustic signature driven multi-scale CNN (MSCNN) framework for identifying unknown main shaft crack defects. Methods A double threshold energy to zero-crossing (EZR) segmentation method is introduced to construct acoustic feature maps that capture both transient and steady-state crack characteristics, enhancing detection sensitivity and specificity. The MSCNN architecture automatically extracts multi-scale temporal features without manual feature engineering, while a novel segmentation strategy decomposes complex or unknown cracks into identifiable components for quantitative assessment. Results The proposed EZR-driven MSCNN framework achieves an average recognition accuracy of 90%, representing a 6.73% improvement over extreme learning machine (ELM) and a 3.36% improvement over single scale CNNs. Cross platform testing confirms robust adaptability, with accuracy ranging from 83.9% to 87.2% across different turbine models. Visualization analysis demonstrates improved separability of crack related acoustic features compared to conventional single-scale or handcrafted feature baselines. Discussion This work provides a practical and effective solution for wind turbine crack detection with enhanced capability for detecting diverse and previously unseen crack types in data scarce scenarios. The proposed framework demonstrates superior recognition stability and supports practical condition monitoring and early warning systems for wind turbine maintenance.
Zheng et al. (2026) studied this question.