Bolt tightness evaluation using acoustic emission technique via Hankel transform and residual convolutional networks

Bolted joints are susceptible to loosening under cyclic vibrations, which may compromise structural integrity and operational safety. Acoustic emission (AE) provides a passive and high-frequency sensing approach for monitoring interfacial friction and micro-slip associated with bolt loosening. However, the extraction of robust and discriminative features from raw AE signals remains challenging. To address this issue, this study proposes a deep learning framework for bolt tightness evaluation by integrating the Hankel transform with AE signal analysis and a ResNet-18 convolutional neural network. Specifically, single-period AE signals are extracted through zero-crossing detection of vibration signals. After equidistant down-sampling, each AE segment is reorganised into three Hankel matrices and encoded as a three-channel image, to enable efficient convolutional feature extraction while preserving temporal continuity. The proposed framework is systematically evaluated using the publicly available ORION-AE dataset under both same-domain and cross-domain scenarios. A comprehensive set of comparative studies is conducted, including different data processing methods, varying image sizes and multiple convolutional neural network (CNN) architectures. The comparative results across multiple evaluation aspects demonstrate that the proposed method achieves consistently superior classification performance under both same-domain and cross-domain conditions.