Monitoring bolt looseness in joint structures is vital for their safety and integrity. The relationship between bolt looseness-sensitive acoustic emission (AE) feature selection and damage mechanisms is unclear. This research combines AE monitoring with machine learning to identify bolt looseness levels and wear mechanisms. Using gradient-boosting tree-based ensemble machine learning models, we evaluated bolt looseness in an AE dataset of bolted lap joints under vibrational excitation, focusing on the influence of AE feature selection, dataset size, and sample imbalance. The rise angle-average frequency-energy and K-means++ algorithms were employed to identify AE sources during bolt loosening, and we developed a cumulative damage index. The Light gradient boosting machine model excelled with an 84.9% accuracy rate. Our findings highlight the sensitivity of Renyi entropy and the 200-300 kHz power band to bolt looseness and reveal a nonlinear relationship between training dataset size and model accuracy. The F1-score emerged as a reliable metric for imbalanced samples. Variations in wear particle generation and migration at different tightening force levels are correlated to the AE source for bolt looseness. The cumulative damage index revealed the time-varied process of wear patterns during bolt loosening. These insights enhance the application of AE and machine learning in bolt looseness monitoring and damage mechanism identification.

中文翻译：

---

基于声发射和机器学习的螺栓搭接接头松动监测和损伤识别


监测连接结构中的螺栓松动对于其安全性和完整性至关重要。螺栓松动敏感声发射（AE）特征选择与损坏机制之间的关系尚不清楚。这项研究将 AE 监测与机器学习相结合，以识别螺栓松动程度和磨损机制。使用基于梯度提升树的集成机器学习模型，我们评估了振动激励下螺栓搭接接头的 AE 数据集中的螺栓松动情况，重点关注 AE 特征选择、数据集大小和样本不平衡的影响。采用上升角-平均频率-能量和K-means++算法来识别螺栓松动过程中的声发射源，并制定了累积损伤指数。光梯度增强机模型表现出色，准确率高达 84.9%。我们的研究结果强调了 Renyi 熵和 200-300 kHz 功率带对螺栓松动的敏感性，并揭示了训练数据集大小和模型精度之间的非线性关系。 F1 分数成为不平衡样本的可靠指标。不同拧紧力水平下磨损颗粒生成和迁移的变化与螺栓松动的 AE 源相关。累积损伤指数揭示了螺栓松动过程中磨损模式的时变过程。这些见解增强了 AE 和机器学习在螺栓松动监测和损坏机制识别中的应用。