Integrating machine learning with electromechanical impedance for non-destructive detection of bolt looseness in steel structures

Structural integrity of bolted joints remains critical for the safety and longevity of steel structures. Bolt looseness is one of the failures in steel structures typically caused by cyclic loading, wear, and environmental influences. In the field of structural health monitoring, Surface-bonded PZT sensors enable the Electromechanical Impedance (EMI) method to detect damage in structures effectively because of its high sensitivity and accurate detection capabilities. However, The EMI technique does not establish a clear numerical connection between torque levels and its damage detection parameters including RMSD, MAPD, CCD, and Peak Frequency. To address this limitation, the present study proposes a Machine Learning models for the quantitative prediction of torque values in bolted joints of a steel truss structure. 15 machine learning algorithms were analyzed through regression evaluation of R², MAE, RMSE and EVS metrics. The results show that Extra Trees Regressor and XGBoost emerged as superior predictive analytical models compared to all other techniques. Results from Explainable AI techniques like SHAP and LIME. established Peak Frequency as the key positive predictor along with negative influence between RMSD and MAPD against torque behavior in accordance with EMI-based principles. The results were improved by conducting hyperparameter optimization through GridSearchCV. A user-friendly software application integrating the top models to enable real-time torque prediction through an intuitive Graphical User Interface, providing practical utility for engineers. In conclusion, the integration of EMI with ensemble-based ML models offers a reliable, interpretable, and scalable solution for bolt looseness detection and torque assessment in structural health monitoring applications.