Robustness of CNN models for EMI‐based bolt preload prediction

In a previous study, the combination of electromechanical impedance (EMI) and convolutional neural networks (CNN) was shown to enable accurate prediction of bolt preload. The models were trained in a position‐specific manner and applied to the trained bolt under controlled laboratory conditions, resulting in robust prediction performance. The high level of accuracy obtained in that study provides the basis for further development of the approach with regard to generalisability and robustness for practical application. Key questions concern the transferability of a reference model to different bolt positions and the reduction of training effort through modelling at the bolt‐type level. This study therefore systematically investigates position and bolt transferability, as well as the prediction of temperature information from EMI spectra. A compact CNN was trained on data from a single bolt and validated using measurements from different positions and bolt specimens. The results demonstrate high prediction accuracy ( R 2 = 0.95, MAE = 5.8 % F p,C ) across different positions. In addition, temperature values could be reconstructed precisely from the EMI spectra (MAE = 0.09 °C), providing a basis for future temperature compensation strategies. The findings highlight the potential of the EMI‐based CNN approach to evolve from a laboratory application towards a robust, transferable, and scalable solution for practical structural health monitoring (SHM).