A combined vibration-machine learning method for structural performance assessment of modular supporting structure

Modern modular and prefabricated supporting structures, such as those used in wind energy systems, require reliable assessment methods to ensure long-term structural performance. In modular structures, bolted joints typically exhibit much higher local stiffness than adjacent structural components. Consequently, minor stiffness reductions induced by bolt looseness often result in only subtle variations in global modal characteristics, making joint damage difficult to localize accurately using conventional modal-parameter-based approaches. To address this challenge, this study proposes a combined vibration-machine learning method for detecting and quantifying joint damage in wind turbine support structures. A hybrid deep learning architecture integrating a one-dimensional convolutional neural network (1D CNN) and a long short-term memory (LSTM) network is developed to extract spatial features and to model long-range dependencies between frequency components, thereby enhancing sensitivity to subtle joint stiffness changes. Experimental investigations on a lab-scale modular structure with multiple damage locations and severity levels are conducted to validate the proposed approach. The results demonstrate that the proposed method can accurately estimate joint degradation and reliably identify the damaged joint under noise and uncontrolled excitations. Comparative analyses further demonstrate that the proposed framework achieved improved performance compared with two existing 1D CNN models.