Compound damage detection using wavelet transform and deep neural network trained on healthy and single damage states: Validation on a laboratory-scale offshore jacket model

Structural health monitoring is vital for the early detection of damage, enabling effective life cycle management of structures. Detecting compound damage, where multiple types of damage occur simultaneously in different sections of a structure, is particularly challenging, especially when some of the damages are subtle or minor. Existing methods typically treat compound damage as a distinct category, separate from single damage types. This paper introduces a novel approach to damage detection based solely on vibration responses, combining the wavelet transform with a deep convolutional neural network with multiple interference (MIDCNN). In this approach, the MIDCNN is trained using time-frequency data from healthy and single damage states, intentionally excluding compound damage data from the training phase. During testing, the MIDCNN model accurately distinguishes between healthy, single damage, and untrained compound damage states when the output probabilities meet predefined conditions. The method is validated using data from a laboratory-scale offshore jacket structure. The results demonstrate the method’s ability to extract relevant features and accurately classify different structural states, including healthy, single, and compound damage.