Enhancing load prediction for structures with concrete overlay using transfer learning of time–frequency feature-based deep models

This study proposes a method to predict applied loads on bonded structures via non-destructive testing results. Various types of concrete, such as high-performance alkali-activated slag concrete (HPAASC) and regular Portland cement concrete (OPCC) are simultaneously tested via ultrasound testing and bi-surface shear for extracting features and target values, respectively. Variational Mode Decomposition is employed to extract useful features from ultrasound signals. In the first part, selected features using a PCA-based method are utilized to train a set of deep-learning models to estimate the imposed mechanical load on the tested specimens. The most effective models from the first step, i.e., a CNN-LSTM model and LSTM models, are then fine-tuned in the second step to estimate loading conditions in another set of specimens. Results indicate the superior performance of the CNN-LSTM model. The proposed algorithm highlights the effectiveness of ultrasound features in accurately evaluating structural loading conditions for efficient monitoring of various structures. •A novel technique for evaluating loading condition of bonded structures is proposed.•The method is demonstrated on laboratory HPAASC and OPCC specimens.•Specimens were simultaneously tested via bi-surface shear and ultrasound testing.•VMD is employed for feature extraction from Ultrasound signals.•A trained CNN-LSTM model is fine-tuned to generalize predictions on new specimens.