Evaluating the thickness of thin-walled structures from ultrasound signals based on CNN-BILSTM model

Ultrasonic methods are widely utilised for the thickness evaluation of structures. However, when the structure’s thickness is smaller than the wavelength of the ultrasonic waves, the incident and reflected signals may overlap, complicating the extraction of accurate thickness features. In this study, convolutional neural networks (CNN) and bidirectional long short-term memory (BiLSTM) networks are combined to extract thickness features from overlapping ultrasonic signals. Ultrasonic immersion experiments were conducted on wedge-shaped specimens to acquire A-scan signals across various thicknesses, which served as training samples. To augment the training dataset for ultra-thin structures, a finite element model for ultrasonic bulk waves was developed. Additionally, aluminium plates with blind holes were tested for thickness evaluation, and the performance of the CNN-BiLSTM model is compared with traditional machine learning methods and other deep learning models. The results demonstrate that CNN-BiLSTM outperforms traditional machine learning methods in accuracy, with a thickness prediction error below 0.05 mm in very thin regions, and offers faster computation than other advanced deep learning models. Finally, the influence of the sampling duration of the time series fed into the CNN-BiLSTM model was analysed, revealing that at least half of the excitation signal’s duration is required for accurate thickness evaluation.