Prediction of stress distribution around crack tips from ultrasound images using convolutional neural networks

Structural health monitoring (SHM) is critical for the safety, reliability, and long-term performance of engineering structures. This study introduces an innovative machine learning-driven framework for predicting key fracture parameters, including crack length, stress fields around the crack tip, and crack mouth opening distance, directly from ultrasound images. The ultrasonic data were acquired using a permanently installed, low-cost, low-profile ultrasonic array, which provides a much lower signal-to-noise ratio than conventional systems and therefore presents additional challenges for data interpretation. Various types of machine learning models were built and trained using total focusing method images paired with corresponding stress and strain fields obtained via digital image correlation. This pairing provided explicit crack-related information, enabling predictions of critical fracture metrics. Substantial parameter tuning was undertaken to identify optimal model configurations and explore the relative advantages and trade-offs of various architectures. These comparisons provided valuable insights into how model structure and hyperparameters influence predictive performance. Experimental validation confirmed the reliability of this framework. By integrating low-cost ultrasonic imaging and machine learning, this work underscores the potential of leveraging data-driven techniques to transform SHM practices. The findings pave the way for scalable, cost-effective, and computationally efficient solutions to monitor structural integrity in real-time, offering significant benefits for infrastructure management and maintenance.