With the advancement of machine learning, many predictions and measurements in visual tasks can be achieved by convolutional neural networks (CNNs). Solidification hot cracking is a significant defect in laser beam welding, commonly encountered in practical applications. Existing theories indicate that the formation of cracks is closely related to strain accumulation near the solidification front. In this paper, we first leverage supervised regression networks to design CNNs that achieve real-time average strain estimation for each frame in the collected welding videos. Two different architectures are proposed and compared: the first model stacks two frames at a set interval and feeds them into the network, while the second model extracts image features individually and predicts the results by calculating the correlation between them. Each network has its own advantages in terms of computational efficiency and accuracy. Finally, we further train a multilayer perceptron (MLP) classification model that can detect the occurrence of cracks based on the predicted strain behaviors.
