Full-Length Hardness Prediction in Wire Rod Manufacturing Using Semantic Segmentation of Thermal Images

As an essential steel product, wire rods have specific requirements regarding their physical properties. Especially for wire rods for automotive springs, it is important to ensure consistent hardness throughout the product. Because traditional hardness testing methods are destructive and sample-based, they have the potential to overlook the non-uniformity of wire rod hardness. This paper presents the application of a convolutional neural network (CNN) to thermal imaging to address these issues. The model segments the thermal image of a wire rod after cooling, separating the temperature of the wire rod and the background on a pixel-by-pixel basis. This temperature data is used to calculate the cooling rate and helps to predict the hardness of the wire rod along its entire length. Experimental results show that the U-Net-based model outperforms a simple FCN model in the segmentation task. This approach provides a more comprehensive quality inspection of wire rod, bringing both economic and quality benefits to the steel industry.