Estimation of Mill-Scale Thickness by Back-Propagation Neural Network for Pickling

In the production of hot-rolled steel strip, the formation of mill scale, a surface oxide layer, results. This byproduct, if not properly removed for subsequent processing, can lead to coating failure resulting in rapid corrosion. Removal, often by pickling in an acid bath, generates considerable toxic waste, and over pickling can also damage the steel surface. Knowledge of the mill-scale thickness is crucial for optimizing pickling, and consequently to ensure the longevity and integrity of steel products, while minimizing toxic-waste production. To obtain a nondestructive measurement of mill-scale thickness, we use the technique of terahertz time-of-flight tomography. In some cases due to the optically thin nature of the mill scale in the terahertz frequency range, we utilize a back-propagation neural network applied to the raw experimental data to rapidly and accurately estimate the mill-scale thickness. In this work, two neural-network approaches are implemented: one for regression and one for classification. Both networks take in the terahertz time-of-flight tomography data and output an estimation on the thickness of the mill scale, which ranges from \(\sim \) $\sim$ 5 to 15 \(\mu \) $\mu$ m. The regression network has the ability to estimate the thickness of mill scale with RMSE error around 1.6 \(\mu \) $\mu$ m. The classification network is able to classify the samples into three categories according to their thickness range with an accuracy of 87 pct on the test measurements.