Recommender system for Weld break prediction

원문 초록

In this paper, we propose an approach to Weld break prediction challenge. In a production line, a steel strip comes to the cold rolling machine in the form of a coil. When a coil is about to be finished, its end is welded with the head of the next coil. This results in a weld in the steel strip. The weld is smoothed in what is known as a scarfing process. The resulting steel strip is pickled on the pickling line and then thrown onto the tandem mill. At the tandem mill, the strip is pressurized to reduce the thickness. In this particular scenario, we address a weld break prediction for a company that welds the end of steel coils together in their production process. Often their production line succeeds in doing this. However, there are cases after the welding, in a subsequent step of the production line, the weld breaks. This is unfortunate because the company has to stop the production line, get the steel coil out of a machine and fix the broken steel. Visualizing one csv file which corresponds to one coil that might or might not have broken and the additional description files; one can understand that each file contains static parameters and a temporal information of dynamic variables measured during the welding process. We can also understand each row is a one-time step and each time step contains measurements from different sensors. The names corresponding to the values can be found at the top of each CSV file. The “OK” folder containing measurements of healthy coils has 2970 items where the “WB” folder containing measurements of broken coils has 200 items, providing a total of 3170 samples. The provided dataset is highly skewed (exhibits a large imbalance) in the distribution of the target classes. Project implementation can be found here https://github.com/Bekyilma/Weld-break-prediction