Towards a More Reliable Identification of Non-Conformities in Railway Cars: Experiments with Mask R-CNN, U-NET, and Ensembles on Unbalanced and Balanced Datasets.

In every business, equipment requires repair services. Over time, equipment wears out; however, with well-conducted and guided maintenance, this degradation can be controlled, and failed equipment can be restored to operational status. Preventive maintenance allows this concept to be applied, given the great advantages for large companies in reusing equipment and machinery, always putting the worker's health and safety first. Rail transport has several pieces of equipment that can be reused if they are in a regular and well-defined maintenance cycle. In this sense, this article sought to create a method using real data for identifying cracks in wagons. Through the use of computer vision algorithms to prepare the data, along with several machine learning classification algorithms to locate cracks in train cars, the classification used properly annotated images and obtained great results, with a best case 98.10% hit-rate when wagons had a crack problem.