An Ensemble of Convolutional Neural Networks for Unbalanced Datasets: A case Study with Wagon Component Inspection

Railway component inspection is a technique widely used for maintenance because defective components pose safety issues. Nevertheless, finding defective components is a hard task because they are normally hidden by dusty, which poses hard problems for the image segmentation algorithms. To approach this problem, manual inspection by humans is normally used, but it is time consuming, expensive and sometimes dangerous. Meanwhile, automatic approaches that uses machine learning algorithms are also difficult because the datasets are strongly unbalanced. Such datasets usually induce biased classification models that identify new instances as members of the class with the greatest abundance of examples in the training data. In this paper, we propose a new method that combines the use of Convolutional Neural Networks (CNN) with imbalanced learning to address the challenge of using machine learning to identify defective components. Our method was tested with realworld data from images used for wagon component inspection. Moreover, we compared our method with an ensemble of MLP networks using features extraction, such as the LeNet, and a CNN network without ensemble learning. Results indicate that our proposed method produced the higher overall accuracy compared to the other methods.