Imbalanced Data Classification with Fuzzy Logic and Universal Image Fusion for Gearbox Defect Detection

Finding data with different appearance defects and imaging them on a large scale is a laborious and costly affair. In this paper, an image-based model for detecting and classifying defects on various surfaces such as metal, wood, carbon fiber, concrete and fabric structures is provided. The model is designed to work with limited and class-imbalanced data. In this model, the Universal Image Fusion (UIF) block is embedded. This block gives a comprehensive view of the distribution of defects, their dimensions, and their location on the surfaces. To make fake defective images, the defects are cropped from the defective images and fused according to the distribution map, with gradient masks on the defect-free surfaces. Next, extracting texture features from images was improved with the help of Fuzzy Inference Systems (FIS) with Gaussian membership function and Sobel operator. Images were classified into two classes, defective and non-defective, with the participation of three networks, VGG-16, InceptionV3, and Resnet-50. The presented model was implemented on a dataset of gearbox components with imbalance data and was able to achieve 97.87% accuracy, 98.59% precision, 98.55% specificity, 97.90% F1 score, 97.22% sensitivity, and 0.9577 informedness (Youden's J statistic). The demo is available via https://github.com/DeepCar/Gearbox-Defect-Detection.