Bearing defects detection based on standardized sample split

Surface defects seriously affect the quality and service life of bearings. In recent years, deep learning has played an important role in defects detection, but for bearing detection, the collection of defects samples is time-consuming and labor-consuming. The bearing inner diameter was chosen as the detection object, a method of standardized sample split based on the symmetry of bearing was proposed, which could greatly increase the number of samples. Different sample processing methods were used respectively, and then the bearing defects detection model was trained by ResNet network to carry out several comparative experiments. The experimental results show that the detection effect is worse when the original images are directly used for training, and the area under the curve (AUC) of the model is only 0.558 0; after the samples are split, the trained model detection effect is better, and the model AUC is improved to 0.632 6; after the samples are corrected by four point perspective transformation, the detection effect is better, and the model AUC is increased to 0.661 3; after the original images are corrected by perspective transformation and the standardized samples are split, the detection effect is the best, and the model AUC is increased to 0.849 6.