A novel CNN ensemble framework for bearing surface defects classification based on transfer learning

Bearing surface defect detection and classification methods based on machine vision have been widely used in bearing quality inspection due to their high-speed, high-precision, and non-contact advantages. However, traditional machine vision algorithms have low reusability and their development processes are often expensive and time consuming. Several deep learning-based bearing surface defect detection methods have been proposed. However, these deep learning models often require a large number of datasets, which is often difficult to achieve in the actual industry. Transfer learning provides a promising solution to the small sample difficulties associated with deep learning. However, the complexity of the illumination conditions and the huge differences between bearing dataset and ImageNet dataset make it impossible to use the current single model-based transfer learning for bearing defect detection. In this study, we propose a novel transitive transfer learning convolutional neural network (CNN) ensemble framework for classifying bearing surface defects. Only small-scale datasets are needed in this framework. A transfer path and transfer method selection strategy for transitive transfer learning is then proposed to train the deep learning models, which enhances the feature extraction ability of the CNN models on the basis of multiple illuminations. Ablation experiments are conducted to verify the effectiveness of the proposed method. Experimental results show that the proposed transitive transfer learning CNN ensemble framework has the accuracy rate of 97.51%. The average time for detecting each bearing is 155 ms, which can meet the requirements of industrial online detection.