LAUSR

Abstract Wear particle classification is an important approach to analysing wear faults. However, it has always been a challenging task for automatic recognition of wear debris in complex ferrography images. To this end, a convolutional neural network (CNN) model called FECNN is proposed in this study to identify wear particles in complex ferrography images. Different from the architecture of ordinary CNN models, one-dimensional (1-D) convolutional operations and a dropout operation are incorporated into the FECNN model to make it more suitable for ferrography image processing. The 1-D convolutional operations reduce the number of model parameters, and the dropout operation makes the model for learning the incomplete features of wear debris. A large number of wear particles are obtained from mechanical parts made of steel to test the FECNN model. The results show that the FECNN model can automatically extract the features of wear particles from the ferrography images containing much background noise. Furthermore, the FECNN model obtains more accurate classification results than that of the previous method with a few training samples, which alleviates the workload of sample collection. The FECNN model’s powerful capability of feature extraction and wear particle classification indicates the FECNN model is a promising tool for wear particle recognition.