Augmented Transfer CNN For Fault Diagnosis Of Roller Bearing Under Frequent Working Condition Variation

Intelligent fault diagnosis (IFD) has been widely investigated and achieved massive successes in past years. Particularly, Transfer learning (TL)-based IFD has been a hot topic recently, and also widely validated effective in dealing with working condition variation. However, owing to its requirement on access to target dataset in network training, its application perspective is limited in machine with frequent working condition variation. Hence, to tackle this limitation, Augmented Transfer CNN (ATCNN) is proposed to diagnose fault more conveniently, which accesses only raw vibration signals from source domain in training, but applicable on several target datasets with anonymous working condition directly, which has better generalization ability even than domain generalization. Firstly, several data augmentation methods are designed and applied to obtain ample and various data and form fake target datasets. Then, IDCNN is designed to extract sharing fault features from raw vibration signals in source and augmented fake target datasets. Apart from common classification tasks, a confusion objective is designed and inserted to prompt confusion in recognizing raw signals and augmented signals via transfer learning mechanism. Experimental validations on a public roller bearing dataset validate its function under frequent working condition variation.