An intelligent fault diagnosis framework for raw vibration signals: adaptive overlapping convolutional neural network

Abstract

Intelligent fault diagnosis methods are promising in dealing with mechanical big data owing to their efficiency in extracting representative features. However, there is always an undesirable shift variant property embedded in raw vibration signals, which hinders the direct use of raw signals in fault diagnosis networks. A convolutional neural network (CNN) is a widely used and efficient method to extract features in various fields for its excellent sparse connectivity, equivalent representation and weight sharing properties. However, raw CNN is time-consuming and has a marginal problem. Heuristically, we construct a fault diagnosis framework called adaptive overlapping CNN (AOCNN) to deal with one dimension (1D) raw vibration signals directly. The shift variant problem is dealt with by the adaptive convolutional layer and the root-mean-square (RMS) pooling layer, and the marginal problem embedded in the CNN is relieved by employing the overlapping layer. Meanwhile, the AOCNN is also characterized by adopting different convolutional strides and diverse activation functions in feature extraction network training and usage. Furthermore, sparse filtering is embedded into the AOCNN, and experiments on a bearing dataset and a gearbox dataset are conducted to verify the validity of the proposed method separately. When compared with other state-of-the-art methods this method reveals its superiority.