Self-stacking random weight neural network with multi-layer features fusion

Randomized methods are practical and efficient for training the connectionist models. In this paper, we contribute to develop a self-stacking random weight neural network. Two different methods of feature fusion are proposed in this paper. The first one inter-connects the coarse and high level features to make the classification decision more diverse by using the proposed hierarchical network architecture with dense connectivity. On the other hand, the different decisions all-throughout the network are incorporated by a novel non-linear ensemble learning in an end-to-end manner. Through experiments, we verified the effectiveness of random features fusion, and even if each hierarchical branch in the network has very unfavorable accuracy, the proposed ensemble learning presents the impressive performance to boost the classification results. Moreover, the proposed connectionist model is applied to address one practice engineering problem of gearbox fault diagnosis, and the simulation demonstrates that our method has better robust to the noise in vibration signal of working gearbox.