A Membership-based Multi-dimension Hierarchical Deep Neural Network Approach for Fault Diagnosis

Accurate fault prognosis of machine component is important to maintain industry operation system. Faults analysis can be very helpful in fault early warning and reducing maintenance cost. The goal of our work is to design an integrated approach of machine faults analysis. A method widely used is Fuzzy Neural Networks (FNNs), but such method lacks of flexibility. We present a Membership-based Multi-dimension Hierarchical (MMH) neural network model to jointly include new feature selection approaches and generalized membership operators. MMH model is an adaptive model that employs modified KPCA and Back Propagation algorithm respectively. By introducing optimized KPCA we can extract features of higher importance that are appropriate for fault diagnosis. Our prediction model is inspired by the traditional fixed membership. In our approach, an observing value will be segmented into multiple dimensions where each dimension captures deep structural information in the network. The transformation is updated by back propagation. The proposed approach takes advantage of membership thinking and benefits from large learning capacity of deep neural networks (DNNs). This is aiming to take advantage of membership thinking and neural network deep learning abilities. Experimental results on public datasets demonstrate the superiority of our model that has the character of faster convergence, which also improving the accuracy by an average of 5% for fault prediction.