Discriminative Deep Belief Networks with Ant Colony Optimization for Health Status Assessment of Machine

On-line health status monitoring, a key part of prognostics and health management, provides various benefits, such as preventing unexpected failure and improving safety and reliability. In this paper, a data-driven approach for health status assessment is presented. A novel method based on discriminative deep belief networks (DDBN) and ant colony optimization (ACO) is used to predict health status of machine. DDBN is a new paradigm that utilizes a deep architecture to combine the advantages of deep belief networks and discriminative ability of back-propagation strategy. DDBN works through a greedy layer-by-layer training with multiple stacked restricted Boltzmann machines, which preserves information well when embedding features from high-dimensional space to low-dimensional space. However, selecting the parameters of DDBN is quite challenging. To address the problem, ACO is introduced to DDBN in this paper. By optimization, the structure of DDBN model is determined automatically without prior knowledge and the performance is enhanced. To evaluate the proposed approach, two case studies were carried out, which shows that it can achieve a good result. The performance of this model is also compared with support vector machine. It is concluded that the proposed method is very promising in the field of prognostics.