Online Monitoring for Non-Stationary Operation via a Collaborative Neural Network

Abstract

Online monitoring has become a challenging issue because of the changing knowledge in non-stationary operation. It requires an appropriate monitoring algorithm to timely self-adjust based on a finite number of online data. This study proposes a collaborative neural network (CoNN) for detecting physical intrusions in perimeter security monitoring through real-time adaptation. The proposed CoNN method consists of two half-shared neural networks for online learning with finite measurements through preserving previous knowledge and absorbing the new knowledge. Specifically, the shared feature layers can learn latent domain-invariant representations through weight sharing and adapt CoNN to prior knowledge changes. On this basis, the classification layers use concept-exclusive parameters to model the changes in posterior knowledge. Furthermore, the objective of CoNN is formulated as a linearly constrained non-convex minimization problem with coupled functions, and an alternating direction method is used to solve it with theoretically guaranteed convergence. The feasibility and effectiveness of the proposed method are validated through a real-field experiment of a perimeter security system.