Physical intrusion monitoring via local-global network and deep isolation forest based on heterogeneous signals.

• An extensible local-global network is proposed for sensor heterogeneity in physical intrusion monitoring based on DOFS for the first time. • A deep isolation forest (DIF) is proposed to improve performance of local classifiers in extremely imbalanced, large-scale and high-dimensional cases. • The semi-sharing strategy and parallel computing framework of proposed network are introduced to reduce the computation costs in training and monitoring. • The characteristics of the proposed method are detailedly discussed in theory, and the effectiveness is verified through real-field experiments. This paper proposes a local–global semi-sharing network (LGSSN) for heterogeneous signals in physical intrusion monitoring. The signals are collected by heterogeneous distributed optical fiber sensor (DOFS). The local classifier of LGSSN is constructed via a hybrid model deep isolation forest (DIF). It can extract the dominant representations of original high-dimensional signals through deep autoencoders (DAE). Then, an isolation forest (IF) is added to the last layer to obtain local classification for extremely imbalanced cases. In addition, the network is simplified by semi-sharing strategy, and a parallel computing framework is presented for accelerating the process. Further, the final decision on monitoring state is acquired by a Bayesian inference-based global integrated monitor (GIM) with enhanced classification accuracy. The proposed strategy is tested on a monitoring application along the Nanjing Metro Line S7, Jiangsu Province, China. Comparative experimental results illustrate the feasibility and effectiveness of proposed strategy. [ABSTRACT FROM AUTHOR]