Structural Monitoring with Distributed-Regional and Event-based NN-Decision Tree Learning Using Mobile Multi-Agent Systems and Common Java Script Platforms. (2016)

• Record Type: Journal Article
• Authors: Bosse, Stefan
• Is Part Of: Procedia technology. Volume 26(2016)
• Page Start: 499
• Page End: 516
• Publication Date: 2016
• Subjects: Structural Health Monitoring -- Ubiquitous Perception -- Sensor Network -- Mobile Agent -- Distributed Learning -- Heterogeneous Networks -- Embedded Systems; Technology -- Congresses; Technology -- Periodicals; Engineering -- Congresses; Engineering -- Periodicals; Engineering; Technology; Conference proceedings; Periodicals; 605
• DOI: 10.1016/j.protcy.2016.08.063
• Languages: English
• ISSNs: 2212-0173

Abstract

Abstract: Among the Internet-of-Things, one major field of application deploying agent-based sensor and information processing is Structural Load and Structural Health Monitoring (SLM/SHM) of mechanical structures. This work investigates a data processing approach for material-integrated and mobile ubiquitous SHM and SLM systems by using self-organizing mobile multi-agent systems (MAS), executed on a highly portable JavaScript-based Agent Processing Platform (APP), and optimized Machine Learning (ML) methods providing load class recognition from a set of sensors embedded in the technical structure. Machine learning approaches usually require a large amount of computational power and storage resources and ML is commonly performed off-line, not suitable for resource constrained sensor network implementations. Instead, a novel distributed-regional on-line learning is applied, with on-line distributed sensor processing and learning performed by the agent system. The APP provides ML as a service, and the agent itself only collects training and analysis data passed to the APP, finally returning a learned model that is saved by the agent in a compact format (and is available on any other location). A case study shows that the learning algorithm is suitable (stable) for noisy and time varying sensor data. Spatial global learning is reduced and mapped on local region learning with global voting.