Distributed Inference Condition Monitoring System for Rural Infrastructure in the Developing World

Remote condition monitoring systems for rural infrastructure lack "intelligent" analysis and advanced insights offered by recent Internet of Things devices. This is because the extreme and inaccessible operating locations necessitate the conservative use of limited resources, such as battery life and data transmission. Present implementations are often limited to usage data loggers, which are informative of general usage but post-processed advanced insights lag real-time system changes. A lightweight novelty filter is implemented onboard rural handpumps to identify subsets of data as potential infrastructure failure. The "intelligent" summaries of these data subsets are sent to a cloud-based system, where more advanced machine learning approaches are applied to increase the fidelity of potential failure predictions. The proposed method was tested on three independent data sets and found that the on-pump novelty filter could predict failure with up to 61.6% in situ. Incorporating more advanced machine learning methods on the cloud-based platform increased the classifiers' positive predictive value by at least an additional 10%-73%. This novel method has proven that it is possible for rural operating, resource-constrained devices to use lightweight, onboard machine learning approaches to perform anomaly detection in the embedded system. Distributed inference between the embedded system at the rural node and powerful cloud-based machine learning algorithms offers robust information without the need for expensive hardware or sensors embedded in situ-making the possibility of a large-scale (and perhaps even continent-wide) monitoring system feasible.