Reliable diagnostics using wireless sensor networks

Monitoring activities in industry may require the use of wireless sensor networks, for instance due to difficult access or hostile environment. But it is well known that this type of networks has various limitations like the amount of disposable energy. Indeed, once a sensor node exhausts its resources, it will be dropped from the network, stopping so to forward information about maybe relevant features towards the sink. This will result in broken links and data loss which impacts the diagnostic accuracy at the sink level. It is therefore important to keep the network's monitoring service as long as possible by preserving the energy held by the nodes. As packet transfer consumes the highest amount of energy comparing to other activities in the network, various topologies are usually implemented in wireless sensor networks to increase the network lifetime. In this paper, we emphasize that it is more difficult to perform a good diagnostic when data are gathered by a wireless sensor network instead of a wired one, due to broken links and data loss on the one hand, and deployed network topologies on the other hand. Three strategies are considered to reduce packet transfers: (1) sensor nodes send directly their data to the sink, (2) nodes are divided by clusters, and the cluster heads send the average of their clusters directly to the sink, and (3) averaged data are sent from cluster heads to cluster heads in a hop-by-hop mode, leading to an avalanche of averages. Their impact on the diagnostic accuracy is then evaluated. We show that the use of random forests is relevant for diagnostics when data are aggregated through the network and when sensors stop to transmit their values when their batteries are emptied. This relevance is discussed qualitatively and evaluated numerically by comparing the random forests performance to state-of-the-art PHM approaches, namely: basic bagging of decision trees, support vector machine, multinomial naive Bayes, AdaBoost, and Gradient Boosting. Finally, a way to couple the two best methods, namely the random forests and the gradient boosting, is proposed by finding the best hyperparameters of the former by using the latter. (C) 2018 Elsevier B.V. All rights reserved.