Low-power wireless mesh networks (WMNs) have been widely deployed to connect sensors, actuators, and controllers in industrial facilities. As industrial WMNs become increasingly heterogeneous and complex, recent research has reported that resorting to advanced machine learning techniques to configure WMNs presents significant performance improvements compared to traditional methods. However, it is costly to collect sufficient data to train good network configuration models in many industrial facilities. In such scenarios, the benefits of using learning-based methods that depend on a large amount of data are outweighed by the costs. Recently there have been growing interests in using simulations to configure WMNs because simulations can be set up in less time and introduce less overhead. Unfortunately, recent studies show that the network configuration selected from a simulated network may not be able to help its corresponding physical network achieve desirable performance due to the simulation-to-reality gap. In this paper, we formulate the network configuration prediction as a multi-source domain adaptation problem and introduce a novel solution. Experimental results show that our solution effectively closes the simulation-to-reality gap and provides 80.45% prediction accuracy when it uses cheaply generated simulation data and 440 data traces collected from the physical network for training. As a comparison, the deep neural network (DNN) model trained without using simulation data requires 3,080 costly physical data traces to achieve 80.39% prediction accuracy.
