Data-Driven Monitoring for Distributed Sensor Networks: An End-to-End Strategy Based on Collaborative Learning

Abstract

This article proposes an end-to-end event monitoring method for a distributed sensor network with non-concentrated label noises based on a collaborative soft-label network (CSLN). The proposed CSLN framework is composed of a fundamental learner (FL) and a collaborative label modification. FL is designed based on a bidirectional gated recurrent unit (BiGRU) with attention to tackle the indistinct boundary problem by modeling multiscale dependencies. BiGRU can learn latent representations through capturing local dependencies in a bidirectional time flow. The attention mechanism is capable of modeling long-range dependencies through assigning learnable weights to the latent representations. Then, the collaborative label modification is established to reduce label noises by combining a truncated loss function and a dual-space smoothing technique. The truncated loss function can prevent FL from overfitting to noisy labels by isolating them in optimization. The dual-space smoothing technique can generate soft labels based on the local similarities. Furthermore, the proposed CSLN method is optimized by a bi-loop recursive strategy to reduce label noises gradually through alternatively training FL and generating soft labels. The feasibility and effectiveness of the proposed method are validated through real-field experiments of perimeter security applications based on distributed optical fiber sensors (DOFSs).