This paper presents a hierarchical autoencoder framework for multi-stream sensor data compression in resource constrained wireless sensor networks (WSNs). The proposed framework aims to manage computational consumption and bandwidth usage, while ensuring acceptable task performance despite channel errors. The proposed compression process is divided into two phases: intra-stream and inter-stream compression. In the first phase, a neural encoder compresses each individual data stream separately to eliminate intra-stream temporal redundancies, producing their compact individual representations. In the second phase, those representations of individual streams are combined and further compressed by another neural encoder to eliminate redundancies across streams and produce a single compressed latent representation for all the streams, which is sent to the receiver through a noisy channel. The neural decoder at the receiver uses a similar hierarchy for reconstructing individual original streams, and any subsequent task execution on those streams. To address the vulnerability of compressed data to channel errors, the framework integrates an error-resilient and learning-based transmission coding scheme. The proposed approach is demonstrated on an environmental monitoring task in a greenhouse setting and further validated on a multimodal physiological sensing scenario. Extensive simulations confirm the framework's effectiveness in balancing transmission cost, energy consumption, and task accuracy.
Gao et al. (Thu,) studied this question.