A MTTFF-Oriented Optimization to Guarantee Reliable Inference of Distributed Deep Systems in Industrial IoT Systems

The distributed deep learning architecture between front-deployed sensors and edge-deployed gateways attracts increasing interest. However, the inference performance of distributed deep models is also impacted by the delivery loss of intermediate representation in the wireless link, especially in the harsh industrial fading environments. Traditional communication systems usually focus on transmission errors at bit level, which treat all bits in the packets equally and fail to suit the varying importance in distributed deep models, which urges the essential evolution of the communication method to form a joint co-design paradigm for distributed deep models. This article then proposes to optimize the Mean Time To First Failure (MTTFF) of wireless link instead of traditional bit error rate, which enables a guaranteed transmission window. This paper first derives the analytical model of MTTFF under MIMO systems, then utilizes the kernel mixture distribution to obtain a closed-form solution of MTTFF, which forms a optimization algorithm minimizing the transmitted power while achieving the aiming MTTFF. Extensive real-life experiments show more than 70% satisfaction rate of MTTFF, which leads to more than 10 times higher inference accuracy than the original deep model.