The rapid evolution of Industrial Cyber-Physical Systems (ICPS) with cloud-fog automation calls for the deployment of Deep Neural Networks (DNNs) on edge devices to enable intelligent and autonomous decision-making. However, the resource constraints, heterogeneity, and dynamic nature of edge devices pose limitations to the efficient deployment of DNNs. Federated Learning-based Neural Architecture Search (FL-NAS) has been proposed to address these limitations, but achieving an effective balance between the generalized global model and personalized local models remains a non-trivial task due tosuboptimal aggregation of homogeneous neural blocks, knowledge waste of heterogeneous neural blocks, and high communication and energy overhead. In this paper, we proposeF²NAS, an energy-friendly federated neural architecture search framework tailored for ICPS. The fine-grained aggregation strategy adapts weights for each device during aggregation, enhancing the global and personalized local models. The bidirectional knowledge transfer mechanism leverages heterogeneous neural blocks, promoting knowledge sharing among local and global models. The adaptive communication strategy optimizes interactions between edge devices and the cloud server based on model performance, reducing energy costs while maintaining effective model collaboration. Extensive experiments demonstrate thatF²NASoutperforms baselines by up to 30.31% in accuracy on edge devices, 38.75% on the cloud server, and achieves a 65.2% reduction in energy consumption. When applied to the surface defect detection task in ICPS,F²NASsurpasses other baselines by up to 13.38% and 302.46% for edge devices and cloud servers, respectively, and reduces energy consumption by 44.8%.

The rapid growth of edge computing calls for fine-tuned deep neural network (DNN) deployment that emphasizes energy-efficient implementation, due to the resource constraints of edge devices. Traditional Federated Learning-based Neural Architecture Search (FL-based NAS) has been instrumental in the complexities of this deployment, particularly in addressing constraints posed by device heterogeneity, limited resources, and privacy preservation. However, it is hindered by issues such as suboptimal aggregation of homogeneous neural blocks, significant knowledge waste in disregarding heterogeneous neural blocks, and excessive communication energy consumption. This paper introduces F2NAS in green edge computing, a novel energy-efficient approach that addresses these limitations by ensuring flexible and energy-efficient model design and training for edge devices. Firstly, F2NAS introduces an innovative aggregation strategy that enhances the integration of homogeneous neural blocks by using inter-block distances to optimize weight allocation. Further, it employs a unique parameter extraction technique that recaptures valuable insights from previously overlooked heterogeneous neural blocks. Finally, F2NAS meticulously calibrates communication energy consumption by balancing loss function and model interaction, setting and refining an upper limit for model communication. Experimental results reveal F2NAS enhances model accuracy by 2.8% to 4.7%, simultaneously reducing the energy consumption by nearly 50% through optimizing the communication cost.

Deep neural networks are deemed to be the cornerstone of a series of mobile intelligent systems, and their inference processes bring about a mass of computation-intensive tasks. To migrate the burden of inference computation from resource-constrained mobile devices, device-edge cooperative inference in mobile edge computing provides a fine-grained processing method. However, the geographical dispersion of resources and the mobility pattern of devices pose technical issues in the scheduling of co-inference systems, which have not been fully considered. In this paper, we propose a learning-based scheduling framework for such device-edge systems to improve the pipeline time of model inference. First, we consider a resource provisioning strategy based on the number of devices and a pre-fetching service migration setting in the environment of multiple mobile devices and edge nodes. Next, we propose an algorithm based on proximal policy optimization for each device to make the decision independently. Further, we adopt long short-term memory in the algorithm to capture the temporal characteristics of the system state. Experiments using a real-world network and computing trace demonstrate that the proposed algorithm can efficiently sense the mobile system to make decisions at various system scales and two mobility scenes. The average pipeline time of the proposed algorithm is only 67.63% of that of local processing, which is 97.50% of that of the omniscient algorithm.