Joint Task Offloading, DNN Pruning, and Computing Resource Allocation for Fault Detection With Dynamic Constraints in Industrial IoT

In this paper, we investigate a joint task offloading, deep neural network (DNN) model pruning, and edge computing resource allocation (JOPA) problem for supporting a fault detection service on industrial washing machines in layered industrial Internet-of-Things (IIoT) systems. Specifically, we aim to maximize the overall network resource utilization while guaranteeing diverse and time-varying task processing delays and accuracy requirements for generated processing/computing tasks for the fault detection service. To capture the network dynamics, we formulate a stochastic optimization problem to maximize the long-term network resource utilization with per-time-slot constraints on the end-to-end (E2E) task latency and accuracy. Considering the network state transitions and the relations between network states and policies, we transform our problem to a Markov reward process (MRP) formulation where the state transitions are characterized independent of the actions taken. To deal with the large problem size and dynamic quality-of-service (QoS) constraints (e.g., E2E delay and accuracy constraints), we design a deep-reinforcement-learning (DRL) solution framework based on a refined soft actor-critic (SAC) algorithm, where the main SAC algorithmic components (i.e., actor networks, critic networks, and target networks) are customized to accommodate hybrid actions (mixed discrete and continuous actions), achieve a robust evaluation of state-action policies, and stabilize the training process. Extensive simulation results are provided to demonstrate the effectiveness of the proposed scheme and the advantages over benchmark approaches in terms of 1) achieving high network resource utilization, 2) balancing the trade-off between resource utilization and QoS satisfaction, and 3) adapting to the network load variation and dynamic QoS requirements.