Energy-efficient adaptive sampling for IIoT via deep perception and reinforcement learning

To mitigate resource waste and information loss in fixed-frequency sampling for the Industrial Internet of Things (IIoT), this paper proposes a collaborative framework integrating dynamic perception with adaptive sampling. By incorporating the multi-dimensional state perception results of the dynamic perception model into the reinforcement learning state space, the proposed framework aims to facilitate a transition from a signal-driven to a cognition-driven sampling approach. Specifically, we develop the Learnable Decomposition Dual-branch Network (LD-DualNet), which decouples time-series signals via learnable decomposition into trend components reflecting long-term global variation laws and seasonal components capturing short-term local periodic fluctuations, and employs a dual-branch architecture combining Transformer with CNN-BiLSTM for feature extraction. Furthermore, a Deep Q-Network (DQN) optimizes sampling strategies via a semantic-signal dual feedback mechanism integrating semantic-level state criticality and signal-level data fluctuation feedback, optimizing high-level semantic planning while satisfying low-level signal constraints. Experimental results on three public datasets show LD-DualNet achieves 97.59% classification accuracy, with adaptive sampling reducing energy consumption by 36.6%–65.2%. The coefficient of determination (R²) exceeds 0.90, and the ratio of sampling density in critical and non-critical states is over 2.0, which reflects a tendency toward targeted sampling resource allocation. Practical validation on a rubber production line indicates a 45% data reduction does not notably compromise the performance of downstream forecasting and anomaly detection tasks, which provides empirical support for the industrial applicability of the proposed approach.