A novel hierarchical temporal refinement diffusion network for data augmentation of industrial soft sensors.

The inherent difficulty in acquiring high-frequency process data and sparse information representation in complex industrial environments often leads to small-data problems, which degrade the generalization ability and prediction reliability of data-driven soft sensors. To address this challenge, a Hierarchical Temporal Refinement Diffusion Network is proposed for high-fidelity soft sensing data augmentation. By systematically decomposing industrial time-series patterns into multi-scale temporal dependencies, this method realizes coarse-to-fine generation of virtual samples through progressive diffusion refinement. The hierarchical structure effectively preserves both the macro-scale process trends and micro-scale fluctuation characteristics of the generated data. This architecture also integrates a dedicated noise prediction network, enabling simultaneous global correlation modeling and local feature extraction. Furthermore, a dynamic weighting strategy is developed for the joint training of hybrid datasets, which adaptively coordinates the learning of real-sample characteristics and virtual-sample regularities. The effectiveness of the proposed method is verified through two industrial application cases.