MPSS: A spatiotemporal-decoupled massively pretrained soft sensor for general industrial scenarios with heterogeneous data robustness

Abstract

Data-driven soft sensors are widely used for estimating key quality variables in industrial processes. However, most existing models are task-specific and lack generalization, limiting their applicability in complex multi-task scenarios. Moreover, constraints on model and input capacity often lead to insufficient representational ability and degraded performance under sample-scarce settings. To address these, in this paper, a three-stage massively pretrained soft sensor (MPSS) is proposed for general industrial applications. Specifically, a spatial–temporal decoupled self-supervised learning framework and two distinct masked reconstruction strategies for representation learning are introduced for pretraining, aiming to acquire universal temporal dependency and spatial-variable coupling representations. To enhance model capacity and adaptivity to heterogeneous temporal-variable patterns, two sparsely structured routing modules—dual-branch temporal-aware routing (DTAR) and adaptive channel-aware routing (ACAR) are proposed, achieving adaptive allocation and specialized processing of heterogeneous inputs. Additionally, a prefix-enhanced time series embedding strategy is proposed, which encodes key statistical information as learnable conditional prefixes, increasing input information density and strengthening generalization. For downstream tasks, MPSS freezes pretrained parameters and integrates lightweight, task-specific adapters via parameter-efficient fine-tuning (PEFT), enabling plug-and-play adaptation across diverse soft sensing tasks. Experiments on four datasets demonstrate MPSS’s strong generality, transferability, and state-of-the-art (SOTA) performance under both full-data and few-shot settings.