A spatial multi-scale and adaptive filtering attention framework for industrial soft sensor

Abstract

Soft sensors are vital for industrial monitoring but face severe challenges arising from complex spatiotemporal multi-scale dynamics and prevalent information redundancy. Existing deep learning methods often struggle to capture multi-scale dependencies and suffer from weak or redundant interactions introduced by fully connected interactions. To address these issues, this paper proposes SSTAFormer (spatial–scale and top-K filtering attention former), a novel framework based on hybrid attention mechanisms. First, a spatial multi-scale fusion attention module is designed to capture heterogeneous variable dependencies via parallel multi-receptive field convolutions. Second, a top-K filtering attention module is introduced to suppress lower-scoring interactions. By employing a content-dependent top-K filtering strategy, this module retains the most discriminative dependency paths while filtering weak interactions. Extensive experiments on the SRU and Debutanizer Column datasets demonstrate the competitive overall predictive performance of SSTAFormer compared with representative baseline methods.