Soft sensor for multi‐operating‐condition processes based on a multi‐branch quality‐relevant adaptive fusion network

Abstract To address the challenges of heterogeneous feature fusion, quality relevance modelling, and adaptive feature interaction in multi‐operating‐condition soft sensing, this study proposes a multi‐branch quality‐relevant adaptive fusion network (MB‐QAFN). The proposed framework employs a multi‐branch feature extraction module (MB‐FEM) with multiple isomorphic but parameter‐independent branches to learn diverse latent representations from process data. A residual aggregation quality‐relevant module (RA‐QRM) is further introduced to evaluate the contribution of branch features to quality prediction by transforming branch‐level prediction residuals into quality relevance information and adaptive factors. Based on these factors, a multi‐feature interaction adaptive fusion module (MFI‐AFM) utilizes cross‐attention to dynamically regulate interactions between target and auxiliary branch features, enhancing quality‐relevant information while reducing the influence of less informative features. The proposed method integrates feature representation learning, quality relevance evaluation, and adaptive fusion into a unified framework. Experimental results on industrial sulphur recovery and thermal power generation processes demonstrate that MB‐QAFN achieves improved quality prediction performance under complex multi‐operating‐condition scenarios, validating its effectiveness for industrial soft sensor modelling.