Abstract: Recently, just-in-time learning (JIT) based soft sensors have been widely used to provide online estimates of difficult-to-measure variables in process industry. However, conventional JIT soft sensors rely only on a single learning configuration while ignoring the diversity of JIT learning, leading to poor prediction performance. To address this issue, we propose a multimodal perturbation-based ensemble JIT learning soft sensing algorithm, referred to as diverse subspaces, and similarity measures based on ensemble locally weighted partial least squares (DSS-ELWPLS). This method achieves diversity creation of JIT learning by exploiting input feature and similarity perturbations. Then, we generate a set of base JIT learners that are accurate but diverse through evolutional multi-objective optimization, with locally weighted partial least squares as the base learner. Finally, we combine all base JIT learners using the stacking strategy. The effectiveness and superiority of the proposed DSS-ELWPLS soft sensor method are demonstrated through its application to the penicillin fermentation and industrial rubber mixing processes.