A new local-based data regression technique named locality preserving regression (LPR) is developed and applied for soft sensor modelling in the present study. By taking the local variation obtained by locality preserving projections into consideration, the regression algorithm LPR is employed to construct a soft sensor model and applied to industrial case. Furthermore, to deal with the time-varying behaviour of the process variables, just-in-time learning is also integrated to regularly update the soft sensor. Two case studies on a fermentation process for penicillin concentration prediction and the Tennessee Eastman process for output component prediction are provided to demonstrate the performance of the proposed method. Finally, the effectiveness and robustness of the proposed local-based technique for soft sensor modelling are assessed and compared with the global-based soft sensors based on the mean square error and the coefficient of determination. Experimental results showed that the novel soft sensor model could estimate the output with higher accuracy and generalization ability than the general soft sensor based on the global information.