Soft-sensor technique is often used to estimate key variables in industrial manufacturing, of which the commonly used approaches as the mechanism modeling and data-driven modeling both have their limitations. To take full advantage of the modeling methods and overcome the problems of nonlinearity, unmodeled dynamics and unlabeled data in industrial manufacturing, a hybrid modeling method combining the mechanism with the semi-supervised learning is developed in this paper. In the framework of this hybrid model, the model can be divided into mechanism view and data view. In the mechanism view, the unmodeled dynamics in the mechanism model are obtained by an improved data-driven semi-supervised weighted probability partial least squares regression (SWPPLSR). In the data view, the present SPPLSR can solve the problem of the noise disturbance and output absent. On this basis, the locally weighted is adopted to handle the nonlinearity. Moreover, aiming at the imperfection of similarity measurement, varying working conditions and model redundancy, ensemble just-in-time learning and moving window techniques are combined to obtain an improved SWPPLSR. Finally, the real-world data in the roller hearth kiln of ternary cathode material manufacturing is applied for simulation to verify the validity of the model. The results have practical guiding significance. (c) 2020 Elsevier Ltd. All rights reserved.