Aiming at the problem that changes in working conditions in the process industry can easily lead to mismatches between the current sample and the historical sample distribution, and the inaccuracy of traditional soft-sensing model, considering the impact of industrial data's time series, dynamics and process drift on modeling, this paper proposes a partial least squares regression soft sensor method based on transfer subspace learning. First, the regression framework adopts the nonlinear iterative partial least squares method and applies a domain adaptation regular term based on subspace reconstruction to the objective function of solving the mapping vector. During the shooting process, it is guaranteed that each sample in the current working condition can be linearly reconstructed by the historical working condition sample. On this basis, a low-rank and sparse representation is applied to the reconstruction matrix, while maintaining the data structure, the reconstruction matrix has a block structure to deal with process drift characteristics. The proposed method is tested on a numerical case and three different multi-condition data sets and compared with the existing domain adaptive regression methods. Experiments have demonstrated that this method can effectively improve the prediction accuracy of the model under cross-conditions, and reduce the impact of the difference in data distribution between conditions on the model performance.