Clinker f-CaO content is an important indicator of cement quality. Considering the production characteristics (strong coupling, time-varying delay) in the cement process industry, a soft sensor model was developed by combining various methods. First, a new decoupling method is proposed to deal with the strong coupling between variables, which achieves data decoupling between process variables through the attention mechanism and the LSTM network. Second, a novel time-series matching technique is proposed to handle the time-varying delays, which utilizes a window selection mechanism to adaptively select the time period in which each process variable influences the target variable. Third, the critical features of the variables are extracted by a one-dimensional convolution network(1D-CNN). Last, a combination of the data decoupling method, window selection mechanism, and 1D-CNN is applied to develop a soft sensor model (ADM-WGM-CNN), which implements the measurement of f-CaO content. The experimental results demonstrate that the ADM-WGM-CNN model has better measurement performance.