Dynamic Canonical Correlation Analysis (DCCA), as an extension variant of canonical correlation analysis, it employs the autoregressive exogenous(ARX) time series extension method to extend the CCA model to dynamic scenarios, facilitating dynamic feature extraction and prediction of key performance indicators. However, in DCCA, the process variable space variables may contain components that are either irrelevant or have a negligible contribution for predicting quality, while the residual space may contain components with significant fluctuations. To overcome these limitations, a Dynamical-Improved canonical correlation analysis (DICCA) model is proposed. Firstly, the principal components were seperated into two blocks, i.e., quality-related subspace and quality-unrelated subspace. Secondly, a PCA decomposition was performed to separate the quality-related components in the residual space. Thirdly, the statistics were designed for each of the four spaces and conducted the online testing, representively. Finally, the effectiveness of the DICCA algorithm is demonstrated through its application to the Tennessee-Eastman (TE) process experiment.