Abstract: A dymanic data modeling method called modified dynamic slow feature analysis (MDSFA) is proposed for fault detection about insufficient consideration of high-order temporal correlation characteristics of dynamic industrial process data in two-steps dynamic slow feature analysis (TS-DSFA). At first, on the basis of TS-DSFA algorithm, the derivative estimation formula of slow features with respect to time is redescribed to design the optimal objective function which satisfies the extracted features with the slowest change and the highest autocorrelation. Then, with the constraints that dynamic latent variables can capture some change information and be orthogonal to each other, a set of potential features that change slowly with time and have an explicit dynamic autoregressive representation are extracted. Finally, a fault detection model for linear dynamic processes is constructed based on the proposed algorithm, and the statistics and corresponding control limits are calculated to realize the steady-state fault detection and dynamic fault detection of real-time dynamic processes. It has been proven that the proposed algorithm outperforms the exciting fault detection algorithms such as dynamic internal principal component analysis and TS-DSFA through simulation verification of numerical system and Tennessee-Eastman (TE) process.