This paper proposes a novel intelligent fault detection method, i.e., combining the subspace identification technique (SIT) with the data-model interactive compensation mechanism for fault detection. More precisely, the design process is divided into two primary stages: offline learning process and real-time fault monitoring. In the offline learning phase, the fault isolation algorithm is employed, along with parameter identification using the subspace identification technique. Then the online data undergo normalization processing and are subjected to online fault detection using a trained offline model. In addition, considering the accuracy of fault detection, this paper designs the dynamic optimal threshold based on the fault detection rate (FDR) and the false alarm rate (FAR). In the end, the case studies on PT700 and Tennessee Eastman (TE) process data are provided to confirm that the proposed method achieves the fault detection objectives.