This study proposes a subject-independent stress monitoring model based on temporal feature disentanglement, aimed at improving the accuracy and generalizability of stress detection. Existing stress prediction models often perform poorly with new participants due to differences in individual physiological and psychological responses. Addressing this issue, we develop a method that effectively extracts and removes individual-specific features while retaining stress-sensitive features. Using physiological signals and public datasets, we demonstrate that temporal disentanglement enhances model performance and generalization. Experimental results show significant accuracy improvements for detecting stress in new participants. Additionally, we introduce the use of ID labels for segment disentanglement, effectively reducing individual feature interference and increasing detection stability. This research lays the groundwork for advanced stress monitoring systems, promoting employee health and sustainable corporate development through real-time, accurate detection.