In public safety scenarios, extreme individual behaviors are often sudden, harmful, and highly uncertain. Timely identification of relevant abnormal states in a non-intrusive manner is therefore of great significance for risk warning and early intervention. To address the limitations of existing studies, including the lack of dedicated datasets, insufficient unimodal representations, difficulty in jointly modeling short- and long-term dynamics, and unstable fusion caused by missing modalities and signal quality fluctuations, this paper proposes an individual psychological abnormal state detection method based on multi-scale temporal representation and mask-adaptive fusion. First, a transfer learning scheme is designed by pretraining on a public mental health dataset and fine-tuning on a self-built dataset, so as to leverage transferable psychological state knowledge, improve model initialization, alleviate undertraining and overfitting under small-sample conditions, and enhance adaptation to the target task. In model design, multi-source information from video, audio, and rPPG is integrated, and a multi-scale temporal representation method is introduced to better capture the coexistence of short-term fluctuations and long-term evolution in psychological abnormal states, thereby improving collaborative modeling of cross-modal dynamic cues. Furthermore, a mask-adaptive fusion mechanism is proposed to address modality absence, signal quality variation, and unequal modality contributions in real-world scenarios, thus improving model stability and robustness under complex conditions. Experimental results show that the proposed method achieves an MAE of 5.14 and an RMSE of 6.37 on the AVEC2014 dataset, and an accuracy of 0.76 and an F1-score of 0.82 on the self-built dataset, demonstrating its effectiveness.