| 摘 要: 针对高校心理健康预警中部署资源受限、风险识别精度不足问题,设计基于多源异构数据融合的轻量化校园心理风险预警系统。系统融合消费、门禁、网络、学业、家庭和心理六类数据,通过分层递进融合构建可解释时序特征；提出融合注意力机制的轻量化BiLSTM模型,结合通道压缩与瓶颈分类器,将参数量压缩至8.97万、体积控制在0.34MB。系统已集成智慧平安校园平台,形成数据采集至分级干预的全链路闭环。实验表明,模型宏平均F1值达97.14%,单样本推理延迟11.36ms,可满足资源受限场景实时应用需求。 |
| 关键词: 心理风险预警 轻量化部署 多源异构数据融合 BiLSTM 智慧平安校园 |
|
中图分类号:
文献标识码:
|
| 基金项目: 2024年度江西省教育厅科学技术研究重点项目(GJJ2408611)；2021 年度江西省高校人文社会科学研究项目（JY21108）；江西泰豪动漫职业学院重点项目(THDMZD-25-1） |
|
| Lightweight-Deployable Campus Psychological Risk Early Warning System |
|
zhaobingbing1, Luo Hui2, Zhang Yuanlai3, Ji Xiaojie2, Chen Zhilin1
|
1.Nanchang Technology Vocational University;2.East China Jiaotong University;3.East China Normal University
|
| Abstract: To address the problems of limited deployment resources and insufficient risk identification accuracy in college mental health early warning, a lightweight campus psychological risk early-warning system is designed based on multi-source heterogeneous data fusion. The system fuses six types of data including consumption, access control, network, academic, family and psychology data, and constructs interpretable temporal features through hierarchical progressive fusion. A lightweight BiLSTM model combined with attention mechanism is proposed, which compresses parameters to 89,700 and controls model size at 0.34 MB using channel compression and bottleneck classifier. The system is integrated into the smart safe campus platform, forming a full-chain closed loop from data collection to graded intervention. Experiments show that the model achieves a macro-average F1 score of 97.14% and a single-sample inference delay of 11.36 ms, which can meet the real-time application requirements in resource-limited scenarios. |
| Keywords: psychological risk warning lightweight deployment multi-source heterogeneous data fusion BiLSTM smart safe campus |
