安全与环境学报

针对城市轨道交通工程施工人员不安全行为监管中多源信息割裂、评价主观静态及管控闭环缺失等问题，本文提出一种融合多模态感知与动态量化评估的智能管控方法。构建涵盖视觉、生理与环境要素的协同数据采集架构，采用动态时间规整与Daubechies小波阈值去噪实现异构数据时空对齐与噪声抑制。设计融合通道与时序注意力机制的卷积神经网络-长短期记忆网络(Convolutional Neural Network - Long Short-Term Memory，CNN-LSTM)混合识别模型，通过门控策略实现多模态特征自适应加权融合。提出基于风险累积理论的动态量化评价模型，引入时间衰减因子与敏感系数，将不安全行为与生理异常映射为连续可追踪的安全分值，研发闭环管控平台并在实际工程中开展实证。结果表明，该模型对典型不安全行为的识别准确率达94.6%，精确率93.8%，召回率92.5%，F₁分数93.1%，系统累计识别93起不安全行为并处理7起高等级预警事件，项目月度安全违规率由基线期的4.39%降至3.79%，降幅达13.7%。该方法有效突破了单一模态在复杂工况下的识别瓶颈，实现了施工人员安全风险的动态感知与主动干预，为施工安全管理的数字化与精细化转型提供了可复用的技术范式。

To address the issues of fragmented multi-source information， subjective and static evaluation， and lack of closed-loop control in the supervision of unsafe behaviors of construction workers in urban rail transit projects， this study proposes an intelligent control method integrating multimodal perception and dynamic quantitative assessment. A collaborative data collection architecture covering visual， physiological， and environmental elements was constructed to capture multi-source heterogeneous data. Dynamic time warping and Daubechies wavelet threshold denoising were applied to achieve spatiotemporal alignment and noise suppression of the collected data. A CNN-LSTM hybrid recognition model integrating channel and temporal attention mechanisms was designed with a dual-stream parallel encoding and fusion decoding architecture. Specifically， an improved ResNet-50 with Squeeze-and-Excitation modules was utilized to extract visual spatial features， while a three-layer 1D-CNN was employed to encode sensor sequences. Dual-branch LSTM networks were then constructed to model the temporal dependencies of both visual and sensor streams. A gating fusion strategy was introduced to dynamically adjust the fusion weights of multimodal features via a Sigmoid function. Furthermore， a weighted cross-entropy loss function with a high-risk penalty coefficient of 3.0 was adopted to effectively handle the severe class imbalance problem. A dynamic quantitative evaluation model based on risk accumulation theory was proposed， incorporating a time decay factor (λ=0.95) for safety behavior scoring and a sensitivity coefficient (0.05) for physiological safety scoring. This evaluation model maps unsafe behaviors and physiological abnormalities into continuous and traceable safety scores divided into three risk levels. Finally， a closed-loop control platform was developed and empirically validated in a real-world urban rail transit project. Results show that the proposed model achieved an accuracy of 94.6%， precision of 93.8%， recall of 92.5%， and F₁ score of 93.1% for typical unsafe behaviors. The system cumulatively identified 93 unsafe behaviors and successfully handled 7 high-level warning events. The monthly safety violation rate of the project decreased from 4.39% in the baseline period to 3.79% in the pilot period， representing a significant reduction of 13.7%. The proposed method effectively overcomes the recognition bottleneck of single modality in complex working conditions， realizing dynamic perception and proactive intervention of construction workers' safety risks. It provides a reusable technical paradigm for the digital and refined transformation of construction safety management.