基于深度学习的电子商务物流仓储火情监控
E-Commerce Logistics Warehouse Fire Monitoring Based on Deep Learning

随着电子商务物流规模的持续扩大，仓储环境的安全防控面临严峻挑战，传统火灾检测方法在复杂场景下存在误报率高、实时性不足等缺陷。本文提出一种基于改进YOLOv10的电商物流仓储火情实时监测算法(Fire-YOLOv10)，通过神经能量驱动注意力机制与多模态特征融合实现火焰与烟雾的高精度检测。首先，针对仓储环境中火焰与烟雾的小目标、动态扩散特性，设计多尺度特征增强网络，融合浅层细节信息与深层语义特征以提升检测灵敏度；其次，引入无参数SIMAM注意力模块，基于神经能量理论动态抑制货架纹理等背景干扰，增强火情区域的特征显著性；同时，结合数据增强策略与迁移学习优化模型泛化能力，适配仓储货架遮挡、光照变化等复杂场景。为满足实时性需求，利用SIMAM的零参特性与深度可分离卷积协同压缩计算开销，实现边缘设备(如Jetson Nano)的低延迟部署。实验表明，基于自建电商仓储火情数据集(EC-Fire Dataset，涵盖10类场景、超10万帧标注图像)，Fire-YOLOv10在检测精度(mAP@0.5达95.6%)与推理速度(1080p视频流下68 FPS)上均显著优于YOLOv7、Faster R-CNN等基线模型。消融实验进一步验证了SIMAM模块在复杂背景抑制方面的有效性(误检率降低19.7%)。实际部署中，该系统可联动消防喷淋装置与声光报警模块，实现火灾的早期预警与快速响应，为电商物流安全提供可靠保障。
With the continuous expansion of e-commerce logistics scale, the security prevention and control of warehousing environment is facing severe challenges. Traditional fire detection methods have defects such as high false alarm rate and insufficient real-time performance in complex scenarios. This article proposes a real-time monitoring algorithm for e-commerce logistics warehouse fires based on improved YOLOv10 (Fire-YOLOv10), which achieves high-precision detection of flames and smoke through neural energy driven attention mechanism and multimodal feature fusion. Firstly, a multi-scale feature enhancement network is designed to address the small targets and dynamic diffusion characteristics of flames and smoke in storage environments, integrating shallow detail information with deep semantic features to enhance detection sensitivity; Secondly, a parameter free SIMAM attention module is introduced to dynamically suppress background interference such as shelf texture based on neural energy theory, enhancing the feature saliency of fire areas. At the same time, by combining data augmentation strategies with transfer learning to optimize the model’s generalization ability, it can adapt to complex scenarios such as storage shelf occlusion and lighting changes. To meet real-time requirements, the zero parameter feature of SIMAM and depthwise separable convolution are utilized to achieve low latency deployment of edge devices (such as Jetson Nano) through collaborative compression of computational overhead. Experiments have shown that based on the self built e-commerce warehouse fire dataset (EC Fire Dataset, covering 10 scenarios and over 100,000 annotated images), Fire-YOLOv10 achieves high detection accuracy (mAP@0.5 up to 95.6%). It significantly outperforms baseline models such as YOLOv7 and Faster R-CNN