【作者】 李伟；

【导师】 刘彤； 李邵喜；

【作者基本信息】 大连海事大学 ， 交通信息工程及控制， 2023， 硕士

【摘要】 船舶火灾探测一直是船舶安全领域的重中之重,现阶段船舶火灾探测器通过感应燃烧反应的特征参数触发报警,具有延迟性,非可视化,单一探测器兼顾范围小等不足。随着船舶大型化的发展,这些不足之处愈加明显。为了降低船舶火灾探测的延迟,扩大探测的范围,以及获取更多的火灾信息,成为船舶火灾防范亟待解决的技术难题。针对上述问题,本文通过深入研究传统的火灾探测和基于视频的火灾探测,提出基于改进YOLOv5(You Only Look Once version 5)的船舶火灾烟雾检测算法。为了训练出针对船舶火灾烟雾的探测模型,模拟制作了以船舶封闭空间为背景的数据集,验证了构建模拟数据的有效性和改进算法的优势。主要研究工作如下:(1)构建了船舶火灾数据集。数据集对于检测模型的性能具有决定作用,在船舶火灾探测领域,却没有相应的船舶封闭空间火灾数据集。因此,本文使用相关软件模拟制作了船舶封闭空间火焰烟雾数据集,用于训练神经网络并取得良好效果。(2)提出了一种基于改进YOLOv5的船舶火灾烟雾探测方法。在众多卷积神经网络中,YOLOv5算法在检测的实时性,精度等方面具有优势。为了进一步提高精度且更加轻量化以适用于船舶火灾探测,首先通过替换YOLOv5特征提取网络为更轻量化的PP-LCNet网络,减小了模型大小,加快了检测速度。其次调整PP-LCNet网络中SE(Squeeze and Excitation)注意力机制模块的位置,使注意力机制功能得到优化,提高了模型精度。最后在损失函数部分使用更合理的EIOU(Efficient Intersection Over Union)作为新的边界框损失函数,使模型精度进一步得到提高。综上所述,本文构建了以船舶封闭空间为背景的火焰烟雾数据集。提出基于改进YOLOv5的船舶火灾烟雾探测方法,并在构建的数据集上完成训练,经过仿真测试有较好的效果。该基于卷积神经网络的船舶火灾烟雾探测方法,提高了船舶火灾探测效率,弥补了传统船舶火灾探测的不足,为卷积神经网络目标检测技术在船舶火灾探测领域的应用做了新的探索。更多还原

【Abstract】 Ship fire detection is always the top priority in the field of ship safety.At this stage,ship fire detectors trigger alarms by sensing the characteristic parameters of the combustion reaction,which has delays,non-visualization,and a single detector with a small range.With the development of large-scale ships,these shortcomings become more and more obvious.In order to reduce the delay of ship fire detection,expand the detection range,and obtain more fire information,it has become an urgent technical problem to be solved in ship fire prevention.In response to the above problems,this paper proposes a ship fire smoke detection algorithm based on improved YOLOv5 through in-depth research on traditional fire detection and video-based fire detection.In order to train a detection model for ship fire smoke,a data set with the background of the ship’s closed space was simulated to verify the effectiveness of building simulated data and the advantages of the improved algorithm.The main research work is as follows:(1)A ship fire dataset is constructed.The data set plays a decisive role in the performance of the detection model.In the field of ship fire detection,there is no corresponding ship closed space fire data set.Therefore,this paper uses relevant software to simulate and produce a data set of fire and smoke in the enclosed space of the ship,which is used to train the neural network and achieves good results.(2)A ship fire smoke detection method based on improved YOLOv5 is proposed.Among many convolutional neural networks,the YOLOv5 algorithm has advantages in terms of realtime detection and accuracy.In order to further improve the accuracy and make it lighter for ship fire detection,firstly,the YOLOv5 feature extraction network is replaced by a lighter PPLCNet network,which reduces the model size and speeds up the detection speed.Secondly,adjust the position of the SE attention mechanism module in the PP-LCNet network to optimize the function of the attention mechanism and improve the model accuracy.Finally,in the loss function part,a more reasonable EIOU is used as the new bounding box loss function to further improve the model accuracy.To sum up,this paper constructs a data set of fire and smoke with the background of the enclosed space of the ship.A ship fire smoke detection method based on improved YOLOv5 is proposed,and the training is completed on the constructed data set,and the simulation test has a good effect.The ship fire smoke detection method based on convolutional neural network improves the efficiency of ship fire detection,makes up for the shortcomings of traditional ship fire detection,and makes a new exploration for the application of convolutional neural network target detection technology in the field of ship fire detection.更多还原