Research On Initial Fire Identification Method Of Marine Engine Room Based On Semantic Segmentation

Ship intelligence and automation is an important direction for the future development of ships,and ship cabin fire identification is an important branch of ship intelligence.Ship cabin fires are different from other building fires.Due to the closed nature and complexity of the cabin,ship cabin fires spread rapidly and are difficult to find,control and deal with in time.Therefore,it is important to establish an effective ship cabin fire identification method to reduce the occurrence of fire accidents and improve the safety of ships.Firstly,this thesis introduces the current research status of fire identification at home and abroad and the characteristics of ship cabin fires,and establishes the cabin fire data set by simulating ship cabin fires.To address the problem of limited smoke features in the dataset,the real smoke synthesis technique is used to introduce more real-world smoke features into the ship cabin fire scenario,and this method can significantly improve the recognition performance of the model.Through the above method,a more complete cabin fire dataset is obtained,which lays the foundation for further research.Secondly,in order to realize the recognition of ship cabin fires,an improved ship cabin initial fire recognition model Deep Labv3+ IDLKNet is proposed based on Deep Labv3+network,the decoding network of Deep Labv3+ network is reconstructed,the multi-scale feature fusion decoding network structure is designed,and the large convolution kernel is introduced to increase the model perceptual field.Analyzing and drawing on the characteristics of existing loss functions,a new power-squared loss function is proposed,which is flexible enough to tune the parameters according to different tasks and data sets to improve the model’s detection of small flames and sparse smoke.m Io U of the Deep Labv3+ IDLKNet model is improved by 5.8% overall compared with the original Deep Labv3+ model,and for better performance in separating thin smoke and small flames,which can still be accurately identified even when some features are not obvious.Compared with other classical models,the Deep Labv3+ IDLKNet model segmented the best results.The purpose of this thesis is to investigate the impact of these improvements on the model performance and to provide a more accurate method for the prediction and prevention of ship cabin fires.Finally,DeepLabv3+ CNet,a ship cabin fire scene and stranded person identification model,was constructed based on Deep Labv3+ IDLKNet network to ensure the safety of crew members’ lives and properties as much as possible and to make escape and rescue plans based on the fire danger at the scene.The experimental results show that the Deep Labv3+ CNet model still has good recognition ability when only a part of human body is shown in the footage,the scene is more complicated and there are more smoke occlusions,while other classical classification models are prone to false detection.The Deep Labv3+ CNet model and the semantic segmentation model Deep Labv3+ IDLKNet are co-constructed to build a ship cabin fire early warning system in order to quickly formulate escape and rescue plans,so as to ensure the safety of crew members’ lives and properties and realize the fire early warning function.The combined application of these technical methods will effectively enhance the life safety of crew members and the overall safety of the ship.