Research On Harmful Gas Sensing Technology Based On Deep Learning

With the acceleration of industrialization and urbanization,chemical,thermal power and other enterprises will release a variety of harmful gases in the production process,resulting in loss of people's lives and property and damage to the ecological environment.Therefore,it is urgent to use multi-source image acquired by remote sensing and ground-assisted detection technology to realize the omni-directional monitoring of harmful gas release source location and harmful gas.This paper focuses on the method of multi-source image release source localization and harmful gas detection based on deep learning.The main research contents and achievements are as follows.Aiming at the low precision of existing harmful gas release source localization methods caused by insufficient training samples,a method based on relation metric for remote sensing image harmful gas release source location is studied and implemented.Firstly,the convolutional neural network is used to extract the different scale feature information of the small sample support set and the test set image respectively,and the low-level features and the high-level features are merged and sent to the regional proposal networks to generate proposal boxes.Then,the proposal boxes features of different dimensions are selected.The pooling operation of the region of interest is performed to generate a feature map of a fixed dimension size.Finally,the feature information is sent to the relation unit and the bounding box regression respectively,and the relation unit predicts the target category according to the similarity score of the test image and the support set image,and the bounding box regression returns fine position correction is performed on the proposal boxes.Through multiple sets of experiments,and compared with other methods,the method has higher positioning accuracy under the small sample harmful gas release source data set.Aiming at the problem that the current infrared image gas detection method has a poor gas detection effect,a method based on deformable convolution for full convolution network infrared image gas detection is researched and implemented.First,the deformable convolution operation is added to the standard convolution of the full convolutional neural network;secondly,the network pre-training is performed on the public infrared data set;finally,the pre-training network is fine-tuned on the small sample gas infrared data set to improve the network's ability to detect gases.By testing the actual sets of infrared data,it is found that the detection method of this method is better than the other two classical methods.Aiming at the problem that the harmful gas sensing technology processing method is insufficient for computing resource utilization and the operating environment is difficult to be unified,a harmful gas sensing platform based on private cloud architecture is designed and constructed.The platform builds a multi-node distributed cluster based on container technology,creates a mirror image related to harmful gas sensing technology,realizes the rapid deployment of container applications,and provides a good server environment for the operation of high-speed real-time harmful gas sensing technology.