Research On Key Technologies Of Atmospheric Environment Monitoring System Based On Deep Learning

Air quality is closely related to people's production,life and social development,so atmospheric environmental monitoring system plays an irreplaceable role in environmental protection and pollution control.Using the real-time monitoring information to grasp the air pollution situation,evaluate and predict the environmental quality,provide technical support for early warning decision-making,scientific control of air quality and regional joint prevention and control work.Deep learning is a new field in machine learning research.Its purpose is to establish neural network,which can simulate human brain for analysis and learning,and learn the inherent laws and representation levels of sample data.Using the mechanism of human brain to read and interpret data is to enable machines to learn and analyze like humans.Deep learning has achieved great success in computer vision tasks such as image classification,object detection,semantic segmentation and face recognition.This paper analyzes the massive multi-source,multi-dimensional and multi-state data collected by the atmospheric environment monitoring system,and studies the key technologies of the atmospheric environment monitoring system by using deep learning technology.The key technologies in the process of atmospheric environment monitoring are discussed from three aspects: the source of abnormal data of environmental monitoring equipment,the analysis of atmospheric aerosol particle composition and the optimization of air quality numerical prediction model.At the same time,the specific solutions are given:1.Research on abnormal data source diagnosis model based on deep learning.Aiming at the problem that the atmospheric environment monitoring system can't locate the monitoring equipment that produces abnormal data in detail,an improved Faster RCNN model is proposed.The target detection algorithm based on deep learning is applied to train a large number of images of normal and abnormal operation status of environmental monitoring equipment to establish the source diagnosis model of atmospheric environment monitoring abnormal data.With the help of the image collected by the video monitoring system in the environmental monitoring station,the problem of the location of the abnormal data source equipment is transformed into the problem of the target detection of the image collected by the monitoring equipment.Considering the feature attributes of the detected target,the Faster R-CNN model is improved by convolution layer reconstruction,feature fusion,anchor frame reset and data amplification,which improves the accuracy of model abnormal data source diagnosis,and makes the monitoring equipment troubleshooting work for abnormal data source to be unattended.2.Research on automatic classification model of aerosol particles based on deep learning.There are some problems in the existing classification methods of atmospheric aerosol particles,such as the lack of a unified extraction standard and relying on manual experience to name them,which costs a lot of manpower and material resources.At the same time,the monitoring process of aerosol particles based on a single particle will produce too much particle information.Aiming at the current situation and shortcomings of atmospheric aerosol particle monitoring and classification,based on deep learning classification algorithm,an improved Alexnet model is proposed to train and establish the automatic classification model of atmospheric aerosol particles.In this paper,the aerosol particle mass spectra which have been named artificially in the past monitoring activities are used to label the category information manually,and the data set for training the classification model is established.Based on the in-depth study of the characteristics of aerosol particle mass spectrometry,the optimization methods such as adjusting the image resolution,reducing the size of convolution kernel and reducing the number of network layers are adopted to improve the deep learning classification algorithm of alexnet.After deep convolution network extraction,it can reflect the mass spectrum characteristics of particle categories,automatically learn the aerosol particle composition,and generate the automatic classification model of atmospheric aerosol particles.It improves the accuracy of the classification model,realizes the purpose of automatic classification,and achieves the effect of real-time detection.3.Air quality prediction model based on deep learning.The existing air quality numerical prediction system WRF-CMAQ,due to the uncertainty of emission inventory of pollution sources and the inability to fully quantify the physical and chemical changes in atmospheric transport,leads to the deviation of air quality prediction values.In view of the current situation of the existing air quality prediction system,this paper proposes the application of depth confidence network model to mine the relationship between the predicted value and the measured value of the regional numerical prediction model,and establishes the air quality prediction model based on deep learning.The model uses the historical monitoring data of several national monitoring stations in the study area and the corresponding meteorological forecast data,and fully considers the temporal variation and spatial distribution characteristics of air pollutant concentration.In the forecast period,the predicted value of pollutant concentration at any station in the region is modified to improve the effectiveness of the air quality forecast model.