Foggy Weather Monitoring Method Based On Deep Learning And Radio Signal

In foggy weather,water vapour in the air condenses into small droplets suspended in the air,causing reduced visibility at ground level.The impact of fog on aviation,sea and land transportation is becoming increasingly obvious with the rapid development of the transportation industry,which may lead to huge economic losses and serious traffic accidents.In addition,pollutants in the air can combine with water vapour and become difficult to disperse and settle,allowing them to collect at heights where people are active and can damage health.However,the detection of fog is still dominated by manual visual inspection,and meteorological monitoring equipment such as Doppler weather radar,visibility meters and microwave radiometers are only used where precise detection is required,but these monitoring devices have problems such as high installation costs,maintenance difficulties and small detection ranges.Therefore,it is important to study fog monitoring methods at low cost and high spatial and temporal resolution.With the widespread use of radio technology,researchers have discovered that weather such as rain,fog and dust can have an effect on radio signals that is always present and does not disappear with changes in signal frequency.And with the development of artificial intelligence,convolutional neural networks are widely used in deep learning,which can continuously extract features through convolution kernels to improve recognition accuracy.Therefore,this thesis takes advantage of the fact that radio signals are affected by weather phenomena such as rain,fog,sand and dust and leave the characteristics of the medium in the spectrum of the received signal.The deep learning algorithm is used to extract the characteristics of the propagation medium left in the signal,and then monitor the different propagation mediums for the purpose of monitoring different weather.Combining wireless communication with deep learning and using it in the field of fog weather monitoring,a deep learning-based fog weather monitoring method using wireless communication links is proposed.The main research is carried out in the following areas.(1)Building the dataset and pre-processing of the data.The 3900 A radio monitoring receiver was used to collect radio signals in six different frequency bands at four different concentrations of foggy weather to create the radio foggy weather monitoring dataset.All other conditions are held constant and the channel is changed to create a radio foggy weather monitoring dataset under different channels.The data are pre-processed to convert IQ data into spectral waterfall plots,and the accuracy of the two pre-processing methods of IQ data and spectral waterfall plots are compared,and the one with better performance is selected as the final input data.(2)A radio foggy weather monitoring model based on improved ResNet50 network and a radio foggy weather monitoring model based on improved deep residual shrinkage network are proposed.Firstly,the attention mechanism is introduced in the traditional ResNet50 network and the signal features are extracted using the feature fusion method.The accuracy of the improved model and the traditional classification model are compared,and the applicability of the improved model to foggy weather monitoring under different channels is verified.Then,in order to further enhance the model performance and improve the recognition ability of noise-containing signals,a radio foggy weather monitoring model based on an improved deep residual shrinkage network is proposed based on the improved ResNet50 network model.Firstly,a soft thresholding is added to the residual learning unit of the deep residual module to obtain a deep residual shrinkage network,then a wide convolutional layer is added to the network and an improved CBAM attention mechanism is introduced,and the influence of the position relationship between the channel attention module and the spatial attention module on the final recognition results is explored.Finally,a two-layer attention mechanism is embedded to improve the learning ability of key features.By comparing the recognition results with the models of Alex Net,VGG,and the improved ResNet50 network,it is concluded that the improved deep residual shrinkage network model has the best performance and achieves the accuracy of 93.75%.The usability and effectiveness of the proposed method for fog recognition and monitoring are proved.(3)The applicability of the model proposed in this thesis to the monitoring of different weather conditions is investigated.Firstly,The signals in the frequency bands 1268.52 MHz and 1575.42 MHz are collected under dusty,rainy and sunny conditions respectively to build a radio dusty weather monitoring dataset;the signals in the frequency bands 1815 MHz,1850MHz,1870 MHz and 1895 MHz are collected under rainy and sunny conditions respectively to build a radio rainfall weather dataset.The improved deep residual shrinkage network model proposed in this thesis is applied to the radio dust weather monitoring dataset and radio rainfall weather monitoring dataset established in this thesis,and the final accuracy of monitoring dust weather reaches 89.52%,and the accuracy of monitoring rainfall weather reaches 95.83%.The applicability of the monitoring models proposed in this thesis to other weather such as sand and dust and rainfall is demonstrated.