| The radio propagation model,studying on the propagation law of radio waves under complex propagation paths,is used to estimate the path loss or Reference Signal Receiving Power(RSRP)of wireless signals.It applies to the planning,deployment,and optimization of wireless communication system networks.With increasing expansion of mobile communication system coverage and growing number and kinds of users,the propagation of wireless signals in complex urban scenarios is severely affected by great changes in the propagation environment.Under the urban area environment with full coverage,massive connections,and strong blockages,the existing radio propagation model is hard to meet the multidimensional requirements for accuracy,efficiency,and generalization of RSRP estimation,which has become a problem impeding operators to plan and optimize mobile communication network performance.With the rapid development of artificial intelligence technology,the search for high accuracy RSRP estimation method based on artificial intelligence technology is the key to optimize the communication performance of urban scenes.Therefore,it is an important research topic to develop a new deep learning model with high accuracy and generalization capability for radio propagation based on deep learning techniques.The existing radio propagation models can be divided into deterministic and empirical models.The empirical model based on mathematical statistics obtains the radio propagation features in chosen environment by fitting a large amount of actual measurement data,which has a fast calculation speed yet low accuracy of RSRP estimation.The deterministic model based on electromagnetic wave propagation theory can estimate RSRP with high accuracy,but it requires huge computational overhead and high-precision environment modeling.However,both of the two types are hardly applicable for complex urban areas with a large number of building clusters.In this dissertation,the important progress of deep learning techniques in image processing is taken advantage to study the radio propagation deep learning model for urban areas.In the model building process,the dissertation converts all the environmental information(building index,building height,etc.)of the base station coverage area into an electronic environment map(EEM)in addition to considering the physical features(base station height,antenna pitch angle,and other parameters),extracts the environmental features from the EEM through deep neural networks,and finally puts forward an RSRP estimation method that integrates both physical and environmental features.The RSRP estimation problem is converted into an image translation one from the EEM to the RSRP map and propose a cell-level RSRP estimation method,which estimates the RSRP map of the whole cell at once by Generative Adversarial Networks to sense the EEM of the whole cell,effectively improving the RSRP estimation capability and the cross-cell generalization capability of the model.Meanwhile,we propose the offline model training-online deployment-fine turning scheme using a small amount of measured RSRP data with a cold start method based on incremental learning.In view of the difficulty of obtaining the EEM in practical engineering,a method is put forward to estimate the RSRP map of the whole cell by reconstructing the EEM of a cell based on a small amount of measured RSRP data combined with the dual learning structure.The main innovative work of this dissertation is as follows:(1)To address the problem of low accuracy of existing radio propagation models for estimating RSRP in urban areas,a RSRP estimation method based on feature fusion is come up with.In this dissertation,14-dimensional physical features are extracted from the measured dataset provided by HUAWEI Technologies such as base station height and link distance.Next,the cell environment information contained in the dataset is converted into an EEM which automatically extracts the environment features from the EEM by a deep convolutional neural network.Finally,a feature fusion-based RSRP estimation method is put forward by fusing physical features with environment features.After experimental validation on the measured dataset,the root mean square error of the extracted environmental features in RSRP estimation is reduced by 0.95 dB comparing with that via the commonly used machine learning regressors.The experiment verifies the role of the extracted environmental features in improving the RSRP estimation accuracy within complex urban areas.(2)To address the problem that the existing radio propagation model cannot achieve cell-level RSRP estimation,a cell-level RSRP map estimation method is presented in this dissertation.The method treats the RSRP of the whole cell as an RSRP map and transforms the RSRP estimation problem into an image translation one.Drawing on the advantages of deep learning techniques in the field of image translation,a cell-level RSRP map estimation method is suggested based on generative adversarial networks.It "translates" the electronic environment map into the cell’s RSRP map and estimate the RSRP of the entire cell directly.Meanwhile,a residual estimation method is designed to improve the accuracy of RSRP estimation by using the estimation results of empirical models as the prior knowledge to guide and constrain the training process of the deep learning model.The experimental results of the measured dataset show that the proposed residual estimation method can accelerate the training speed of the image translation model.In new cells without any measured RSRP data,the RMSE of the cell-level RSRP estimation method proposed is 6.94 dB,better than the one of current radio propagation model.Given the practical application of the model,a three-stage radio propagation deep learning model deployment method,based on offline training,online deployment,and fine-tuning update,is presented,which solves the cold start problem of deep learning model during training step without historical collection data by incremental learning of the model through the collection of small amount of new scene data in batches,.(3)To address the difficulty of obtaining accurate EEMs,a dual learning-based RSRP map estimation method,which reconstructs the building map and predicts the RSRP map of the whole cell through only part of the measured RSRP data collected in the cell,is presented.The measured data is transformed into an incomplete measured RSRP map based on location coordinates.In this dissertation,a U-Net-based image transformation model,RadioCycle,is constructed to first transform the incomplete measured RSRP map into a building one,and then transform the building map into a complete RSRP one.In this dissertation,we take the mutual transformation between building maps and RSRP ones of cells as a dual task,and use RadioCycle to predict building and RSRP maps simultaneously.Experiments on the simulated dataset demonstrate that RadioCycle achieves a building reconstruction accuracy of 0.95 and an estimated RSRP root mean square error of 0.069 with only 20%measured RSRPs in a cell.RadioCycle achieves state-of-the-art prediction accuracy in both map transformation tasks.The models proposed in this dissertation are validated on both the measured datasets and publicly simulated datasets and comparison has been made with existing models.The motivation,structure,operation mechanism,and training mechanism of the model construction are explained in this dissertation.The experimental method,model usage and measurement results are well generalized and generalized in different scenarios of communication cells,which can provide support and reference for network planning,deployment,and optimization of mobile communication systems and channel simulation,road test analysis,coverage planning,and access performance optimization. |
