Machine Learning-based Recognition Algorithm For Terrain Of Microwave-based Wireless Network

Microwave-based Wireless Network(MWN),a kind of wireless network based on microwave communication links,which has the characteristics of large capacity and long transmission distance,is an important part of national communication networks.Because the performance of a MWN is often affected by terrain,when planning a MWN,we need to consider factors that affect network performance,including long(microwave)link,small angle(between neighboring microwave links),the ratio of links that are eligible to communications,and link cross(between microwave links).For this reason,operators need to accurately identify the terrain characteristics of a MWN before they carry out topological planning for the MWN,which leads to an important research problem,i.e.,the terrain identification of MWN.Although there have been a lot of studies on terrain recognition,the mainstream technology of terrain recognition mainly uses remote sensing data,geographic information system,global positioning system(3S data)and digital elevation model(DEM)to realize partition recognition and analysis oriented to slice areas.The terrain of MWN is a network terrain,facing three challenges in its recognition,including the lack of high information samples,the recognition of non-slice area,and the difficulty of multi-level recognition,which,causes inability to apply the mainstream terrain recognition technology.To resolve the above difficulties,this thesis proposes two novel algorithms for microwave network terrain recognition based on machine learning techniques,which are as follows:First,we propose a MWN terrain recognition algorithm based on the K-means clustering approach.The algorithm randomly grabs a certain number of reference networks similar to the networks to be identified in target area,which include mountains,plains and other terrain.Then digital elevation models of the network terrain are established for the networks to be identified and the corresponding reference networks.Based on the DEM,five terrain features are extracted,including average elevation feature,elevation variance feature,horizontal narrow feature,oblique narrow feature and link ratio feature.Adopting the idea of comparative location recognition,we employ the Elbow Method to determine the number of optimal clustering centers,the maximum and minimum initial clustering center method to determine the initial location of the optimal clustering centers,and the single dimension recognition strategy to solve the problem of interference under different data densities;We finally employ the K-means algorithm to realize unsupervised clustering for network terrain characteristics of each MWN.The clustering results of each dimension of the networks to be identified are considered as the corresponding terrain labels.Through experimental verification,the MWN terrain recognition algorithm based on K-means clustering can effectively identify the terrain of network to be recognized,and has certain advantages in the accuracy and efficiency of recognition compared with human recognition.Moreover,the larger the number of networks to be recognized,the more obvious the advantages are.Second,considering the need of evaluating atypical terrain between two typical terrain categories,we further propose a MWN terrain recognition algorithm based on K-Nearest Neighbors algorithm(KNN).Similar to the previous one,this algorithm also grabs reference networks,builds DEM,and extracts the terrain feature value.The reference networks form samples for the MWN terrain recognition algorithm.The elastic K-value strategy is employed to combine the K-value and the position of the networks to be identified in the feature space,which solves the problem of determining the K-value in the KNN algorithm.The feature ratio output strategy is employed to solve the dividing problem of the MWN to be identified between classes,which effectively reduces the influence of K-value selection on recognition results.The KNN algorithm is used to grab the labels of K reference networks which are the closest to each other in feature space,and the ratio of the number of terrain categories in these reference networks is used as the "feature ratio" to identify and evaluate the networks to be identified.Experimental results show that the MWN terrain recognition algorithm based on KNN classification can effectively evaluate atypical networks between two typical terrain categories.