GSM Network Optimization Based On Self-Organizing Feature Maps Neural Network

GSM network is still the major kind of networks that are operated by mobile carriers and are bearing users and businesses in our country. The responsibility of GSM network optimization is to improve the network quality and customers'experience about the network for enterprise's sustainable development, to create more economic and social benefit under a specific network resource condition. In recent years, the mobile communication market has become more and more popular that resulted in more fierce competition between mobile communication operators. Such fierce competition led to the rapid progress in the GSM network optimization research. Optimization engineers are prompted to enhance network optimization level through new ideas, new methods and new tools.The upgrading of network optimization level lies in the following three aspects: upgrading of the network optimization efficiency, standing out the end-to-end characteristics of network and emphasizing the users'satisfaction, improving the detailed level of optimization.The top priority of network optimization process is to collect network statistics data, test result data, the customer complaint data and so on and to convert it to the reliable data which can be able to guide optimizing work by optimization engineers. This process is also an important segment in enhancing the level of network optimization.Since network optimization is the process during which the optimize solution is realized with the changes of the network, business and the users, this method has strong subjective factors and there is no absolute "correct" or "wrong" solution, but only relative "better" or "poor" solutions. This characteristics of network optimization are in accordance with the finite nonlinear and convex characteristics of the artificial neural network, which suggests us that artificial neural network method can be employed to study and solve some network optimization problems.Artificial neural network, has been greatly developed since being produced, and has been applied in several areas successfully. The artificial neural network study is also constantly becoming breakthroughs and progresses. Taking the Self-Organizing Feature Maps Neural Network as an example, the researchers are enhancing it continuously, based on dynamic neurons in number, the matching neurons strategy, SOFM in combination with other algorithms, SOFM network's own combination, in order to overcome its own limitations and solve all kinds of practical problems more efficiently and accurately.The author regarded a GSM manufacturer's equipment air interface measurement report data as the research object of network optimization, used the method of neural network SOFM and neural network tools provided by MATLAB, conducted the cluster analysis on the network elements which attributes are indicated by portion of measurement report data, and found several kinds of problems existing in different classes of TRXs. Experimental results show that this method is successful in mass data batch processing and is able to clustering network units with different characteristics.The author used the drop-call indicator to verify the clustering results. Clustered bad TRXs also have a poor performance in drop-call statistics. Thus, the conclusions are as follows:For a first kind of TRX, even when their RxLev higher than -66dBm, the quality of the voice begin to go worse. Furthermore when the RxLev is higher than -80dBm, the quality has evident trend of getting worse. Former experience of network optimization shows that this kind of TRXs needs careful attention, and the operator should look for their problems within hardware and interference from outside of the network.For a second kind of TRX, when their RxLev lower than -81dBm, their voice quality begin quickly goes down. This kind of TRXs takes 14 percents among all the TRXs. Former experience tells us that such kind of TRXs always have troubles in over-covering or the victim of another cell which have the over-covering problem. We should solve the problems by optimizing the structure and coverage of each cell in the network.For a third kind of TRXs, when their RxLev become lower than -91dBm,their voice quality begin to get worse apparently. This kind of TRXs takes 16 percents. If these TRXs belong to cells in the urban, in a common sense, the area must involve in problem of bad background noise. To solve the problem, network optimizers should tune the frequency policy of the area.Another kind of TRXs, to which the network optimizer should pay little attention, their voice quality keep ideal level through all the RxLev sections. This kind of TRXs takes 30 percents. This kind of TRXs is the most valuable part of the network and we should take effort to enlarge their percentage within the network.From the result of clustering, we can also find out that the point of RxLev from which the whole network mean voice quality begins to be even worse, is between -81dBm and -86dBm. In this case, the result shows that the customers'experience of network begin getting unpleasant. The problems such as noise, discontinuous voice will appear. The network optimizer should find a way of planning the network structure to raise the average quality of the mobile network.At the end of this paper, the author discussed the training times, cluster quantity and clustering stability of the SOFM neural network in the network optimization application, and drew some conclusions. Finally, the author brought up prospects of the combinatorial data analysis, improving the clustering effects and strategies.