Advanced Analysis Methods for 3G Cellular Networks

Intelligent Database Systems Lab

國立雲林科技大學National Yunlin University of Science and Technology

Advanced Analysis Methods for 3G Cellular Networks

Advisor : Dr. Hsu

Presenter : Chih-Ling Wang

Author : Jaana Laiho, Kimmo Taivio, Pasi

Lehtimaki, Kimmo Hatonen, Olli

Simula

2 Intelligent Database Systems Lab

N.Y.U.S.T.

I. M.Outline Motivation

Objection

Introduction

SOM

Network analysis using SOM

Conventional analysis of WCDMA cellular network

Usage of clustering in optimization

Conclusion

My opinion


N.Y.U.S.T.

I. M.Motivation The operation and maintenance of the third generation (3G) mobile

networks will be challenging.

These networks will be strongly service driven, and this approach differs significantly from the traditional speech dominated in the second generation (2G) approach.


N.Y.U.S.T.

I. M.Objection This paper shows that a neural network algorithm called the self-

organizing map, together with a conventional clustering method like the k-means, can effectively be used to simplify and focus network analysis.

It is easier for a human expert to discern different states of the network.

This makes it possible to perform faster and more efficient troubleshooting and optimization of the parameters of the cells.


N.Y.U.S.T.

I. M.Introduction

The mobile communication industry is currently shifting its focus from second generation networks (2G) toward the third generation networks (3G).

There will be a number of new challenges when shifting from the current 2G to the new 3G networks, many of them related to the design and the operation of true multiservice radio networks.

An essential part of the new challenges is related to the provisioning, monitoring and optimization of the services.


N.Y.U.S.T.

I. M.Introduction (cont.) The fact that instead of monitoring GSM voice only, one must conce

ntrate on monitoring multisystem and multiservice environment.

This fact is the main driver pushing development toward advanced network analysis solutions.

Operating a cellular network is an iterative quality cycle process combining the network and service configuration and the related performance measurements.

In this cycle, the overall end-to-end quality target is defined and the quality criteria and thresholds for key performance indicators (KPIs) for each service type are determined.


N.Y.U.S.T.

I. M.Introduction (cont.) This paper concentrates on the analysis and visualization part of th

e quality cycle.

The motivation for the introduction of neural analysis on the network performance data is to provide effective means to handle multiple KPIs simultaneously.

Effective analysis methods reduce operators’ trouble shooting efforts, speed up the cycle, and thus, the network utilization rate increases.


N.Y.U.S.T.

I. M.Introduction (cont.) In this paper, an application of the SOM in analyzing telecommunic

ations networks is presented.

WCDMA: wideband code division multiple access

寬頻分碼多工存取， WCDMA 是第三代行動通訊系統無線傳輸技術的一種。在同一個傳輸通道中， WCDMA 可包含電路交換和分封交換的服務，因此，消費者可以同時利用交換方式接聽電話，然後以分封交換方式存取網際網路，提昇行動電話的使用效率。


N.Y.U.S.T.

I. M.SOM-SOM in Network Analysis A behavior pattern of a cell at a certain instant is a set of

indicator values that have been recorded at that instant.

In network analysis, the SOM can be used to find and show similarities between behavior patterns of cells.

A set of n indicator values form an n-dimensional pattern vector.

During the training phase a set of these vectors is used to train a SOM.

When the SOM is visualized, similarly behaving cells can be spotted close to each other.


N.Y.U.S.T.

I. M.Network analysis using SOM The method consists of the following phases:

─ target selection─ data preprocessing─ clustering analysis─ result interpretation


N.Y.U.S.T.

I. M.Network analysis using SOM-Target selection The first step in the process is the target definition that includes the

selection of the geographical area, network objects (base stations, radio network controllers, routers, etc.), and visualization task specification.

The selection of network objects and the visualization task have a strong impact on the selection of the measurements and KPIs to be analyzed.

Each object in the network has its own specific measurements.

The visualization task can be more generic or problem oriented.


N.Y.U.S.T.

I. M.Network analysis using SOM-Preprocessing

The main objective of the preprocessing phase is to ensure that the analysis methods are able to extract correct and needed information from the data.

In order to get correct information out of network data, the used variables must be balanced by scaling.

The most common method to do the balancing is to normalize the variance of each variable equal to one.


N.Y.U.S.T.

I. M.

Network analysis using SOM-Preprocessing (cont.) After the normalization, the distribution of the data might be skewe

d if there are outliers in the data.

The usual solution is to remove outliers or to replace them with estimated normal or correct values.

If outliers carry interesting information, where they can be signs of network problems that are searched for, it is possible to keep outliers but prevent their large values from dominating the analysis results.

This can be done by using some sort of conversion function like tanh(x) [or log(x) ] before the variance normalization.

Such a function can decrease the effect of outliers and emphasize proper parts of the distribution.


N.Y.U.S.T.

I. M.

Network analysis using SOM-Clustering Analysis In this paper, clustering is also used to find groups of similarly

behaving cells. The data vectors of all the cells are clustered using a combination

of the SOM and the k-means algorithm.─ At first, an SOM with M map units is trained using the data vectors.─ Next, the set of M codebook vectors of the SOM are clustered into

several different numbers of clusters using the k-means algorithm.─ In this algorithm, the SOM is used to quantize the data and to

visualize the cluster structure in the data.


N.Y.U.S.T.

I. M.

Network analysis using SOM-Clustering Analysis (cont.)

In order to analyze a time-series data or a sequence of data over a time period instead of single data points, the frequency of appearance or the number of “hits” in each data cluster is computed for the given sequence of data.

The vector containing these proportions or “hits” over some time period is called a hit-histogram.

The hit-histogram of a cell over a time period provides the characterization of the cell behavior.


N.Y.U.S.T.

I. M.

Network analysis using SOM-Clustering Analysis (cont.) The clustering of the cells is performed by processing the hit-

histograms of the cells computed over consecutive time periods with a similar combination of the SOM and k-means clustering algorithm as in the extraction of the cell features.─ At first, an SOM is trained using the hit-histogram vectors of each

cell.─ Then, the codebook vectors of the SOM are clustered into different

number of clusters.─ Finally, the best clustering is selected according to the Davies-

Bouldin index.


N.Y.U.S.T.

I. M.

Network analysis using SOM-Result of Neural Network The method described above has been used to analyze the uplink

direction in the microcellular network scenario. This scenario was selected since it represents a challenging

environment from the propagation point of view. The WCDMA radio networks used in this paper were planned to

provide 64-kb/s service with 95% coverage probability, and with reasonable (2%) blocking.


N.Y.U.S.T.

I. M.

Network analysis using SOM-Result of Neural Network (cont.) The network layout comprises 46 omnidirectional base station site

s. Due to the lack of measured data from live networks, simulated da

ta produced by a dynamic system simulator is used in the advanced analysis cases.

The network parameters are collected in Table I. The system features used in the simulations are according to 3GPP.


N.Y.U.S.T.

I. M.

Network analysis using SOM-Uplink Results in Microcellular Scenario

The frame error rate values are preprocessed using y=tanh(ax) function because it maps all x>=0 into a range [0,1] and the shape of the mapping can be controlled with the parameter a .

Each component plane shows what kind of values a single variable has in different parts of the map.

The value of the variable is indicated by gray-level and it can be read from the gray-level axis on the right side of the corresponding component plane.


N.Y.U.S.T.

I. M.

Network analysis using SOM-Uplink Results in Microcellular Scenario (cont.)


N.Y.U.S.T.

I. M.

Network analysis using SOM-Uplink Results in Microcellular Scenario (cont.) In Fig. 5(a), a classification of mobile cells using the histogram map

for uplink direction data is shown. The classification is based on histogram features which are

computed from a time window.


N.Y.U.S.T.

I. M.

Network analysis using SOM-Uplink Results in Microcellular Scenario (cont.) The only cell in the bad performance area, that is, the behavioral

cluster 3 described by rules for data cluster 6 and 7, is cell 44.

Characteristics of this cell are high load and high number of users.

The frame error rate (FER) performance of this cell is degraded and, thus, it can be concluded that the cell is operating at the edge of its capability.

It is worth noting that an analysis utilizing conventional means did not identify cell 44 as problematic.

Only the use of SOM and further analysis using expert knowledge proved this.


N.Y.U.S.T.

I. M.

Network analysis using SOM-Uplink Results in Microcellular Scenario (cont.)


N.Y.U.S.T.

I. M.

Conventional Analysis of WCDMA Cellular NetworkA. Analytical Approach for the Network Performance

Data As presented in [20], the uplink load factor can be

calculated as a sum of load factors of all N uplink connections in a cell.

R is the used bit rate, is the signal-to-noise ratio requirement, and W is the WCDMA chip rate, i is other to own cell interference ratio.

UL


N.Y.U.S.T.

I. M.

Conventional Analysis of WCDMA Cellular Network (cont.)B. Traditional Analysis Results for Microcellular Case Using (4), two reference figures for the cell performance can be

calculated based on the input data, namely the loading caused by a user and the number of users a cell can serve.

The used uplink value in this paper was 3.5 dB. The frequently quoted i value is 55%.

Ob NE /


N.Y.U.S.T.

I. M.

Conventional Analysis of WCDMA Cellular Network (cont.)B. Traditional Analysis Results for Microcellular Case


N.Y.U.S.T.

I. M.

Conventional Analysis of WCDMA Cellular Network (cont.)B. Traditional Analysis Results for Microcellular Case For uplink performance evaluation, a simple function f combin

ing the interference control and throughput aspects were generated.

The weighting for each item in the cost function was the same

ThroughputNORM is the normalized throughput. In the normalization, the maximum throughput was the maximum value in a cell.

According to this classification the top 10% performing cells were cells 8, 9, 11, 25, 29, and 44.


N.Y.U.S.T.

I. M.

Evaluation of the SOM-BASED Method-Validity of the SOM Results During analytical analysis (Section IV-B) for the microcell upli

nk case, the best performing cells were 8, 9, 11, 25, 29, and 44. When mapping these cells on the performance spectrum of Fig.

5(a), it is very interesting to note that cells 8, 9, 11, and 29 are all in the same behavioral cluster (i.e., 6).

Cell 25 is in behavioral cluster 2, owing to the fact that as an edge cell (dominance area mostly on water) it has a lot less users than the other cells.

Traditional means are not capable of finding the performance problems of cell 44.

As a conclusion, it can be said that traditional means support the conclusions of the cell classification performed by the SOM.

6

2


N.Y.U.S.T.

I. M.

Evaluation of the SOM-BASED Method-Convenience and Usability of the SOM Based Analysis Current network-performance monitoring and analysis tools

are not capable of meeting the needs and requirements of service-driven networks.

Each measurement is presented separately and the end user is responsible for correlating the different measurements.


N.Y.U.S.T.

I. M.

Evaluation of the SOM-BASED Method-Convenience and Usability of the SOM Based Analysis (cont.) Current methods rely strongly on averaging or cumulating

values over longer period of time, most often one day. KPI values are analyzed as snapshots representing one period. This approach loses details such as the form of value

distribution of a KPI inside the period. The approach can be enhanced by dividing the period into sub-

periods and calculating averages or cumulative sums over those.


N.Y.U.S.T.

I. M.

Evaluation of the SOM-BASED Method-Convenience and Usability of the SOM Based Analysis (cont.) The method proposed in this paper is highly visual and it can

be used to combine different types of information like time and geography with the cell behavior profile.


N.Y.U.S.T.

I. M.

Evaluation of the SOM-BASED Method-Scalability of the Method As mentioned earlier, the network performance is currently handl

ed by visualizing the KPIs separately on the screen. As a result, the number of graphs to be visualized increases with r

espect to the number of services, the number of network elements and the number of KPIs used in the analysis.

However, the number of required visualizations does not increase so radically in the SOM-based approach presented in this paper.

The number of symbolic rules per data cluster increases with the number of used variables only if the new set of variables is able to separate the data clusters better than the old set of variables.

The SOM-based method is more scalable with respect to the changes in the problem setting than the currently used methods.


N.Y.U.S.T.

I. M.

Evaluation of the SOM-BASED Method-Computational Complexity of the Method The method consists of two independent algorithms (SOM, k-

means) whose computational complexity is dependent on the amount of analyzed data.

The computational complexity of the SOM is proportional to NMd, where N is the number of data samples in the data set, M is the number of map units in the SOM, and d is the number of variables (dimension) of the data set.

The computational complexity of the basic k-means clustering algorithm is said to be proportional to , where N is the number of data samples and Cmax is the maximum number of clusters in which the data is to be clustered.

max

2

C

kNk


N.Y.U.S.T.

I. M.

Evaluation of the SOM-BASED Method-Computational Complexity of the Method (cont.) When the k-means clustering is applied to the codebook

vectors of the SOM trained with the data set instead of the original data set, the computational cost of clustering is reported to reduce to

max

2

C

kMkNM


N.Y.U.S.T.

I. M.Usage of clustering in optimization

The analysis method presented and used in this article consists of two different phases: the clustering of single data points consisting of several KPIs and clustering of sequences of these data points.

The first phase is actually the traditional way to use the SOM. The basic clustering can be analyzed further in order to see ho

w an average behavior of cells has developed during a longer period of time.

In the case of behavioral clusters, the method takes into account not only the average behavior pattern based on all selected KPIs, but also the recent variation in behavior pattern.


N.Y.U.S.T.

I. M.

Usage of clustering in optimization (cont.)


N.Y.U.S.T.

I. M.Conclusion This paper proposes the use of the SOM, a neural network method,

in the analysis phase of the high-level telecommunication network optimization process.

The strength of the SOM-based analysis methods is in the fact that multiple measurements are used in the analysis at the same time.

The output can be provided in a descriptive format to ease the operator decisions.


N.Y.U.S.T.

I. M.My opinion Advantage: Combine SOM and k-means to analysis the network.

Disadvantage: Some steps in the paper didn’t explain clearly.

Apply:…

Advanced Analysis Methods for 3G Cellular Networks

Documents

Transcript of Advanced Analysis Methods for 3G Cellular Networks