• Share
  • Email
  • Embed
  • Like
  • Save
  • Private Content
Advanced Analysis Methods for 3G Cellular Networks

Advanced Analysis Methods for 3G Cellular Networks






Total Views
Views on SlideShare
Embed Views



0 Embeds 0

No embeds



Upload Details

Uploaded via as Adobe PDF

Usage Rights

© All Rights Reserved

Report content

Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

  • Full Name Full Name Comment goes here.
    Are you sure you want to
    Your message goes here
Post Comment
Edit your comment

    Advanced Analysis Methods for 3G Cellular Networks Advanced Analysis Methods for 3G Cellular Networks Document Transcript

    • Advanced Analysis Methods for 3G Cellular Networks Jaana Laiho, Kimmo Raivio, Pasi Lehtim¨ki, Kimmo H¨t¨nen, Olli Simula a ao Abstract— The operation and maintenance of the 3G mo- works is manifold compared to that of today’s networks. bile networks will be challenging. These networks will be This fact is the main driver pushing development towards strongly service driven, and this approach differs signifi- cantly from the traditional speech dominated 2G approach. advanced network analysis solutions, like the one presented Compared to 2G, in 3G the mobile cells interact and inter- in this paper. Such functionality is located in the network fere with each other more, they have hundreds of adjustable and service management layers of the Telecom Operations parameters, and they monitor and record data related to several hundreds of different variables in each cell. Map (TOM) framework [1]. This paper shows that a neural network algorithm called Operating a cellular network is an iterative quality cycle the Self-Organizing Map (SOM) together with a conven- process combining the network and service configuration tional clustering method like the k-means can effectively be and the related performance measurements. In this cy- used to simplify and focus network analysis. It is shown that these algorithms help in visualizing and grouping similarly cle, the overall end-to-end quality target is defined and the behaving cells. Thus, it is easier for a human expert to dis- quality criteria and thresholds for key performance indi- cern different states of the network. This makes it possible cators (KPI) for each service type are determined. Net- to perform faster and more efficient trouble shooting and work performance data is gathered from Network Man- optimization of the parameters of the cells. The presented methods are applicable for different radio access network agement Systems (NMS), drive tests, protocol analyzers technologies. and/or customer complaints. The actual measured service Keywords— Network management, 3G cellular system, performance is analyzed and the results are compared with WCDMA, radio access network, artificial neural network, the set targets. In case of conflict, corrective actions to the Self-Organizing Map, k-means clustering, data mining. network configuration are carried out. Effective analysis methods are prerequisites for dynamic I. Introduction and successful operations. This paper concentrates on the analysis and visualization part of the quality cycle. Data- The mobile communication industry is currently shifting driven algorithms and data mining methods provide effi- its focus from second generation networks (2G) towards cient tools for exploratory data analysis. Algorithms based the third generation networks (3G). The shift is not only on Artificial Neural Networks (ANNs) have proved to be related to the evolution of the access technology, but also especially suitable in highly complex and data intensive to the vision of the development of service provisioning applications. The Self-Organizing Map [2] is one of the and service demands, customer expectations and customer most popular neural algorithms due to its efficient visu- differentiation. alization properties. The motivation for the introduction While current wireless networks still evolve and ser- of neural analysis on the network performance data is to vice providers bring new internet packet data services provide effective means to handle multiple KPIs simulta- into the markets, an increasing number of operators and neously. Furthermore, effective analysis methods reduce other wireless communication professionals are becoming operators’ trouble shooting efforts, speed up the cycle, and familiar with the wideband code division multiple access thus, the network utilization rate increase. (WCDMA) technology and prepare themselves for 3G ser- Furthermore, the quality as experienced by the end user vices and networks. There will be a number of new chal- (QoE) becomes increasingly important owing to the fact lenges when shifting from the current 2G to the new 3G that over provisioning quality is inefficient and expensive. networks, many of them related to the design and the op- On the other hand, customers suffering from low quality eration of true multi-service radio networks. An essential tend to change to the competitor’s network. Determin- part of the new challenges is related to the provisioning, ing the QoE is about collection and combination of infor- monitoring and optimization of the services. The number mation from different domains (access networks, interfaces of network counters and measurements shall increase owing and core network etc.). to the fact that instead of monitoring GSM voice only, one In this paper, an application of the SOM in analyzing must concentrate on monitoring multi system and multi- telecommunications networks is presented. Mobile cells of service environment. The information space of future net- a WCDMA network are classified according to their perfor- mance [3], [4]. SOM applications in a cellular environment J. Laiho and K. H¨t¨nen are with Nokia Group, Finland a o K. Raivio, P. Lehtim¨ki and O. Simula are with the Neural Net- a are new and thus the results are verified using analytical works Research Centre, Helsinki University of Technology, Espoo, results and expert knowledge. The Self-Organizing Map Finland is introduced in Section II. Radio access network analysis The study has been financed by Nokia Networks, Nokia Mobile Phones and National Technology Agency of Finland (TEKES) which methods based on the SOM and the results of the analy- is gratefully acknowledged. sis are presented in Section III. In Section IV, the same
    • network is analyzed using conventional methods. Methods data samples and prototype vectors in the input space. The and results of the novel SOM based analysis are evaluated update step can be performed by applying in Section V. Finally, the usability of the new method is discussed in Section VI. mi (t + 1) = mi (t) + α(t)hc (t, i)[x(t) − mi (t)] (2) II. The Self-Organizing Map where α(t) is the learning rate and hc (t, i) is the neighbor- hood function of the algorithm. The last term in the square A. SOM in data mining brackets is proportional to the gradient of the squared Eu- Data mining and exploration is an emerging area of re- clidean distance d(x, mi ) = ||x − mc ||2 . search in artificial intelligence and information manage- The learning rate α(t) ∈ [0, 1] is usually a monotonically ment. The objective of data mining is to extract relevant decreasing function of time. A good candidate is α(t) = information from large amounts of data. Data mining and α0 (1 − t/T ), where α0 is the initial value for the learning analysis tasks typically include clustering, classification, rate and T is the total number of training iterations. and regression of data, determining parameter dependen- A very frequently used form for the neighborhood func- cies, and finding various anomalies from the data. Artificial tion hc (t, i) is the Gaussian one centered on the winner map Neural Networks provide efficient tools in data mining and unit c: they have been successfully used in various data engineer- ||ri − rc ||2 hc (t, i) = exp − (3) ing applications. 2σ(t)2 The Self-Organizing Map (SOM) is a widely used neural where rc depicts the coordinates of the winner unit c, and network algorithm [2]. It maps a high-dimensional data ri denotes the coordinates of an arbitrary unit i on the manifold onto a low-dimensional, usually two-dimensional, discrete output lattice of the map and σ(t) is the width of grid or display. The SOM has several beneficial features the neighborhood. It is necessary that hc (t, i) → 0 when which make it a useful tool in data mining and exploration. t → ∞ for the algorithm to converge. During learning, the The SOM follows the probability density function of the learning rate and the width of the neighborhood function data and is, thus, an efficient clustering and quantization are decreased, typically in a linear fashion. The map con- algorithm. However, the most important feature of the verges to a stationary distribution, which approximates the SOM in data mining is the visualization property. The density of the data. topology preserving property of the SOM mapping results One step of the training algorithm of the SOM is illus- in a display inherently visualizing the clusters in the data. trated in Fig. 1. The size of the SOM is 16 units, which SOM based methods have been applied in the analysis of have been arranged into a two-dimensional grid of 4 by 4 process data, for example, in steel and forest industry [5], units. A data sample is marked with a cross; the black cir- [6], [7], [8]. cles are the values of the prototype vectors before, and the gray circles after updating them towards the data sample. B. SOM algorithm This kind of an update step is repeated iteratively during The basic SOM consists of a regular grid of map units the training process. or neurons. They are connected to the neighboring units There exist two freely available software packages that using, for instance, a rectangular or hexagonal neighbor- include implementation of the SOM: SOM-PAK [9] and hood. Each map unit, denoted here by i, is represented by SOM Toolbox for Matlab [10]. a prototype vector mi . The dimension of prototype vec- tors is equal to the dimension of the input data. First, the prototype vectors are initialized, for instance, to random values. Then, during the training the values of the pro- totype vectors are adapted to follow the properties of the input data. Training of the SOM is divided into two alternating steps, typically thousands of times each. First, one data vector x from the training data set is randomly selected and the corresponding best-matching (winner) unit (BMU) c is determined. The prototype vector of the BMU, denoted by mc , is the one that is nearest to the data sample. In other words, it minimizes the Euclidean distance between x and mc : Fig. 1. An illustration of the SOM training. c = arg min ||x − mi || (1) i In the next step, the prototype vectors of the winner and C. SOM in Network Analysis its neighbors are moved towards the data vector. It should A behavior pattern of a cell at a certain instant is a be noted that the neighborhood is defined in terms of the set of indicator values that have been recorded at that in- lattice structure, not according to the distances between stant. In network analysis, the SOM can be used to find 2
    • and show similarities between behavior patterns of cells. common method to do the balancing is to normalize the These indicators can include, for example, any subset of variance of each variable equal to one. After the normal- those indicators that are used in the traditional analysis of ization, the distribution of the data might be skewed if a network. Indicators that have been used in this paper, there are outliers in the data. If the average normal be- are shown in Table II. A set of n indicator values form an havior is studied, the usual solution is to remove outliers n-dimensional pattern vector. During the training phase or to replace them with estimated normal or correct val- a set of these vectors is used to train a SOM. During the ues. If outliers carry interesting information, for example, training, a SOM approximates the distribution of pattern as is the case in our study, where they can be signs of net- vectors in the n-dimensional space so that everywhere in work problems that are searched for, it is possible to keep the space, where there are vectors, there are nodes of a outliers but prevent their large values from dominating the SOM map as well. A prototype vector can be a BMU for analysis results. This can be done by using some sort of several data samples that are almost similar. The SOM conversion function like tanh(x) (or log(x)) before the vari- gives a topology preserving mapping of the input. SOM ance normalization. Such a function can decrease the effect nodes that are direct neighbors in a SOM grid pull them- of outliers and emphasize proper parts of the distribution. selves towards each other. So when the training has been For example, the tanh function emphasizes small values at done, most of the SOM nodes neighboring each other in the the expense of large values. grid are placed next to each other also in the input space. Therefore, cell behavior patterns that have been mapped C. Clustering analysis to neighboring SOM nodes usually resemble each other. In general, clustering is the grouping of similar samples When the SOM is visualized, similarly behaving cells can together. In this work, clustering is also used to find groups be spotted close to each other. of similarly behaving cells. The features used in mobile cell clustering and classification are computed with the method III. Network analysis using SOM shown in Fig. 2. The data vectors of all the cells are clus- In this section, a neural method for classification of mo- tered using a combination of the Self-Organizing Map and bile cells is presented. The method consists of the follow- the k-means algorithm. At first, a SOM with M map units ing phases: target selection, data preprocessing, clustering is trained using the data vectors. Next, the set of M code- analysis and result interpretation. book vectors of the SOM are clustered into several different numbers of clusters using the k-means algorithm [11]. The A. Target selection clustering process can be repeated several times for differ- The first step in the process is the target definition. This ent values of k, for example, 100 clusterings for each value √ includes the selection of the geographical area, network ob- of k, 2 ≤ k ≤ M . The k-means clustering has to be jects (base stations, radio network controllers, routers, etc.) run several times for each k because the algorithm gives and visualization task specification. The selection of net- different results depending on the initialization. The best work objects and the visualization task have a strong im- clustering for each k minimizes the sum of squared errors. pact on the selection of the measurements and KPIs to be Several values of k have to be tested because the correct analyzed. Naturally each object in the network has its own number of clusters is not known. The optimal value of specific measurements. The visualization task can be more k is defined using some clustering validity index, like the generic or problem oriented. General performance analy- Davies-Bouldin index [12]. Such an index is able to tell us sis requires a different set of measurements than a specific the best number of clusters for the current data set. In trouble shooting case. this algorithm, the SOM is used to quantize the data and to visualize the cluster structure in the data because it is B. Preprocessing much faster to find the clusters of SOM codebook vectors Neural network methods are multivariate methods that than the clusters of original data directly [13]. study the combination of variables, i.e., their joint distribu- tion. Before they can be applied to the data, the data has Training Classification SOM Clustering to be prepared for the analysis in a preprocessing phase. Data set Labeling The main objective of the preprocessing phase is to ensure that the analysis methods are able to extract correct and Fig. 2. Two phase clustering with classification. needed information from the data. Preprocessing can filter out noise, handle the problem of missing values and bal- In order to analyze a time-series data or a sequence of ance different variables and their value ranges. The action data over a time period instead of single data points, the required stems from the current information need. For ex- frequency of appearance or the number of “hits” in each ample, in network analysis one can either be interested in data cluster is computed for the given sequence of data. bad cells with abnormal indicator values in order to be able The vector containing these proportions or “hits” over some to fix them or in the behavior of the best cell in order to time period is called a hit-histogram. The hit-histogram of copy its configuration to other corresponding cells. a cell over a time period provides the characterization of In order to get correct information out of network data the cell behavior and is later used in clustering of the cells the used variables must be balanced by scaling. The most into similarly behaving groups. 3
    • The clustering of the cells is performed by processing variables are given in Table II. The frame error rate val- the hit-histograms of the cells computed over consecutive ues are preprocessed using y = tanh(ax) function because time periods with a similar combination of the SOM and k- it maps all x ≥ 0 into a range [0, 1] and the shape of the means clustering algorithm (see Fig. 2) as in the extraction mapping can be controlled with the parameter a. In this of the cell features (hit-histogram computation). At first, a way, the user has the possibility to focus on a certain range SOM is trained using the hit-histogram vectors of each cell. of FER values defined as interesting by the user, thus avoid- Then, the codebook vectors of the SOM are clustered into ing the possible dominance of the uninteresting phenomena different number of clusters. Finally, the best clustering is in the data. selected according to the Davies-Bouldin index. It should TABLE II be noted here, that the hit-histogram vectors are considered Key performance indicators as general feature vectors and the distance measure used in SOM training, SOM clustering and Davies-Bouldin index evaluation is the Euclidean distance measure. nUsr Number of users ulANR Uplink average noise raise D. Results of neural analysis ulFER Uplink frame error rate The method described above has been used to analyze the uplink direction in the microcellular network scenario First, the structure of the data has been visualized using (see Fig. 5). This scenario was selected since it represents a SOM with 2D hexagonal grid of size 10×15. In Fig. 3 the a challenging environment from the propagation point of component planes of the SOM are shown. Each component view. Furthermore, the high capacity requirements of data plane shows what kind of values a single variable has in services require a small cell environment. The WCDMA different parts of the map. The value of the variable is radio networks used in this study were planned to provide indicated by gray-level and it can be read from the gray- 64kbps service with 95% coverage probability, and with level axis on the right side of the corresponding component reasonable (2%) blocking. A Ray tracing model was used plane. For example, the variable nUsr has values roughly for the propagation loss estimation [14], [15]. The network between [0, 8] as can be seen from the gray-level axis on layout comprises 46 omnidirectional base station sites. The the right side of the component plane corresponding to the selected antenna installation height was 10 meters in aver- variable nUsr. The high values of nUsr are represented by age. Due to the lack of measured data from live networks, dark gray-levels (as can be seen from the gray-level axis) simulated data produced by a dynamic system simulator and are located close to the lower right corner and the [16] is used in the advanced analysis cases. During simu- lower left corner of the map. Similarly, the top of the map lations, the multipath channel profile of the ITU Outdoor represents data samples in which the values of nUsr are to Indoor A channel from [17] was assumed. The network much lower (represented by light gray-levels). parameters are collected in Table I. The system features used in the simulations are according to 3GPP [18]. The nUsr ulANR ulFER analysis results are presented in the following sections. 7.69 20.7 0.11 TABLE I The network parameters during simulations. 3.87 10.5 0.043 Chip rate 3.84 Mchip/s 2.92 0.012 Base station (BS) 37 dBm 0.0364 d d d maximum transmit power MS maximum transmit power 21 dBm Fig. 3. SOM component planes. MS minimum transmit power -44 dBm As mentioned earlier, the codebook vectors of the SOM MS speed 3 km/h have been clustered using the k-means algorithm in order BS antennas Omni, 11.0 dBi to obtain a clustering for the data set. The properties of MS antennas Omni, 0.0 dBi each data cluster can be analyzed using the component Propagation model Ray tracing, in-building plane representation shown in Fig. 3 or using a set of au- loss 12 dB tomatically generated rules in order to find a quantitative Propagation Outdoor to description for the data clusters [19]. In Fig. 4, the clus- channel profile Indoor A [17] tered SOM (a) and the descriptive rules for the correspond- ing clusters are shown (b). The map consists of 7 clusters, each shown using a different gray-level. In addition, the map units are labeled according to the cluster in which E. Uplink results in microcellular scenario the map units belong to. On the right are shown the de- For the analysis of the uplink direction in the microcel- scriptive rules that indicate the kind of data samples in the lular scenario, three variables have been selected. Selected different clusters. From the rules it can be seen that, for 4
    • example, data cluster 6 represents data samples with an tion into behavioral cluster 3. In Fig. 5b, the geographical unacceptably high ulFER. It should be noted here that the locations of mobile cells and their corresponding classifica- automatically generated rules represent the same informa- tion results are shown. The dominant behavioral clusters tion as the component plane representation shown in Fig. 3 in terms of the number of classified base stations are 1, 2 although in quite different form. and 6. Behavioral clusters 1 and 2 are described by rules for data clusters 3 and 5, and for behavioral cluster 6 the rules for data clusters 2, 3 and 5 of Fig. 4b apply. Typical of these behavioral clusters is the relatively low number of users, good quality and low uplink noise raise value. The explanation for the low number of users is the relatively small cell dominance area, which is typical of microcellular networks. When comparing the geographical area of cells in behavioral clusters 1, 2 and 6, high correlation can be found with uplink loading. (See Fig. 7 produced by the traditional analysis. The light areas in the figure indicate low loading.) 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 FER performance of this cell is degraded and thus, it can be concluded that (a) (b) 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 (see Sec. IV-B). Only the use of SOM and further analysis using expert knowledge proved this. In Fig. 6, the trajectories of mobile cells 8, 14 and 44 that can be used for mobile cell monitoring purposes are shown. Cell 8 has been chosen as an example to demon- strate the behavior of an ”average”cell, which is surrounded from all directions with interfering cells (see Fig. 5b). Com- pared to the situation with cell 44, the difference is that this cell is better isolated from the surrounding cells, and thus, the performance tends to be better. Cell 44 is sur- (c) (d) rounded by water areas and interfering signals can freely Fig. 4. (a) K-means clustering of the map of data vectors, (b) the propagate. Whereas in the case of Cell 8, the street canyon single variable rules for data clusters and (c-d) K-means clustering of structure and buildings isolate Cell 8 from the interfering the histogram map. sources. Cell 8 starts from behavioral cluster 2 with almost all measured data points in data cluster 5 indicating very In Fig. 4, the clusters of the histogram map (c) and the low ulANR and ulFER. Then the operation point moves hit-histogram prototypes stored in each map unit of the into behavioral cluster 6 due to an increase in nUsr and histogram map are shown (d). The histogram map con- ulANR, because the number of data samples in data clus- sists of 8 behavioral clusters (behavior characterized by a ters 2 and 3 has increased. Behavioral cluster 8 represents histogram), each indicated by a different gray-level. The another increase in the number of data samples with the map units describing the cell behavior are labeled accord- number of users up to 6-10 and ulANR between 2.4-9.1 dB. ing to the behavioral cluster in which they belong. From Then the operation point visits behavioral cluster 1 with the figure it can be seen that behavioral cluster 3 on the data samples mostly in data clusters 3 and 5, which in turn histogram map consists of feature vectors in which most of indicates operation with low number of users. In the end, the single data samples are located in data cluster 6. This the operation point on the histogram map returns back to can be seen from the sixth bar of the hit-histograms in the behavioral cluster 6. The behavior in this cell as a func- corresponding behavioral cluster. Since data cluster 6 rep- tion of time changes rapidly, but the number of users and resented data samples with high ulFER (see Fig. 4b), the loading are strongly correlated, and thus, the interference behavioral cluster 3 characterizes a cell behavior as unde- from other cells does not dominate the loading. The load- sirable. ing is generated by the traffic in cell 8 itself, and thus, no In Fig. 5a, a classification of mobile cells using the his- capacity is wasted. togram map for uplink direction data is shown. The classi- Cell 14 operates mostly in behavioral clusters 1 and 2 fication is based on histogram features which are computed with almost all of the data samples in data clusters 3 and from a time window. From the figure it can be seen that 5. As mentioned earlier, these data clusters represent data mobile cell 44 has quality problems due to its classifica- samples that have low nUsr and ulANR. The behavior of 5
    • cells. Cell 8 Cell 14 Cell 44 Fig. 6. Trajectories of the cells. In Sec. IV-B, the cells with interference problems were identified by traditional means. These cells are 3, 6, 7, 18, 24, 28, 42 and 43. When checking the position of these cells in Fig. 5a, one can see that they are dominantly lo- cated in behavioral clusters 2, 4 and 7. Behavior in clusters 4 and 7 is rather similar as it is characterized with relatively (a) high noise rise, but moderate number of users. This find- ing supports that the cause is the interference from other 1 1 cells. Behavioral cluster 2 is characterized by a low num- 1 1 8 ber of users, and thus, the other cells’ interference is again dominating in the load factor. 7 6 2 6 1 IV. Conventional analysis of WCDMA cellular 7 6 8 1 1 1 2 network 6 1 6 8 4 1 6 A. Analytical approach for the network performance data 5 1 2 6 6 2 3 As presented in [20], the uplink load factor ηUL can be 4 7 4 6 1 calculated as a sum of load factors of all N uplink connec- 2 6 7 2 7 6 tions in a cell. The effect of the multicell environment is 4 1 6 taken into account by multiplying the single cell case with 1 the term (1 + i), where i is other to own cell interference (b) ratio. Thus, the loading for a single service, fixed mobile station (MS) speed, multicell case is Fig. 5. (a) Mobile cell clustering and (b) locations of classified cells. 1 ηUL = N W (1 + i) (4) 1 + ρR cell 14 can be explained with the geographical position of this cell. It is at the edge of the analyzed area and thus, it in which R is the used bit rate, ρ is the signal-to-noise ratio has less interfering neighbors than the other cells. Low level requirement and W is the WCDMA chip rate. of interaction between the cells and low number of users ex- plain the rather static behavior of this cell. The behavioral B. Traditional analysis results for microcellular case clusters in question are characterized as ones having rela- Using Eq.(4), two reference figures for the cell perfor- tively low loading and good quality performance. mance can be calculated based on the input data, namely Cell 44, located on an island off the coast of the city, first the loading caused by a user and the number of users a operates in behavioral clusters 7 and 8 with data samples cell can serve: The used uplink Eb /No value in this study in data clusters 1, 2, 4 and 7. This is the lower left corner of was 3.5 dB. During the simulations loading was set to 0.95. the map in Fig. 4b which represents data samples with dif- Frequently quoted i value is 55%. The theoretical capacity ferent loads but low ulFER. However, the operation point values are presented in Table III. moves into behavioral cluster 3 due to an increase in the In addition to these analytical values the network was number of samples in data cluster 6, that is, the cluster analyzed with a static radio network planning tool. These with the highest values for ulFER. Characteristic of this results from the radio network planning tool and especially cell is that the loading ranges (both uplink and downlink) the format used in visualization of the results is what the from moderate to high, but the number of users remains network management systems at their best can offer to- insignificant. The low number of users and high loading in- day: visualization support of individual KPIs and textual dicates interference problems due to interference from other reports containing measurements and alarms. 6
    • TABLE III seen that the situation is not straightforward. Typically, Theoretical capacity values. admission control limits the number of users if the loading caused by active users is high. In this case, loading seems Number of users, upper bound 26 not to be the reason for blocking. Blocking is caused by Loading/user, upper bound 0.03621 downlink power outage, i.e. the maximum power allowed Number of users, i included 16 for one individual user. This explanation is based on the Loading/user, i included 0.05613 fact that microcellular scenarios are often downlink limited. In Table IV example results for the microcellular case are introduced. Only some example cells are chosen, the cells are the same as in the SOM trajectory analysis, i.e., BS14 BS1 cells 8, 14, 44. Full set of results can be found in Appendix BS2 BS13 BS15 in Table V. BS3 BS12 BS25 TABLE IV BS35 BS40 Example results from traditional analysis. BS4 BS16 BS24 BS33 BS26 BS34 BS39 BS11 BS23 BS32 BS5 BS10 BS17 BS36 Microcells BS38 BS18 BS27 BS31 BS37 BS41 BS42 BS44 BS46 Cell Id 8 14 44 BS6 BS9 BS19 BS22 BS28 BS30 BS Txp Traffic [W] 0.40 0.25 0.39 BS43 BS45 BS8 Loading 0.88 0.75 0.88 BS7 BS20 BS21 BS29 Other to own cell 0.16 0.13 0.15 interference ratio, i Users UL 22 18 22 (a) Throughput UL [kbit/s] 1408 1152 1408 BS14 BS1 BS2 BS13 All of the selected microcells have relatively high loading, BS15 and significantly low i, compared to the often used value BS3 BS12 BS25 BS35 BS40 55%. This indicates very good isolation of the cells. For BS4 BS16 BS24 BS33 all of these cells the loading per user is 0.04, which is very BS11 BS26 BS34 BS39 close to the upper bound value in Table III. In general, all BS5 BS10 BS17 BS23 BS32 BS36 the microcells have a well-controlled interference situation, BS18 BS27 BS31 BS38 only 8 cells out of 46 had i higher than 55%. These cells are: BS37 BS42 BS6 BS9 BS41 BS44 BS46 3, 6, 7, 18, 24, 28, 42, and 43. For uplink performance PUL BS22 BS19 BS28 BS30 BS43 BS45 evaluation a simple function f combining the interference BS8 BS7 BS21 control and throughput aspects were generated, see Eq.(5). BS29 BS20 The weighting for each item in the cost function was the same. PUL = f (ThroughputN ORM , i, ηUL ) (5) Cell loading % In Eq.(5), ThroughputN ORM is the normalized through- put. In the normalization, the maximum throughput was 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 the maximum value in a cell. According to this classifica- (b) tion the top 10% performing cells were cells 8, 9, 11, 25, 29 Fig. 7. (a) Used network scenario and (b) Uplink loading. Site dis- and 44. tance is couple of hundreds of meters. In (a) users with transmission power problems are denoted with ⋄ and ⋆ for uplink and downlink V. Evaluation of the SOM based method respectively. In (b) the lighter the gray-level the lower is the uplink loading. A. Validity of the SOM results During analytical analysis (Sec. IV-B) for the microcell uplink case, the best performing cells were 8, 9, 11, 25, In Fig. 7b, the uplink loading is visualized. The darker 29 and 44. When mapping these cells on the performance the gray-level the higher is the loading in a cell. In Fig. 7a, spectrum of Fig. 5a it is very interesting to note that cells the location of the mobile stations that suffer from power 8, 9, 11 and 29 are all in the same behavioral cluster (i.e. outage are depicted. When uplink and downlink informa- 6). Cell 25 is in behavioral cluster 2, owing to the fact tion is compared, it can be noted that the locations tend to that as an edge cell (dominance area mostly on water) it correlate. Thus, uplink performance and downlink perfor- has a lot less users than the other cells. Traditional means mance are well in balance. When the correlation between are not capable of finding the performance problems of cell loading and the number of uplink users is tested it can be 44. As noted earlier in Sec. III, this cell is performing at 7
    • the edge of its capabilities. Similar results can be found solutions that actually disturb the overall quality of the in the macrocellular and microcellular case when both the cell. uplink and the downlink direction are analyzed [21]. As a conclusion, it can be said that traditional means support the conclusions of the cell classification performed by the SOM. B. Convenience and usability of the SOM based analysis Current network performance monitoring and analysis tools are not capable of meeting the needs and require- ments of service driven networks. The reason for this is the increasing number of measurements that one should process simultaneously. The introduction of each new ser- vice or service class will increase the amount of measure- ments the operator needs to collect from the network el- ements. Fig. 8 represents an example of a typical KPI output. Each measurement is presented separately and the end user is responsible for correlating the different mea- surements. Furthermore, there is also a physical limitation Fig. 9. Measurement example from NMS with different time reso- lutions. For the data a minute, an hour and daily average filter has as to how many measurements can be visualized at once. been used. The end user is responsible for combining the information from different domains. Deriving the QoE from solely objective measurements requires advanced intelligence and tools, like the SOM. 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. This is different from the traditional analysis where only a few KPIs are analyzed separately at a time. C. Scalability of the method As mentioned earlier, the network performance is cur- rently handled by visualizing the KPIs separately on the screen. As a result, the number of graphs to be visual- ized increases w.r.t the number of services, the number of network elements and the number of KPIs used in the anal- ysis. Thus, the optimization of the total performance over a large set of network elements, covering also all of the most Fig. 8. An example of measurement visualization in network man- important quality aspects, requires the analysis of a large agement system. For each measurement own window is needed. number of separate graphs, and the combination of their re- sults to form a single big picture of the current state of the Current methods rely strongly on averaging or cumulat- network. However, the number of required visualizations ing values over longer period of time, most often one day. does not increase so radically in the SOM-based approach KPI values are analyzed as snapshots representing one pe- presented in this paper. For example, the increase in the riod. This approach loses details such as the form of value number of network elements increases only the number of distribution of a KPI inside the period. The approach can cell trajectories (as in Fig. 6). If the number of services or be enhanced by dividing the period into sub-periods and the number of network measurements (variables) increases, calculating averages or cumulative sums over those. This the number of component plane visualizations as in Fig. 3 easily generates an amount of data so large that details van- also increases. If only the most descriptive combination of ish into them and the simultaneous analysis of KPI combi- the single variable rules for each data cluster is shown, the nations or performance of cell group becomes even harder. number of symbolic rules per data cluster increases with Fig. 9 presents an example of NMS level KPI visualization. the number of used variables only if the new set of vari- The same KPIs are presented with minute, hour and daily ables is able to separate the data clusters better than the average filters. When a problematic KPI is detected, and it old set of variables. Thus, the number of rules per cluster is further analyzed, it can happen that the actual behavior does not necessarily grow at the same rate as the number pattern of the cell and its development is not detected at of variables used in the analysis. The SOM-based method all but the operator ends up optimizing or fixing only one is more scalable with respect to the changes in the problem aspect of the cell’s behavior. This can lead to sub-optimal setting than the currently used methods. 8
    • D. Computational complexity of the method The method consists of two independent algorithms (SOM, k-means) whose computational complexity is depen- dent on the amount of analyzed data. According to [10], the computational complexity of the SOM is proportional to N M d, 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. Thus, the training of relatively small SOMs (less than thousand map units) is computationally efficient even for large data sets. In [13], the computational complexity of the k-means algorithm and its application to the clustering of the SOM codebook vectors is described. The computational com- plexity of the basic k-means clustering algorithm is said to be proportional to Cmax N k, where N is the number of k=2 data samples and Cmax is the maximum number of clusters in which the data is to be clustered. However, 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 re- Cmax duce to N M + k=2 M k, thus making it computationally tractable solution for the analysis process described in this paper. As a reference, the computation time for a single network scenario was less than an hour, including the basic computation, visualization and report generation. VI. 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 se- quences of these data points. The first phase is actually Fig. 10. Example of utilization of cluster information in optimization the traditional way to use the SOM. It is able to show process. Selection for cells to be optimized/autotuned. what types of KPI value combinations there are in general. The basic clustering can be analyzed further in order to see how an average behavior of cells has developed during clustered based on their traffic profiles and density, prop- a longer period of time. In the case of behavioral clusters, agation conditions, cell types, radio resource management the method takes into account not only the average behav- functionality performance etc. When homogeneous mean- ior pattern based on all selected KPIs but also the recent ingful groups have been found, all the cells in one group can variation in behavior pattern. By doing this, the method share a set of configuration parameter values. This kind of is able to detect the collapse of an excellent behavior much grouping based on multiple criteria is more accurate and earlier than by analyzing only the average behavior. Cell the operation of the network will benefit from this: the op- clusters that have been found by the proposed methods timization process is greatly simplified, improved and made can be used as a starting point for a more detailed analysis. less error prone and more efficient. This is especially suitable for trouble shooting type of tasks. In behavioral clustering, the performance change over For example, in Fig. 10, it can be assumed that the lower time can be shown on the SOM (see Fig. 6). The method left corner of the SOM in the upper figure indicates cells actually combines two sources of information: change over that have a certain defect. These cells can automatically time and type of behavior at each separated point of time. be selected as optimization targets. More detailed analysis This type of analysis can be performed using data averaged can be provided on them and probably also actions to fix over various time periods, ranging from tens of seconds to them can be suggested. The ability to combine the result days. One could, for example, follow one cell movement in provided by the SOM and the geographical locations of the the SOM during peak traffic hours, assuming that networks cells is essential for an operator. The geographical location are able to report cell performance frequently enough. information can provide support to problem resolution. The sequence clustering method can be applied to follow The cell clusters can also be used to simplify the task up the optimization actions made in the network. When of parameter provisioning and optimizing. It is not feasi- the network element configuration is changed, the operator ble technically nor timewise to optimize the network cell normally wishes to see the effect of the change to perfor- by cell. Cell clustering can be used to increase the oper- mance. The method can be used to show how the changed ational efficiency of an operator’s tasks. The cells can be cells change their places on the SOM. This kind of infor- 9
    • mation combination will be essential in the optimization Acknowledgment of WCDMA radio networks. The complexity of the radio The authors wish to thank Albert H¨glund, Jukka Hen- o networks is growing as well as the size of networks them- riksson, Ari H¨m¨l¨inen, Mikko T. Toivonen, Hannu Mul- a aa selves. Operators will need means to rapidly analyze the tim¨ki and other colleagues from Nokia and Sampsa Laine a changes in the network, given the high number of cells, sev- from Helsinki University of Technology for their valuable eral services with different QoS criteria and a large amount comments. of collected performance data. References VII. Conclusion [1] Telemanagement Forum, “Enhanced telecom operations map This paper proposes the use of the Self-Organizing Maps eTOM, The Business Process Framework for the Information and Communications Services Industry, Version 3.0,” Tech. Rep. (SOM), a neural network method, in the analysis phase GB921, June 2002. of high level telecommunication network optimization pro- [2] T. Kohonen, Self-Organizing Maps, Springer-Verlag, Berlin, cess. The results provided by the SOM were verified with 1995. [3] K. Raivio, O. Simula, and J. Laiho, “Neural analysis of mo- the combination of traditional means and expert knowl- bile radio access network,” in IEEE International Conference edge. It can be stated that the results show a good agree- on Data Mining, San Jose, California, USA, November 29 - De- ment. Being able to understand the SOM results with tra- cember 2 2001, pp. 457–464. [4] P. Lehtim¨ki, K. Raivio, and O. Simula, “Mobile radio access a ditional means increases confidence in the novel analysis network monitoring using the self-organizing map,” in Euro- and its applicability in the area of cellular networks. pean Symposium on Artificial Neural Networks, Bruges, Bel- During the course of this work it was noticed that the tra- gium, April 24 - 26 2002, pp. 231–236. [5] T. Kohonen, E. Oja, O. Simula, A. Visa, and J. Kangas, “Engi- ditional analysis as such is not adequate enough to provide neering applications of the self-organizing map,” Proceedings of as enhanced demonstration of the network performance as the IEEE, vol. 84, no. 10, pp. 1358–1384, October 1996. [6] E. Alhoniemi, J. Hollm´n, O. Simula, and J. Vesanto, “Process e the SOM provides. Current performance analysis methods monitoring and modeling using the self-organizing map,” In- include combinations of planning tool analysis data and tegrated Computer Aided Engineering, vol. 6, no. 1, pp. 3–14, real network measurements. The information format in a 1999. [7] O. Simula, P. Vasara, J. Vesanto, and R.-R. Helminen, Industrial planning tool is a static snapshot of the network situation, Applications of Neural Networks, chapter The Self-Organizing and thus, the aspect of time is not present. The KPIs of- Map in Industry Analysis, pp. 87–112, CRC Press, 1999. fered by the NMS are averaged values. Certain amounts of [8] O. Simula, J. Ahola, E. Alhoniemi, J. Himberg, and J. Vesanto, “Self-organizing map in analysis of large-scale industrial sys- information of the past behavior of the KPI is also avail- tems,” in Proceedings of the Workshop on Self-Organizing Maps, able. In the case of KPI information, no prior knowledge Espoo, Finland, July 1 - 3 1999, pp. 375–387, (invited paper). on the correlation of the different KPIs is available. The [9] T. Kohonen, J. Hynninen, J. Kangas, J. Laaksonen, and K. Torkkola, “SOM PAK: The self-organizing map program correlation analysis and the combination of these two in- package,” Tech. Rep. A31, Helsinki University of Technology, formation sources is done manually and it strongly relies Laboratory of Computer and Information Science, 1996, Avail- on the expertise of the person performing the task. able: http://www.cis.hut.fi/research/som pak/. [10] J. Vesanto, J. Himberg, E. Alhoniemi, and J. Parhankan- The strength of the SOM based analysis methods is in gas, “SOM toolbox for Matlab 5,” Tech. Rep. the fact that multiple measurements are used in the anal- A57, Helsinki University of Technology, Laboratory of ysis at the same time. Furthermore, the output can be Computer and Information Science, 2000, Available: http://www.cis.hut.fi/projects/somtoolbox/. provided in a descriptive format to ease the operator deci- [11] B.S. Everitt, Cluster Analysis, Arnold, 1993. sions. The SOM based application in the analysis of cellular [12] D.L. Davies and D.W. Bouldin, “A cluster separation measure,” networks is not widely spread, and it is worth noting that IEEE Transactions on Pattern Analysis and Machine Intelli- gence, vol. 1, no. 2, pp. 224–227, April 1979. the SOM based analysis has been successfully applied to [13] J. Vesanto and E. Alhoniemi, “Clustering of the self-organizing GSM data to detect anomalous behavior of base stations, map,” IEEE Transactions on Neural Networks, vol. 11, no. 3, see [22] for more details. pp. 586–600, May 2000. [14] K. Heiska and A. Kangas, “Microcell propagation model for In the future, the operation of cellular networks will be network planning,” in Personal, Indoor and Mobile Radio Com- strongly service driven. Compared to the current situation munications, 1996, vol. 1, pp. 148–152. [15] J. Rajala, K. Sipil¨, and K. Heiska, “Predicting in-building cov- a with provisioning of voice and simple best effort data ser- erage for microcells and small macrocells,” in IEEE Vehicular vices only, the change is enormous. Effective analysis of Technology Conference, 1999, vol. 1, pp. 180–184. 2G networks’ voice service is currently challenging enough [16] S. H¨m¨l¨inen, H. Holma, and K. Sipil¨, “Advanced WCDMA a aa a radio network simulator,” in Personal, Indoor and Mobile Radio because of large volumes of data collected from network Communications, Osaka, Japan, September 12-15 1999, vol. 2, elements. The evolution towards 3G systems will further pp. 951–955. increase the amount of data. The operators’ task is to fil- [17] “Guidelines for evaluation of radio transmission technologies for IMT-2000,” 1997, Recommendation ITU-R M. 1225. ter the relevant information to a level on which it can be [18] “http://www.3gpp.org/,” . easily handled. Furthermore, the data set must include all [19] M. Siponen, J. Vesanto, O. Simula, and P. Vasara, “An ap- the essential parts needed to conclude the service quality. proach to automated interpretation of SOM,” in Advances in Self-Organizing Maps, N. Allinson, H. Yin, L. Allinson, and Operators can benefit from the introduced neural methods J. Slack, Eds. 2001, pp. 89–94, Springer. already today. The full gain and potential of the advanced [20] J. Laiho, A. Wacker, and T. Novosad, Eds., Radio Network analysis methods can be reached when multiple end-user Planning and Optimisation for UMTS, John Wiley & Sons Ltd., 2001. services are provided, and the quality perceived by the cus- [21] J. Laiho, K. Raivio, P. Lehtim¨ki, K. H¨t¨nen, and O. Sim- a a o tomers need to be monitored and optimized. ula, “Advanced analysis methods for 3G cellular networks,” 10
    • Tech. Rep. A65, Helsinki University of Technology, Laboratory of Computer and Information Science, 2002. [22] A.J. H¨glund, K. H¨t¨nen, and A.S. Sorvari, “A computer host- o a o based user anomaly detection system using the self-organizing map,” in Proceedings of the International Joint Conference on Neural Networks, 2000, vol. 5, pp. 411–416. Appendix 11
    • TABLE V Results from traditional analysis Cell Id 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 BS Txp 0.39 0.24 0.46 0.56 0.28 0.15 0.15 0.40 0.41 0.23 0.28 0.46 0.32 0.25 0.30 0.24 Traffic [W] Loading 0.84 0.72 0.89 0.89 0.70 0.65 0.65 0.88 0.89 0.89 0.89 0.89 0.89 0.75 0.89 0.73 Other to own cell 0.32 0.51 0.88 0.20 0.28 0.65 0.59 0.16 0.07 0.29 0.12 0.53 0.28 0.13 0.40 0.32 interf. ratio, i Own cell loading 0.63 0.47 0.47 0.74 0.55 0.39 0.41 0.76 0.83 0.69 0.79 0.58 0.69 0.66 0.64 0.55 ∆ loading caused by 24 34 47 17 22 40 37 14 7 23 11 35 22 12 29 24 other cells [%] Users UL 19 12 16 22 15 11 12 22 25 21 23 17 21 18 19 15 Loading/user 0.044 0.060 0.055 0.040 0.047 0.059 0.054 0.040 0.035 0.042 0.039 0.053 0.042 0.042 0.047 0.049 Own cell 0.033 0.040 0.029 0.033 0.036 0.035 0.034 0.035 0.033 0.033 0.034 0.034 0.033 0.037 0.033 0.037 interf./user Throughput 1216 768 1024 1408 960 704 768 1408 1600 1344 1472 1088 1344 1152 1216 960 UL [kbit/s] Cell Id 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 BS Txp 0.32 0.14 0.40 0.38 0.40 0.24 0.39 0.33 0.51 0.21 0.19 0.11 0.37 0.49 0.43 0.14 Traffic [W] Loading 0.89 0.70 0.89 0.74 0.76 0.85 0.88 0.89 0.87 0.72 0.81 0.60 0.86 0.87 0.88 0.52 Other to own cell 0.23 0.68 0.18 0.27 0.21 0.36 0.19 1.20 0.13 0.39 0.28 0.89 0.12 0.28 0.27 0.46 interf. ratio, i Own cell loading 0.73 0.42 0.76 0.58 0.63 0.62 0.74 0.40 0.77 0.52 0.63 0.32 0.77 0.68 0.69 0.36 ∆ loading caused by 18 40 15 21 18 27 16 55 12 28 22 47 11 22 21 31 other cells [%] Users UL 22 12 22 16 18 18 21 16 22 15 18 9 23 20 21 10 Loading/user 0.040 0.058 0.041 0.046 0.042 0.047 0.042 0.056 0.039 0.048 0.045 0.067 0.038 0.044 0.042 0.052 Own cell 0.033 0.035 0.034 0.037 0.035 0.035 0.035 0.025 0.035 0.035 0.035 0.035 0.033 0.034 0.033 0.036 interf./user Throughput 1408 768 1408 1024 1152 1152 1344 1024 1408 960 1152 576 1472 1280 1344 640 UL [kbit/s] Cell Id 33 34 35 36 37 38 39 40 41 42 43 44 45 46 BS Txp 0.38 0.22 0.23 0.28 0.34 0.36 0.22 0.17 0.35 0.12 0.16 0.39 0.29 0.30 Traffic [W] Loading 0.72 0.80 0.79 0.89 0.79 0.87 0.68 0.64 0.89 0.46 0.75 0.88 0.77 0.85 Other to own cell 0.31 0.26 0.20 0.28 0.53 0.37 0.31 0.51 0.23 0.85 0.63 0.15 0.30 0.21 interf. ratio, i Own cell loading 0.56 0.63 0.66 0.69 0.51 0.63 0.52 0.42 0.73 0.25 0.46 0.76 0.59 0.70 ∆ loading caused by 23 20 17 22 35 27 24 34 19 46 39 13 23 17 other cells [%] Users UL 14 18 19 20 15 18 12 12 22 6 13 22 16 21 Loading/user 0.052 0.044 0.042 0.044 0.052 0.048 0.056 0.053 0.041 0.076 0.057 0.040 0.048 0.040 Own cell 0.040 0.035 0.035 0.035 0.034 0.035 0.043 0.035 0.033 0.041 0.035 0.035 0.037 0.033 interf./user Throughput 896 1152 1216 1280 960 1152 768 768 1408 384 832 1408 1024 1344 UL [kbit/s] 12