Seediscussions,stats,andauthorprofilesforthispublicationat:https://www.researchgate.net/publication/281586621

AMachineLearningFrameworkforDetectionofSleepingCellsinLTENetwork

ConferencePaper·January2014

CITATION

1

READS

61

4authors,including:

AliImran

UniversityofOklahoma

45PUBLICATIONS214CITATIONS

SEEPROFILE

ArsalanSaeed

UniversityofSurrey

13PUBLICATIONS40CITATIONS

SEEPROFILE

AdnanAbu-Dayya

QatarMobilityInnovationsCenter

59PUBLICATIONS198CITATIONS

SEEPROFILE

Allin-textreferencesunderlinedinbluearelinkedtopublicationsonResearchGate,

lettingyouaccessandreadthemimmediately.

Availablefrom:ArsalanSaeed

Retrievedon:22July2016

A Machine Learning Framework for Detection ofSleeping Cells in LTE Network

Ahmed Zoha, Ali Imran, Adnan Abu-DayyaQMIC

Qatar Science and Technology ParkDoha, Qatar 210531

Email: ahmedz@qmic.com

Arslaan SaeedCCSR

University of SurreyGuildford, UK, GU2 9PL

Abstract—The rapid advancements in telecommunication sys-tems leads to growing data volume and high customer expecta-tions in terms of cost and quality of service. The changing dy-namics of radio network usage poses challenges for the operatorsin terms of optimizing and maximizing network efficiency whilereducing maintenance and operational expenditure. Automaticdetection of sleeping cell (SC) (i.e. a cell which is not providingnormal services to the users) in the network is one way oflowering maintenance cost and improving network performance.This paper presents an intelligent machine learning frameworkthat make use of minimize drive testing (MDT) functionality togather key performance indicators (KPI’s) of the LTE network.These measurements are further projected to a low-dimensionalembedding space and are used in conjunction with state of theart learning models to automate the SC detection process.

Keywords—Sleeping Cell, LTE, Anomaly Detection, Low-Dimensional Embedding

I. INTRODUCTION

The increased demands of high throughput, coverage andto guarantee quality of service (QoS) incur additional chal-lenges for the network operators. One such challenge is theoptimization and maintenance of network performance in acost-efficient manner. Self-organizing Network (SON) con-cepts [1] have emerged in the last years, with an aim tointroduce intelligent automation in the network. Automationof the network management process through SON conceptsas specified in 3GPP Release 10 standards, increases therobustness and efficiency of LTE network, while minimizingthe cost of operation. One of the main objective proposedin Self Healing functional block of SON is to automate thedetection of hardware (HW) or software (SW) failures toenhance the reliability of the system. Conversely, the clas-sical monitoring and fault detection methods are resourceconsuming and requires human labor. Manual drive testing isrequired in order to collect radio coverage measurements formonitoring the performance of the network. This is not onlytime and resource consuming but also have reachability issuesas the drive testing is limited to outdoor environments only.To address this limitation while reducing the expenditure ofdrive testing, minimize drive testing (MDT) functionality [1]is specified in SON by 3GPP.

MDT functionality offers a user equipment (UE) assisteddata gathering solution in which a UE is configured to reporteither periodic or event triggered measurements to the E-UTRAN NodeB (eNB). These measurements includes key per-

formance indicator (KPI’s) from the serving and neighboringcells, in addition to time and location information. This userperceived measurements can be leveraged to learn the systembehavior and accordingly identify any unexpected deviationsin an automated fashion. Upon detection of abnormality inthe network behavior, timely compensation actions can betriggered to resolve any issues.

Motivated by this, we propose a machine learning frame-work to automate the detection of specific network failure,named as Sleeping Cell (SC) in LTE network. SC is a situationwhen Base Station (eNB) failure is not recognized by theoperator as there is no alarm triggered. This situation mayoccur because of the HW, SW failures at eNB. Failure canfurther be classified into logical(i.e. failure of random accesschannel procedures) and hardware (i.e. break down of eNBcomponent). Such failures are responsible for a cell to becomedegraded, crippled or catatonic. In this study, we have lookedat bidirectional antenna gain failure as a HW fault that causesa cell to become catatonic. In a real world scenario, suchfaults might occur due to the malfunctioning of transmittingand receiving modules at eNB. We have adopted a model-driven approach to automatically identify such situations re-lying on the MDT measurements forwarded to eNB. In ourproposed framework a normal cell behavior is profiled intwo stages. Firstly, the UE reported KPI’s from a fault-freeoperating scenario are acquired and further embedded into alower dimensional subspace. In the next stage, the embeddedmeasurements are then used to train anomaly detection modelsfor identifying abnormal network behavior. In the SC detectionstage, the trained models leverage the intrinsic characteristicsof embedded representation to finely differentiate betweennormal and abnormal instances.

Typical method that addresses the problem of SC detectionare either based on quantitative models [2] which requiresdomain expert knowledge, or simply rely on performancedeviation metrics for detection [3]. In particular, the problemof detecting catatonic sleeping cells has been addressed usingNeighbor Cell List (NCL) reports [4]. Until recently, we see anincreased interest in applying machine learning methods suchas Bayesian Networks [5] and clustering algorithms [6] fordetecting cell outages. Although, all the previously mentionedstudies tries to address the problem of detecting abnormality inthe cell behavior, however our study differs in various aspect.Firstly, we in particularly focus on LTE cellular networks,simulated in accordance to existing 3GPP standards. Secondly,

we compare the performance of global and local anomalydetection models for developing a normal cell profile. To thebest of our knowledge, we are not aware of any comparativeevaluation of such methods for detecting SC in LTE networks.

The organization of the paper is as follows: we discuss ourproposed framework for SC detection in Section II. This alsoincludes a brief discussion on two state of the art anomalydetection models namely k-nearest neighbor and Local OutlierFactor based Anomaly detector. In Section III, we provide thedetails of our simulation setup and further report and analyzethe results. Finally, we conclude the paper in Section IV.

II. OUR SLEEPING CELL DETECTION FRAMEWORK

Our three step framework for SC detection in LTE networkshas been shown in Figure (1). It consists of acquiring MDTmeasurements, followed by detection and localization.

1) Measurement: In the measurement phase, the train-ing and test datasets are independently collectedfrom the simulation environment. The training datasetconsists of KPI’s from the normal fault-free operatingscenario. The validation and test datasets are acquiredfrom a scenario representing the faulty operation ofthe network. There are two possible approaches foracquiring the KPI’s from the simulation environmentas discussed in Section II-A.

2) Detection In the detection phase, we first projectthe training data into a low-dimensional space usingMulti-dimensional Scaling (MDS) method [7] as dis-cussed in Section II-B. The MDS embedding of theKPI vectors maximizes the variance in the datasetby increasing the distance between the dissimilarobservations and vice versa. Hence, the abnormalKPI’s samples lie far from the normal KPI’s in thelow-dimensional space which results in accurate pro-filing of normal cell behavior. Therefore, embeddedmeasurements are further used to train global and alocal anomaly detection algorithms namely k-NearestNeighbor and Local Outlier Factor, respectively. Abrief overview of the algorithms are provided inSection II-C.

3) Localization In the localization phase, the locationof the sleeping cell is identified based on the classifi-cation performed by employed algorithms as furtherexplained in Section II-D.

A. MDT Measurements for SC Detection

Minimization of Drive Testing (MDT) use case for SONwere introduced by NGMN during 2008. The idea is tominimize the cost of manual drive testing by enabling theUE to report the coverage measurements. In LTE Release10 the MDT measurement and reporting schemes have beendefined. The MDT measurement functionality allow operatorto collect measurements either periodically or event based. Asubset of events which generates a MDT report is listed inTable I. The measurement reports consists of cell identificationand radio-measurement data as tabulated in Table II. In thisstudy, we have employed A2 and A3 event triggered MDTmeasurements for SC detection. The simulation parameters arelisted in Table III.

Fig. 1. An three step framework of Sleeping Cell Detection and Localizationin LTE Network

Events Description

A2 Serving becomes worst than a thresholdA3 Neighbor becomes offset better than serv-

ingRLF Link quality falls below a certain thresh-

old and an interruption in service happens.

TABLE I. EVENTS EMPLOYED FOR SC DETECTION

B. Low-Dimensional Embedding

After acquiring MDT measurement reports, the data isfurther preprocessed by cleaning and scaling it. The last fourmeasurements listed in Table II are combined into singleaugmented feature vector as shown in Equation 1

V = {RSRPS , RSRPN1, RSRPN2, ...RSRPN3,

...RSRQS , RSRQN1...RSRQN3, CQI} (1)

where S and N stands for serving and neighboring cells,respectively. The 9-dimensional feature vector is further em-bedded to only two dimensions in the Euclidean space usingMDS method. In context of SON, the dimensionality reductionis a crucial step as a high-dimensional KPI database poseschallenges for network engineers as well for experts. The realnetwork is complex and dynamic in nature, and it is often notpossible to identify few KPIs that really capture the behaviorof the system. On the other hand projecting the data onto fewerdimensions of maximum variance uncovers the true structurewhich ultimately aids the cell profiling process. Moreover, lesscomputational effort is required which consequently leads tolow detection delays.

C. Sleeping Cell Detection

The embedded KPI representation is then used togetherwith state of the art anomaly detection algorithms, which aretrained to reject any abnormal test observations which donot conform to the normal network behavior. In one class

Features Description

Time and Location Time stamp and longitude and latitudeinformation

Serving Cell info Cell Global Identification (CGI)RSRP Reference Signal Received Power in dBmRSRQ Reference Signal Received Quality in dB

Neighboring Cell Information Three Strongest intra-LTE RSRP, RSRQinformation

CQI Serving Cell Channel Quality Indicator

TABLE II. STRUCTURE OF MDT MEASUREMENTS

classification framework, normal observations are used to trainthe anomaly detection models so that test instances can beclassified as either belonging to a normal class or vice versaby computing a threshold ’θ’ based on a certain dissimilaritymeasure ’D’ between the two:

f(xi) =

{Normal, if D(xi, Dtrain) ≤ θAbnormal, if D(xi, Dtrain) ≥ θ

(2)

where Dtrain is a subset of dataset D. The dissimilarity criteriaused by k-Nearest Neighbor and Local Outlier Factor basedmodels are briefly summarized as follows:

1) k-Nearest Neighbor based Anomaly Detector (k-NNAD):Let xi be the test instance, and k be the kth neighbor in thetraining set Dtrain. To label xi as normal or abnormal, theKNNAD computes a DKNNAD based on Equation 3

DKNNAD(xi, k, D) =1

Ntr

Ntr∑i=1

I(dt ≤ di) (3)

The Ntr =| Dtrain |, and dt is the distance of xi from itskth nearest neighbor in the training set D, whereas di is thedistance between i and its kth nearest training object in Dtrain.Equation 3 represents a global anomaly detection score asproposed in [8], which is compared against θ to mark the testinstance as anomalous or otherwise.

2) Local Outlier Factor based Anomaly Detector (LOFAD):The LOFAD [9] tries to compare the local density ρ ofthe object to that of its k neighbors. It constructs a localneighborhood of an instance xi and defines its distance to kthnearest neighbor NN(xi, k):

db(xi, k) = d(xi, NN(xi, k)) (4)

The db(xi, k) is used to construct a neighborhood N (xi, k)by including all those points in the neighborhood fulfilling thefollowing criteria: d(xi, xj) ≤ db(xi, k). Formally, reachabilitydistance dr is defined to estimate the ρ(xi, k) as follows:

dr(xi, k) = max{db(xj , k), d(xj , xi)} (5)

and ρ can be defined as

ρ(xi, k) =| N (xi, k) |∑

xj∈N (xi,k)dr(xi, xj , k)

(6)

The dr(xi, xj , k) ensures that instances that lie farther awayfrom xi have lesser impact on ρ(xi, k). Finally the D can becalculated by comparing the ρ of xi to its N (xi, k), formallydefined as:

DLOFAD(xi, k,Dtrain) =∑xj∈N (xi,k)

ρ(xj ,k)ρ(xi,k)

| N (xi, k) |(7)

The DLOFAD will be 1 if xi lie inside a cluster or else itreceives a higher value which can be compared against θ tolabel it as an anomaly.

1000500 2000

1000

2000

500

US

ER

PO

SIT

ION

Y

User POSITION X

ANOMALIES NORMAL

Fig. 2. SC localization based on UE reported position information

D. Localization of SC

After the employed detection algorithms classify the testmeasurements, the UE reported position information is utilizedto perform localization of SC. As shown in Table II, the MDTmeasurement reports also contains time and location informa-tion. These two are not used at the SC detection stage as shownin Equation 1. However, based on the coordinate information,the classified measurements can be further mapped to networktopology. As a result, the cell which corresponds to the highestnumber of abnormal measurements can be easily identified asshown in Figure 2.

III. PERFORMANCE EVALUATION

A. Simulation Setup

A full dynamic system tool is employed to simulate theLTE network based on 3GPP specifications. Two referencescenarios are generated to collected normal and abnormal mea-surements. In a SC scenario, the antenna gain of a problematiccell is degraded to −50 dBi as indicated in Table III.

B. Analysis of Detection Performance

To train and evaluate the accuracy of our detection models,as pointed out earlier the training observations are acquiredfrom a normal operating scenario. However, the test dataset isdivided into (30%) validation and (70%) test sets. We combinethe training and validation sets and has applied 10-fold cross-validation to optimize the parameters (i.e. k = 1, 2, ...10) fork-NNAD and LOFAD, (α = 0.1, ...1.0). A standard evaluationmethod, namely Area Under the Curve (AUC) measure as-sociated with Receiver receiver operating characteristic (ROC)curve [10] is adopted to evaluate the performance of our target

Parameter Values

Cellular Layout Macro World 27 sitesSectors 3

User Distribution UniformPath Loss L[dB] = 128.1 + 37.6log10(R)

Antenna Gain (Normal Scenario) 15 dBiAntenna Gain (SC Scenario) −50 dBi

Slow Fading Std 8 dBSimulation Length 100s (1 time step = 71.43µs)

BS Tx Power 46 dBmNetwork Synchronization Asynchronous

HARQ Asynchronous, 8 SAW channels, Maxi-mum Retransmission = 3

Cell Selection Criteria Strongest RSRP defines the target cellLoad 20 users/cell

Traffic Model Infinite BufferHO Margin 3dB

HO trigger time 256 ms

TABLE III. SIMULATION PARAMETERS

Model Approach AUC scorek-NNAD Global 85± 2.3LOFAD Local 76 ± 3.8

TABLE IV. PERFORMANCE OF TARGET ANOMALY DETECTIONMODELS FOR SC DETECTION

models. We report the performance of each model in Table IV,which is based on optimal parameter setting of the algorithms.The main difference between the employed algorithms are theirmethod to compute the D, which determines their approach aslocal or global. Local anomalies are localized to a small spatialregion (i.e. local density) or a neighborhood whereas globalanomalies are bounded by entire dataset (i.e. global densities).It has been observed that the KPI’s from the normal scenariowhen projected to an Euclidean space are grouped into denseneighborhood, whereas the measurements obtain from the ab-normal or sleeping cell scenario lie far from this neighborhood.This is because MDS tries to maximize the variance betweenthe data points and dissimilar points are projected far fromeach other. Thus, the abnormal measurements acts as globalanomalies. This explain the reason that k-NNAD being a globalanomaly detector has outperformed LOFAD, and achieve highdetection accuracy. Moreover, we found out that an averagedifference of 35 dBm is observed between the RSRP valuesof a normal and a sleeping cell.

The density value of abnormal measurements is low,however some of them do overlap with micro clusters ofnormal measurements. These micro clusters lie at the borderof dense neighborhood and LOFAD wrongly treat them aslocal outliers, and thus achieve low detection scores. Theresults clearly indicate that global models are well suited fordetecting abnormal behavior of the network in the embeddedspace in comparison to local models. MDS embedding ofthe KPI measurements offers a clear advantage, as regularoccurring observations form a clear grouping in the lower-dimensional space. This aids the anomaly detection algorithmsin isolating normal network measurements from the abnormalones with high accuracy. Once, we automatically classify themeasurements as normal or abnormal , they can further bemapped to network to localize the position of sleeping cell asdiscussed in Section II-D.

IV. CONCLUSION AND FUTURE WORK

In this paper, we present a machine learning frameworkfor automating the sleeping cell detection process in an LTEnetwork. Our proposed approach first acquire key performancemeasurements from the fault-free operating network. The datais further embedded into a lower-dimensional space. Theembedded measurements are used to build a normal profile ofthe network by training the k-NNAD and LOFAD detectionmodels. The models are later used to automatically detectabnormal measurements from the test scenario. The detectionaccuracy of k-NNAD is found to be much higher than LOFADdue to its global detection approach. Finally the UE reportedcoordinate information is employed to localize the position ofsleeping cell. This is a preliminary study and in future we aimto extend our research for heterogeneous networks.

ACKNOWLEDGMENT

This work was made possible by NPRP grant No. 5-1047-2437 from the Qatar National Research Fund (a member ofThe Qatar Foundation). The statements made herein are solelythe responsibility of the authors.

REFERENCES

[1] S. Hamalainen, H. Sanneck, C. Sartori et al., LTE Self-OrganisingNetworks (SON): Network Management Automation for OperationalEfficiency. John Wiley & Sons, 2012.

[2] R. Barco, V. Wille, and L. Dıez, “System for automated diagnosis incellular networks based on performance indicators,” European Trans-actions on Telecommunications, vol. 16, no. 5, pp. 399–409, 2005.

[3] B. Cheung, S. G. Fishkin, G. N. Kumar, and S. A. Rao, “Method ofmonitoring wireless network performance,” Sep. 21 2004, uS PatentApp. 10/946,255.

[4] C. M. Mueller, M. Kaschub, C. Blankenhorn, and S. Wanke, “A celloutage detection algorithm using neighbor cell list reports,” in Self-Organizing Systems. Springer, 2008, pp. 218–229.

[5] R. M. Khanafer, B. Solana, J. Triola, R. Barco, L. Moltsen, Z. Alt-man, and P. Lazaro, “Automated diagnosis for umts networks usingbayesian network approach,” IEEE Transactions on Vehicular Technol-ogy, vol. 57, no. 4, pp. 2451–2461, 2008.

[6] Y. Ma, M. Peng, W. Xue, and X. Ji, “A dynamic affinity propagationclustering algorithm for cell outage detection in self-healing networks,”in Proceedings of IEEE Wireless Communications and NetworkingConference (WCNC). IEEE, 2013, pp. 2266–2270.

[7] T. F. Cox and M. A. Cox, Multidimensional scaling. CRC Press, 2010.[8] M. Zhao and V. Saligrama, “Outlier detection via localized p-value

estimation,” in In Proceedings of 47th Annual Allerton Conference onCommunication, Control, and Computing, 2009. IEEE, 2009, pp.1482–1489.

[9] M. M. Breunig, H.-P. Kriegel, R. T. Ng, and J. Sander, “Lof: identifyingdensity-based local outliers,” in ACM Sigmod Record, vol. 29, no. 2.ACM, 2000, pp. 93–104.

[10] A. P. Bradley, “The use of the area under the roc curve in the evaluationof machine learning algorithms,” Pattern recognition, vol. 30, no. 7, pp.1145–1159, 1997.