Sie sind auf Seite 1von 5

Troubleshooting of 3G LTE Mobility Parameters

using Iterative Statistical Model Refinement


Moazzam Islam Tiwana, Berna Sayrac, Zwi Altman Tijani Chahed
Orange Labs, RESA/NET TELECOM et Management SudParis
38-40 Rue du General Leclerc 9 rue C. Fourier
92794 Issy-Les-Moulineaux Cedex 9, France 91011 Evry CEDEX, France
Email: {moazzamislam.tiwana,berna.sayrac,zwi.altman}@orange-ftgroup.com Email: tijani.chahed@it-sudparis.eu

Abstract—This paper presents a new troubleshooting method- Once the problem has been identified, there can be various
ology for 3G Long Term Evolution (LTE) networks based on ways which can be used to improve the performance of
a closed-form expression between Radio Resource Management the cells that exhibit a degraded performance. The steered
(RRM) and Key Performance Indicator (KPI) parameters, using
statistical learning. This methodology aims at locally optimising optimization technique is one such method. It has been used
the RRM parameters of the cells with poor performance in in [6] to identify the most coupled eNodeBs (eNBs), denoted
an iterative manner. The optimization engine uses the closed- as influence sets, with the problematic cell, based on the
form relationship to calculate the optimized RRM parameters for interference matrix approach. Indeed, knowledge of the inter-
these cells. The main advantage of this methodolgy is the small action between any couple of stations in terms of interference,
number of iterations required to achieve convergence and the
QoS objective. A troubleshooting application scenario involving macrodiversity and load difference allows one to accurately
mobility in LTE networks is considered. Numerical simulations identify sectors with poor performance on the one hand and
illustrate the benefits of our proposed scheme. to suggest corrective measures on the other hand [7].
Keywords: Statistical learning, linear regression, automated In this paper, we use a statistical approach based on data
troubleshooting, handover margin, LTE. measurement to construct closed-form expressions, which we
denote by the model, between the Radio Resource Manage-
I. I NTRODUCTION ment (RRM) parameters and a selected set of KPIs. Owing to
The release 8 of the Third Generation Partnership Project the linearity of the model [8], Linear Regression (LR) will be
(3GPP) specifications known as Long Term Evolution (LTE) used to construct it. Optimization techniques, such as linear
has been standardized in March 2009 [1]. LTE is a set of programming or combinatorial heuristics, can then be used
enhancements to the Universal Mobile Telecommunications to solve the optimization problem and determine the optimal
System (UMTS); the resulting architecture is called Evolved RRM parameters that best fit the KPI objectives.
Packet System (EPS) and comprises Evolved Universal Ter- The remainder of this paper is organized as follows. Section
restrial Radio Access Network (E-UTRAN) on the access side II describes the proposed troubleshooting scheme block dia-
and Evolved Packet Core (EPC) on the core side. gram. Section III presents the details of the system model for
With the growth of mobile wireless networks, the trou- mobility in 3G LTE. Section IV contains the troubleshooting
bleshooting costs of the network are increasing and the algorithm description. Section V shows the 3G LTE simulation
growing competition requires the efficient utilisation of the scenario and results. Section VI eventually concludes the
network resources and personnel. The three main tasks of paper.
troubleshooting are: detecting problems, identifying the cause
(diagnosis) and solving the problems [2]. Troubleshooting II. T ROUBLESHOOTING B LOCK D IAGRAM
may be manual, involving human experts that detect alarms It is assumed that the initial two steps of the troubleshooting
and monitor the statistical performance of the network, or process, namely fault detection and fault diagnosis, have
automated wherein these tasks are carried out by specialized already been done. Let’s suppose that the diagnosis system
modules in the Operation and Management Center (OMC). determined that the fault cause was a RRM parameter whose
Bayesian Network is an efficient way to represent the value has degraded the performance of the eNB or the first
relationship between causes and symptoms and can thus be tier neighbours of the eNB. Under these circumstances, the
advantageously used for the diagnosis in automated trou- proposed system uses a local optimization process in order
bleshooting. It has been studied in [3] and [4] for the GSM to find the RRM parameter value that solves the problems.
network. Therein, a diagnostic model is constructed from Thus, the troubleshooting process should be accomplished in
relevant data obtained from the network: Key Performance a minimum number of iterations.
Indicators (KPIs), alarms and configuration data. A Bayesian- The block diagram of the automated troubleshooting, also
based diagnosis has been studied for the UMTS network in termed healing, is shown in Figure 1.
[5] It is composed of four blocks:

978-1-4244-5661-1/09/$26.00 ©2009 IEEE


of an eNB and its first tier neighbouring eNBs have
a conflicting tendency with respect to change in RRM
parameter and also because the functional form of the
relationship between the RRM parameters and KPIs,
captured using LR, is linear. Hence, the weighted av-
erage gives the RRM parameter value which satisfies the
required QoS constraint. The weights are chosen with
respect to the relative importance of the KPIs under
consideration for troubleshooting. It is noted that the
data is noisy due to the random nature of the traffic
and of the radio channel, but also due to imprecisions in
measurements. The automated healing process is iterative.
At each new iteration, on average, the model precision
improves and is used by the optimization block to find a
Figure 1. Block diagram of the proposed troubleshooting scheme
better value for the RRM parameter. This approach allows
optimization convergence in few iterations and allows fast
reaction to problems in the network.
• Initialization block
The Initialization block provides the initial RRM param- III. S YSTEM M ODEL FOR M OBILITY
eter of the faulty eNB to the Network/Simulator block We now consider user mobility between neighbouring cells.
and to the Statistical Learning block. The 3G LTE mobility parameter considered in our study is
• Network / Simulator block Handover Margin (HM). It is the main parameter that governs
The Network/Simulator block represents the real network the Handover (HO) algorithm between the two eNBs and refers
or the network simulator. It captures, in the case of a to the minimum difference in power between the neighbouring
real network, or calculates, in the case of a network cell and the current one necessary for the mobile to make the
simulator, a set of KPIs of an eNB and of its neighbours HO.
for each new value of a RRM parameter introduced The LTE standard has adopted hard HO wherein a mobile
by the Initialization or the Optimization blocks. These terminal will not be simultaneously connected to the current
KPIs allow to assess the performance of the eNB and cell and the new cell [1]. Hard HO is implemented here using
of its neighbours. They are forwarded to the Statistical a similar algorithm to the one used in GSM systems. This
Learning block. handover is based on the comparison of the received signal
• Statistical Learning block strength from the serving cell and from the neighbouring cells.
The Statistical Learning block processes the data com- The corresponding algorithm is given below.
prising RRM-KPIs couples using the LR method to The handover algorithm: While in communication, the
extract closed-form functional relationships relating the mobile periodically measures the received power from its
KPIs to the RRM parameters. These functional relation- serving eNB and from the neighbouring eNBs. The mobile,
ships constitute the statistical model which we term the initially connected to eN Bi , triggers a HO to eN Bj if the
model. LR is a general method for estimating/describing following conditions are satisfied:
the association between a continuous outcome (depen- 1) The Power Budget Quantity (PBQ) is higher than the
dent) variable and one predictor (explanatory variable) HM: a mobile connected to eN Bi triggers a HO to an
[9]. It has been used in [8] to write the KPI (response adjacent eN Bj if:
variable) as a function of RRM parameter p (explanatory
P BQ = Pj − Pi ≥ HMij + Hysteresis (2)
variable) as follows:
where Pj is the received power from eN Bj expressed
KP Ij = fj (p) j = 1, 2, ....J (1) in dB and HMij is the outgoing HM of eN Bi towards
eN Bj . Hysteresis is a constant independent of the eNBs
where J is the number of KPIs of interest (the number and of the mobile stations, and is fixed here to 0.
of dependent variables) and fj (p) is linear with respect 2) The power being received from the target cell must be
to the regression coefficients. LR is chosen due to its higher than a given threshold.
simplicity and its validity in most engineering problems 3) There are sufficient number of resources/chunks in the
[10]. target cell. Otherwise, the mobile is bounced back to the
• Optimization block original cell.
The Optimization block calculates the optimal RRM
value using the current statistical model by taking the IV. T ROUBLESHOOTING A LGORITHM D ESCRIPTION
weighted average of RRM parameter values correspond- We now consider the adaptation of the troubleshooting
ing to the maximum allowable thresholds of the KPIs algorithm to the mobility parameter of the 3G LTE network
under consideration. This is possible because the KPIs by optimizing the HM parameter of the degraded eNBs.
Let eN Bc (c standing for central) or its first tier neigh- V. C ASE S TUDY
bouring eNBs T1 experience a degraded performance. We,
A. Simulation Scenario
again, assume that the cause of the degraded performance
has been diagnosed and that it is related to a bad mobility A LTE network comprising 19 eNBs in a dense urban
parameter HMi . We have: HMcj = HMkj = HMi , j ∈ T1 environment is depicted in Figure 2.
and k ∈ T2 , where T2 are the second tier neighbours of eN Bc .
Also, HMjc = HMjk = HMmax − HMi , where HMmax is
the maximum HM value.
Two KPIs are next used in the troubleshooting process:
Block Call Rate (BCR) and Drop Call Rate (DCR). Our
optimization objective is to determine the optimized HMi as
P
t∈{BCRc ,DCRc ,BCR,DCR} ωt HMt
HMi = P (3)
ωt

where BCR and DCR are the mean BCR and mean DCR
of T1 eNBs, ωt is the weight given to KPI t, denoted
as KP It and HMt is the HM value corresponding to the
maximum allowable threshold for KP It , determined using
the relationship:
KP It = ft (HMi ) (4)

where ft is obtained using LR on vector Pk consisting of k Figure 2. The network diagram of the simulated system
data points. The kth data point pk is given as:

pk = (HMi , BCRc , DCRc , BCR, DCR)k (5) We consider downlink transmissions. The simulation pa-
rameters are listed in Table I. A MATLAB LTE simulator
described in [11] has been used.
Now according to the requirements in [9], the initial number
of data points necessary to obtain regression coefficients must Parameters Settings
be two or greater. The initial number of data points, before System bandwidth 5MHz
applying LR and consequently optimization, is chosen as three. Cell layout 19 eNBs, single sector
Maximum eNB transmit power 32 dBm
One extra data point (more than the minimum required data Inter-site distance 1.5 to 2 KM
points) is chosen to compensate the effect of errors in the Subcarrier spacing 15 kHz
KPIs. The smaller the number of points, the more sensitive PRBs per eNB 15
Path loss L=128.1 + 37.6 log10 (R), R in kilo-
the regression is to errors. Furthermore, a badly estimated meters
initial model of ft can cause the optimization problem to Thermal noise density -173 dBm/Hz
be erroneous and the new points found by the optimization Shadowing standard deviation 6 dB
Traffic model FTP
module to get stuck in a non-optimal region. File size 5700 Kbits
The complete troubleshooting algorithm is as below: PRBs assigned per mobile 1 to 4 (First-come, first-serve basis)
Mobility of mobiles 90%
Initialization: Mobile speed 15 m/s
HMmax 12dB
1. Compute an initial set Pk of k data points by ap-
plying initially chosen HMi values one by one to net- Table I
T HE SYSTEM LEVEL SIMULATION PARAMETERS .
work/simulator and obtaining the corresponding KP It
values
Repeat until convergence:
The simulator performs correlated Monte Carlo snapshots
2. Compute the statistical model ft using Pk
with time steps of one second. The simulator performs corre-
3. Compute the new HMi value using (3)
lated snapshots to account for the time evolution of the traffic.
4. Apply HMi in the network/simulator and observe
At the end of each time step of one second, new mobile
corresponding KP It values. Compute new data points
positions are updated, HO events are processed, new users
using (5)
are admitted according to the conditions of access and some
5. Update Pk+1 : Pk+1 = Pk ∪ pk+1
other users leave the network (end their communications or are
6. k=k+1
dropped). The simulations are run for 3300 time steps, with a
End Repeat
fixed HM value, and the KPIs are averaged using the interval
between 500 and 3300 seconds to account for transient effects.
Reference Solution: An optimal default value for HM is HMi (dB) BCRc DCRc BCR DCR
chosen as 6dB for all eNBs in the network and will serve as

Phase I
6 0.040 0.016 0.82 0.041
the reference (default) solution. This reference solution will be 4 0.012 0.010 0.095 0.056
8 0.063 0.028 0.069 0.038
used as a starting point for the automated healing process. The
8.49 0.055 0.032 0.067 0.031
default HM value is determined by varying it uniformly from 8,60 0.062 0.023 0.065 0.031

Phase II
0.05 to 12 in steps of 0.15 for all the eNBs. For each HM value, 8.64 0.063 0.025 0.064 0.030
the network performance is assessed in terms of the mean Ping 8.56 0.063 0.033 0.059 0.028
8.41 0.044 0.021 0.064 0.029
Pong Rate (PPR), mean BCR and mean DCR. If these three 8.47 0.068 0.024 0.064 0.035
KPIs are aggregated as shown in Figure 3, we observe that the 8.47 0.068 0.024 0.066 0.037
global optimum HM value occurs around HM =6dB. Hence, Table II
the value of HM = 6dB is selected as the reference (default) C ONVERGENCE OF HMi DURING OPTIMIZATION PROCESS AND THE
CORRESPONDING KPI S OF TROUBLESHOOTING ALGORITHM
HM value for eNBs in the network.

Figure 3. The aggregated network KPIs (mean PPR, mean BCR and mean
DCR) as a function of uniform (default) HM
Figure 4. The KPIs in troublshooting as a function of HMi after first
optimization iteration
B. Troubleshooting Scenario I : Equal weights to KPIs
With reference to the network with 19 eNBs shown in Fig-
ure 2, eN B1 is the central eNB which we denoted by eN Bc .
Its 6 first tier eNBs T1 are given as eN B2 , eN B3 , eN B4 ,
eN B5 , eN B6 and eN B7 . While the 12 second tier neighbours
T2 are eN B8 , eN B9 , eN B10 , eN B11 , eN B12 , eN B13 ,
eN B14 , eN B15 , eN B16 , eN B17 , eN B18 and eN B19 .
We consider the case where  equal importance is given to
all the KPIs, i.e., ωt =1, t ∈ BCRc , DCRc , BCR, DCR .
Initially, the system is working with default HMi = 6dB.
Table II shows the convergence of these parameters start-
ing from the three initial points (Phase I), corresponding to
HMi = 6, 4, 8dB, and seven iterations of the optimization
algorithm (Phase II). Figures 4, 5 and 6 show the KPIs as
a function of HMi after the first, second and third iterations,
respectively. As explained earlier, the KPI curves are estimated
from corresponding KPI data points using LR.
We set the thresholds for BCR and DCR to 8% and 4%, Figure 5. The KPIs in troublshooting as a function of HMi after second
optimization iteration
respectively. We observe from Table II that obtained BCR
and DCR values corresponding to the initial working point
exceed these thresholds. process (when phase-II starts) we get the next HMi =8.49dB.
Now, we apply the troubleshooting algorithm and collect the This iteration corresponds to Figure 4.
initial data points corresponding to HMi =6,4 and 8dB (Phase- After the second and third optimization iterations (shown in
I). We can see that after the first iteration of the optimization Figures 5 and 6, respectively), the HMi values converges to
Figure 6. The KPIs in troublshooting as a function of HMi after third
optimization iteration Figure 7. Convergence of HMi in troubleshooting algorithm for different
KPI weights, ωt

8.64. The corresponding BCR improves from 8.2% to 6.4%


and DCR improves from 4.15% to 3%, which are within the This troubleshooting methodology can be used for other
maximum allowable thresholds of 8% and 4%, respectively. RRM parameters such as those related to interference man-
BCRc and DCRc of the central eNB are also within the agement.
thresholds. Eventually, we can see from Table II that even ACKNOWLEDGMENT
over the next 7 optimization iterations during Phase-II, there
The authors would like to thank Mr. Ridha Nasri who
is not much variation in the HMi value.
developed the Matlab simulator.
C. Troubleshooting Scenario II : Different weights to KPIs
R EFERENCES
The troubleshooting methodology is now verified using
 [1] 3GPP TS 36.300 V8.4.0: Evolved Universal Terrestrial Radio Access
different ωt values, t ∈ BCRc , DCRc , BCR, DCR as (E-UTRA) and Evolved Universal Terrestrial Radio Access Network (E-
given below: UTRAN); Overall description (2008-03)
[2] T. Halonen, J. Romero, J (eds). Melero: GSM, GPRS and EDGE Perfor-
1) ωBCRc =1, ωDCRc =1, ωBCR =1, ωDCR =1 mance. Wiley: Chichester, 2003
2) ωBCRc =2, ωDCRc =1, ωBCR =1, ωDCR =1 [3] R. Barco, V. Wille, L. Diez: System for automatic diagnosis in cellular
3) ωBCRc =3, ωDCRc =1, ωBCR =1, ωDCR =1 networks based on performance indicators, European Trans. Telecommu-
nications, vol. 16, no. 5, Oct. 2005, pp. 399-409.
4) ωBCRc =1, ωDCRc =1, ωBCR =1, ωDCR =2 [4] R. Barco, L. Nielsen, R. Guerrero, G. Hylander, S. Patel: Automated
5) ωBCRc =1, ωDCRc =1, ωBCR =1, ωDCR =3 troubleshooting of a mobile communication network using bayesian
For each set of wt weights, we apply our troubleshooting networks, in Proc. IEEE International Workshop on Mobile and Wireless
Communications Networks (MWCN02), Stockholm, Sweden, Sept. 2002,
algorithm to the network shown in Figure 2. We repeat the pp. 606-610.
scenario eight times and store each time the seven optimization [5] R.M. Khanafer, B. Solana, J. Triola, R. Barco, L. Moltsen, Z. Altman,
iterations as above. P. Lazaro: Automated Diagnosis for UMTS Networks Using Bayesian
Network Approach, in Proc. IEEE Transactions on Vehicular Technology,
We now take the variance between the eight HMi value July 2008, pp. 2451-2461.
calculated during each iteration as the measure of the con- [6] S. Ben Jamaa, Z. Altman, J.-M. Picard, A. Ortega: Steered Optimization
vergence. Figure 7 shows that after the third iteration, HMi Strategy for Automatic Cell Planning of UMTS Networks, IEEE 61st
Vehicular Technology Conference, 2005 Spring.
converges to an optimum fixed value. [7] S. Ben Jamaa, Z. Altman, A. Ortega, B. Fourestie: UMTS design strategies
based on indicator matrix approach, IEEE Int. Conf. on Communications,
VI. C ONCLUSION ICC 2004, Paris, France, June 20-24, 2004.
[8] M. I. Tiwana, B. Sayrac, Z. Altman: Statistical Learning for Automated
We presented in this paper a new troubleshooting algorithm RRM: Application to eUTRAN Mobility, IEEE Int. Conf. on Communi-
that models the functional relationship between KPIs and cations, ICC 2009, Dresden, Germany, June 14-18, 2009.
RRM parameters, using statistical learning. The optimization [9] D. C. Montgomery, E. A. Peck, G. G. Vining: Introduction to Linear
Regression Analysis, 4th Edition, John Wiley, 2006.
engine uses this model to obtain the optimized HM values [10] C. J. Leggetter, P. C. Woodland: Maximum likelihood linear regression
for the cells with degraded performance. A case study of for speaker adaptation of continuous density hidden Markov models,
eUTRAN HO algorithm has been considered and shows that Comput. Speech Lang., vol. 9, 1995, pp. 171-185.
[11] R. Nasri, Z. Altman: Handover Adaptation for Dynamic Load Balancing
our model improves after each iteration and that the RRM in 3GPP Long Term Evolution Systems., MoMM’2007-The Fifth Inter-
parameter converges to the optimal value in a limited number national Conference on Advances in Mobile Computing and Multimedia,
of iterations. 3-5 December 2007, pp. 145-154.

Das könnte Ihnen auch gefallen