Self-Diagnosing Low Coverage and High Interference in 3G/4G Radio

Access Networks based on Automatic RF Measurement Extraction

M. Sousa

2,3

, A. Martins

1,3

and P. Vieira

1,2

Instituto de Telecomunicac¸

oes (IT), Lisbon, Portugal

Instituto Superior de Engenharia de Lisboa (ISEL), ADEETC, Lisbon, Portugal

CELFINET, Consultoria em Telecomunicac¸

oes Lda., Lisbon, Portugal

Keywords:

Wireless Communications, SON, Self-Diagnosis, Coverage Detection, Interference Control.

Abstract:

This paper presents a new approach for automatic detection of low coverage and high interference scenarios

(overshooting and pilot pollution) in Universal Mobile Telecommunications System (UMTS) /Long Term Evo-

lution (LTE) networks. These algorithms, based on periodically extracted Drive Test (DT) measurements (or

network trace information), identify the problematic cluster locations and compute harshness metrics, at clus-

ter and cell level, quantifying the extent of the problem. Future work is in motion by adding self-optimization

capabilities to the algorithms, which will automatically suggest physical and parameter optimization actions,

based on the already developed harshness metrics. The proposed algorithms were validated for a live network

urban scenario. 830 3

Generation (3G) cells were self-diagnosed and performance metrics were computed.

The most negative detected behaviors regards high interference control and not coverage veriﬁcation.

1 INTRODUCTION

The increasing network complexity, in terms of num-

ber of monitored parameters and parallel operation of

Generation (2G), 3G and 4

Generation (4G), is

increasing dramatically, besides the ever growing traf-

ﬁc volume and service diversity. In the third quar-

ter of 2015, an average monthly data trafﬁc of 4,700

petaBytes was registered, with an increase of 65%

compared with the third quarter of 2014 (Ericsson,

2015). This increases the network Operating Expense

(OpEx), forcing mobile operators to pursue strategies

for reducing it. Self-Organizing Networks (SON) al-

gorithms have been seen as the solution, and a way to

automatically operate the current and beyond mobile

networks.

This paper focus on the automatic-diagnosing al-

gorithms of low coverage and high interference sce-

narios, used to trigger optimization processes in self-

optimizing functions (within SON). Problem detec-

tion is based on periodically extracted DT measure-

ments or geo-positioned network traces (Vieira et al.,

2014). The DT data provides measurements of re-

ceived signal strength and quality for the different pi-

lot or reference signals that reach a certain location.

Moreover, speciﬁc ﬁltering is applied to identify the

data that denotes either, coverage issues or interfer-

ence problems, such as overshooting or pilot pollu-

tion.

SON and self-diagnosing is a hot research topic

and recent work has been done leading to the cur-

rent research stage (Duarte et al., 2015),(Sousa et al.,

2015), (Sallent et al., 2011). The paper contribu-

tion is incremental to previous works. Firstly, the

used cell’s service area approximation based on prop-

agation modelling is more accurate when compared

with existent research. Secondly, a new cluster par-

tition approach using the auto-correlation distances

for shadow fading (Kysti et al., 2007), to limit clus-

ter size, results in a more coherent data analysis. Fi-

nally, this work address the DT minimization target of

Generation Partnership Project (3GPP) by allow-

ing trace based inputs.

The aim of the present work is to be the cor-

nerstone in a fully self-optimization algorithm. This

work corresponds to the network data analysis phase,

which will report the detected areas with low perfor-

mance issues and all relevant data for the further opti-

mization algorithm. Additionally, harshness metrics

are already computed, accessing the cell’s detected

problems.

This paper is organized as follows. Section 2 de-

ﬁnes the scope embedding the present work. In sec-

tion 3 the concept of DT reliability is presented. Sec-

Sousa, M., Martins, A. and Vieira, P.

Self-Diagnosing Low Coverage and High Interference in 3G/4G Radio Access Networks based on Automatic RF Measurement Extraction.

DOI: 10.5220/0005958300310039

In Proceedings of the 13th International Joint Conference on e-Business and Telecommunications (ICETE 2016) - Volume 6: WINSYS, pages 31-39

ISBN: 978-989-758-196-0

tion 4 presents the detailed self-diagnosing process

for each algorithm. The results are shown in section 5

and ﬁnally, in section 6, conclusions are drawn.

2 SELF-OPTIMIZATION

Self-Optimization, aims to maintain network quality

and performance with a minimum of manual inter-

vention. It monitors and analyses, either Key Per-

formance Indicator (KPI)s, DT measurements, traces

or other sources of information, triggering automated

actions in the considered network elements. The

self-optimization process for a network under perfor-

mance enhancement is shown in Figure 1. This work

researches the Self-Diagnosis process block, hence

future work will complete the self-optimization func-

tion by implementing the optimization and Opera-

tional Support System (OSS) script blocks.

Figure 1: Global self-optimization ﬂowchart.

The self-diagnosis module uses either DT mea-

surements or trace data as in Figure 2. Thus, with an

introduction of a reliability index R for the input data,

it accomplishes not only a control mechanism for the

self-optimization process, but also increases quality

in the self-diagnosis output. Only when the available

data, is signiﬁcant to access the cell’s performance,

reﬂected by a value of R higher than a certain thresh-

old, min value, the self-diagnosis is executed. Oth-

erwise, the available information is not sufﬁcient to

retain any strong conclusion. Furthermore, regarding

the trace reliability block, it is under development.

The self-diagnosis module is composed by three

different algorithms, coverage holes, overshooting

and pilot pollution, see Figure 2. For all of them, the

concerning data in the network optimization scenarios

is grouped in clusters. Moreover, a harshness index

cell

is calculated accessing the problem’s severity in

the cell.

Figure 2: Self-diagnosis ﬂowchart.

WINSYS 2016 - International Conference on Wireless Networks and Mobile Systems

3 DRIVE-TEST DATA

CLASSIFICATION

There is an undeniable correlation in the assertiveness

of the DT based algorithm results and the complete-

ness of a DT campaign. As well, the harshness met-

rics will be as accurate as the DT is complete or sig-

niﬁcant. This highlights the importance of a quality

metric associated with the DT used in any algorithm.

Hence, an assessment of the DT quality is executed

before the use of DT data itself, giving origin to a DT

reliability index R (Sousa et al., 2015). It evaluates the

data collected in a cell’s service area, taking into ac-

count the percentage of measurements collected and

its spatial distribution.

3.1 DT Spatial Distribution

The quadrant method is a statistical spatial analysis

method used as a mean to test a population point pat-

tern (random, clustered and dispersed). In this con-

text, the method is not used to classify the DT mea-

surements on its geographical distribution pattern, but

as a relative measurement of the dispersion level. This

allows to quantify how well distributed are the DT

measurements in the cell’s service area.

For that purpose, the service area is divided into

quadrants and the number of DT measurements on

each accounted. Figure 3 shows a cell’s service area

divided into equal areas, before applying the method.

Figure 3: Cell’s geographical area.

Considering the quadrant method data,

[Y (A

)]

= [x

, ..., x

] (1)

where x

is the number of measurements in the m

quadrant, for all A

quadrants covering the service

area P. Using this data set, a parameter that quantiﬁes

the DT data pattern is calculated.

The parameter Variance to Mean Ratio (VMR)

allows to identify the distribution pattern of a data set

and it is given by,

V RM =

¯x

(2)

where s

is the variance of the number of measure-

ments and ¯x the average measurement number in each

quadrant. When the VMR value is higher than one,

the pattern is clustered. Below the unit is dispersed

and if it’s equal to one, the pattern is random. Due

to the direct correlation between DT measurements

and the own existence of roads, the pattern can’t be

random and hardly will ever be dispersed. So, the rel-

evant information is how much clustered might be.

This knowledge is reﬂected in the Dispersion In-

dex D

for a cell i. Using Equation (2) over the data

set given by Equation (1) the DT data dispersion in

the cell’s service area is calculated (V RM). The Dis-

persion Index D

, results from the normalization of

the DT data V MR, given by,

= 1 −

V RM

max

(3)

where V RM

max

is calculated using Equation 2 in the

case of all measurements being in one single area A,

thus resulting in a relative value of the data distribu-

tion.

3.2 Road Filling Ratio for DT

Measurements

The aim is to calculate the road ﬁlling ratio, P

, which

represents the percentage of road/street covered in the

service area of cell i. In order to proceed, the road net-

work information is fetched using an Application Pro-

gramming Interface (API). The obtained points are

linearly interpolated, for resolution purposes, result-

ing as shown in Figure 4.

An important piece of information, is the treat-

ment that the original geospatial positioned measure-

ments, enroll. They are subsequently aggregated on

geographical areas of ten by ten meters, called bins.

This term refers to several measurements from dif-

ferent cells aggregated in the same area, and will be

mentioned throughout this paper. To refer to a sin-

gle cell measurement contained in a bin, we refer as

a cell sample. Bearing in mind this information, the

same procedure must be applied to the retrieved road

points. Only then P

is calculates as,

Max

(4)

Self-Diagnosing Low Coverage and High Interference in 3G/4G Radio Access Networks based on Automatic RF Measurement Extraction

Figure 4: Street data information.

where M is the number of collected bins and Max

the equivalent of bins for all the road/street extension

in the service area of a cell i.

3.3 Calculating the DT Reliability Index

The DT Reliability Index R

, gets values in the zero to

one range, and it merges the previous two metrics, in

the following way,

= α

+ α

(5)

where α

is the weight of the dispersion index D

for cell i and α

is the weight of the percentage of

road covered P

, in cell i. Regarding the weights of

both factors (α

and α

) they resulted from empirical

knowledge, through a statistical analyze of ﬁfty DT

classiﬁcation inquiries, performed to Celﬁnet’s expe-

rienced radio engineers (Sousa et al., 2015).

4 THE SELF-DIAGNOSIS

PROCESS

The underlying process for the identiﬁcation of either

coverage holes, overshooting or pilot pollution sce-

narios has a common ground. The goal is always, re-

garding the speciﬁc algorithm, to diagnose prevalent

under performance situations in the form of clusters.

Furthermore, to attribute a harshness metric based on

a statistical analysis to each cluster, and a harshness

level of the cell, due to all clusters found, as presented

in Figure 5.

4.1 Self-Diagnosis Modules

As stated before, in this work three detection algo-

rithms are presented. This subsection characterizes

Figure 5: General detection ﬂowchart.

each algorithm, in terms of the data identiﬁcation pro-

cess.

4.1.1 Detecting Coverage Holes

A coverage hole is deﬁned as an area where the pi-

lot (or reference) signal power is in between the low-

est network access threshold and the lowest value re-

quired for assigning full coverage. Users in this area

tend to suffer poor voice or data user experience and

possibly dropped calls or high latency.

From the cell’s footprint, a coverage hole sample

must accomplish two conditions. Firstly, it must be

best server or within a small power interval to the best

server of the respective bin. To these, we will refer

as serving samples. Note that best server is the mea-

surement that corresponds to the best cell, and that

should serve the mobiles at that point. Secondly, the

best server measurement from the respective serving

sample bin, must be below the coverage threshold,

power

< T hr

(6)

where BS

power

is the best server value and T hr

the minimum serving power value, which is tecnhol-

ogy (3G 4G) and use case dependent.

4.1.2 Detecting Overshooting

An overshooting situation occurs when the cell’s cov-

erage reaches beyond what is planned. Generally, oc-

WINSYS 2016 - International Conference on Wireless Networks and Mobile Systems

curs as an “island” of coverage in another cell’s ser-

vice area. Overshooted areas may also suffer from

call drops and bad quality of experience.

From the cell’s DT measurements, the ones con-

sidered in overshooting, must be located beyond the

cell’s service area. Then, to be considered overshoot-

ing, it must comply with the following conditions, re-

garding the power value measured,

(

pwr

> T hr

, if CS is best

pwr

> Best

pwr

− ∆

pwr

, if CS is not best

(7)

where CS

pwr

is the cell sample power value, T hr

is the minimum power value to be considered over-

shooting, Best

pwr

is the power value of the bin best

server cell and ∆

pwr

is used to deﬁne a power range to

consider overshooting. Two conditions are presented,

for the case when the corresponding sample of the cell

in analysis is the best server cell and for the opposite

case.

Also, the quality of the cell’s measurements is

evaluated, using the same conditions as Equation 7

but regarding the quality measurments,

(

qual

> QT hr

, if CS is best

qual

> Best

qual

− ∆

qual

, if CS is not best

(8)

Not all overshooting situations are necessarily

damaging to the network or non-intended. Even

though effectively being far from the normal service

area, it might happen that, due to terrain proﬁle, it

still might be the cell in best condition to serve in

that area. In that sense, in case of the majority of the

overshooting bins, containing the best server as the

analyzed cell and a delta value of power and quality

to the second best server, these will also be marked.

They will still be identiﬁed as overshooters and con-

tinue the process, but containing an observation that

optimizing this overshooting area might reduce the

overall network performance.

4.1.3 Detecting Pilot Pollution

Pilot pollution remarks a scenario where too many pi-

lots (or reference signals in the case of LTE) are re-

ceived in one area. Besides the excess of pilots, it

lacks a dominant one. These areas are highly inter-

fered, resulting in a poorer user experience.

The cell’s measurements inducing pilot pol-

lution, are in a serving ranking bellow the

ideal maximum number of cells serving in that

area. To the cell’s footprint where the previ-

ous is veriﬁed, the following conditions, must

be complied to be in a pilot pollution situation,

(

pwr

> Best

pwr

− ∆

pwr

qual

> Best

qual

− ∆

qual

(9)

where CS

pwr

is the cell sample power value, Best

pwr

is the power value of the bin best server cell and ∆

pwr

is used to deﬁne a pilot pollution power range. The

second condition is equal to the ﬁrst, but regarding

quality measurements.

4.2 Cluster Partitioning

The radio mobile channel is uncertain due mainly to

the effects of fading and multipath. So, the values

caught on DT measurements, which are reported at

one single instance of time, may not correspond to

the average behavior of the radio channel in that point.

In order to mitigate this variability, the detection pro-

cess is executed at the cluster level and not at bin

level. This enables to detect prevailing under perfor-

mance areas and not simply variations, normal and

non-correlated with network issues. Therefore, us-

ing the auto-correlation distances for shadow fading

(Kysti et al., 2007), to limit cluster size, this gives

more assertiveness in the detected results.

Regarding the cluster division process itself, it

is accomplished using a dendrogram structure (Izen-

man, 2008). It is a tree diagram that, in this applica-

tion, translates the distance relation between all DT

measurements detected, as can be seen in Figure 6.

The tree diagram building process, aggregates succes-

Figure 6: Dendrogram tree structure.

sively the closest bin/cluster pair until all of the bins

form a unique cluster. At each aggregation, is con-

structed a different cluster division possibility.

The next step is to deﬁne which cluster arrange-

ment is best. Several algorithms accomplish this pur-

pose, and throw different metrics and approaches. But

overall what they evaluate is how well concentrated

are the points in clusters, when the distance similarity

(Zheng and Xue, 2009) is high. The approach used

was the silhouette method (Witten and Frank, 2005).

It provides a metric s(i),

b(i) −a(i)

max{a(i), b(i)}

(10)

Self-Diagnosing Low Coverage and High Interference in 3G/4G Radio Access Networks based on Automatic RF Measurement Extraction

where a(i) is the average dissimilarity to the other

points of the cluster, b(i) the lowest average dissimi-

larity of i to any other cluster. It evaluates how well

any given point lies within its cluster. Thus, an s(i)

close to one means that the bin is appropriately clus-

tered. This approach is only applied to the cluster di-

vision possibilities that respect the maximum cluster

size due to the correlation distance.

4.3 Cluster Statistical Analysis

A statistical analysis is conducted to each detected

cluster. The purpose is to classify the harshness

(severity of the problem) and rank by it. The result

is a harshness index, H

cluster

, per cluster, given by,

cluster

∑

i=1

U(c(i))

∑

i=1

, 0 ≤ H

cluster

≤ 1 (11)

where β

is the weight for condition i and the U(c(i)),

U(c(i)) =

(

1, if c(i) is full ﬁeld

0, otherwise

(12)

, are the evaluated conditions. Once again these con-

ditions are dependent on the behavior that is being

analyzed.

The coverage hole algorithm calculates it’s H

cluster

index (11) using the following conditions:

C(1) : Prob(Power

≤ Power

cov

) ≥ T hr

(13)

C(2) : Prob(Power

≤ Power

cov1

) ≥ T hr

(14)

where Power

is the cell’s power measurement

value, Power

cov

is a coverage threshold and T hr

is the minimum percentage of data samples that are

below the coverage value, so that the condition is ful-

ﬁlled. The variable Power

is the power measurement

value of a neighbor cell, who also serves that area.

The Power

cov1

is just another power value to be com-

pared with. These conditions allow to classify the

harshness of a coverage hole in two forms. The con-

dition from Equation (13) concerns only to the power

value of the source cell. Using different Power

cov

val-

ues and different percentages T hr

Power

a cluster will

be classiﬁed more or less harsh. The second condi-

tion, in Equation (14), evaluates the existence of other

fallback cell in the cluster.

The coverage hole harshness index H

cell

repre-

sents a percentage of clusters in coverage hole versus

the clusters of the cell footprint.This metric might be

devious, especially if the DT data are low in number.

In that case, the metric will exceed the true percent-

age value. That’s why the DT reliability index R is so

important in terms of interpreting the results.

The overshooting cluster harshness evaluation

proceeds with the following conditions:

C(1) : Prob(Power

≥ Power

) ≥ T hr

(15)

C(2) : Prob(Power

≥ Power

OS1

) ≥ T hr

(16)

C(3) : Prob(Qual

deg

≥ Qual

degOS

) ≥ T hr

(17)

where Power

and Power

OS1

are different power val-

ues to be compared with. The conditions from Equa-

tion (15) and Equation (16) are exactly the same,

but using different Power

values, evaluating the

power level of the overshooting, allowing more res-

olution in distinguishing overshooting clusters in se-

vere terms. The condition from Equation 17, evalu-

ates the quality degradation caused by the existence of

an overshooting cell (Sanchez-Gonzalez et al., 2013),

in which Qual

deg

is the quality degradation caused

and Qual

degOS

is a quality degradation threshold.

The overshooting harshness index H

cell

repre-

sents the average percentage of overshooting clusters

against the victim cells footprint, divided also in clus-

ters. Once again, the importance of the DT reliability

index must be highlighted.

Concerning the pilot pollution algorithm, the con-

ditions to evaluate the harshness level are the follow-

ing:

C(1) : Prob(Power

best

≥ Power

) ≥ T hr

(18)

C(2) : Prob(Qual

deg

≥ Qual

degPP

) ≥ T hr

(19)

where Power

best

is the signal power of the best server

measurement and Power

is a threshold value to be

compared with. The condition presented in Equation

(18), reﬂects that, a pilot pollution scenario is as dam-

aging to the network as higher is the power of the

correspondent best servers. The second condition (in

Equation (19)), as seen in the overshooting, gauge the

quality degradation that the source cell induces on the

best servers.

In case of the pilot pollution algorithm, being a

scenario where cells affect another cell performance,

the approach for the pilot pollution harshness index

cell

is the same as described in the overshooting

module.

5 APPLYING THE ALGORITHM

TO A LIVE NETWORK

The developed algorithms were applied in a live net-

work, and for an urban scenario. 830 3G cells oper-

ating at different frequencies were self-diagnosed and

performance metrics were computed. Extensive DT

data corresponding to this area was used.

WINSYS 2016 - International Conference on Wireless Networks and Mobile Systems

5.1 Overview Details

Regarding to the thresholds used in section 4.1) to

identify the intended network behaviors, they are ex-

tremely dependent on mobile operator policies, re-

quirements, type of service, etc., so a set of default

parameters was used, as detailed in Table 1. The over-

all results are shown in Table 2. The columns ”CH”,

”OS” and ”PP” refer to coverage holes, overshooting

and pilot pollution algorithms, respectively.

Table 1: Algorithms thresholds and inputs.

CH OS PP

RSCP

threshold [dBm]

-100 -95 -95

RSCP

delta [dB]

3 6 6

Min EcN0

threshold [dB]

N/A -10 N/A

EcN0

delta [dB]

N/A 6 4

Number

of bins / cluster

10 10 10

Max

cluster radius [m]

57,5 57,5 57,5

Pilot

pollution window

N/A N/A 3

The results on Table 2, reveal a considerable num-

ber of cells with performance malfunctions, which are

currently under more detailed evaluation by radio op-

timization teams. In terms of the coverage hole index

cell

, it is the highest. The DT reliability index R is

Table 2: Self-diagnosis detected scenarios.

CH OS PP

Analyzed cells 830

Detected cells 13 35 64

Average clusters [#] 3 2 3

Average index H

cell

[%] 38 11 11

Average index R [%] 37 45 55

the lowest, though. Which admits that the DT did not

retrieve data from all cell service area, leading to an

overestimated H

cell

value. With regard to the interfer-

ence detection algorithms, the average cell index H

cell

was 11% with more quality of the respective DT data.

5.2 Coverage Holes

One of the detected coverage holes is illustrated in

Figure 7. The blue cell is the diagnosed cell, with the

red points corresponding to the cell’s measurements

in coverage hole, grouped in one cluster.

Figure 7: Coverage hole scenario.

The detailed Radio Frequency (RF) metrics are

displayed in Table 3. It can be observed the -103 dBm

Table 3: Coverage hole measurements details.

Number

of bins

Average

RSCP

[dBm]

Average

Ec/No [dB]

Cluster 1 14 -103 -9

low signal power level, on average, for the 14 detected

measurements.

Figure 8: Coverage hole cluster terrain proﬁle.

Using software with elevation proﬁling and 3D

building modulation, see Figure 8, it can be asserted

that the coverage hole area is in Non-Line-of-Sight

(NLoS) due to building obstruction and terrain eleva-

tion.

In relation to the index H

cell

, for this occurrence,

it was 25%, meaning that 25% of the clusters are in

coverage hole. Nevertheless, as the DT reliability in-

dex R for the cell was only 23%, it may indicate that

the index S be off the real percentage.

Self-Diagnosing Low Coverage and High Interference in 3G/4G Radio Access Networks based on Automatic RF Measurement Extraction

5.3 Overshooting

Concerning the overshooting, Figure 9, illustrates an

overshooting in an overlap area between two cells.

Figure 9: Overshooting scenario.

Again, the blue cell is the diagnosed cell, and in

this case, it was detected an overshooting cluster. Fur-

thermore, the cells with orange and purple, were iden-

tiﬁed as the victim cells, in the sense that, the over-

shooting cluster is located in their service areas. The

measurement stats are shown in Table 4. The blue cell

is reaching the other two cell service areas with a high

power, causing interference.

Just for additional information, see Figure 10,

where the terrain proﬁle between the source cell and

the respective overshooting cluster is illustrated, con-

ﬁrming the ”high pass”.

Table 4: Overshooting measurements details.

Number

of bins

Average

RSCP

[dBm]

Average

Ec/No [dB]

Cluster 1 31 -69 -9

Figure 10: Overshooting cluster terrain proﬁle.

It exists Line-of-Sight (LoS) between the source

cell and the overshooting cluster which enables the

overshooting scenario. In regard to the overshooting

cell index, H

cell

, it was obtained a value of 9%. This

represents, in this particular case, that the overshoot-

ing clusters affected, on average, 9% of the victim

cells footprint divided in clusters. Concerning the DT

reliability index R, it was obtained an average value

of 41% for the two victim cells.

5.4 Pilot Pollution

Regarding the pilot pollution detection algorithm, an

example is illustrated in Figure 11. The diagnosed

cell (in blue) is reaching an area, within the pilot pol-

lution conditions (4.1.3), where the red, green and

light purple cells are the serving (best) cells in that

area. The detected pilot pollution bins, were arranged

in one valid cluster.

Figure 11: Pilot pollution scenario.

The pilot pollution average metrics are presented

in Table 5. Even for the high average power detected,

the average quality level is very low, exhibiting how

interfered is this area.

Table 5: Pilot pollution measurements details.

Number

of bins

Average

RSCP

[dBm]

Average

Ec/No [dB]

Cluster 1 17 -77 -17

It can be seen in Figure 12 that the NLoS scenario

is not enough to reduce the source power received.

Due, mainly, to the relative small distance between

the cluster and the source cell, besides that, this cell

operates in the 900 MHz band.

In terms of the cell severity index H

cell

, for this

case the reported value, was 14%. It shows that the

pilot pollution clusters are in average 14% of the de-

tected serving cells footprint. In the matter of the DT

reliability index of the victim cells, was retrieved an

average value of 62%, which empowers greatly this

WINSYS 2016 - International Conference on Wireless Networks and Mobile Systems

Figure 12: Pilot pollution cluster terrain proﬁle.

pilot pollution H

cell

index, compared to the overshoot-

ing harshness index.

By deﬁnition, pilot pollution areas display an ex-

cessive number of pilot signals. Sometimes, this num-

ber is high enough, that enables the algorithm to sug-

gest which cell(s) are being victimized. In this sce-

nario, the detection is done, but insufﬁcient informa-

tion retains the algorithm from the harshness evalua-

tion and the affected cell identiﬁcation.

6 CONCLUSIONS

This paper presented a new approach for automatic

detection of low coverage and high interference sce-

narios (overshooting and pilot pollution) in UMTS

/LTE networks. These algorithms, based on period-

ically extracted DT measurements, identify the prob-

lematic cluster locations and compute harshness met-

rics, at cluster and cell level, quantifying the extent of

the problem.

The proposed algorithms were validated for a live

network urban scenario. 830 3G cells were self-

diagnosed and performance metrics were computed.

The results showed that for an urban area and with

a high site density, the main optimization efforts rely

on the interference mitigation and not coverage opti-

mization. Moreover, the pilot pollution scenarios are

the most prevalent.

The cluster division, with a RF correlation dis-

tance limitation, provides a simpliﬁcation for the an-

tenna physical parameter optimization algorithms, in

the sense that, the optimization process can be re-

duced to the cluster’s centroids.

Future work is in motion by adding self-

optimization capabilities to the algorithms, which will

automatically suggest physical and parameter opti-

mization actions, based on the already developed

harshness metrics.

ACKNOWLEDGEMENTS

This work was supported by the Instituto de

Telecomunicac¸

oes (IT) and the Portuguese Founda-

tion for Science and Technology (FCT) under project

PEst-OE/EEI/LA0008/2013.

REFERENCES

Duarte, D., Vieira, P., Rodrigues, A. J., and Silva, N. (2015).

A new approach for crossed sector detection in live

mobile networks based on radio measurements. In

Wireless Personal Multimedia Communications Symp.

- WPMC, volume 1.

Ericsson (2015). Ericsson mobility report. Technical report,

ERiCSSON.

Izenman, A. J. (2008). Modern Multivariate Statistical

Techniques: Regression, Classiﬁcation, and Manifold

Learning. Springer Publishing Company, Incorpo-

rated, 1 edition.

Kysti, P., Meinil, J., Hentil, L., Zhao, X., Jms, T., Schnei-

der, C., Narandzi, M., Milojevi, M., Hong, A., Ylitalo,

J., Holappa, V., Alatossava, M., Bultitude, R., Jong,

Y., and Rautiainen, T. (2007). Ist-4-027756 winner ii

d1.1.2 v1.2. Technical report, EBITG, TUI, UOULU,

CU/CRC, NOKIA.

Sallent, O., Perez-Romero, J., Sanchez-Gonzalez, J.,

Agusti, R., Diaz-Guerra, M., Henche, D., and Paul,

D. (2011). Automatic detection of sub-optimal perfor-

mance in umts networks based on drive-test measure-

ments. In Network and Service Management (CNSM),

2011 7th International Conference on, pages 1–4.

Sanchez-Gonzalez, J., Sallent, O., P

erez-Romero, J., and

Agust

ı, R. (2013). A multi-cell multi-objective self-

optimisation methodology based on genetic algo-

rithms for wireless cellular networks. Int. Journal of

Network Management, 23(4):287–307.

Sousa, M., Martins, A., Vieira, P., Oliveira, N., and Ro-

drigues, A. (2015). Caracterizacao da ﬁabilidade

de medidas r

adio em larga escala para redes auto-

otimizadas. In 9. Congresso do Comit

e Portugu

es da

URSI - ”5G e a Internet do futuro”.

Vieira, P., Silva, N., Fernandes, N., Rodrigues, A. J.,

and Varela, L. (2014). Improving accuracy for otd

based 3G geolocation in real urban/suburban environ-

ments. In Wireless Personal Multimedia Communica-

tions Symp. - WPMC, volume 1.

Witten, I. H. and Frank, E. (2005). Data Mining: Practi-

cal Machine Learning Tools and Techniques, Second

Edition (Morgan Kaufmann Series in Data Manage-

ment Systems). Morgan Kaufmann Publishers Inc.,

San Francisco, CA, USA.

Zheng, N. and Xue, J. (2009). Statistical Learning and Pat-

tern Analysis for Image and Video Processing. Ad-

vances in Pattern Recognition. Springer.

Self-Diagnosing Low Coverage and High Interference in 3G/4G Radio Access Networks based on Automatic RF Measurement Extraction