See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/309365201

A machine learning approach for predictive maintenance for mobile phones

service providers

Conference Paper · January 2017

DOI: 10.1007/978-3-319-49109-7_69

CITATIONS READS
3 3,535

4 authors, including:

Anna Corazza Francesco Isgrò

University of Naples Federico II University of Naples Federico II
85 PUBLICATIONS 1,020 CITATIONS 99 PUBLICATIONS 1,240 CITATIONS

SEE PROFILE SEE PROFILE

Roberto Prevete
University of Naples Federico II
89 PUBLICATIONS 1,138 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Computational models of neuromodulation in the control of plastic behaviours View project

automatic measurement of the nuchal translucency thickness from ultrasound imagery View project

All content following this page was uploaded by Francesco Isgrò on 15 November 2017.

The user has requested enhancement of the downloaded file.

A machine learning approach for predictive
maintenance for mobile phones service providers

A. Corazza, F. Isgrò, L. Longobardo, R. Prevete

Abstract The problem of predictive maintenance is a very crucial one for ev-
ery technological company. This is particularly true for mobile phones service
providers, as mobile phone networks require continuous monitoring. The ability
of previewing malfunctions is crucial to reduce maintenance costs and loss of cus-
tomers. In this paper we describe a preliminary study in predicting failures in a
mobile phones networks based on the analysis of real data. A ridge regression clas-
sifier has been adopted as machine learning engine, and interesting and promising
conclusion were drawn from the experimental data.

1 Introduction

A large portion of the total operating costs of any industry or service provider is
devoted to keep their machinery and instruments up to a good level, aiming to ensure
a minimal disruption in the production line. It has been estimated that the costs of
maintenance is the range 15-60% of the costs of good produced [14]. Moreover
about one third of the maintenance costs is spent in not necessary maintenance; just
as an example, for the U.S. industry only this is a $60 billion each year spent in
unnecessary work. On the other hand an ineffective maintenance can cause further
loss in the production line, when a failure presents itself.

Anna Corazza
DIETI, Università di Napoli Federico II e-mail: [email protected]
Francesco Isgrò
DIETI, Università di Napoli Federico II e-mail: [email protected]
Luca Longobardo
DIETI, Università di Napoli Federico II e-mail: [email protected]
Roberto Prevete
DIETI, Università di Napoli Federico II e-mail: [email protected]
 Predictive maintenance [14, 10] attempts to minimise the costs due to failure
via a regular monitoring of the conditions of the machinery and instruments. The
observation will return a set of features from which it is possible in some way to
infer if the apparatus are likely to fail in the near future. The nature of the feature
depend, of course, on the apparatus that is being inspected. The amount of time in
the future that the failure will arise also depends on the problem, although we can
state, as a general rule, that the sooner a failure can be predicted, the better is in
terms of effective maintenance.
In general the prediction is based on some empirical rule [23, 17, 19], but over the
last decade there has been some work devoted to apply machine learning [6, 22, 5]
techniques to the task predicting the possible failure of the apparatus. For instance,
a Bayesian network has been adopted in [9] for a prototype system designed for the
predictive maintenance of non-critical apparatus (e.g., elevators). In [12] different
kind of analysis for dimensionality reduction and support vector machines [7] have
been applied to rail networks. Time series analysis has been adopted in [13] for
link quality prediction in wireless networks. In a recent work the use of multiple
classifiers for providing different performance estimates has been proposed in [18].
An area where disruption of service can have a huge impact on the company sales
and/or the customer satisfaction is the one of mobile phone service providers [8, 4].
The context considered in this work is a the predictive maintenance of national
mobile phone network, that is being able to foresee well in advance if a cell of
the network is going to fail. This is very important as the failure of a cell can have
a huge impact on the users’ quality of experience [11], and to prevent them makes
less likely that the user decides to change service provider.
In this paper we present a preliminary analysis on the use of a machine learning
paradigm for the prediction of a failure on a cell of a mobile phones network. The
aim is to predict the failure such in advance that no disruption in the service will
occur, lets say, at least a few hours in advance. A failure is reported among a set
of features that are measured every quarter of an hour. The task is then to predict
the status of the feature reporting the failure within a certain amount of time. As for
many other predictive maintenance problems given we are dealing with a very large
amount of sensors [15].
The paper is organised as follows. Next section describes the data we used and
reports some interesting properties of the data that have been helpful in designing
the machine learning engine. The failure prediction model proposed is discussed in
Section 3, together with some experimental results. Section 4 is left to some final
remarks.

2 Data analysis

To predict failure, considered data is obtained by monitoring the state of base

transceiver stations (also known as cells) in a telecommunication network, during a
1 month (31 days) time-span. A cell represents the unit for a telecommunication net-
work in the tackled case study. Cells are grouped into antennas, so that one antenna
can contain several cells. The goal for the problem is to predict a malfunctioning
(pointed out by an alarm signal originated from cells) in a cell.
Furthermore, information about the geographical location of the cell can be rele-
vant. When the Italian peninsula is considered, the total number of cells amounts to
nearly 52, 000. For instance, when considering the total number of measurements,
we get more than 150 millions of tuples.
Several kinds of statistical analysis were implemented to explore the data, and
some interesting key-points and critical issues emerged from this analysis.
First of all, more than the 60% of the cells did not show any alarm signal. This
is a quite usual behavior, as the system works smoothly for most of the time. Even
when such cases are excluded, the average number of alarms per cell is only 3 in
a month. In order to obtain a data set which is meaningful enough for a statistical
analysis, only cells with at least 6 alarms have been kept: in this way the number
of cells is further reduced to less than 2, 000. Moreover, among the remaining cells
the proportion between alarm tuples and non-alarm tuples still remains high, as the
former represent barely 1% of the total. However, we considered it acceptable, as
malfunctioning must be considered unlikely to happen. In the end, this unbalance
strongly influence the pool size of useful input data, and must faced with an adequate
strategy.
Another critical issue regards the presence of several non-numeric values spread
among the tuples. There are four different undefined values, among whose, INF
values are the most frequent. Indeed, INF is the second most frequent value among
all fields. All in all, discarding tuples containing undefined values in some of the
fields would cut out 80% of data, leaving us with a too small dataset. We therefore
had to find a different strategy to face the problem.
As already stated, another problematic issue regards the temporal dimension of
data. Time-span is only one month, which on a time series related problem is not
very much, to begin with the very basic problem of properly splitting data into
training and test sets.
Another key-point regarding the data was found by looking at scatter plot dia-
grams between pairs of features. These diagrams highlighted two aspects: the first
one is that alarm occurrences seem related to the values of some specific features.
The second one is that there are two identical features. Since we don’t have infor-
mation about the meaning of the various features, we can’t tell if this is supposed to
be an error on the data supplied.
In addition to these, other statistical analysis were performed, focusing mainly
on the values assumed by the features. Average values are summarized, along with
standard deviations in Figure 1.
Inspecting Figure 1 we can see that FEATURE 7 and FEATURE 9 show a sig-
nificant difference in term of average value between alarm and non-alarm events and
thus, can trace a good starting point for a machine learning approach. Moreover, we
can split features in three different groups.
Fig. 1 Average values for the features in a stable or alarm situation. The line on every bar represent
the standard deviation

Fig. 2 Pearson correlation coefficient between features and alarm indicator

1. The first group is composed of: FEATURE 4, FEATURE 8, and FEATURE 9.

These features have constant values in all the three conditions but also a relatively
high standard deviation.
2. The second group is composed of: FEATURE 6, FEATURE 3, FEATURE 5,
and FEATURE 1. Also in this case, the features tend to have constant values in
all three situations, but with a relatively lower standard deviation.
3. The third group is composed of: FEATURE 7 and FEATURE 2. These features
show a large difference in terms of both average value and standard deviation
between alarm and non-alarm situations.
To better analyse the differences between alarm and non-alarm situations, Pear-
son correlation coefficients have been calculated between each feature and the alarm
indicator. Results are shown in Figure 2, and confirm that FEATURE 2 and FEA-
TURE 7 appear to be more related with an alarm occurrence, although the correla-
tion value is always lower than 0.2.
Last but not least, the alarm propagation effect has been analyzed, to check if
an alarm occurring in a cell is correlated to alarms in nearby cells. The results in
Figure 3 show that this is the case only for cells belonging to the same antenna. In
general, when the distance increases and cells of different antennas are considered,
the probability of cooccurrent alarms drops close to 0. We can therefore conclude
that, according to our data, there is no propagation effect.
3 Failure prediction

Alarm prediction is approached as a binary classification in a vector space model [6].

In simple words, the input to the classifier is given by a vector of features measured
at times t,t + 1, . . . ,t + d in a given cell, while the output is positive or negative,
depending whether an alarm in the same cell is foreseen or not for the time t + ∆ ,
with ∆ > 0. Such an approach is very general, and can be ported to similar problems
in different domains.
More specifically, we chose a classification model based on Tikhonov regular-
ization [21] (also known as Ridge regression). This is a direct classification model,
based on the assignment of binary label to a tuple of features in accordance with the
presence or the absence of an alarm signal after a chosen amount of time.
The representation of the features we use for classification is important for the
system performance. In fact, we have a time series and how the temporal information
is represented is a crucial point. Our approach consists of the following two steps:
1. Feature expansion;
2. Feature selection.
Feature expansion is obtained by calculating new features derived from the ex-
isting ones. An example of this is the application of some aggregation measures like
means, standard deviations and/or variances. Another used technique for increasing
the dimensionality is related to the use of convolution filters on the variables. More
specifically, we used a set of filters belonging to the family of wavelets. One single
filter was used multiple times with different parameters. The output for this step are
tuples with a much larger dimensionality. We switched, in some extreme cases, from
a 9 dimensions problem to a 1200 dimensions problem. Such number of variables

Fig. 3 Illustration of the probabilities of getting an alarm for the cells close to a cell signaling an
alarm, after 15 and 180 minutes
is vastly overabundant. For this reason a further step for feature selection becomes
necessary.
A process of automatic feature selection was chosen to increase portability and
maintain a data-oriented approach. In particular we used an algorithm for L1L2
regularization implemented in the “l1l2py”1 Python package. This algorithm com-
bines the classic shrinkage methods given by Ridge Regression and Lasso Regres-
sion [20].
We consider a regression problem where the output y is reconstructed from fea-
tures xi , i ∈ [1, p] by combining them with coefficients β = {βi }. Ridge Regression
uses a L2 norm in order to force a constraint on the regression coefficients size by
reducing their absolute value:
 
 p 2 p 
β̂ridge = arg min y − ∑ x j β j + λ ∑ β j2 (1)
β j=1 j=1
 

On the other hand, Lasso Regression instead uses a L1 norm forcing sparsity in
data and the annulment of some of the coefficients:
 
 p 2 p 
β̂lasso = arg min y − ∑ x jβ j + λ ∑ β j (2)
β j=1 j=1
 

The final step is the effective experimental assessment of the classifier. First of
all, we have to decide how to solve the critical issues emerged from data analysis
and pointed out in the preceding section: how to mandage undefined values and
how to split data into training and test set while reducing unbalancing of positive
and negative examples.
With relation to the issue of undefined values we decided to operate a fixed sub-
stitution of the most frequent of such values ( INF ) based on the average value
assumed by the considered feature, according to the scheme in Table 1. The number

Table 1 Substitution of undefined values in the features which assume such value.
Feature Substitution value
FEATURE 6 120
FEATURE 3 120
FEATURE 5 120
FEATURE 1 120
FEATURE 7 -10

of occurrences of the other undefined values is relatively negligible and the tuples
containing those values were simply dropped.

1 http://slipguru.disi.unige.it/Software/L1L2Py/
 In order to fix the balance between positive and negative samples we kept all the
available positive samples, which were the ones with a minor number of occurrences
N p , and randomly choose Nn = 4N p negative examples.
The splitting of data into training and test set has been solved by a temporal
based partitioning: we selected the first 2/3 of the month for the training, and the
remainder of the data was used as test set. We could act like that because positive
examples have a nearly uniform distribution in data. Therefore, even if we applied
the split without considering the frequencies of positive and negative examples, we
obtained an acceptable balance for both sets. Furthermore, we want to underline how
it is fundamental to operate an accurate sampling of data composing the training set,
because including tuples related in some way with the occurrence of an alarm results
in a hike of performance.
In the experiments different time shifts ∆ have been considered. An analysis of
the results showed some few interesting points.
First of all, we tested both generic and location-based models. The former does
not consider geographical information, while the latter is a location-based model.
Classification results have shown how the geographical information is crucial to
the classification, while a single generic model for the whole area fails to catch the
different variety of underlying key factors specific to each geographic subarea. One
example is illustrated in Figure 4, where we compare the results, in terms of ROC
curve, from a sample of such two models. Training strongly geo-localized models
resulted, in some of the best performance, with AUC values (the area under ROC
curve) of 0.7 − 0.8.
Another point regards the inverse proportionality between classification perfor-
mance and the time shift: performance decreases while the time shift between obser-
vations and alarm increases. In fact, a regular loss in performance can be observed
when the time shift raises from a quarter of an hour up to 6 − 7 hours; after that
performance fundamentally go close to a random guess.
Last, we run some tests to analyse how performance changes in relation to the
introduction of automatic feature selection. Models built directly using all the fea-
tures produced by the feature expansion phase and models where we an l1l2py step
of feature selection was applied have been compared. The performance of the sys-
tem with feature selection shows a constant (although relatively low) improvement
with respect to the one without it. One example of this is showed in Figure 5.
We noticed that the set of features chosen by the feature selection step changes
depending on the location of the considered cell. However, the final features are
always correlated with the two which showed the largest coefficient of linear cor-
relation with the output, that is: FEATURE 7 and FEATURE 2. Such analysis can
also help the service provider to analyse which are the most likely causes of mal-
functioning.
Fig. 4 Comparison between ROC curves generated on test set by a model trained and tested on the
whole Italian area (top) and another specific to a single location (bottom). Time shift for prediction
is 3 hours.

4 Conclusions and future work

In this paper we described a possible strategy to tackle a problem of alarm predic-

tion in a domain where time series of features are available. From a first analysis of
the dataset some issues have been raised, including the problem of undefined val-
ues. The alarm prediction has been defined as a classification problem which was
solved by ridge regression. From the experimental results, we can conclude that ge-
ographical localization is important for the performance and that it is only possible
to preview alarms occurring in a few hours.
Future directions of this analysis should explore models able to exploit the mutual
position among cells, or, in general, to better exploit data. Among these, one of the
most promising is the cascade based model. This approach is not focused on a single
defined model to use, but on the chaining of various models for data classification
which are applied sequentially. It aims to reduce, through several steps, the range
of variables and dataset size. In a way which reminds of the decision tree approach,
every stage is built to reduce the pool of data that is given as input to the next stage
or the dimensionality of the problem.
The idea had origin from the implementation of AdaBoost [16] algorithm, which
consists of a sequence of steps aiming to reduce the number of elements to classify.
Fig. 5 Comparison between ROC curves obtained from models in a case without using feature
selection (top) and in a case with using it (bottom)

The main key-point is to use a sequence of models increasing in complexity in order

to cut out as many samples as possible in each stage. The early stages are typically
the ones where the biggest cut is operated, while the latest stages are reserved to
more refined models which have to take the hardest decision on the most difficult
data.
In our case, this approach could not be applied because of the relativly small
size of the sample dataset which remained after removing all non-relevant cells and
balancing data. In fact, such multi-stage system is effective when it can exploit a
really large dataset. On the other hand, in actual cases, a nearly unbounded quantity
of data can be collected and therefore this approach could express its potentialities.

Acknowledgements

The research presented in this paper was partially supported by the national projects
CHIS - Cultural Heritage Information System (PON), and BIG4H - Big Data Ana-
lytics for E-Health Applications (POR).
 References

1. Amato, F., De Pietro, G., Esposito, M., Mazzocca, N.: An integrated framework for securing
semi-structured health records. Knowledge-Based Systems 79, 99–117 (2015)
2. Amato, F., Moscato, F.: A model driven approach to data privacy verification in e-health sys-
tems. Transactions on Data Privacy 8(3), 273–296 (2015)
3. Amato, F., Moscato, F.: Exploiting cloud and workflow patterns for the analysis of composite
cloud services. Future Generation Computer Systems (2016)
4. Asghar, M.Z., Fehlmann, R., Ristaniemi, T.: Correlation-Based Cell Degradation Detection
for Operational Fault Detection in Cellular Wireless Base-Stations, pp. 83–93. Springer Inter-
national Publishing, Cham (2013)
5. Barber, D.: Bayesian Reasoning and Machine Learning. Cambridge University Press (2012)
6. Bishop, C.M.: Pattern recognition and machine learning. Springer (2006)
7. Cortes, C., Vapnik, V.: Support-vector networks. Machine learning 20(3), 273–297 (1995)
8. Damasio, C., Frölich, P., Nejdl, W., Pereira, L., Schroeder, M.: Using extended logic program-
ming for alarm-correlation in cellular phone networks. Applied Intelligence 17(2), 187–202
(2002)
9. Gilabert, E., Arnaiz, A.: Intelligent automation systems for predictive maintenance: A case
study. Robotics and Computer-Integrated Manufacturing 22(5), 543–549 (2006)
10. Grall, A., Dieulle, L., Berenguer, C., Roussignol, M.: Continuous-time predictive-maintenance
scheduling for a deteriorating system. IEEE Transactions on Reliability 51(2), 141–150 (2002)
11. Jain, R.: Quality of experience. IEEE MultiMedia 11(1), 96–95 (2004)
12. Li, H., Parikh, D., He, Q., Qian, B., Li, Z., Fang, D., Hampapur, A.: Improving rail network ve-
locity: A machine learning approach to predictive maintenance. Transportation Research Part
C: Emerging Technologies 45, 17 – 26 (2014). Advances in Computing and Communications
and their Impact on Transportation Science and Technologies
13. Millan, P., Molina, C., Medina, E., Vega, D., Meseguer, R., Braem, B., Blondia, C.: Tracking
and predicting link quality in wireless community networks. In: 2014 IEEE 10th International
Conference on Wireless and Mobile Computing, Networking and Communications (WiMob),
pp. 239–244 (2014)
14. Mobley, R.K.: An introduction to predictive maintenance, 2nd edn. Butterworth-Heinemann
(2002)
15. Patwardhan, A., Verma, A.K., Kumar, U.: A Survey on Predictive Maintenance Through Big
Data, pp. 437–445. Springer International Publishing, Cham (2016)
16. Schapire, R.E.: Explaining AdaBoost, pp. 37–52. Springer Berlin Heidelberg, Berlin, Heidel-
berg (2013)
17. Scheffer, C., Girdhar, P.: Practical machinery vibration analysis and predictive maintenance.
Elsevier (2004)
18. Susto, G.A., Schirru, A., Pampuri, S., McLoone, S., Beghi, A.: Machine learning for predictive
maintenance: A multiple classifier approach. IEEE Transactions on Industrial Informatics
11(3), 812–820 (2015)
19. Swanson, D.C.: A general prognostic tracking algorithm for predictive maintenance. In:
Aerospace Conference, 2001, IEEE Proceedings., vol. 6, pp. 2971–2977. IEEE (2001)
20. Tibshiriani, R.: Regression shrinkage and selection via the Lasso. Journal of the Royal Statis-
tical Society. Series B 58(1), 267–288 (1996)
21. Tychonoff, A., Arsenin, V.: Solution of ill-posed problems. Winston & Sons, Washington
(1977)
22. Vapnik, V.: The Nature of Statistical Learning Theory. Springer, New York (1995)
23. Zhou, X., Xi, L., Lee, J.: Reliability-centered predictive maintenance scheduling for a con-
tinuously monitored system subject to degradation. Reliability Engineering & System Safety
92(4), 530–534 (2007)

View publication stats