Application of probabilistic modeling and machine

learning to the diagnosis of FTTH GPON networks

Stephane Gosselin, Jean-Luc Courant, Serge Romaric Tembo Mouafo,
Sandrine Vaton

To cite this version:

Stephane Gosselin, Jean-Luc Courant, Serge Romaric Tembo Mouafo, Sandrine Vaton. Application
of probabilistic modeling and machine learning to the diagnosis of FTTH GPON networks. ONDM
2017 : 21st Conference on Optical Network Design and Modeling, May 2017, Budapest, Hungary. pp.1
- 3. �hal-01573963�

HAL Id: hal-01573963

https://hal.archives-ouvertes.fr/hal-01573963
Submitted on 11 Aug 2017

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 Application of Probabilistic Modeling and Machine
Learning to the Diagnosis of FTTH GPON Networks

Stéphane Gosselin, Jean-Luc Courant Serge Romaric Tembo

Orange Labs ActiveEon
Lannion, France Sophia Antipolis, France
stephane.gosselin, [email protected] [email protected]

Sandrine Vaton
IMT Atlantique
Brest, France
[email protected]

Abstract—This paper presents insights on the promises of decision rule engine. This rule-based system reaches high
probabilistic modeling and machine learning for fault diagnosis diagnosis performance, but maintenance and creation of new
in optical access networks. A Bayesian inference engine, called rules is difficult and requires the analysis of large amounts of
Probabilistic tool for GPON-FTTH Access Network self- data. Due to the dynamic, distributed and non-deterministic
DiAgnosis (PANDA), is applied to fault diagnosis of Gigabit nature of networks, fault propagation is a complex
capable Passive Optical Networks (GPON). PANDA approach phenomenon and available rules do not cover all possible fault
has been assessed on real diagnosis data, showing very configurations. This leads to situations in which the expert
satisfactory alignment with an operational rule-based expert system is unable to diagnose the root causes of network faults.
system. Furthermore, it provides diagnosis conclusions for all
tested cases, even if some monitoring data is missing or Other diagnosis methods reproduce fault and alarm
incomplete. Finally, an expectation maximization algorithm propagation through a model of the considered system [3],
allows to finely tune the probabilistic model. whereas some blind methods rely on machine learning
techniques, without system modeling [4].
Keywords—Network management; optical access network; fault
management; model-based approach; Bayesian inference; machine This paper describes principles and outcomes of a modular
learning; expectation maximization approach to fault diagnosis in optical access networks,
benefiting from both probabilistic modeling and machine
learning. All implementation details of the probabilistic model
I. INTRODUCTION and of the machine learning algorithm used are given in
[5][6][7]. This paper provides a synthesis of main principles,
Significant efforts are made by telecommunication outcomes and insights of this work. The reader can refer to
operators to provide broadband services with high availability. above mentioned papers for all implementation details.
This includes improvements of reliability and robustness of
network infrastructures, through automation of complex
II. PROBABILISTIC MODELING OF A GPON SYSTEM
network management operations such as fault diagnosis [1].
Model-based approaches for fault diagnosis rely on a
Fault diagnosis aims at detecting and locating any type of representation of the network structure and its behavior, in
failures that degrade the quality of service or lead to service particular alarm propagation and correlation [3][8]. Compared
unavailability. A fault occurrence typically triggers other faults to rule-based expert systems, model-based approaches cover
and alarms, which might result in complex and distributed more diverse fault configurations than expert systems and can
patterns of failures and alarms. In particular, a given fault may manage unforeseen issues. When the model is probabilized,
lead to several alarms, and a given alarm may result from e.g. with a Bayesian network formalism [9], it gets the ability
various faults. Fault diagnosis consists in finding the root to deal with uncertainty resulting from non-deterministic fault
causes which have triggered a given set of alarms in specific propagation. It also becomes robust to missing data [7]. This is
network conditions. The latter are derived from monitoring particularly interesting in network management situations, for
data which is gathered by the network management system. which collected monitoring data is often incomplete and
Fault diagnosis requires expert knowledge which is depends on the specific network conditions which have led to
typically involved in designing rule-based expert systems [2]. alarms.
As a matter of fact, the current diagnosis tool used by Orange A modular approach for fault diagnosis has been proposed
for its FTTH networks relies on an expert system based on a and fully described in [5]. It relies on a 3-layer probabilistic
modeling of the considered system with Bayesian networks. A first derived from the GPON standard [11][12] and expert
Bayesian network [9] is a directed acyclic graph the vertices of knowledge. This first PANDA implementation was run on a
which are random variables (i.e. alarms, counters, intermediate dataset of 10611 diagnosis cases collected from Orange FTTH
and root causes in our diagnosis problem) and the edges of network, so as to test its alignment with the operational rule-
which represent conditional dependencies between these based expert system. The dataset contains alarms typically
random variables. A very simple example of Bayesian network collected during and after customer calls to the hot line of the
is illustrated in Fig. 1. access provider.
Table I. gives a confusion matrix crossing diagnosis
conclusions obtained with both tools on the 10611 cases [7].
The rows of the table give the numbers of occurrences of
diagnosed root causes obtained from the rule-based expert
system, whereas the columns give results from the first
PANDA implementation.

TABLE I. DIAGNOSIS CONCLUSIONS OF RULE-BASED EXPERT SYSTEM

(ROWS) AND FIRST PANDA IMPLEMENTATION (COLUMNS) OVER 10611 CASES

Root causes 1 2 3 4 5 6 7 8 9

1. No default 7210 183 39 17

2. Faulty ONT 3
Fig. 1. Example of Bayesian network.
3. ONT 0
configuration
The 3-layer probabilistic model detailed in [5] has been mistake
applied to fault diagnosis of a Gigabit capable Passive Optical 4. Drop fiber 72 18
Networks (GPON) system, composed of Optical Line attenuated
Termination (OLT) in the central office, Optical Network 5. Drop fiber 1463
Terminations (ONT) at customer premises and a (passive) broken
Optical Distribution Network (ODN) in-between, as illustrated 6. ONT power 2 780
in Fig. 2. The model was implemented in Python and called supply failure
Probabilistic tool for GPON-FTTH Access Network self- 7. Feeder fiber 0
DiAgnosis (PANDA). PANDA tool does not include attenuated
Residential Gateways (RG) connected to ONTs, but the 8. Feeder fiber 1 57
modularity of the approach would easily allow this scope broken
extension. Also, a refinement of the model, not addressed in 9. Unknown 716 4 19 27 0
root cause
this paper, includes several splitter stages in the ODN. This
refinement and related detailed results are presented in [10].
A first important result is that PANDA approach derives a
Based on collected monitoring data (e.g. alarms, counters, diagnosis for all tested cases, even if some monitoring data is
optical powers), the PANDA model infers missing data and missing or incomplete. However, it is worth noting that the 716
then the most probable root causes of the network fault. The cases of row 9 and column 1 (i.e. unknown root cause for the
robustness of the model to missing observations is one of its rule-based system and no default for PANDA) can be partly
key characteristics compared to a rule-based expert system. explained by the much larger scope of the rule-based expert
system, which encompasses equipment beyond the GPON
system itself. Nevertheless, the ability of PANDA to deal with
uncertainty and missing observations and to manage
unforeseen situations is a key property.
Another outcome from table I. is that diagnosis decisions of
PANDA and of the rule-based expert system are aligned in
more than 97% of all the cases effectively diagnosed by the
rule-based engine (the first 8 rows of the confusion matrix).
The analysis of the remaining 3% cases (off-diagonal numbers
in the confusion matrix, excluding the last row) provides key
insights. 183 cases detected as ONT configuration mistakes by
Fig. 2. Structure of a GPON system. PANDA are not managed by the expert system, which has no
dedicated rule for these situations. PANDA also diagnoses 39
III. FIRST PANDA IMPLEMENTATION cases of drop fiber attenuation for which no default is found by
the rule-based expert system. In those cases, a low received
The structures of the Bayesian networks composing the optical power was measured either at the OLT or at the ONT
model and rough estimates of conditional dependencies were under study, whereas the received optical power in the reverse
direction was missing. Additionally, in those 39 cases, optical indeed characterize a low power level which can result in high
powers related to all ONT neighbors in the PON tree were bit error rate and then a loss of frames: machine learning thus
measured and nominal. In those cases, in spite of one missing allowed more consistent decisions in those cases. In 148 other
optical power measurement, PANDA is capable to derive a cases, the decision is changed from "feeder fiber attenuated" to
diagnosis decision which is fully compliant with the conclusion a more clear-cut "faulty ONT" decision. In those cases, only
that would be drawn by a human expert. This capability is one ONT is connected to the OLT port, upstream and
allowed by a global analysis of the PON which is performed by downstream "loss of signal" alarms are observed, but no optical
the Bayesian inference engine [7]. power measurement could be collected. The "faulty ONT"
decision is fully reasonable for a human expert, although those
IV. IMPROVEMENTS WITH MACHINE LEARNING situations could also result from fiber attenuation [6].
The conditional dependencies of PANDA model have then
been finely tuned by machine learning thanks to an Expectation V. CONCLUSION
Maximization (EM) algorithm which is detailed in [6]. This A 3-layer probabilistic model based on Bayesian networks
EM algorithm was run on 5121 diagnosis cases collected from was applied to fault diagnosis of GPON systems. With model
Orange FTTH network and used as a learning dataset. The parameters roughly estimated by skilled humans, very
principle is to mine the dataset so as to learn the conditional satisfying diagnosis results are obtained on operational GPON
dependencies which will afterwards enable Bayesian inference FTTH data. The tuning of the model thanks to machine
on other diagnosis cases. Expectation Maximization learning leads to even better results in some cases, while some
implements maximum likelihood estimation for an incomplete uncertain situations give rise to more clear-cut diagnosis.
dataset, and leads to model parameters that "best match" the Outcomes and perspectives of this work are further detailed in
dataset [6]. The initial model parameters used for running the [5][6][7] and will be presented at the conference.
EM algorithm were those of the first PANDA implementation
presented in the previous section. REFERENCES
This second PANDA model, tuned by machine learning,
has been assessed on another dataset (test dataset) of 5490 [1] M. Steinder and A. S. Sethi, "A survey of fault localization techniques,"
diagnosis cases and compared to the initial PANDA model in computer networks. Science of Computer Programming, 53:165–194,
without machine learning. Table II. gives the confusion matrix January 2004.
crossing diagnosis conclusions obtained with both PANDA [2] L. F. Pau, "Survey of expert systems for fault detection, test generation
and maintenance," Expert Systems, 3:100–110, April 1986.
versions on the 5490 cases of the test dataset [7].
[3] R. D. Gardner and D. A. Harle, “Alarm correlation and network fault
resolution using the Kohonen self-organising map,” Global
TABLE II. DIAGNOSIS CONCLUSIONS OF FIRST (ROWS, NO LEARNING) Telecommunications Conference (GLOBECOM 1997), pp. 1398–1402,
AND SECOND (COLUMNS, WITH LEARNING) PANDA IMPLEMENTATIONS OVER 1997.
5490 TEST CASES [4] J. R. A. Goel and P. Sadayappan. Towards a ’neural’ architecture for
abductive reasoning. IEEE International Conference on Neural
Root causes 1 2 3 4 5 6 7 8 9
Networks, pages 681–688, 1998.
1. No default 4030 6 7 9 [5] S. R. Tembo, J. L. Courant and S. Vaton, "A 3-layered self-
2. Faulty ONT 0 reconfigurable generic model for self-diagnosis of telecommunication
3. ONT 183 networks," 2015 SAI Intelligent Systems Conference (IntelliSys),
configuration London, 2015, pp. 25-34.
mistake [6] S. R. Tembo, S. Vaton, J. L. Courant and S. Gosselin, "A tutorial on the
4. Drop fiber 56 EM algorithm for Bayesian networks: Application to self-diagnosis of
attenuated GPON-FTTH networks," 2016 International Wireless Communications
5. Drop fiber 14 602 1 and Mobile Computing Conference (IWCMC), Paphos, 2016, pp. 369-
376.
broken
[7] S. R. Tembo, S. Vaton, J. L. Courant, S. Gosselin, M. Beuvelot, "Model-
6. ONT power 402 Based Probabilistic Reasoning for Self-Diagnosis of Telecommunication
supply failure Networks: Application to a GPON-FTTH Access Network," Journal of
7. Feeder fiber 148 32 Network and Systems Management, pp. 1-33, Dec. 2016
attenuated [8] S. Kätker and K. Geihs, "A generic model for fault isolation in
8. Feeder fiber 0 integrated management systems," Journal of Network and Systems
broken Management, 5(2):109–130, 1997.
9. Unknown 0 [9] J. Pearl, “Bayesian networks: A model of self-activated memory for
root cause evidential reasoning,” in 7th Conference of the Cognitive Science
Society, 1985.
[10] S. R. Tembo, J. L. Courant, S. Vaton and S. Gosselin, "A highly
The tuning of the PANDA model through machine learning adaptable probabilistic model for self-diagnosis of GPON-FTTH access
changes the diagnosis decisions in 3% of the cases, leading to networks," 2016 24th International Conference on Software,
more consistency in some cases, and more clear-cut decisions Telecommunications and Computer Networks (SoftCOM), Split, 2016,
in other cases. In 14 cases, PANDA tuning changed the pp. 1-5.
decision from "drop fiber broken" to "drop fiber attenuated". In [11] Telecommunication Standardization Sector of ITU, G.984.3
Recommendation. ITU-T, 2008.
all those cases, a "loss of frame" alarm was observed and no
[12] Telecommunication Standardization Sector of ITU, G.988
optical power measurement could be collected. Those cases Recommendation. ITU-T, 2012.