Page 56 - Proceedings of the 2018 ITU Kaleidoscope
P. 56
2018 ITU Kaleidoscope Academic Conference
Insight
Anomaly detection Diagnose labels
Insight
Raise the most important anomalies
Diagnosis
Anomaly
event CBR Active
pattern diagnosis learning Diagnosis
KB
Figure 4 – An overview of active learning in the diagnosis process
during operation. As analyzed anomaly patterns are added,
the quality of the automated diagnoses improves. The 5.3 Transfer Learning
collection and maintenance of a sufficient knowledgebase
can be, however, expensive and time-consuming. To Because the fault states can be very rare, even with the
mitigate this process, active learning methods can be used. augmented diagnosis it can be difficult and time-consuming
In active learning, the system raises those cases to be to create a comprehensive diagnosis knowledgebase. It
analyzed and labeled by the human expert that are the hardest would also be highly desirable to be able to diagnose and
for the diagnosis function to diagnose or that would improve quickly remedy or even prevent problems that have never
the quality of the diagnosis knowledgebase the most. This occurred in the system before. One way to enable this is to
way it can guide the human expert diagnosis process to use transfer learning to share diagnosis knowledge between
analyze the anomalies that are the most meaningful for the different networks.
automated diagnosis.
[10] describes a framework for sharing diagnosis knowledge
Figure 4 shows an overview of active learning in the and presents an example using topic modeling and Markov
diagnosis process. The black dotted lines depict the flow of Logic Networks (MLNs). It defines three components for a
data and the continuous blue lines the insights shared diagnosis cloud: Central, Gateway and Local Diagnostic
between the machine analytics and human expert. The Agents (CDA, GDA and LDA). The GDA is an agent
iterative process works as follows: mapping models between the CDA’s central storage of
models and the local models in an LDA. Sharing knowledge
1. Initially, the detected anomalies are diagnosed by
the CBR-based diagnosis. between diagnosis knowledgebases enables fast
“bootstrapping” of completely new self-healing
2. The diagnosis results are fed to the active learning deployments, or updating an existing one, for example when
component, which analyzes, which diagnosed managed network functions go through major upgrades. It
anomalies are the most relevant to be raised to the raises also the question, how such diagnosis knowledge can
human operator, i.e. the ones where the automatic best be shared. Standardized knowledge sharing methods
diagnoses are the most unreliable or the ones which may be required in addition to sharing data.
are on the border of different diagnoses.
6 HOLISTIC SELF-HEALING
3. The human expert analyzes the raised cases and
provides the analysis results as new labeled and
diagnosed anomaly patterns into the diagnosis Another recurring principle in resilient systems is holism. In
component. a complex system, improving the resilience of only one part
or level of organization can sometimes (unintentionally)
4. Steps 1-3 are repeated for: introduce fragility in another. To improve the resilience, it is
a. Refining the labelling of existing often necessary to work in more than one domain and scale
anomalies at a time. [3]
b. Incorporation of newly detected In mobile network management, this means we cannot
anomalies in the labelling consider different management domains and levels in
– 40 –