Maintworld 1/2020

TECHNOLOGY
Industrial AI’s capacity to analyze very large amounts of
high-dimensional data can change the current maintenance
paradigm and shift from preventive maintenance systems to
new levels. The key challenge, however, is operationalizing predictive
maintenance, and this is much more than connecting
assets to an AI platform, streaming data, and analyzing those
data. By integrating conventional data such as vibration, current
or temperature with unconventional additional data, such
as audio and image data, including relatively cheap transducers
such as microphones and cameras, Industrial AI can enhance
or even replace more traditional methods. AI’s ability to predict
failures and allow planned interventions can be used to
reduce downtime and operating costs while improving production
yield. For example, AI can extend the life of an asset beyond
what is possible using traditional analytics techniques by
combining data information from designer and manufacturer,
maintenance history, and Internet of Things (IoT) sensor data
from end users, such as anomaly detection in engine-vibration
data, images and video of engine condition. This information
fusion during the lifecycle of the asset is called product lifecycle
management (PLM).
Explainable AI in Maintenance
Advances in AI for maintenance analytics are often tied to
advances in statistical techniques. These tend to be extremely
complex, leveraging vast amounts of data and complex algorithms
to identify patterns and make predictions. This
complexity, coupled with the statistical nature of the relationships
between input data that the asset provides, makes them
difficult to understand, even for expert users, including the
system developers, Figure 4. This makes explainability a major
concern.
Figure 4: Engineers and data scientist must co-create the AI
solution for maintenance together
While increasing the explainability of AI systems can be
beneficial for many reasons, there are challenges in implementing
explainable AI. Different users require different
forms of explanation in different contexts, and different contexts
give rise to different needs. To understand how an AI
system works in the maintenance domain, users might wish to
know which data the system is using, the provenance of those
data, and why they were selected; how the model and prediction
work, and which factors influence a maintenance decision;
and why a particular output is obtained. To understand what
type of explanation is necessary, careful stakeholder engage-
ment and well-thought-out system design are both necessary
as can be seen in figure 5.
Figure 5: Architecture of explainable AI systems for Maintenance
Decisions
There are various approaches to creating interpretable systems.
Some AI is interpretable by design; these systems tend to
be kept relatively simple. An issue with them is that they cannot
get as much customization from vast amounts of data as
more complex techniques, such as deep learning. This creates a
performance-accuracy trade-off in some settings, and the systems
might not be desirable for those applications where high
accuracy is prized. In other words, maintainers must accept
more black boxes.
In some AI systems – especially those using personal data or
those where proprietary information is at stake – the demand
for explainability may interact with concerns about privacy. In
areas such as healthcare and finance, for example, an AI system
might be analyzing sensitive personal data to make a decision
or recommendation. In determining the type of explainability
that is desirable in these cases, organizations using AI will
need to take into account the extent to which different forms
of transparency might result in the release of sensitive insights
about individuals or expose vulnerable groups to harm.
In the area of maintenance, when the AI recommends a
maintenance decision, decision makers need to understand
the underlying reason. Maintenance analytics developers need
to understand what fault features in the input data are guiding
the algorithm before accepting auto-generated diagnosis
reports, and the maintenance engineer needs to understand
which abnormal phenomena are captured by the inference algorithm
before following the repair recommendations.
One of the proposed benefits of increasing the explainability
of AI systems is increased trust in the system. If maintainers
understand what led to an AI-generated decision or recommendation,
they will be more confident in its outputs. But the
link between explanations and trust is complex. If a system
produces convincing but misleading explanations, users might
develop a false sense of confidence or understanding. They
might have too much confidence in the effectiveness or safety
of systems, without such confidence being justified. Explanations
might help increase trust in the short term, but they do
not necessarily help create systems that generate trustworthy
outputs or ensure that those deploying the system make trustworthy
claims about its capabilities.
REFERENCES:
Galar, Diego, Pasquale Daponte, and Uday Kumar. Handbook of Industry 4.0 and SMART Systems. CRC Press, 2019.
Galar, Diego. Artificial intelligence tools: decision support systems in condition monitoring and diagnosis. Crc Press, 2015.
Galar, Diego, Uday Kumar, and Dammika Seneviratne. Robots, Drones, UAVs and UGVs for Operation and Maintenance. CRC Press, 2020.
50 maintworld 1/2020