Efficient Plant April 2019 - 12

feature | operating strategies

It would be
physically
impossible
for a human
analyst to
process the
quantity of
data that AI,
ML, and IIoT
advancements
create.

easily identify any among them that are abnormal.
 Classification can be used to identify a specific defect
and provide a diagnosis. For example, classification
algorithms can go beyond simply telling personnel that
there is an anomaly to actually predicting cavitation in
a pump.
For classification to work, you need to acquire a
body of historical performance data, which can require
considerable time and money. In the case of cavitation,
this information would be identified as data from actual
instances of pump cavitation and then used to train the
predictive model to recognize conditions that signal
future cavitation incidents..

WHY (AI) NOW?
Khushraj acknowledges that anomaly detection and
classification aren't new concepts. "Vibration-analysis
software," he explained, "had rules-based expert systems to codify knowledge and diagnose problems 15
to 20 years ago." But rules-based systems have several
issues: They're painstaking to develop. They're only
as good as the few people codifying the rules, and
as, Khushraj put it, "They're really hard to maintain,
especially when presented with new failure patterns.
As a result, most plants didn't implement rules-based
systems. Moreover, not enough data was collected at a
scale to truly test their efficacy."
Machine-learning-based systems eliminate such
problems by minimizing the need for extensive
human intervention. Unsupervised anomaly detection doesn't require any human intervention. Even
supervised classification systems only need an adequate amount of labeled data upfront.
"Once you get this data," Khushraj said, "you
can re-purpose the models across a wide variety of
use-cases."
As for the question of "why (AI) now" Khushraj
pointed to three trends driving the explosion of interest in machine learning for PdM:
 Decreased sensor costs means it is possible to equip
more assets with permanent sensors and collect substantial data at scale.
 Cloud technology makes it affordable to securely
process, collect, and store data.
 Machine learning can be applied to make sense of all
data that's being generated.

| EFFICIENTPLANTMAG.COM

THE HUMAN FACTOR
Technology by itself won't change anything if there
isn't support in implementing it. Success or failure of
any initiative depends on the human dimension, so it
is important to understand the perceptions of people
who will be working with it.
Khushraj said his organization has found that the
three main reasons people are reluctant to try MLbased technology are all based on the human element.
The most common reason is associated with the possibility of the system missing failures. "There's been a
lot of hype around ML for PdM," he explained. "And
people are skeptical of promises that ML will solve
everything. As a result, it can be difficult to tell what is
reality, and no one wants to put his or her job on the
line for technology that isn't going to work. "
Another common fear is that of "we will be
replaced by computers." In manufacturing today,
anything that seems like it could result in headcount
reduction is sensitive. This often is expressed in statements along the lines of, "We're already doing PdM
using walkarounds and it's working for us." According
to Khushraj, that approach may, in fact, be working
well, but there could still be opportunities to improve
through more frequent data collection or automated
analysis.
A third reason people are reluctant to implement
ML-based systems for PdM involves the required
skills. Most plants don't have data scientists on staff,
and the hurdle to retrain for these jobs is large.

ML CAN HELP
Today, most plants are conducting PdM by using
in-house personnel or third-party service providers to
collect data on a monthly or quarterly basis. Khushraj
points out that, whether done in-house or outsourced,
the interval between readings can result in failures
that develop more quickly than the intervals. "The
falling cost of wireless sensors," he said, can help to
address this by providing more-frequent data."
But, Khushraj continued, more-frequent data
also means more human analysis, which often isn't
available. He described the experience of electric
utility Arizona Public Service (APS), Phoenix (aps.
com), which implemented a wireless PdM program
and, within the first six months, collected more than

APRIL 2019

http://www.EFFICIENTPLANTMAG.COM

Efficient Plant April 2019

Table of Contents for the Digital Edition of Efficient Plant April 2019