Health Assessment of Pump Stations using Time Series
Anomaly Detection: Deploying AI on the Industrial Edge
Abhishek Murthy
1
, Babak Afshin-Pour, Willem Malloy and Vasileios Geroulas
Schneider Electric
1
abhishek.murthy@se.com
Supervisory Control and Data Acquisition (SCADA) is
widely used to manually monitor and manage distributed
physical assets. Supporting infrastructure was designed and
optimized for that need. Specialized communication proto-
cols are utilized for applications which span large geograph-
ical deployments. These protocols ensure data robustness
and consistency in variable-quality network environments.
However, the resulting data, while forming enterprise data
pipelines, lacks granularity and has irregular time spacing,
making it unsuitable for machine-learning-based predictive
maintenance applications.
We present a hybrid cloud-to-edge health monitoring solution
for assets connected to SCADA or other legacy control sys-
tems. Our solution uses a modbus-based polling system on
the edge, to collect data at a much higher granularity than
the adjacent SCADA system, letting us detect even subtle
and acute patterns in the data. Note that no new sensors are
needed, as we connect to the same registers as the existing
SCADA system. The high granularity data is assessed at the
edge for anomalies, using time series anomaly detection algo-
rithms. We then synthesize the predictions into a health index
that quantifies the recency and the frequency of the detected
anomalies for the asset. The health index is then transmit-
ted to a web-based application, where the user can configure
thresholds for generating alerts based on the criticality of the
asset.
We demonstrate our solution in a case study, where the ap-
plication was deployed using Schneider Electric’s Customer
First Digital Hub, to monitor a sewage pump station for
blockages and other subtle deviations in operating patterns.
1. INTRODUCTION
Wastewater management involves a network of intercon-
nected assets whose purpose is to reliably move wastewater
from homes and workplaces back to centralized processing
facilities. Failures in these operational assets must be very
Abhishek Murthy et al. This is an open-access article distributed under the
terms of the Creative Commons Attribution 3.0 United States License, which
permits unrestricted use, distribution, and reproduction in any medium, pro-
vided the original author and source are credited.
rapidly addressed to ensure the continuity of critical services
to our homes and cities. There are several failure modes in
the complex industrial assets. Blockages in pipes may lead
to overflows, as there is no redundancy in the infrastructure.
As a critical active component of wastewater networks, mo-
tors and pumps keep wastewater flowing. Pumps may be in-
stalled as redundant pairs, or as single units. Their failure can
lead to significant financial costs related to repair and restora-
tion. The regulatory climate for compliance and safety has
also motivated the various stakeholders to maintain wastew-
ater treatment facilities with the latest technology and over-
all effectiveness. The community has conventionally relied
on scheduled or surveillance-based maintenance. Such tradi-
tional approaches are often more expensive and may not even
enable timely interventions. To this end, the wastewater treat-
ment industry has untapped potential to benefit significantly
from machine-learning-based predictive maintenance.
Wastewater is hazardous prior to treatment. Any untreated
discharges, caused by failure events that are not addresses in
time, can harm the environment and the public health. On a
human and cost front, failures can occur unpredictably, mak-
ing it difficult for organizations to plan effectively. On a
human level, people must respond at any hour of the day.
They are often recompensed via overtime, which ultimately
the users of the system pay for.
Predicting failure generally relies on proactive identification
of leading indicators. In complex systems, those leading in-
dicators are not easily identifiable either to human operators
or to the simplistic SCADA alarming mechanisms. Each as-
set produces a constant stream of data from interdependent
variables read from multiple sensors. This inherent complex-
ity does not lend itself to automated machine learning-based
modeling and processing.
Moreover, wastewater treatment facilities present some
unique challenges for predictive maintenance. The makeup of
the waste water changes constantly with time. As the assets
age, the types of failures also change. The existing SCADA
systems that are used for monitoring them are woefully ill-
equipped for supporting advanced predictive maintenance
1
