IEEE Computational Intelligence Magazine - February 2023 - 11

automatic systems to assure high repeatability,
which is a key feature of the
Zero-Defect Manufacturing Paradigm.
In the manufacturing market and,
in general, in all industrial markets
including food and agriculture, many
AI-based solutions are developed to
automate different tasks like robotarm
guidance, automatic product
picking, and quality control of complex
shaped objects. Today, AI and
machine vision can solve many tasks
autonomously.
Prognostics and Health Management
(PHM) for industrial assets to reduce/eliminate
unplanned maintenance events.
High-value assets, such as turbines,
medical scanners, and aircraft engines
must maintain their operational performance
over time, improving their
utilization and availability, while minimizing
their maintenance cost. We
can achieve this by reducing or eliminating
unplanned maintenance events
using PHM-enabled condition-based
maintenance. This is critical for the
efficient execution of contractual service
agreements offered by Original
Equipment Manufacturers to their
customers.
PHM has two main components:
Health Assessment,which is the
evaluation and interpretation of the
asset's current and future health states,
based on descriptive and predictive analytics;
and Health Management,
which is the control, operation, and
logistic plans to be executed in
response to such assessment, based on
prescriptive analytics [2].
Health Assessment: (1) Anomaly
Detection; (2) Diagnostics; and (3)
Prognostics
1) Anomaly Detection (AD). Platformdeployed
sensors collect and preprocess
the assets data via segmentation,
filtering imputation, and validation.
This data is summarized by extracting
feature subsets that provide a
more compact, robust representation
of its information content. The
features are analyzed by anomaly
detection (AD) models to assess the
degree ofabnormal behavior ofeach
asset in the fleet. Such assets are
identified, recording the time of
their first anomaly. AD leverages
unsupervised learning techniques,
such as clustering, to extract the
underlying structural information
from the data, to define normal
structures and regions, and to identify
departures from such regions.
Ensembles of unsupervised models
are usually deployed to prevent the
occurrence offalse positives.
2) Diagnostics (Dx). The information
generated by the AD models allows
Diagnostic models to focus on the
assets' relevant subsystems, analyzing
key variables and matching their
patterns with a library of signatures
associated with faults or incipient
failure modes. The result is usually
an ordered list of probable faults.
Diagnostics leverages supervised
learning techniques (classification)
by mapping a feature space into a
labeled fault space. Label generation
in the training set is one of the
critical tasks for the development of
diagnostics models.
3) Prognostics. Prognostics is the prediction
of the asset's Remaining Useful
Life (RUL), once the anomaly
detection and diagnostics modules
identified an incipient failure
through its preceding faults. This
incipient failure changes the trajectory
ofRUL versus time from a linear,
normal-wear trajectory to an
exponentially decaying one. The
fault time and incipient failure mode
determine the inflection point in
such a curve and the deterioration
steepness, respectively. These estimates
are usually in units of time or
utilization cycles and have associated
uncertainty, such as a probability
density curve around the actual estimate.
This uncertainty (RUL confidence
interval) increases as the
prediction horizon is extended.
Operators can choose a confidence
level that allows them to incorporate
a risk level into their decision-making.
They can change operational
characteristics, such as load, which
may prolong the life of components
at risk. They can also account for
upcoming maintenance and set in
motion a logistics process to support
a smooth transition from faulted
equipment to fully functioning.
Prognostics leverages supervised learning
techniques (regression) to forecast
RUL from time-series data. Prognostics
is the most difficult of the PHM
functions because RUL depends on an
estimate of the current deterioration
FIGURE 1 The six verticals of the IEEE Conference on Artificial Intelligence (CAI) 2023.
FEBRUARY 2023 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE 11
IEEE Computational Intelligence Magazine - February 2023

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - February 2023
