Instrumentation & Measurement Magazine 26-1 - 20

instrumentationand
measurementsystems continued
Fig. 3. A general architecture of a selective always-on wake-up detector based on spectro-temporal decomposition.
understanding of the underlying
physics. In the
recent couple of years, a
new data processing
paradigm called physicsinformed
machine learning
has raised attention in the
research and engineering
community as a promising
way of bridging the gap between
the worlds of physics
and data science [10].
Industrial and
Emerging
Applications
Fig. 4. Detection of AE in the ultrasonic range based on piezoelectric MEMS resonator.
as it usually requires active circuitry based on operational
amplifiers.
MEMS Signal Processing: In the recent years, MEMS technology
has been successfully employed in electronic filtering
applications, mainly for use in communication circuits that
operate in the RF range. Micro-mechanical piezoelectric
resonator arrays have been actively studied as a promising solution
for passive mechanical filtering of acoustic signals in the
ultrasonic frequency range (~100-500 kHz) [9] (Fig. 4). In a context
of AE wake-up detection, MEMS resonators can be used
instead of analog filter banks to further reduce power and increase
the autonomy of the whole I&M system.
AI-empowered Data Processing: Let us now turn back to the
main problem of AE measurements: getting quantitative and
structured information on the object or process under inspection
from raw acoustic signals. Once all the signal processing is
done, a new question arises: how to effectively transform the reconstructed
spectrograms into interpretable information? With
high-performance embedded processing platforms becoming
increasingly available, novel data processing paradigms can be
explored. On one side, data-driven approaches that rely on advanced
statistical clustering, features extraction and machine
learning may provide a powerful set of tools for capturing the
statistical nature of the problem. The other approach is to employ
analytical and numerical modeling in conjunction with
inverse optimization procedures to deconvolve measured
AE signals and the system transfer function, localize emission
sources more accurately, and in general, gain a deeper
20
Continuous monitoring
of industrial machinery
for detection of anomalous
operating modes and
malfunctions has traditionally
been one of the main
application domains for AE systems. As new AI-enabling technologies,
such as those based on neuromorphic computing
hardware, are becoming more widely available, a new generation
of industrial AE systems has emerged. Such system may
take the form of a single rugged, machine-mountable device
that is able to learn from recorded acoustic data without any
prior knowledge on failure modes [11] (Fig. 5). In the training
phase, the device learns the number and key features of characteristic
operating modes of the machine from real-time data
clustering analysis. The feature vectors corresponding to identified
data clusters can then be attributed to " conditionally
good " states of the equipment. During normal operation of the
AE device, new acoustic data are processed, and the obtained
feature vectors are compared to the ones generated during
training. If the features differ too much, a conclusion is made
about the presence of a new anomalous mode of operation.
Fig. 5. An industrial AI-powered machine condition monitoring device,
adapted from [11].
IEEE Instrumentation & Measurement Magazine
February 2023

Instrumentation & Measurement Magazine 26-1

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 26-1