Instrumentation & Measurement Magazine 24-9 - 52

anomalous samples from the entire set of ObsBox data in
view of more detailed analyses. At this stage, each bunch is
considered as independent, implying that phenomena like
coupled-bunch instabilities are disregarded, which greatly
simplifies the cluster identification as this then applies to a
univariate time series. After having determined the clustering,
one can proceed with a Hierarchical Clustering Algorithm
[49] to identify iteratively and link clusters in a form of a
dendrogram.
The experience gathered so far is that signals with similar
patterns are, for the most part, clustered together, although
some issues have been observed, linked with the fact that the
designed clustering algorithm does not allow for a partial
match of the time series. In summary, there is an enormous potential
for ML techniques to correctly identify false triggers,
thus allowing for a more efficient and detailed study of the relatively
few data sets that contain instability data.
Pressure Readings and Heating
Detection
Charged particles stored in a high-energy, high-intensity accelerator
ring may generate heating in the surrounding
equipment due to electron cloud effects [50], particles lost on
the beam surroundings [51], synchrotron radiation effects [52],
and beam-induced RF heating due to impedance effects [53].
In a high-vacuum environment, temperature increase may
lead to outgassing [54] that appears as a pressure increase in
vacuum gauges, and Beam-induced heating can be directly
measured by means of temperature probes. However, standard
vacuum monitoring is more widespread and systematic
than temperature monitoring, e.g., more than 1200 vacuum
gauges are distributed along LHC ring circumference [56].
Pattern analysis of the vacuum gauges readings after each
LHC fill to identify abnormal behavior is a time-consuming
and heavy task. Furthermore, there is no reliable technique to
convert the vacuum measurements into equivalent temperature
values.
The pressure readings produced by the vacuum gauges are
analyzed by a newly developed automatic classification algorithm
in view of detecting heating patterns from an anomalous
pressure increase. Fig. 7 shows the time evolution of the pressure
of a Beam 2 vacuum gauge situated in Sector 4-5 of the
LHC, in the vicinity of the stand-alone magnets D4 and Q5.
Sudden changes in pressure readings are clearly visible and
may be linked with outgassing generated by a temperature
rise. Note that we disregard the underlying physical phenomenon
inducing the temperature rise as only its effects are
considered.
To speed up finding the anomalous gauge readings, the
classifier selects a subset of all gauges containing such signals,
i.e., the classifier should aim at a high recall score, defined as
the fraction of true positives detected from the total number
of positive cases.
More than 700 vacuum gauge readings were labelled as
anomalous through human expert supervision, and a data set
containing 700 time series of 3000 time steps each was created.
52
Fig. 7. Typical data (pressure readings and beam intensity) used to estimate
beam heating effects from pressure readings of a Beam 2 vacuum gauge
situated in Sector 4-5 of the LHC, in the vicinity of the stand-alone magnets D4
and Q5, during a physics fill of the 2015 run. The sudden pressure increases
may indicate a beam-induced heating effect close to the vacuum gauge (from
[36]).
A PCA was then performed, resulting in the retention of
only 12 features that retain no particular physical meaning but
explain 99.9% of the full data set variance, hence without leading
to a significant loss of information. A K-Nearest Neighbour
Classifier (KNN) followed by a Multi-Layer Perceptron (MLP)
were then trained on the resulting data set containing only
these 12 features (see [36] for details).
The performance of the KNN and MLP classifiers was evaluated
by means of a 4-fold cross-validation approach [57],
applied when training each model. The 4-fold technique implemented
Stratified splitting [58], the folds being built by
preserving the percentage of samples for each class.
Fig. 8a shows the results of the parameter set scan maximising
the recall for the KNN. The red dots, corresponding to
recall=1 in the KNN classifier, are parameter sets for which the
algorithm overfits the training data, i.e., it behaves extremely
well on training data, but poorly on the test data.
At index 7 in the plot, one can find the best parameter set,
i.e., the first set of parameters that does not overfit the training
set, for which recall=0.60 ± 0.09. The results of the MLP
scan over the network parameters is shown in Fig. 8b. The best
parameter set is achieved at index 0, corresponding to a network
made of 2 hidden layers with 176 neurons per layer. For
the neural network, recall=0.86 ± 0.10, an improvement of 35%
with respect to the KNN case. Hence, a rather simple Neural
Network features interesting recall scores, thus motivating the
test of more refined ML techniques on this task, such as Convolutional
Neural Network and ensemble methods (see [36]).
Digression: ML Learning Applied to Simulated Measurements
of Dynamic Aperture
This section is devoted to the presentation and discussion
of applications of ML to numerical simulations, hence not
beam measurements, of nonlinear beam dynamics. Among
various concepts that can be used to describe nonlinear beam
dynamics, that of Dynamic Aperture (DA) is particularly
useful. It is the radius of the smallest sphere inscribed in the
connected volume in phase space where bounded motion over
a given time interval occurs [59].
Tracking simulations are used to estimate the DA by probing
the evolution of a certain set of initial conditions, usually
uniformly distributed in polar coordinates in normalized
IEEE Instrumentation & Measurement Magazine
December 2021

Instrumentation & Measurement Magazine 24-9

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 24-9

Instrumentation & Measurement Magazine 24-9 - Cover1
Instrumentation & Measurement Magazine 24-9 - Cover2
Instrumentation & Measurement Magazine 24-9 - 1
Instrumentation & Measurement Magazine 24-9 - 2
Instrumentation & Measurement Magazine 24-9 - 3
Instrumentation & Measurement Magazine 24-9 - 4
Instrumentation & Measurement Magazine 24-9 - 5
Instrumentation & Measurement Magazine 24-9 - 6
Instrumentation & Measurement Magazine 24-9 - 7
Instrumentation & Measurement Magazine 24-9 - 8
Instrumentation & Measurement Magazine 24-9 - 9
Instrumentation & Measurement Magazine 24-9 - 10
Instrumentation & Measurement Magazine 24-9 - 11
Instrumentation & Measurement Magazine 24-9 - 12
Instrumentation & Measurement Magazine 24-9 - 13
Instrumentation & Measurement Magazine 24-9 - 14
Instrumentation & Measurement Magazine 24-9 - 15
Instrumentation & Measurement Magazine 24-9 - 16
Instrumentation & Measurement Magazine 24-9 - 17
Instrumentation & Measurement Magazine 24-9 - 18
Instrumentation & Measurement Magazine 24-9 - 19
Instrumentation & Measurement Magazine 24-9 - 20
Instrumentation & Measurement Magazine 24-9 - 21
Instrumentation & Measurement Magazine 24-9 - 22
Instrumentation & Measurement Magazine 24-9 - 23
Instrumentation & Measurement Magazine 24-9 - 24
Instrumentation & Measurement Magazine 24-9 - 25
Instrumentation & Measurement Magazine 24-9 - 26
Instrumentation & Measurement Magazine 24-9 - 27
Instrumentation & Measurement Magazine 24-9 - 28
Instrumentation & Measurement Magazine 24-9 - 29
Instrumentation & Measurement Magazine 24-9 - 30
Instrumentation & Measurement Magazine 24-9 - 31
Instrumentation & Measurement Magazine 24-9 - 32
Instrumentation & Measurement Magazine 24-9 - 33
Instrumentation & Measurement Magazine 24-9 - 34
Instrumentation & Measurement Magazine 24-9 - 35
Instrumentation & Measurement Magazine 24-9 - 36
Instrumentation & Measurement Magazine 24-9 - 37
Instrumentation & Measurement Magazine 24-9 - 38
Instrumentation & Measurement Magazine 24-9 - 39
Instrumentation & Measurement Magazine 24-9 - 40
Instrumentation & Measurement Magazine 24-9 - 41
Instrumentation & Measurement Magazine 24-9 - 42
Instrumentation & Measurement Magazine 24-9 - 43
Instrumentation & Measurement Magazine 24-9 - 44
Instrumentation & Measurement Magazine 24-9 - 45
Instrumentation & Measurement Magazine 24-9 - 46
Instrumentation & Measurement Magazine 24-9 - 47
Instrumentation & Measurement Magazine 24-9 - 48
Instrumentation & Measurement Magazine 24-9 - 49
Instrumentation & Measurement Magazine 24-9 - 50
Instrumentation & Measurement Magazine 24-9 - 51
Instrumentation & Measurement Magazine 24-9 - 52
Instrumentation & Measurement Magazine 24-9 - 53
Instrumentation & Measurement Magazine 24-9 - 54
Instrumentation & Measurement Magazine 24-9 - 55
Instrumentation & Measurement Magazine 24-9 - 56
Instrumentation & Measurement Magazine 24-9 - 57
Instrumentation & Measurement Magazine 24-9 - 58
Instrumentation & Measurement Magazine 24-9 - 59
Instrumentation & Measurement Magazine 24-9 - 60
Instrumentation & Measurement Magazine 24-9 - 61
Instrumentation & Measurement Magazine 24-9 - 62
Instrumentation & Measurement Magazine 24-9 - 63
Instrumentation & Measurement Magazine 24-9 - 64
Instrumentation & Measurement Magazine 24-9 - 65
Instrumentation & Measurement Magazine 24-9 - 66
Instrumentation & Measurement Magazine 24-9 - 67
Instrumentation & Measurement Magazine 24-9 - 68
Instrumentation & Measurement Magazine 24-9 - 69
Instrumentation & Measurement Magazine 24-9 - 70
Instrumentation & Measurement Magazine 24-9 - 71
Instrumentation & Measurement Magazine 24-9 - 72
Instrumentation & Measurement Magazine 24-9 - 73
Instrumentation & Measurement Magazine 24-9 - 74
Instrumentation & Measurement Magazine 24-9 - 75
Instrumentation & Measurement Magazine 24-9 - 76
Instrumentation & Measurement Magazine 24-9 - 77
Instrumentation & Measurement Magazine 24-9 - 78
Instrumentation & Measurement Magazine 24-9 - 79
Instrumentation & Measurement Magazine 24-9 - 80
Instrumentation & Measurement Magazine 24-9 - Cover3
Instrumentation & Measurement Magazine 24-9 - Cover4
https://www.nxtbookmedia.com