Instrumentation & Measurement Magazine 24-9 - 54

away or in lower density from other clusters of normal points,
so they are expected to work better overall. The post-processing
method following DBSCAN, represents an improvement
as it reduces the number of false positives, whilst retaining the
same TP and FN rates. The two examples in Fig. 9 are those
where the classification was performed using a post-processing
after DBSCAN.
As a final remark, we would like to mention that ML
Fig. 10. Outcome of the anomaly detection performed with SVM, DBSCAN,
LOF, a binary OR between DBSCAN and LOF, and post-processing following
DBSCAN methods. TP = True Positives (anomaly correctly detected), TN = True
Negatives (normal point correctly detected), FP = False Positives, FN = False
Negatives (from [36]).
with β*=15 cm, Q'
=15, and strong powering of the Landau octupoles,
with no beam-beam effects. The right plot refers to the
optics configuration of the LHC for the 2016 proton run at injection
energy, Q'
=8, and strong Landau octupoles to combat
electron-cloud effects.
It is rather uncommon that for a given angle the stable amplitude
features considerable variations from seed to seed, so
that its distribution over seeds features outliers. For our purposes,
the outlier identification is carried out in steps [36]. Two
types of ML approaches for the automatic detection of outliers
have been tested. When using SL, the task of outlier detection
is treated as a classification problem and a SVM is trained to
distinguish between normal and abnormal points.
Two UL approaches have also been considered, namely
the Density-Based Spatial Clustering of Applications with
Noise (DBSCAN) method [60] and the Local Outlier Factor
(LOF) [61] algorithm. Fig. 10 reports the comparison between
SVM, DBSCAN, and LOF algorithms. The labels predicted by
DBSCAN and LOF were combined through a binary OR operation
to create a fourth set of labels. A fifth set of labels is
generated by detecting and removing false positives using a
statistical method [36]. An iterative approach is performed
with this post-processing, namely the algorithm starts at the
minimum (maximum) point, and moves outwards (inwards),
recalculating the statistical variables of the regular points at
each step.
The results indicate that the unsupervised methods have
a performance an order of magnitude better than SVM in
terms of false positives. However, they are worse in terms of
false negatives, particularly when using LOF. The main explanation
of this observation is in how the algorithm works. In
supervised learning, the classes are assumed to be balanced
between normal and abnormal points, as the algorithm tries
to learn a decision boundary that separates the two sets of
points. An imbalance induces a bias in the position of the decision
boundary, so false positives are more likely. On the other
hand, the unsupervised methods try to identify points further
54
approaches have recently been applied to the problem of improving
the quality of the model describing the evolution of
DA with time [62]. In this case, Gaussian processes have been
very successfully used to generate synthetic points representing
the DA at a given number of turns. These synthetic points
have been added to those generated by the numerical simulations,
and the whole data set has been used to determine the
model describing the time evolution of the DA. In this way, a
sizable improvement in the quality of the fit has been shown
[62].
Conclusions and Outlook
A selection of ML applications from a variety of domains
linked with beam measurements for the CERN LHC have
been presented and discussed in detail. In all applications, ML
is seen as an efficient way to perform various types of classification
processes on large data sets. One of the first LHC ML
applications originated from the quest to improve optics measurement
and correction by detecting faulty BPMs. Anomaly
detection together with clustering techniques were instrumental
to achieving successful data cleaning. This was followed by
basic Neural Network implementation for optics correction
which is already producing very promising results. Further
improvements are anticipated using an autoencoder network
for improving the quality of betatron-phase measurements
and performing noise reduction on turn-by-turn data.
For the automated alignment of collimators, excellent results
have been obtained thanks to ML techniques which were
used to discriminate between true beam-loss spikes and spurious
events. This has resulted in a remarkable shortening of
the setup time for the collimation system, with the ML implementation
now fully operational. The next step is to move to
the analysis of crosstalk effects between the loss signals of the
two beams, as this could open the possibility to perform parallel
collimator alignment with two circulating beams.
Promising results have also been obtained by using a ML
model for optimizing LHC beam lifetimes. However, rather
than using only the operational data, this necessitated that the
parameter space was properly explored by means of dedicated
machine experiments. Although the ML-based model was
found to describe operational configurations rather well, nonstandard
configurations cannot be predicted, which indicates
a certain lack of predictive power. Hence, exploring alternative
ML approaches is a future line of research.
Detection and classification of beam instabilities by means
of Unsupervised Learning has been applied to the data collected
by the LHC ObsBox system, with the goal of selecting
interesting instability events from a huge collection of less
IEEE Instrumentation & Measurement Magazine
December 2021
Instrumentation & Measurement Magazine 24-9

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