Process Quality Monitoring Through a LSTM Network Derived from a Rule-Based Approach

Abstract

The railway infrastructure condition is a crucial factor for the
safe and efficient operation of trains. Regular maintenance
is inevitable as the track geometry degrades over time due
to traffic and environmental effects. To restore the ideal position
and provide sufficient durability of ballasted track so
called tamping machines are used. These machines lift the
track, correct the longitudinal level and the alignment of the
track panel and tamp the ballast. During the tamping process
the tamping tines penetrate the ballast bed, fill voids and compact
the ballast underneath the sleepers by a squeezing movement
with superimposed vibration. A detailed description of
the tamping cycle can be found on section 2. Monitoring and
evaluating this tamping process is essential for maintaining
process quality. This can be achieved through a variety of
sensors, such as incremental encoders, angle encoders, temperature,
pressure, and acceleration sensors, coupled with a
measurement unit (DAQ and edge device) to collect, locally
store and transmit the data to a cloud. This paper explores the
development of a rule-based algorithm for assessing the quality
of the tamping process execution in reference to its nominal
chronological sequence. The focus is on identifying tamping
occurrences and classifying them into acceptable (OK)
or non-acceptable (NOK) categories. This involves selecting
relevant measurement parameters and processing them,
considering the inherent imprecision in real-world processes.
Empirical thresholds are established to differentiate between
good and bad outcomes. The classification approach has to
be sufficiently generic in order to cover a high variety of customized
tamping machine types. As each machine is individually
designed, the process of generalization is challenging
and complex. The paper demonstrates the accuracy and universal
applicability of the developed rule set across different
tamping machines. The model’s effectiveness is validated using
the Hold-Out-Test-Set method. Furthermore, the rule-set-
achieved outcomes are compared with results gained from an
LSTM network. Both the rule-based approach and the neural
network demonstrate precision, but the latter requires significantly
more effort.