IV. MONITORING SYSTEM
The data from sensors is treated as either a time series,
where data are produced continuously or periodically, or a
sequence of readings where data is generated ad hoc, for
example, generated every time a train passes. The data can
be monitored by searching for thresholds (triggers), known
problem signatures (classification), identifying unknown events
(short-term analysis using outlier detection), or identifying drift
over a longer period of time (long-term outlier detection).
Condition monitoring can be performed continuously or
periodically. Continuous monitoring should detect a problem
straight away but it is often expensive; energy hungry, which
is a problem for WSNs where the network components need
power; and the sensor data are very noisy, which requires
careful preprocessing to ensure accurate diagnostics. Periodic
monitoring is cheaper, uses less energy, and allows time for data
cleaning and filtering but a problem will only be diagnosed at
the next processing run. This may be acceptable for some situations
that change slowly such as cracks developing in bridges
but for time critical scenarios, then continuous monitoring is
necessary.
In basic condition monitoring, the system is only able to
distinguish between normal and abnormal conditions (no fault
or fault). Beyond this, systems exhibit increasing levels of monitoring
sophistication as outlined by the staircase of structural
health monitoring of Lopez-Higuera et al. [78] described in
Section I. A Level 5 system will comprise complex hardware,
custom algorithms, and software to allow the diagnosis and/or
the prognosis and even the solutions.
The topology of WSNs often varies over time. One very
important factor in this topological variation is the mobility of
the sensor nodes. Lorestani et al. [79] subdivided the communications
network for their WSN into two: the fixed network
relates to sensor nodes in fixed locations such as bridges, tunnels,
and special points, whereas the movable network relates
to sensor nodes attached to locomotives or rail wagons. The
data for the movable network are logged with accompanying
GPS coordinates. Movable (on-board) sensors can monitor the
whole track length travelled by the train but only monitor the
sections of the train where the sensor nodes are attached [84].
In contrast, track-mounted (fixed) sensor nodes can measure
the whole train as it passes but only at specific points where
the nodes are mounted on the track. This tradeoff needs to be
considered when designing the node placements. The following
sections analyze existing WSNs for condition monitoring on
the railways and conducts analyzes from the perspective of
fixed (static) and movable (dynamic) monitoring [122]. In each
category, the systems are grouped according to the object(s)
being monitored.