Page 73 - Read Online

Page 166 Ji et al. Intell Robot 2021;1(2):151-75 https://dx.doi.org/10.20517/ir.2021.14

architectures [72,86,97,116,118] . Fifth, the trained deep learning model is put in production with the trained
parameters for real-world applications. Due to the criticality of rail track condition monitoring, redundancy
of inspections by human operators could be provided to double confirm the accuracy. Finally, the efficiency
and effectiveness of the deep learning models are reviewed and enhanced for improved performances.

Various deep learning methods are reported to produce promising results. With more data available from
the rail industry, breakthroughs of deep learning methods, and more advanced and cheaper hardware, deep
learning methods will only become more popular and useful for rail track condition monitoring. Deep
learning models performed well for feature extraction and data classification tasks. The image processing
requirements and the man-made feature extraction efforts are low for deep learning methods, which make
the application economical. However, the nature of rail operations causes the distribution of rail track image
data to be uneven and extremely disproportional, which could cause class imbalance problems in deep
learning applications. The extremely high safety requirement of rail operations and the considerably black-
box nature of deep learning models contradict each other and might cause some trust issues, which is
demonstrated by the redundancies applied to rail track inspections.

Data pre-processing: removing outliers, normalizing data, and applying image process techniques to
enhance the images are common pre-processing techniques. Fourier transforms such as Fast Fourier
Transforms and Short Time Fourier Transform have been applied to transform sequential data (e.g.,
acoustic emission) to 2D spectrograms, which can then be applied to CNN models [65,82,108] .

4. DISCUSSIONS
Condition monitoring and anomaly detection and classification are important to a productive rail
maintenance operation. There are four main types of maintenance strategies: corrective maintenance,
preventive maintenance, proactive maintenance, and predictive maintenance. Deep learning methods can
support the maintenance strategies depending on the tasks it performs. For example, detection of a certain
type of defect will normally trigger corrective maintenance actions to rectify the rail track. The prediction
tasks supported by deep learning methods can be used for predictive maintenance strategies, and they are
becoming more and more popular.

Internet of Things (IoT) technologies can be implemented to support rail maintenance operations. Sensor,
networking, and application layers in IoT can collect big data, which can then be analyzed by deep learning
techniques for application services. In rail track condition monitoring, various types of sensors are being
deployed across the rail network to collect data that need enablers such as deep learning techniques to
unleash their full potential.

Deep learning methods can be more effective when they can be used in real time in the field by the
technicians. Integration of deep learning, IoT, mobile technologies, and edge computing has the potential to
develop useful applications that support the daily rail track maintenance operations.

Deep learning models are normally trained with high computing powers. In order for the deep learning
models to be used in the field, light deep learning models need to be developed so that less powerful but
more accessible devices such as mobile phones can be used with deep learning techniques to support the rail
maintenance operations.