Page 9 - Read Online

Ji et al. Intell Robot 2021;1(2):151-75 https://dx.doi.org/10.20517/ir.2021.14 Page 153

[4]
method is particularly suitable to inspect steel rails even under high temperatures . Due to the harsh
[5]
environment, late hours, and tired patrollers, the inspection accuracy might be affected . Rail inspection
vehicle and sensor technologies are being deployed as an efficient and cost-effective data collection
technology solution to support rail maintenance operations and can capture vast amounts of data. Data-
driven automatic condition monitoring and detection and classification of rail track anomalies have been
attracting attention from researchers at universities and railway institutes.

Deep learning methods are producing successes in various applications with the recent advances of the
techniques. Deep learning has neural networks as its functional unit to mimic how the human brain solves
complex problems based on data. Methods such as long-short-term memory (LSTM) as a type of recurrent
neural network (RNN) and convolutional neural network (CNN) propelled the development of deep
learning and the field of artificial intelligence and have been reported with convincing performances in
monitoring conditions for tools, machines, and turbines. The performances in prediction and learning of
[6-8]
these methods are improving with the increasing amount of data available . Deep learning methods have
been adopted for rail track condition monitoring and anomaly detection and classification.

Some research questions are formulated to guide our review with clear purposes. Subsequently, we select the
publications for the detailed review.

• What types of deep learning models are available for rail track condition monitoring?

• What deep learning techniques can be useful for applications in rail track condition monitoring?

• What types of rail track anomalies are more commonly chosen to identify?

• What types of data are collated for the deep learning applications?

• Where are the specific objectives of applying deep learning models to rail track condition monitoring?

• What are the deep learning data pre-processing methods adopted?

• What are the challenges that researchers face?

• How does the application in rail track condition monitoring correspond to the evolution of deep learning
techniques?

• What are the trend and insights for future directions of research and practice?

Figure 1 shows the review framework that is proposed by this paper to address the list of research questions.
The importance, the types of defects, and the existing manual inspection techniques of rail track monitoring
are presented to give the context and introduction of this study. The shortcomings of manual inspection
techniques partially provide the need to adopt deep learning methods. We then review the deep learning
methods available and their relevance by briefly discussing the evolution of the deep learning field and
describing the deep learning models for the ease of selecting suitable models for tasks. The studies applying
deep learning methods to rail track condition monitoring are then reviewed where summaries are made
according to the trend over time, the region of study, the raw data type, the pre-processing data, the purpose