Page 70 - Read Online
P. 70

Ji et al. Intell Robot 2021;1(2):151-75  https://dx.doi.org/10.20517/ir.2021.14     Page 163






















                                           Figure 7. Single Shot multibox Detector architecture.

               • PSPNet utilizes a pyramid parsing module to exploit global context information by aggregating different
                                  [59]
               region-based contexts . A pre-trained CNN with the dilated network strategy is used to extract the feature
               map, on top of which the pyramid pooling module gathers context information. The final feature map size
               is one-eighth of the input image. Using a four-level pyramid, the pooling kernels cover the whole, half of,
               and small portions of the image. They are fused as the global prior, which is then concatenated with the
               original feature map in the final part. It is followed by a convolution layer to generate the final prediction
               map. The local and global clues together make the final prediction more reliable.


               3. REVIEW OF RAIL TRACK CONDITION MONITORING WITH DEEP LEARNING
               The authors systematically searched published peer-reviewed journal articles and papers found in Google
               Scholar. Combinations of keywords such as “rail”, “surface”, “rail track”, “defect”, and “deep learning” were
               used as search keys to find research works published in the application of deep learning techniques to rail
               track condition monitoring and anomaly detection and classification. The review covers work from 2013 to
               2021. In total, we identified 62 relevant research publications to review.


               The trend over time: a clear increasing trend can be observed of the popularity of deep learning approaches
               in rail track condition monitoring applications. Table 1 summarizes the findings. The number of papers
               surged in 2018. Before 2018, machine learning techniques other than deep learning approaches were more
               widely adopted. The rail industries are adopting deep learning methods with growing interests. An upwards
               trend of publication number is observed. There is also a gap of a few years from the invention of a deep
               learning model to its adoption by the rail industry.


               Regions of study: fourteen regions are represented by the papers identified. Among them, China has the
               highest number of papers, which indicates the popularity of rail-related research work corresponding to the
               expanding rail networks across the country. Papers from China surged in 2018 and kept a high number in
               the following years. Table 2 summarizes the distribution of papers over regions.

               Raw data type: it is observed that 70% of studies used image-type raw data for the deep learning models.
               Nevertheless, acoustic emission signals [65,71,100,103,108] , defectogram [96,109] , speed accelerations , concatenated
                                                                                           [98]
                                        [101]
               vector of curve and numbers , current signal , maintenance records [80,99] , synthetic data from generative
                                                       [89]
                                                                   [60]
                                                                                                       [119]
                                                                                   [87]
               model , time-frequency measurement data , time-series , geometry data , and vibration signal
                    [63]
                                                       [82]
               could all be possible input data sources as well.
   65   66   67   68   69   70   71   72   73   74   75