EXPERIMENTAL INVESTIGATION FOR THE CLASSIFICATION OF RAIL DEFECTS AND OBSTACLES IN RAILWAYS USING DEEP LEARNING

• Main Author: Sakhinatul Putri, Yurixa
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/47136
• Institution: Institut Teknologi Bandung

The increase in the number of train users in Indonesia has not been accompanied by the increase in security and good safety so that the number of accidents still quite frequent every year. Some of the factors that cause the accident to occur are due to damage to the railroad tracks, incomplete rail components, obstacles in the railroad tracks, etc. What is usually done in a railroad safety management system is to inspect rails with a human visual. Rail checks conducted by human visuals are considered to be no longer practical so that it can be replaced using a digital camera sensor. In train system, digital cameras are generally only used to monitor and record railroad conditions. However, the monitoring results cannot detect any damage and obstacles in front of the railroad in real time. Therefore, it is necessary to have an object detection system that serves to provide prevention to the train to avoid accidents. This research has designed a miniature LEGO train system to detect damage and obstacles in railway with deep learning-based. Detection is using a camera sensor. This camera sensor able to recognize 3 class of condition, there are ‘broken’ condition (rails broken at the connection), ‘curved’ condition (rails bent or wavy) and ‘erosion’ condition (rail barrier in the form of sand). Preventive action is also done by controlling the speed of the train motor when an object is detected based on the value of confidence interval through wireless media. Taking of image training is carried out as many as 300 images with each label has 100 images. It is done with variations in position and angle, variations in the width of the rail gap cut off at the connection, and variations in the height of the sand above the sleepers. The results of image testing are done by calculating several performance criteria based on the confusion matrix. The average values of precision are 0.85, recall 0.93, accuracy 0.80 and F1 Score 0.88. These values can be said to be quite good because they are close to 1 which shows the good performance of a model. The designed of the train motor speed control system was successfully implemented in real-time based on the value of the level of confidence. When the camera sensor does not detect the object, the sensor will send a control signal to maximize the speed of the train motor based on the PWM value. When the camera sensor detects an object with a confidence level above 20-70%, the sensor will send a control signal to reduce train speed. However, when the sensor detects an object with a confidence level above 70%, the sensor will send a control signal to minimize the speed of the train motor so that the train stops. Control testing is done by varying the distance of the train to the camera's perspective and calculated the distance of the train to the object when it stops. It shows that the average value of the distance of a train to an object while stop is very important to determine the value of a train's distance to the camera's perspective. The farther the distance of the train to the camera's perspective, the further the train is to the object when stopped. its because the system is able to detect objects from a distance so that the control will process it early.