A New Form of Interlocking Developing Technology for Level Crossings and Depots with International Applications

Fayyaz, Muhammad Asad Bilal (2021) A New Form of Interlocking Developing Technology for Level Crossings and Depots with International Applications. Doctoral thesis (PhD), Manchester Metropolitan University in collaboration with OSL Limited.

Preview

Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (8MB) | Preview

Abstract

There are multiple large rail infrastructure projects planned or currently being undertaken within the United Kingdom. Many of these projects aim to reduce the continual issue of limited or overcapacity service. These projects involve an expansion of Rail lines, introducing faster lines, improved stations in towns and cities and better communication networks. Some major projects like Control Period 6 (CP6) are being managed by Network Rail; where projects are initiated throughout Great Britain. Many projects are managed outside Great Britain e.g., Trans-European Transport Network Program, which is planning for expansion of Rail lines (almost double) for High-Speed Rails (category I and II). These projects will increase the number of junctions and Level Crossings. A Level Crossing is where a Rail Line is crossed by a road or a walkway without the use of a tunnel or bridge. The misuse from the road users account for nearly 90% of the fatalities and near misses at Level Crossings. During 2016/2017, the Rail Network recorded 6 fatalities, about 400 near-misses and more than 77 incidents of shock and trauma. Accidents at Level Crossings represent 8% of the total accidents from the whole Rail Network. Office of Rail and Road (ORR) suggested that among these accidents at Level Crossings 90% of them are pedestrians. Such high numbers of accidents, fatalities and high risk have alarmed authorities. These authorities found it necessary to invest time and utilise given resources to improve the safety system at a Level Crossing using the safer and reliable interlocking system. The interlocking system is a feature of a control system that makes the state of two functions mutually independent. The primary function of Interlocking is to ensure that trains are safe from collision and derailment. Considering the risk associated with the Level Crossing system, the new proposed interlocking system should utilise the sensing system available at a Level Crossing to significantly reduce implementation cost and comply with the given standards and Risk Assessments. The new proposed interlocking system is designed to meet the “Safety Integrity Level- SIL” and possibly use the “2oo2” approach for its application at a Level Crossing, where the operational cycle is automated or train driver is alarmed for risk situations. Importantly, the new proposed system should detect and classify small objects and provide a reasonable solution to the current risk associated with Level Crossing, which was impossible with the traditional sensing systems. The present work discusses the sensors and algorithms used and has the potential to detect and classify objects within a Level Crossing area. The review of existing solutions e.g Inductive Loops and other major sensors allows the reader to understand why RADAR and Video Cameras are preferable choices of a sensing system for a Level Crossing. Video data provides sufficient information for the proposed algorithm to detect and classify objects at Level Crossings without the need of a manual “operator”. The RADAR sensing system can provide information using micro-Doppler signatures, which are generated from small regular movements of an obstacle. The two sensors will make the system a two-layer resilient system. The processed information from these two sensing systems is used as the “2oo2” logic system for Interlocking for automating the operational cycle or alarm the train drive using effective communication e.g., GSM-R. These two sensors provide sufficient information for the proposed algorithm, which will allow the system to automatically make an “intelligent decision” and proceed with a safe Level Crossing operational cycle. Many existing traditional algorithms depend on pixels values, which are compared with background pixels. This approach cannot detect complex textures, adapt to a dynamic background or avoid detection of unnecessary harmless objects. To avoid these problems, the proposed work utilises “Deep Learning” technology integrated with the proposed Vision and RADAR system. The Deep Learning technology can learn representations from labelled pixels; hence it does not depend on background pixels. The Deep 3 | P a g e Learning technology can classify, detect and localise objects at a Level Crossing area. It can classify and differentiate between a child and a small inanimate object, which was impossible with traditional algorithms. The system can detect an object regardless of its position, orientation and scale without any additional training because it learns representation from the data and does not rely on background pixels. The proposed system e.g., Deep Learning technology is integrated with the existing Vision System and RADAR installed at a Level Crossing, hence implementation cost is significantly reduced as well. The proposed work address two main aspects of training a model using Deep Learning technology; training from scratch and training using Transfer Learning techniques. Results are demonstrated for Image Classification, Object Detection and micro-Doppler signals from RADAR. An architecture of Convolutional Neural Network from scratch is trained consisting of Input Layer, Convolution, Pooling and Dropout Layer. The model achieves an accuracy of about 66.78%. Different notable models are trained using Transfer Learning techniques and their results are mentioned along with the MobileNet model, which achieves the highest accuracy of 91.9%. The difference between Image Classification and Object Detection is discussed and results for Object Detection are mentioned as well, where the Loss metrics are used to evaluate the performance of the Object Detector. MobileNet achieves the smallest loss metric of about 0.092. These results clearly show the effectiveness and preferability of these models for their applicability at Level Crossings. Another Convolutional Neural Network is trained using micro-Doppler signatures from the Radar system. The model trained using the micro-Doppler signature achieved an accuracy of 92%. The present work also addresses the Risk Assessment associated with the installation and maintenance of the system using Deep Learning technology. RAMS (Reliability, Availability, Maintainability and Safety) management system is used to address the General and Specific Risks associated with the sensing system integrated with the Deep Learning technology. Finally, the work is concluded with the preferred choice, its application, results and associated Risk Assessment. Deep Learning is an evolving field with new improvements being introduced constantly. Any new challenges and problems should be monitored regularly. Some future work is discussed as well. To further improve the model's accuracy, the dataset from the same distribution should be gathered with the cooperation of relevant Railway authorities. Also, the RADAR dataset could be generated rather than simulated to further include diversity and avoid any biases in the dataset during the training process. Also, the proposed system can be implemented and used in different applications within the Rail Industry e.g., passenger census and classification of passengers at the platform as discussed in the work.