Inertial Measurements and Neural network aided enhancements for railway positioning system

The accurate estimation of train and individual coach position and speed is crucial for the safe and efficient operation of high-speed train systems. The use of precise positional data can minimize train and passenger delays, optimize capacity utilization, and enable the development of new customer services while keeping maintenance costs low. The objective of this thesis is to propose novel methods for mitigating or eliminating time-dependent errors in Inertial Navigation Systems (INS) that utilize low-cost inertial sensors, particularly in train positioning applications. This is particularly relevant in scenarios where the Global Navigation Satellite System (GNSS) signals are unavailable, such as inside tunnels or in areas with heavy multipath, such as urban environments. The objective is accomplished through the detection and integration of contextual information, referred to as pseudo-measurements, as well as the introduction of a new type of measurements known as semi-nominal measurements. This study proposes the generation and incorporation of semi-nominal measurements into the train navigation system. To achieve this, suitable algorithms are proposed to detect and extract contextual information and integrate it into the navigation solution. Specifically, Convolutional Neural Networks (CNN) and a Generalized Likelihood Ratio Test (GLRT) are developed to identify the stationary and turning states of the train. Additionally, hysteresis approaches are suggested to eliminate detrimental errors and differentiate between switches and turning nodes. To evaluate the performance of the algorithms, real-world data recorded from multiple train trips is utilized. The experimental results provide insights into the effectiveness and efficiency of the proposed methods, demonstrating their ability to enhance the accuracy and reliability of the train navigation system.