It is important to study accidents and their underlying causes, in order to generate recommendations for improving system safety. A range of methods have been developed in various industries, to understand how accidents have occurred, as well as identify potential human errors in systems. Theories of accident causation, and the development of safety models and methods have evolved over the last few decades. However, the majority of accident analysis methods fail to account for the increasing complexity of socio-technical systems (Hollnagel, 2004 and Lindberg et al. 2010). Much of the previous research has taken a safety I perspective, which considers successful performance as reducing the number of adverse outcomes to as low as possible (Hollnagel, 2014). According to Hollnagel (2014) however, it is important to understand how operators actually carry out work (‘work-as-done’), rather than as it should be carried out (‘work-as-imagined’), to understand how normal variabilities and flexibilities in performance contribute towards both successful and unsuccessful performance. Understanding how work is normally carried out is essential for understanding how it can go wrong. This includes understanding how success is obtained, for example how people adjust their performance in the face of changing conditions and demands, and limited resources (such as time and information). Although variability and flexibility in performance are prerequisites for success and productivity, these can also explain why things can go wrong (Hollnagel, 2014). Understanding normal work (or ‘work-as-done’) is the basis of the safety II perspective, which views safety as increasing the number of things that go right. So far however, there seems to be little application of this safety II perspective in the rail industry. Research in this thesis addresses this gap, by examining whether understanding normal performance in rail engineering contexts contributes towards identifying how incidents occur, and measures for improving safety, compared to the use of existing methods. A range of different methods were used to address the aims of this thesis. Rail incident reports were analysed to understand sources of human errors in rail contexts. Observations were also conducted of operators carrying out work, to understand the opportunities for human errors associated with rail engineering processes. To understand cognitive demands and strategies associated with normal work, a cognitive task analysis was carried out. FRAM (Functional Resonance Analysis Method) (Hollnagel, 2012) wasalso used to determine how incidents may develop, and whether everyday performance can contribute towards successful and unsuccessful performance. Participants in semi-structured interviews and workshops were asked to identify strengths and limitations of various human reliability assessment methods, and offer opinions on their practical applicability. Benefits of understanding normal work included a greater understanding of how human errors can occur (by identifying cognitive demands that contribute towards the occurrence of different error types), and how cognitive strategies can reduce human errors and contribute towards acceptable performance. It was demonstrated how variabilities and flexibilities in performance can contribute towards successful and productive performance, as well as explain why things can go wrong (supporting Hollnagel, 2014). This is especially important to consider, since human errors were not easily identified from rail incident reports and observations of operators carrying out work. System safety can therefore be improved by increasing things that can go right, rather than just decreasing the things that can go wrong (Hollnagel, 2014). Participants in a workshop, however, identified that FRAM may be time consuming to apply, especially for more complex systems. Further research is recommended for the development of a toolkit, from which both practitioners and researchers can choose from a range of different methods. To further understand factors affecting acceptable performance, it is recommended that further data are collected to determine whether varying levels of cognitive demands affect performance, and whether these influence the implementation of cognitive strategies.