This thesis formally proposes a new class of mechanical equipment referred to as single throw mechanical equipment (STME). The initial approach aims at developing an Fault Detection and Isolation (FDI) system for a rotary door operator leading to a neural networks based approach using multiple sensors. The steps involve selection of suitable sensors (throw timing, angular-displacement, airflow and air pressure) from the operator's reliability-centred maintenance worksheets, hardware and software development for data acquisition. The residual generation is achieved using radial basis function neural network models and adaptive thresholds are applied for fault detection to reduce the false alarm rates at low pressure operating points where fault sensitivity is highest. Self-organising maps are employed for fault classification and the output of the winning neurons is associated to the operator's database. This system achieved a reasonable sensitivity to various system and sensor faults with less development time due to the black box approach but the cost-effectiveness in terms of sensor cost and processing power remains debatable.

This preliminary work illustrates the need for a more generic solution that could exploit certain common properties of similar equipments such as point machines and emergency brake-activation mechanisms (train-stops). The first step in this regard was the development of a generic test environment for reliable and safe data acquisition in terms of thousands of unmanned operations with the capability of storing the acquired data in the form of hierarchical data sets. The database has open connectivity with most mathematical analysis and data visualisation software packages.

Data sets from various STMEs are analysed to identify their commonalties such as input, output, throw timing, energy conservation, residual sensitivity and operational characteristics. The terminology used for a STME is formally defined. Based on these characteristics an FDI framework is set out using an array of limit switches placed at the boundaries of regions of constant acceleration during the transition. The estimates of spring and damping coefficient are used to identify expected residuals behaviour. The residual statistical properties and sensitivity are used to develop suitable adaptive threshold functions. The structured residual generation process is used for isolating system, input-sensor and output-sensor faults. The pros and cons of using an exponential model for larger operating regions are discussed. The possibilities of using a parameter estimation approach for FDI are also explored. The thesis includes five case studies illustrating all steps of this FDI framework. The equipments used in these case studies are the train-stops, train-doors and point machines. A model for a distributed embedded FDI system (EFS) for remote monitoring utilising state of the art communication technology is finally proposed.