Fuzzy systems in real-time condition monitoring and fault diagnosis, with a diesel engine case study
2014-12-15T10:37:03Z (GMT) by
Diesel engines have become a common source of power, both for vehicles and for static equipment because they are fuel efficient, robust and reliable. It is important that diesel engines run in their correct condition and properly controlled in order to maintain efficiency, low emissions levels and high reliability.;The following thesis aims to assess the application of fuzzy systems in real-time condition monitoring and fault diagnosis. A 65kW diesel powered generator set has been purchased 'off the shelf' as an example of a typical application which may benefit from the development of CMFD techniques. As a test case, the diesel engine is appropriate as its sub-systems are complex, non-linear and subject to noise and uncertainty.;A diagnostic structure comprising fuzzy systems in three distinct roles has been proposed. Fuzzy reference models, incorporating heuristics and approximate non-linear mathematical relationships, are used for the generation of residuals by comparison with signals from a small number of low cost transducers. The residuals are classified and the diagnosis is inferred from the pattern of residual classes using a fuzzy rule-base. The diagnostic results obtained for three diesel engine sub-systems, show this to be a powerful technique for CMFD system design which may generalised, both for other types of plant and other forms of reference model.;This fuzzy model-based approach to fault diagnosis is shown to have benefits over other techniques by way of its transparency, ease of development, performance under variable engine load conditions, high level output and the lack of any requirement for fault data in the development process.;The robustness of the fuzzy reference models to certain fault conditions remains a key issue. The fuzzy models were generally effective at generating residuals where deviations from the normal condition are small. For larger deviations, robustness of models is not guaranteed or expected. A number of techniques were successfully deployed to reduce the number of misclassifications caused by this lack of robustness.