Abstract
The analysis of the mechanisms and causes of failures is a vital consideration in the safe and economic operation of power plants, but experts with the appropriate knowledge and background are becoming scarce. In consequence there is a need to develop an expert system which can encapsulate the human expertise that is needed in the investigation of power plant failures. In many cases more than one branch of technology and expertise must be brought to bear in an investigation, and a good expert system must reflect this. The failure analysis of power plant superheaters forms a suitable basis for the development of an expert system, as superheaters are susceptible to failure by various mechanisms, including creep, overheating, creep-fatigue, fireside corrosion, erosion, and weld failures. In addition, the root causes of such failures result from the way a plant has been designed, built, or managed, as well as resulting from shortcomings in superheater materials of construction. The focus on the shortcomings of P91 steel shows how the material of construction might contribute to the causes of failure. P91 is a 9Cr-1Mo martensitic alloy whose characteristics are different to the older low alloy carbon steels. This makes it a good test for questioning the assumptions underlying the formulation of the materials oriented If...Then rules in an expert system. Major issues with P91 are the absence of metallographic changes during component life, change in behaviour between medium and long term testing, and the severe affect that steam-side oxide scale has on raising superheater tube temperatures. Methods for formulating If...Then rules are described. The need to incorporate adjectival and adverbial phrases into such rules is shown to be necessary to give qualitative, but consistent estimates of how reliable are conclusions. This technique is applied to a fictitious superheater failure, in which the operation of the flue gas dampers in plant is shown to be the root cause of premature failure, rather than the failure being induced by a changes in the type of coal or to a feed heater failure. This “adjectival/adverbial” approach is itself novel, but the idea is taken much further to show how such phrases may be quantified using Bayesian probabilistics, to indicate, in numeral terms, how evidence from different sources can be used to support conclusions. This Bayesian method is applied to a superheater failure in a Japanese power plant, (reported in the literature), in which the failure mechanism was identified as being that of creep. The Bayesian method was also applied to the identification of the root cause which was shown to be due to a high rate of growth of the steam side oxide, in conjunction with a high rate of superheater heat transfer. This overall conclusions are therefore is that the work has shown that information and knowledge on a complex domain can be systematically formulated into sets of If...Then rules, which will require, as a minimum, the use of adjectival and adverbial phrases which are of a semi quantitative type, to ensure consistency. More importantly Bayesian probability theory can be applied to this type of rule set to quantify the certainty of decisions taken by the expert system, and enable the expert system to bring together evidence from diverse sources.