Feedwater heater system fault diagnosis during dynamic transient process based on two-stage neural networks

At present, researches on power plant fault diagnosis are mostly for steady-state work conditions and can not well adapt to the load-changing dynamic process, which greatly limits the practical application of a fault diagnosis system. Thus, a transient fault diagnosis approach based on two-stage neural networks was put forward for power plant thermal system fault diagnosis. An Elman recurrent neural network with time-delay inputs was applied to predict the expected normal values of the fault feature variables, and a BP neural network was used to identify the fault types. To improve the diagnostic effect for faults of varying severity under transient conditions, fault symptom zoom optimization technique was also used. Taking the high-pressure feedwater heater system of a 600MW supercritical power unit as the object investigated, the predictive model was built, trained and validated with large amount of historical operating data. The BP network fault diagnosis model was trained with the fault fuzzy knowledge library including typical fault samples. The real-time fault diagnosis program was then developed with MATLAB software. By communicating with the power plant simulator, intensive fault diagnosis tests were carried out. It was shown the suggested method can achieve good diagnosis results for the power plant thermal system under load-varying transient process.