Practical fault diagnosis of a thermal system is very important in ensuring safe and reliable operation of a power plant. However, it is a difficult task due to the structural complexity of a thermal system, varying degree of severity of a fault, and the wide range of operation of the power generating unit. An artificial neural network combined with optimal zoom search is proposed in this paper for recognizing varying degrees of faults in a power plant thermal system operating at different load level. The zoom search technology is based on the similarity rules of the feature variables to a same fault with different severity when the system topological structure does not change with fault or with different loading conditions. Two different types of symptoms, a trend symptom and a semantic symptom, are calculated and jointly used for on-line fault recognition, which results in a faster and more stable fault diagnosis. A feedforward neural network structure is adopted and an improved training method is introduced. A high-pressure feedwater heater system is taken as a target system for investigation. Several simulation tests for diagnosing a multidegree fault under different operating conditions are carried out on a 300-MW power plant simulator to demonstrate the validity of the method.