The excitation system is an important control equipment of synchronous generator, which plays an important role in the stable operation of synchronous generator and the whole power system. For this target, the article at first gives out the structure and working principle of the excitation system. After that, a classification method of faults which is established on wavelet packet decomposition is being studied. Furthermore, many signal handling methods are utilized by people to obtain the feature properties of malfunction signals that are connected with generator excitation system wrong operations. Further, a deep belief network model based on genetic algorithm is established, which improves the noise resistance through the binary features of neurons in each layer of the DBN, and at the same time, utilize the global optimization searching of the genetic algorithm to identify the optimized structure parameters of the Deep Belief Network (DBN) with consideration of the input data feature collection. This behavior is conducted for the implementation of the fault detection of the generator excitation system. Through the comparison of the fault diagnosis results got from the three algorithms, we can see that the whole accuracy of the GA-DBN algorithm arrives at a surprising 99%. By comparison, the traditional DBN model possesses a total accuracy of 95 percent when we take the count of wrong diagnoses into consideration. When compared with the traditional shallow intelligence diagnosis method, the method which this paper puts forward can diagnose the fault modes of the generator excitation system more stably and possess a higher degree of accuracy.
