The invention discloses a high-voltage circuit breaker mechanical fault diagnosis method. Opening and closing coil current obtained by a high-voltage circuit breaker fault simulation experiment is used as a target data sample for fault diagnosis; data obtained through simulation by establishing an opening and closing coil mathematical model serve as an auxiliary data sample, deep mining and self-adaptive extraction of sample data features are achieved through a deep belief network (DBN), and information matching of the auxiliary data and target data is achieved in combination with a transfer learning method. According to the method, transfer learning and a deep belief network are combined; the deep belief network is utilized to carry out deep mining and self-adaptive extraction on the current time domain signal of the opening and closing coil of the circuit breaker, the transfer learning method is combined to solve the problem that the actual fault training sample data size is small, and the generalization capability of the fault diagnosis model is improved.