A healthy assessment system has been developed for high voltage shunt reactor, which has the functions such as signal acquisition, storage, and analysis based on vibration signal. Due to the spectral of vibration signals in a single observation point do not represent the health status of the shunt reactor comprehensively and its poor extensibility, we do not take the single-position vibration signal spectrum as feature vectors, but the total discrete spectrum multiple as features into the machine learning model, where the total discrete spectrum is the weighted sum of the signals from all observation points, to achieve a health assessment of the actual operation of 6 750 kV high voltage shunt reactors produced by a same company. Compared with the result of chromatography and ultrasonic partial discharge methods, this new method is consistent with above two conventional methods and the accuracy rate of state assessment is 100%. Therefore, the state assessment based on vibration signal is greatly significant in practical terms.