The invention discloses a converter transformer tap switch condition monitoring system and a non-stationary signal spectrum estimation method. The system comprises a vibration signal sensor which is disposed on a housing of a transformer and is used for monitoring the operation process of a converter transformer tap switch; a digital filter for filtering noise waves; and a vibration data collection function module which is used for eliminating various types of signal noises. The method employs a maximum entropy-spectrum method, carries out solving of data outside a known limited data autocorrelation array through employing an extrapolation method under the condition of the maximum information entropy, and estimates the power spectrum density of a to-be-detected signal. The method achieves the estimation of the tap switch under the non-intruding detection, achieves the system inspection of the converter transformer tap switch, can reduce the incidence rate of faults in two preset correction periods, and employs the modern spectrum estimation method of maximum entropy estimation for the spectrum estimation of a non-stationary signal of an isolation switch, so as to further improve the analysis precision and accuracy.