The invention provides an on-line state monitoring and failure diagnosing method of a wind turbine generator set. The on-line state monitoring and failure diagnosing method comprises the following steps: adopting an unscented Kalman method for processing SCADA data serving as the state variables, input quantity and measuring quantity to acquire the current predicted value of the state variables; calculating the difference value between the current predicted value of the state variables and the measured value of the state variables in the SCADA data to serve as a first difference value; and when the first difference value exceeds the threshold value, judging that the current alarm failure needs to be handled. According to the on-line state monitoring and failure diagnosing method of the wind turbine generator set, the data collection quantity is small, failures of the wind turbine generator set can be detected on line and in real time, the reliability is high, and the operating cost is low.