To realize predictive maintenance of the pitch system in wind turbines, a design idea is presented whereby a pitch anomaly recognition system can be developed on a wind turbines health management platform. In constructing the model algorithm, the frequency domain characteristics of pitch velocity are suggested as classification features, and precise characterization of abnormal symptoms is achieved. The pitch velocity is divided into three categories: high, low and normal frequency with AdaBoost-SAMME as a classification algorithm. Comparing with five algorithms including an artificial neural network, a support vector machine and a random forest, it is found that the AdaBoost-SAMME algorithm is superior in accuracy, precision, recall and G-mean index although a serious disproportion exists in the number of samples. In order to improve the self-learning ability of the recognition system, a new sample judgment method based on Euclidean distance is proposed to automatically expand the training sample size. The application example demonstrates that the pitch anomaly recognition system using the AdaBoost-SAMME algorithm possesses noticeable classifying quality and stability. It solves the technical problem that for the general adaptive law of pitch velocity frequency domain characteristics in different models is implicit, and a conventional logic judgment method cannot be used to identify a pitch system abnormality. The system achieves the early warning function of identifying anomalies before faults. This can guide field personnel to practice predictive maintenance and improve the reliability and utilization rate of wind turbines. This work is supported by National Key Research and Development Program of China (No. 2016YFB1000705).