The intermittency and randomness of wind make the operation state of wind turbine change frequently. As a result, the false positive ration and false negative rate in anomaly detection of equipment are serious. The costs of operation and maintenance in the wind power industry are high. To solve this problem, one kind of Knearest neighbor fault detection method based on dynamic feature matrix is proposed in this work. It constructs a dynamic feature matrix based on mutual information to describe the dynamic characteristics of wind turbine. The weighted knearest neighbor fault detection method is introduced to address the influence of the characteristic contribution and cumulative mutual information in dynamic feature matrix. The dynamic threshold can help reduce false alarm caused by the sudden change of operation state. This paper takes examples of the common sensor faults and actuator faults in the 5MW offshore benchmark of National Renewable Energy Laboratory and the pitch system faults in SCADA system. The fault detection results of the proposed method are compared with PCA, KPCA, FDkNN and PCkNN, respectively. Experimental results demonstrate that the proposed method can accurately detect the fault information. Compared with other methods, it can achieve better fault detection results.