Abnormal yaw positioning during yaw operations induces progressive deviation in yaw alignment accuracy,thereby compromising wind-tracking precision and risking excessive cable twisting that threatens operational safety.Concurrently, frequent position oscillations or repetitive short-duration position holding generate transient control errors,destabilizing the yaw control system. These coupled mechanisms collectively escalate yaw system failure frequency and operational maintenance costs. To proactively mitigate these risks, a data-driven fault diagnosis methodology is proposed for early detection of anomalous yaw positioning in wind turbines. Firstly, a large amount of data in a supervisory control and data acquisition (SCADA) system was processed using a standardized interaction gain and Relief-F (SIG-Relief-F) feature selection algorithm to identify multiple feature parameters with the strongest correlation with the target variable (which in this case may be yaw system failure). The advantage of this method lied in its ability to consider effectively the correlation between features,thus maximizing the retention of relevant features related to yaw system failures and interaction features. Secondly, a back propagation neural network (BPNN) yaw position prediction model was established, and the distribution of residuals was statistically analyzed using a sliding window method to determine the fault threshold. Finally, through empirical verification,the effectiveness and accuracy of the proposed method were demonstrated, and compared with multivariate state estimation technique (MSET) and support vector machine (SVM) algorithms, it was shown to have superior abnormal warning performance. The conclusions drawn can serve as a reference for the fault diagnosis of a practical yaw system.