To effectively identify and eliminate abnormal data in the measured data of wind turbines, an anomaly detection algorithm based on manifold learning is proposed through the analysis of high-dimensional measured data from wind turbines. Firstly, the k-nearest neighbor mutual information algorithm is employed to select feature variables for the wind turbine. Subsequently, an optimized t-distributed stochastic neighbor embedding (t-SNE) algorithm is utilized. This optimized algorithm replaces the sample distance metric with a weighted sum of the Euclidean distance and the local principal component analysis (LPCA) difference, enabling the extraction of low-dimensional features with inherent patterns from the high-dimensional manifold data. This facilitates the distinct separation of data with different distribution characteristics in a visualized two-dimensional space. Furthermore, the density-based spatial clustering of applications with noise (DBSCAN) algorithm is applied to cluster the data within this two-dimensional space. The results demonstrate that, compared to the principal component analysis (PCA) algorithm, locally linear embedding (LLE) algorithm, and the original t-SNE algorithm, the proposed method can effectively achieve visual separation and clustering for data under various complex operating conditions, successfully identifying and eliminating abnormal data.