Anomaly detection of wind turbines is a key technology for achieving intelligent operation and maintenance management. Existing approaches mainly focus on monitoring individual components, requiring separate detection models for each part. This not only reduces overall system efficiency but also neglects interactions among components. Moreover, in the early stage of a fault, it is often difficult to accurately determine the fault type. To address these challenges, an anomaly detection approach based on an autoencoder was proposed. First, a system-wide parameter reconstruction model was built using autoencoders, with the learning ability enhanced by using turbine data under similar operating conditions. Next, dynamic detection threshold curves were designed for each parameter to improve anomaly identification. Furthermore, a correlation-based identification strategy was introduced to locate key parameters most likely to cause anomalies by analyzing out-of-tolerance errors, thereby supporting subsequent operation and maintenance decisions. Finally, real operation data from a Huanghua wind farm in Hebei, China were used for validation. The results demonstrate that the proposed approach achieves reliable performance in anomaly detection and early fault warning of wind turbines.