Anomaly detection (AD) is of great importance to wind turbine (WT) prognostic health management systems. With the rising application of deep learning (DL), both deep regression- and deep reconstruction-based normal behavior modeling (NBM) methods have shown great promise for WT AD. However, massive missing and false alarms still could be witnessed in existing methods due to the lack of effective mining of interdependent relationships between normal and anomalies. Hence, this article proposes a generalized Siamese NBM scheme that can be applied to most existing backbones. By considering fault samples, a parameter-shared backbone and two auxiliary regularization terms are designed to explore characteristics between anomaly and normal. To alleviate the dependence on manual annotation for fault instances, a density-based clustering algorithm is adopted for the predefined outliers chosen. Furthermore, to enhance the trustworthiness of the proposed scheme, we implement a label correction based on temporal neighbor consistency. The experimental results show that the proposed Siamese NBM scheme improves state-of-the-art studies greatly. The outliers filtered by clustering can work as manually labeled samples without large fluctuation, also providing discriminative information. The label correction method can not only improve the reliability of the proposed Siamese NBM scheme but also the comparative deep methods, especially for those weak anomaly detectors.