Data used for wind turbine icing detection are highly susceptible to noise interference, and existing denoising methods often cause excessive signal smoothing and the loss of critical features. To address these shortcomings in d denoising, a novel hybrid method for wind turbine icing detection based on a contrastive blind denoising autoencoder is proposed. First, an autoencoder incorporating a noise contrastive estimation(NCE) mechanism is employed as an adaptive denoising module. By leveraging both reconstruction loss and the NCE mechanism, this module achieves a balance between noise removal and feature preservation; effectively preserving the key temporal features of icing while effectively filtering out noise interference. Notably, this module can perform adaptive blind denoising on raw signals through self-supervised learning, without requiring any prior knowledge of the noise or clean reference data. Building upon this denoising module, a dual-branch detection network is designed, with parallel branches for temporal feature extraction and feature enhancement. The temporal branch consists of a bidirectional long short-term memory (BiLSTM) network and a contract-and-broadcast self-attention (CBSA) module, whereas the enhancement branch comprises an OpenFE-based feature engineering module and a Kolmogorov-Arnold network (KAN). This network jointly captures the dynamic evolution patterns of the data in the temporal dimension and the nonlinear coupling relationships in the feature dimension, thereby enabling accurate identification of icing conditions. Multi-dimensional experiments based on actual wind farm data were conducted to verify the effectiveness of the proposed method. The experimental results demonstrate that this approach effectively suppresses false icing detections caused by interference, improves comprehensive detection accuracy, and maintains stability across varying noise intensities. It demonstrates a certain level of anti-noise robustness and generalization capability, offering reliable support for practical engineering applications.