Identifying ice accumulation on wind turbine blades is increasingly crucial for preventing revenue losses and diminished power output. Traditional methodologies often incur supplementary expenses, while model-centric approaches are heavily reliant on pre-existing domain expertise. Data-driven techniques, especially deep learning paradigms that circumvent the need for laborious manual feature engineering, present a compelling alternative for blade icing detection. However, operational monitoring signals from wind turbines in complex environments typically exhibit intricate and varied characteristics, necessitating sophisticated models for effective data analysis. Furthermore, the inherent imbalance in monitoring data distribution poses challenges for adequate anomaly detection. In this study, we introduce a multilevel convolutional recurrent neural network (MCRNN) specifically designed for blade icing detection. We employ discrete wavelet decomposition to extract multilevel features from both temporal and spectral domains. A parallel architecture integrating an LSTM branch and a CNN branch is implemented at each level to facilitate feature extraction. To mitigate significant data imbalance, we explore two strategies: a data resampling algorithm and a class-balanced loss function. Additionally, we propose a multi-step accumulation strategy to enhance real-time detection accuracy. Comprehensive evaluations demonstrate that our MCRNN achieves F1-score improvements of up to 38.8 and 42.9% over the strongest baseline on balanced datasets processed with data resampling, and 23.9 and 30.6% higher on imbalanced datasets using MCRNN optimized with a class-rebalanced loss function. Real-time detection experiments validate the applicability of our method and indicate that the proposed multi-step accumulation strategy effectively improves icing detection precision.