To address the challenges of high costs, long durations, and limited coverage associated with traditional OPGW icing monitoring methods, this paper proposes a fusion of frequency domain analysis and model self-learning for OPGW icing detection. The proposed method enables continuous sensing and intelligent reasoning of the icing status of transmission lines. Based on DAS technology, the method first obtains and demodulates Rayleigh backscatter interference signals by injecting highly coherent narrow linewidth pulse lasers. To enhance the robustness of frequency detection, the algorithm suppresses misjudgments caused by interference-induced shifts by extracting shared spectral peaks across multiple measurement points. Additionally, FFT is applied to obtain the frequency spectrum and identify the inherent frequency. Furthermore, considering that environmental changes may cause shifts in the inherent frequencies under non-icing conditions, a model self-learning mechanism is developed to dynamically update these frequencies. Lastly, a frequency domain-based icing thickness mechanism model is introduced and combined with a data-driven regression model to adaptively correct the mechanism model's results, improving the accuracy of the thickness estimation. Experimental results demonstrate that the proposed method can accurately detect cable icing thickness, and its effectiveness and robustness are validated in practical scenarios.