Abstract: The accurate identification of microwave rain decay processes is essential for rainfall retrieval using microwave links. This study tackles the limitations of the traditional Moving Standard Deviation (MSD) method in distinguishing between dry and wet periods from rainfall-induced attenuation signals, particularly its reliance on subjective thresholds. We propose a multimodal fusion discrimination method based on time-varying signal characteristics as an alternative. Using microwave link observation datasets collected during typical rainy seasons in Nanchang City, we developed a systematic framework to construct three discrimination models: two unsupervised learning models—K-means clustering and Fuzzy C-means (FCM) clustering—using feature vectors composed of the maximum attenuation rate per one-hour link, the standard deviation of attenuation rates, and the rate of change of the maximum attenuation rate; along with an adaptive MSD threshold method. Evaluation via confusion matrices showed that under prolonged weak precipitation conditions, an integrated voting method combining all three approaches achieved median F1-scores and Matthews correlation coefficients above 0.6 for most links, with discrimination accuracy reaching 85%, leveraging model diversity and complementary mechanisms. In contrast, during short-duration heavy precipitation events, the adaptive MSD threshold method demonstrated superior performance through dynamic window adjustment and adaptive thresholding, achieving 95% discrimination accuracy and a 70% precipitation capture rate, significantly outperforming the clustering algorithms. This study highlights the differential responsiveness of precipitation types to discrimination methodologies and proposes a hierarchical optimization strategy for engineered precipitation monitoring systems.