Abstract: Millimeter-wavelength cloud radar serves as a pivotal instrument for cloud microphysical characterization but is susceptible to clear-air echoes induced by insects, birds, and aerosols, leading to considerable retrieval biases. To enhance detection accuracy and generalizability, this study presents ATNN (Multi-Head Attention Neural Networks), a universal deep learning framework that exclusively employs the reflectivity factor, radial velocity, and relative height for adaptive learning of vertical structural features. Evaluated on observations from Baise, Guangxi, ATNN demonstrates robust performance during the rainy season, achieving a clear-air echo detection rate of 91.98% and a meteorological echo preservation rate of 94.28%; during the non-flood season transition, these metrics are 87.56% and 86.38%, respectively. In contrast, conventional neural networks (NN) achieve marginally higher detection rates but misclassify numerous low-altitude, weak meteorological echoes as clear-air signals, yielding a preservation rate of only 54.96% in non-flood periods. Nationwide validation across 15 operational radar sites further confirms that ATNN attains an average clear-air echo removal rate of 81.17% while preserving 93.89% of meteorological echoes, thereby demonstrating strong regional adaptability.