Abstract: Non-meteorological echoes, represented by clear-air echoes, can interfere with cloud observations by millimeter-wave cloud radar (MMCR). The formation conditions of clear-air echoes are complex and variable, and they can often merge with meteorological echoes, posing a persistent challenge for MMCR quality control (QC). This paper develops a clear-air echo QC model based on the U-net deep learning model, named CR-Unet (Cloud Radar U-net), using minute-by-minute observational data from 49 MMCR stations across multiple provinces in China provided by the China Meteorological Administration from July 2023 to June 2024. Evaluation and validation reveal the following: (1) The CR-Unet exhibits good generalization performance and is applicable to QC of reflectivity factors from most Chinese MMCRs, overcoming the limitation of threshold methods that require individual parameter tuning for each station. (2) CR-Unet employs a four-level downsampling and upsampling architecture, utilizing three MMCR observables—radial velocity (V), linear depolarization ratio (LDR), and radar reflectivity factor (Z)—fused into a three-channel dataset. Data augmentation was applied to MMCR observation images of less frequent special weather phenomena such as snowfall and fog. Using meteorological echo retention rate and clear-air echo identification rate as evaluation metrics, the model achieves a retention rate exceeding 99% and an identification rate of 96–97%. (3) Visual validation using MMCR observation data not included in the training demonstrates that CR-Unet can perform classification and identification of MMCR data from various terrains across China, including plateaus, plains, basins, and coastal areas. It can also effectively identify cases with predominantly single-layer and multi-layer echoes, weak echoes, and missing LDR. The results also indicate that when LDR observations are faulty or significantly different from other MMCR data, it may have adverse effects on CR-Unet identification and classification. CR-Unet deep learning based approach has greater potential and value for application in MMCR image quality control.