Abstract: Millimeter wave cloud radar can be used to detect small particles such as water droplets and ice crystals in clouds, but it will form clear sky echoes when affected by non meteorological targets such as flying insects, birds, dust, or smoke. In order to eliminate clear sky echoes in detection results, this paper proposes and implements a universal deep learning model for millimeter wave cloud radar clear sky echo recognition - ATNN. This study mainly used cloud radar observation data from Baise City, Guangxi for training. To achieve a universal method, three data products, namely reflectivity factor, radial velocity, and relative height, are selected to train the model. The results showed that ATNN had a high score during the flood season, with an average clear sky echo recognition rate of 91.98% and an average meteorological echo retained rate of 94.28%. During the transition from non flood season to flood season, the ATNN score is slightly lower, with an average clear sky echo recognition rate of 87.56% and an average meteorological echo retained rate of 86.38%. Compared with ATNN, NN has a slightly higher recognition rate for clear sky echoes, but it will misjudge a large number of meteorological echoes with low intensity at low altitude as clear sky echoes. During non flood season, the average retained rate of meteorological echoes is only 54.96%. In addition, the evaluation of ATNN was conducted using data from 15 millimeter wave cloud radars in China, and the results showed that ATNN has universality, with a clear sky echo recognition rate of 81.17% and a meteorological echo retained rate of 93.89%.