Modulation recognition of communication signals is a fundamental and crucial work in the field of space electromagnetic signal processing. With the development of emerging technologies such as Radio Frequency Machine Learning (RFML), data-driven approaches have made great progress in this area. However, the challenges of modulation recognition arise from obtaining communication signals, few-shot data, and class-imbalance samples. Based on our previously proposed method called the CNN-LSTM dual-channel model, the transformation of the signal data including rotation and cyclic time-shift is introduced to address the problem of class-imbalance sample. Compared with our previous results, the overall recognition rate is increased by 2.7%, and the number of parameters is reduced by 1 %. The experiment proves that this work is efficient and feasible.