Abstract: Modern TT&C (Tracking, Telemetry and Command) system mostly adopts the composite modulation in a form of “pulse coding/multi-subcarrier internal modulation/external modulation”. This complicated scheme brings great challenges to signal accurate recognition in the absence of prior information and low signal-to-noise ratio (SNR) scenario. The existing composite modulation blind recognition methods based on feature extraction and pattern recognition are sensitive to signal features and sample size, and the whole process becomes even more cumbersome in the case of multiple subcarriers. In this paper, based on the unified carrier system composite modulated signal modeling, a new idea of blind recognition is proposed to train and classify the compressed composite modulated signal frequency domain data by using the inverse residual packet convolutional structure of lightweight neural network. By means of experiment platform construction and Python code designing, the proposed method verification for 10 composite modulated signals in condition of various SNRs is implemented. The results show that the recognition accuracy of the proposed method can reach 94.5% (SNR=0 dB) and 100% (SNR=5 dB) respectively; moreover, the sample size required for equal recognition accuracy is less than the existing statistical features and decision tree-based methods, and both the performance and amount of neural networks parameters used for classification are better than those of the benchmark network.