Abstract: In cognitive radio applications, the current deep learning-based signal modulation style recognition algorithm has the problems of low computing efficiency and high complexity, for which a signal modulation recognition algorithm based on multi-scale timing features is proposed in this paper. The algorithm firstly uses multi-layer convolutional layers to extract temporal data of different scales, fuses the data with features and then uses long and short-term memory networks to extract temporal features, and finally outputs the recognition results from the output layer, and reduces the complexity of the algorithm by designing and optimizing the network structure. Experimental results show that the algorithm achieves recognition accuracy of more than 90% at signal-to-noise ratios of 4 dB and above on the original I/Q signal test set containing 11 modulated signals. The algorithm has lower complexity and shorter inference time on the embedded devices Jetson Nano and Raspberry Pi 4B compared to algorithms with the same recognition accuracy, which has better engineering application value.