Abstract: ‍ ‍Adaptive modulation dynamically adjusts the modulation mode according to the variations in channels to maximize the transmission capacity of channels. Modulation recognition is the key step to identifying the modulation mode and selecting the demodulation algorithm for the demodulator after adaptive modulation. Under a low signal-to-noise （SNR）， the recognition capability of multicarrier composite modulation signals must be enhanced. Traditional modulation recognition algorithms are mostly based on additive Gaussian white noise （AWGN） channel scenarios. However， in many communication scenarios， in addition to Gaussian noise， non-Gaussian noise with pulse characteristics occurs. Currently， the most commonly used noise models are Middleton Class A noise， symmetric stable （symmetric $\alpha$ stable， $\mathrm{S}\mathrm{\alpha }\mathrm{S}$） noise， and Gaussian mixture models， among which the Middleton Class A noise model is the most typical. Existing literature shows that achieving effective modulation recognition under the presence of pulse noise is often difficult. Therefore， this paper primarily studies a modulation recognition scheme for Middleton Class A noise systems. Considering the complex probability density function （PDF） of these types of impulse noise， a modulation recognition algorithm based on a convolutional neural network is proposed. To fully extract the modulation characteristics of signal transmission， this paper adopts a double-layer convolution block structure. Each convolutional block contains four parts： convolutional， normalization， activation， and pooling layers. After two convolutional blocks， the results are sent to a fully connected layer before passing a Softmax layer. Finally， the modulation modes of transmit signals are obtained. In an actual simulation process， 95% of the data samples were selected as the test set， and the remaining samples were used as the training set. A total of 5000 frames were used in the experiment， each consisting of 200 complex signals. The signal in each frame adopts the same modulation mode， whereas different frames adopt different modulation modes. Every time the receiver receives a complete data frame， the real and imaginary parts of 200 complex signals are reconstructed into a real square matrix with a size of （20，20）， which is input into the neural network to determine the recognized modulation mode. Simulation results demonstrate that the modulation recognition accuracy is gradually improved as the training progresses. The modulation recognition accuracy of the proposed scheme is better than those of benchmarking schemes. The training accuracy of different modulation schemes exhibits a similar trend. Under the BPSK， QPSK， and 8PSK modulation modes， the network converges at approximately 8000 steps， whereas 16QAM requires only 500 steps to converge. The training accuracy approaches 100% when the SNR is larger than 1 dB. However， the traditional single convolutional block network requires an SNR of 3 dB to achieve the same effect. This result fully verifies the effectiveness and practicability of the proposed algorithm. Experimental results also show that the recognition accuracy of the proposed scheme is better than other schemes， and it performs well under the other two typical noise models. The results are useful for further research on applying convolutional neural networks to joint adaptive modulation recognition and channel codec. In the future， multilayer convolutional block schemes will be explored further， or the existing algorithm will be improved with compressed sensing technology.