The invention relates to a method for recognizing an underwater acoustic communication modulation mode based on a deep learning technique. The method comprises the following steps: constructing a deeplearning convolutional neural network model; presetting a training set sample recognition accuracy rate T and a testing set sample recognition accuracy rate P; acquiring experimental data or simulation data in different modulation modes; carrying out pretreatment by taking N sampling point data as an original data sample; randomly dividing pretreated data samples into a training set and a testingset; training the data by virtue of a training sample set; judging whether the training set sample recognition accuracy rate reaches a preset value, if yes, switching input into a data sample testingset, and testing by virtue of the data sample testing set; otherwise, continuing to train; judging whether the testing set sample recognition accuracy rate reaches a preset value, and if yes, finishing the model; and other wise, acquiring extra data, mixing with the original data, and repeatedly carrying out the method. According to the method, the difficult extraction of signal features caused due to time variation and space variation of an ocean channel is solved.