Currently, although fuzz testing techniques are widely adopted, there are still some limitations in their methods, such as low automation and poor effectiveness. Further improvements are needed to ensure the comprehensive security of Industrial Internet protocols. Therefore, this paper proposes a fuzz testing method based on deep learning. This method uses convolutional neural networks to extract feature values of Industrial Internet protocols, quantitatively calculates the similarity between these protocols based on the extracted features, and finally determines the similarity threshold for distinguishing each Industrial Internet protocol from others using a large number of samples. This threshold guides the initial seed selection for the AFL tool. Experimental results demonstrate the effectiveness and feasibility of this approach.