Abstract:To address the problems of massive data, the mixture of effective and redundant information in the industrial internet, and the neglect of complex correlations among network traffic by traditional intrusion detection methods and to improve the feature extraction ability and detection accuracy of intrusion detection systems in industrial internet scenarios, this paper proposed an intrusion detection method for the industrial internet stacking circular convolution autoencoder and head attention mechanism. Firstly, a convolutional autoencoder was used to extract the spatial features of traffic data and reduce the redundancy of traffic data. Meanwhile, a two-dimensional convolutional layer and a circular padding method were introduced to expand the receptive field of the convolutional kernel on the data by increasing the dimension and expanding the spatial domain of the original data, thereby improving the feature extraction ability. Secondly, the head attention mechanism was utilized to extract the correlation information among the traffic, and attention head weights were assigned according to the importance of the correlation information extracted by different attention heads to consider the most important correlation information when determining the types of network traffic. Finally, binary classification and multi-classification comparative experiments were conducted with traditional intrusion detection methods and intrusion detection methods based on traditional time-series neural networks. Experimental results show that the intrusion detection method proposed in this paper achieves an accuracy of 99.18% in the binary classification task and 87.36% in the multi-classification task, which outperforms traditional intrusion detection methods. The proposed intrusion detection method can effectively detect network intrusions in industrial internet scenarios.