A self-supervised feature fusion approach to situation assessment

A number of devices in Industrial Internet are various types in recent years. The monitored traffic data from different devices always unlabeled and contain various types of attack traffic. In other words, misjudgments occurring by the ambiguity with these various unlabeled traffic in situation assessment of Industrial Internet need to solve urgently for above complex network scenario. In this paper, a new self-supervised situation assessment method FCVnet (FCM-CNN-ViT Net) is proposed to reduce the misjudgement probability. An enhanced fuzzy c-means clustering method EFCM (Enhanced Fuzzy C-means Clustering), is designed for the unlabelled traffic data. Meanwhile the self-supervised pre-training is carried out by improving initial cluster centre selection to obtain more accurate labels. In order to capture more global features for better feature representation, MCFV (Multi-Convolutional Fusion and Vision Transformer) module combining Multi-Convolutional Neural Network and Vision Transformer (ViT) is designed to capture and fuse features from local details to broader context. Experimental results show that the precision and recall of the proposed FCVnet are improved by 7.51% and 15.16% on average with two data sets.