Anomaly detection method for spatio-temporal correlation of WSN nodes based on graph prompt fine-tuning

Aiming at the existing anomaly detection methods in multi-temporal modal data scenarios, which had the problems of insufficient extraction of spatio-temporal correlation features, high cost of anomaly sample category annotation, and imbalance of anomalous samples, a graph neural network anomaly detection backbone network integrating spatio- temporal correlation features and a multi-task self-supervised training strategy of “pre-training-graph prompting-fine- tuning” was designed. Firstly, the Mamba model was improved by combining the multiscale strategy with the modal fusion mechanism, and the variational graph convolution module was introduced to construct a backbone network that could adequately extract the multi-node and multi-modal spatial and temporal features of wireless sensor network. Secondly, three pre-training sub-tasks of no-negative-case comparison learning, prediction, and reconstruction were designed to learn the generalized features from the unlabeled data, and a “graph prompting-fine-tuning” mechanism to reduce the training cost and enhance the detection generalization performance. The experimental results show that the F1 value of this method reaches 91.30% and 92.31% on the public and actual collected datasets respectively.