A Graph Prompt Fine-Tuning Method for WSN Spatio-Temporal Correlation Anomaly Detection

Anomaly detection of multi-temporal modal data in Wireless Sensor Network (WSN) can provide an important guarantee for reliable network operation. Existing anomaly detection methods in multi-temporal modal data scenarios have the problems of insufficient extraction of spatio-temporal correlation features, high cost of anomaly sample category annotation, and imbalance of anomaly samples. In this paper, a graph neural network anomaly detection backbone network incorporating spatio-temporal correlation features and a multi-task self-supervised training strategy of “pre-training - graph prompting - fine-tuning” are designed for the characteristics of WSN graph structure data. First, the anomaly detection backbone network is designed by improving the Mamba model based on a multi-scale strategy and inter-modal fusion method, and combining it with a variational graph convolution module, which is capable of fully extracting spatio-temporal correlation features in the multi-node, multi-temporal modal scenarios of WSNs. Secondly, we design a three-subtask learning “pre-training” method with no-negative comparative learning, prediction, and reconstruction to learn generic features of WSN data samples from unlabeled data, and design a “graph prompting-fine-tuning” mechanism to guide the pre-trained self-supervised learning. The model is fine-tuned through the “graph prompting-fine-tuning” mechanism to guide the pre-trained self-supervised learning model to complete the parameter fine-tuning, thereby reducing the training cost and enhancing the detection generalization performance. The F1 metrics obtained from experiments on the public dataset and the actual collected dataset are up to 91.30% and 92.31%, respectively, which provides better detection performance and generalization ability than existing methods designed by the method.

💡 Research Summary

Wireless Sensor Networks (WSNs) generate multivariate, multi‑node time‑series data that exhibit complex spatial and temporal correlations. Existing anomaly detection approaches either focus on single‑node multi‑modal or multi‑node single‑modal scenarios, and they struggle to capture long‑range temporal dependencies as well as the intricate spatial‑temporal interplay among sensors. Moreover, labeling costs are high and anomalous events are rare, leading to severe class imbalance and limited supervised learning performance.

To address these challenges, the authors propose a novel graph‑neural‑network backbone combined with a “pre‑training → graph prompting → fine‑tuning” self‑supervised framework. The backbone augments the state‑space Mamba model with multi‑scale dilated convolutions for local periodic pattern extraction and with a cross‑attention module that fuses different modalities (e.g., temperature and humidity) within each node. Spatial relationships across nodes are modeled by a variational graph convolutional network (VGCN), which learns probabilistic node embeddings and captures uncertain inter‑node dependencies.

During self‑supervised pre‑training, three proxy tasks are jointly optimized: (1) contrastive learning without negative samples to enforce discriminative representations, (2) next‑step prediction to encode temporal dynamics, and (3) reconstruction of corrupted inputs to preserve information content. This multi‑task scheme yields a generic feature extractor that does not rely on labeled anomalies.

The pre‑trained model is then equipped with a graph prompt—a set of learnable parameters that condition the graph structure and node features on the downstream anomaly‑detection objective. By fine‑tuning only the prompt (and a small set of downstream parameters), the method dramatically reduces the amount of labeled data required while preserving high detection accuracy.

Experiments on a public WSN dataset and on a real‑world collected dataset demonstrate the effectiveness of the approach: F1 scores of 91.30 % and 92.31 % are achieved, respectively, outperforming state‑of‑the‑art GNN‑based, Transformer‑based, and classical statistical methods by 4–7 percentage points. The method remains robust when the proportion of labeled anomalies is reduced to 5 % or less, and training time is cut by roughly 30 % thanks to the prompt‑based fine‑tuning.

The paper’s contributions lie in (i) a comprehensive backbone that jointly captures multi‑scale temporal patterns, inter‑modal correlations, and spatial graph structure; (ii) a self‑supervised multi‑task pre‑training strategy that alleviates label scarcity; and (iii) the introduction of graph prompting for efficient adaptation to low‑resource anomaly detection scenarios. Limitations include the high memory footprint of the Mamba component and the reliance on domain knowledge to design effective prompts. Future work will explore automatic prompt generation and lightweight graph‑temporal architectures for real‑time WSN monitoring.