A graph embedded in graph framework with dual-sequence input for efficient anomaly detection of complex equipment under insufficient samples

Real-time anomaly detection is essential for the safe launch of some sophisticated equipment, such as liquid rocket engines (LRE), in order to head off disasters. However, the Industrial Internet of Things (IIoT) edge's real-time requirements cannot be addressed by the present methodologies, and the outcome is poor when dealing with a lack of training samples on the device side. We provide a solution for device-side real-time anomaly detection using the architecture known as Graph Embedded in Graph Networks to address these issues (GG-Nets). To fully extract features under insufficient training data, in our method, we learn the temporal relationship of timestamps in the multivariate signal through a time-series graph attention network (T-GAT) and extract features, and use the extracted features to replace the original signal as the information of the sensor attention network (S-GAT) nodes. For efficiency, the signals in the original sample are divided into odd and even sequences, which greatly reduces the number of nodes in the T-GAT. Experiments reveal that the proposed method outperforms other state-of-the-art models on the LRE ignition dataset. Furthermore, the ablation experiment proves that each module of the model improves the effect and extensive discussions explore interpretability. Our code can be found on: https://github.com/yhd-ai/GG-Nets.