Online Anomaly Detection in Industrial IoT Networks Using a Supervised Contrastive Learning-Based Spatiotemporal Variational Autoencoder

As industrial IoT networks evolve, they become increasingly vulnerable to cyberattacks, such as denial-of-service and backdoor attacks, which lead to anomalies in data streams (e.g., unusual spikes or drops in traffic, sudden changes in device behavior, or irregular communication patterns). To address the challenge of detecting these anomalies amidst dynamic data distributions and diverse abnormal patterns, this paper proposes a supervised contrastive learning-based spatiotemporal variational autoencoder (SC-STVAE) for anomaly detection in online data streams. A multi-head graph attention network (MD-GAT) is utilized to capture feature correlations, while a temporal convolution network serves as the hidden layer in the variational autoencoder. This enables SC-STVAE to learn both feature correlations and temporal dependencies. To resolve the issue of ambiguous positive and negative boundaries, supervised contrastive learning is introduced within the STVAE, improving boundary distinction and detection accuracy. To mitigate performance degradation due to data drift, an event-triggered elastic weight consolidation algorithm is introduced, which updates model parameters based on reliability thresholds. Additionally, a fuzzy entropy-weighted anomaly score, which measures the error between reconstructed data and original inputs by computing the weighted sum of the mean squared error across each dimension, is introduced. Experimental results demonstrate superior performance in terms of accuracy, recall, and F1 score compared to benchmark algorithms.