Scarce labels, limited edge-node resources, and degraded data quality are common in industrial Internet of Things (IIoT) multivariate time-series anomaly detection. Under these constraints, existing deep anomaly detection models often struggle to maintain high detection accuracy. At the same time, they find it difficult to ensure strong anomaly discrimination under class-imbalanced conditions and stable performance. To address this problem, a Self-Supervised Quantum Convolutional Neural Network (SQ-CNN) is proposed in this paper. The model is developed for IIoT multivariate time-series anomaly detection. A shared quantum convolutional feature extractor is adopted in the model. A self-supervised branch is constructed for representation learning. A supervised classification branch is constructed for anomaly recognition. A weighted joint loss is used for joint optimization. The parameters of the quantum circuit and the weights of the classical network are jointly updated. Consistent training is thus achieved from representation learning to recognition-oriented decision making. In addition, a classical head and a low-bit quantization strategy are designed. This design is intended to reduce the deployment overhead at the edge side and improve inference energy efficiency. Experimental results show that the proposed SQ-CNN achieves an 18.62% improvement in AUROC and AUPRC over baseline methods. A 19.72% reduction in unit energy consumption is also achieved. In addition, the single-window inference latency is reduced by 13.38%. The throughput reaches up to 392.28 windows/s.
Guo et al. (Fri,) studied this question.