Hybrid Temporal Modeling and Generative Augmentation for Imbalanced Multivariate Time Series Anomaly Detection in Engineering Systems

Abstract

Abstract. Multivariate time series (MTS) data are central to engineering systems, where streams of sensor signals enable trend analysis, anomaly detection, and the discovery of operating patterns—supporting predictive maintenance and quality control. Yet, analyzing MTS remains challenging: faults are rare, severely imbalanced, and embedded in long contexts with overlapping modes. These challenges demand models that capture extended dependencies and focus on the most informative intervals. To address this, we design a hybrid deep learning architecture integrating recurrent units for long-horizon dynamics, attention layers for critical time-steps, and temporal convolutions for multiscale local patterns. We propose GAT-Net (gated recurrent unit (GRU)-attention-temporal convolutional network), a robust end-to-end MTS anomaly detector. Because severe class imbalance remains the dominant barrier, we extend the backbone with a generative augmentation strategy, developing DA-GAT-Net (data-augmented GAT-Net). It uses a segment-aware conditional generative adversarial network (GAN) to synthesize rare fault segments conditioned on local context and inserts them chronologically, preserving temporal continuity and cross-sensor correlations. We evaluate our frameworks on an industrial benchmark dataset. They outperform state-of-the-art baselines, with the generative extension yielding the largest gains in recall and F1-score. We also show that sensitivity rises sharply with a moderate amount of synthetic anomalies but levels off beyond a balanced ratio, underscoring the need for controlled, structure-preserving generation. Overall, this study highlights how attention-guided temporal modeling combined with generative augmentation can enhance fault detection in engineering systems where missed anomalies carry high operational costs.