Abstract
Detecting anomalies in multivariate time series data is crucial to ensure the security and stability of industrial processes. Yet, it remains challenging due to the absence of labeled anomaly data, the complexity of time series data, and the large dataset size. We propose KBJNet, an innovative model incorporating Transformer and Dilated Temporal Convolutional Network (TCN) techniques to address these obstacles. Our model employs a Single TCN-Attention Network, utilizing a single layer of Transformer encoder, making it highly efficient for inference. To further enhance its robustness, we introduce a novel adaptive attention mechanism that dynamically weights temporal context, enabling KBJNet to capture long-range dependencies in time series data effectively. The evaluation of KBJNet on eight publicly available datasets revealed that KBJNet considerably outperforms the most recent methods, enhancing F1 scores by as much as 6%. This result represents a significant contribution to anomaly detection, and we anticipate that our approach will have practical implications for developing next-generation anomaly detection systems in various industrial applications.