CKADformer: Enhanced Transformer with Causal Convolution and KAN for Multivariate Time Series Anomaly Detection

With the rapid advancement of IoT technologies, multivariate time series (MTS) anomaly detection has gained increasing prominence across industrial and financial domains. While Transformer-based models have demonstrated remarkable progress in capturing MTS correlations, they still face critical challenges: conventional self-attention mechanisms struggle to discern feature importance and inadequately amplify distinctions between adjacent anomalous and normal data points. To address this, we propose CKADformer, a novel model incorporating the CKAD-Attention mechanism, which integrates causal convolutions with Kolmogorov-Arnold Networks (KANs) to enhance discriminative feature representation in time series. The proposed mechanism comprises two branches: 1) a Gaussian prior association branch that models adjacent normal temporal patterns using learnable Gaussian kernels, and 2) an enhanced series association branch that extracts KANLinear networks to strengthen global feature interactions. Furthermore, by employing a minimax strategy to dynamically optimize the divergence distribution between prior and series associations, our model significantly amplifies the separability between anomalous and normal data. Experimental results demonstrate that CKADformer outperforms 4 advanced baselines across three real-world datasets. This study establishes a new paradigm for feature enhancement and divergence optimization in MTS anomaly detection, offering broad applicability.