<p>In numerous real-world IoT systems, sensing devices produce vast volumes of multivariate time series data. These infrastructures frequently become targets for cyber-attacks, underscoring the critical importance of anomaly detection in multivariate time series. Previous studies have predominantly addressed non-stationarity through stationarization techniques to facilitate feature extraction. However, this approach can obscure inherent non-stationary patterns, thereby impeding the model’s detection capabilities. Moreover, it has been associated with limitations in nonlinear representation and interpretability. To address these challenges, we introduce MTAD-Kanformer, a novel framework for multivariate time series anomaly detection that integrates the strengths of Transformer architecture with those of the Kolmogorov-Arnold Network (KAN). This combination not only boosts the framework’s capacity to express nonlinear features but also enhances its interpretability. Additionally, we propose the Gaussian De-stationary Attention (GDSA) mechanism, which mitigates the issue of data non-stationarity by integrating learnable Gaussian kernels with de-stationary attention, while simultaneously accentuating the significance of critical time points. Our experimental results demonstrate that MTAD-Kanformer excels on five benchmark datasets for real-world multivariate time series anomaly detection.</p>