COGNOS: Universal Enhancement for Time Series Anomaly Detection via Constrained Gaussian-Noise Optimization and Smoothing

Researchers have introduced COGNOS, a novel, model-agnostic framework designed to enhance time series anomaly detection (TSAD). Current reconstruction-based methods often rely on Mean Squared Error loss, leading to statistically flawed residuals and unstable anomaly scores. COGNOS addresses this fundamental weakness by implementing a two-stage approach. First, it employs a Gaussian-White Noise Regularization strategy during training to constrain model output residuals to a Gaussian white noise distribution. Second, it utilizes an Adaptive Residual Kalman Smoother to denoise raw anomaly scores, acting as a statistically robust estimator. Extensive experiments across multiple benchmarks demonstrate that COGNOS significantly improves the performance of state-of-the-art backbone models. This research validates the efficacy of combining statistical regularization with adaptive filtering, offering a universal solution for more reliable and accurate anomaly detection in various applications. The paper was published on arXiv, highlighting advancements in machine learning and artificial intelligence methodologies for data analysis.