Dyn-ConvNet: Hybrid Time–Frequency Learning for Time Series Anomaly Detection

ICLR 2026 Conference Submission15599 Authors

Keywords: Time Series Anomaly Detection, Multivariate Time Series, Time–Frequency Representation, Deep Learning, Hybrid Neural Network

Abstract: Time series anomaly detection (TSAD) is of critical importance in applications such as industry, finance, and healthcare, yet it remains challenging due to complex temporal dependencies, non-stationarity, and the scarcity of anomalous samples. Existing methods typically operate in a single domain—either time or frequency—limiting their ability to capture diverse anomalies across both global and local patterns. To address this, we propose **Dyn-ConvNet**, a novel hybrid deep learning architecture that systematically integrates time- and frequency-domain representations by combining the Fast Fourier Transform (FFT) for long-range periodic feature extraction with the Wavelet Transform for local and transient anomaly detection. These complementary features are fused through a deep convolutional backbone enhanced with gating mechanisms and residual connections, enabling adaptive learning and robust detection across different scales and anomaly types. Experiments on five popular multivariate time series benchmark datasets show that Dyn-ConvNet outperforms state-of-the-art methods, with even larger gains in complex anomaly scenarios, demonstrating the effectiveness of multi-domain feature integration in enhancing both the performance and generalization capability of multivariate time series anomaly detection.

Primary Area: learning on time series and dynamical systems

Submission Number: 15599