Time series anomaly detection is a core task for ensuring the stability of modern industrial, financial, and operational systems. However, most existing methods model data in the time domain, which presents challenges in capturing global periodic anomalies and subtle structural changes, and they are susceptible to local noise. To address these limitations, we propose a novel unsupervised learning framework for anomaly detection named Multivariate time frequency anomaly detection(MTFAD). The core idea of MTFAD is to shift the main analysis process from the traditional time domain to the frequency domain. Specifically, the framework first maps the input sequence to the frequency domain via Fourier Transform and innovatively introduces a Frequency Patching strategy to decompose the spectrum into multiple frequency patches representing different frequency bands. The model's training is jointly supervised by a time-frequency dual-loss function, ensuring consistency in both the waveform and spectral structure of the reconstructed signal. Our experiments on multiple public datasets show that MTFAD can effectively identify diverse anomaly patterns and surpasses current state-of-the-art methods on several key metrics, providing a new perspective for time series anomaly detection by combining time and frequency domains.

time series, anomaly detection, frequency-domain modeling, unsupervised learning