This paper proposes a novel unsupervised time series anomaly detection method called Frequency-enhanced Conditional Variational Autoencoder (FCVAE). FCVAE integrates both global and local frequency information into the Conditional Variational Autoencoder (CVAE) to improve the reconstruction of normal data and better capture heterogeneous periodic patterns and detailed short-periodic trends. The method uses a "target attention" mechanism to select the most useful frequency information for short-periodic trend construction. FCVAE was evaluated on public datasets and a large-scale cloud system, outperforming state-of-the-art methods in terms of F1 score. The results demonstrate the practical applicability of FCVAE in addressing the limitations of current VAE-based anomaly detection models. The paper also discusses three key challenges in VAE-based anomaly detection: capturing heterogeneous periodic patterns, capturing detailed trends, and handling the noise introduced by numerous sub-frequencies. To address these challenges, FCVAE introduces a target attention mechanism to select the most useful sub-window frequencies. The paper presents the FCVAE model architecture, training and testing procedures, and experimental results on four datasets, showing that FCVAE significantly outperforms existing methods in terms of best F1 and delay F1 scores. The paper also discusses the effectiveness of different types of conditions in CVAE, the comparison between FCVAE and frequency-based VAE, and the role of global and local frequency information in anomaly detection. The paper concludes that FCVAE is a promising approach for unsupervised time series anomaly detection, with the potential to be applied in large-scale cloud systems.This paper proposes a novel unsupervised time series anomaly detection method called Frequency-enhanced Conditional Variational Autoencoder (FCVAE). FCVAE integrates both global and local frequency information into the Conditional Variational Autoencoder (CVAE) to improve the reconstruction of normal data and better capture heterogeneous periodic patterns and detailed short-periodic trends. The method uses a "target attention" mechanism to select the most useful frequency information for short-periodic trend construction. FCVAE was evaluated on public datasets and a large-scale cloud system, outperforming state-of-the-art methods in terms of F1 score. The results demonstrate the practical applicability of FCVAE in addressing the limitations of current VAE-based anomaly detection models. The paper also discusses three key challenges in VAE-based anomaly detection: capturing heterogeneous periodic patterns, capturing detailed trends, and handling the noise introduced by numerous sub-frequencies. To address these challenges, FCVAE introduces a target attention mechanism to select the most useful sub-window frequencies. The paper presents the FCVAE model architecture, training and testing procedures, and experimental results on four datasets, showing that FCVAE significantly outperforms existing methods in terms of best F1 and delay F1 scores. The paper also discusses the effectiveness of different types of conditions in CVAE, the comparison between FCVAE and frequency-based VAE, and the role of global and local frequency information in anomaly detection. The paper concludes that FCVAE is a promising approach for unsupervised time series anomaly detection, with the potential to be applied in large-scale cloud systems.