Network anomaly detection is crucial for maintaining network stability and security. However, the diversity of anomalous behaviors and the similarity of normal behaviors pose significant challenges to anomaly detection tasks. Although deep learning models possess strong feature extraction capabilities, their performance is limited in noise or sparse sample environments due to their reliance on large amounts of labeled data and susceptibility to overfitting. To address these issues, this paper proposes a network time-series anomaly detection model that combines VAE-GAN with fuzzy inference. First, we combine VAE and GAN to use the latent space generation capability of VAE and the adversarial learning mechanism of GAN, enhancing the model's ability to detect anomalies in unknown samples. And the model's ability to capture temporal features more effectively is enhanced by incorporating an improved TCN module into the GAN. Furthermore, we design a fuzzy inference module that uses membership degree calculations and a dynamic adjustment mechanism to flexibly model the fuzzy boundaries between normal and anomalous samples. Experiments conducted on datasets such as Yahoo S5, NAB, and AIOPs demonstrate that the proposed method outperforms existing algorithms in terms of accuracy, precision, recall, and F1 score, validating its effectiveness and robustness in complex network environments.