Large-scale enterprise Wireless Local Area Networks (WLANs) often consist of tens of thousands of access points (APs), rendering manual detection of faulty devices a challenging task. While severe failures are easily identifiable, subtle issues frequently remain undetected until reported by end users. A key challenge in this context is the scarcity of labeled performance data and the highly imbalanced distribution of fault occurrences. To tackle this, we propose a traffic-based anomaly detection framework grounded in unsupervised learning. Leveraging the Variational Auto-Encoder (VAE), our approach learns normal traffic behavior patterns and identifies anomalies without depending on labeled data. Specifically, we integrate dilated convolution into the VAE to capture temporal dependencies across different time scales. For evaluating the proposed method, real-world enterprise WLAN traffic data were collected from a university network. Extensive experiments demonstrate robust detection performance, with a best F1-score of 0.88, outperforming other unsupervised methods.