In complex industrial environments, anomalies in system behavior are often driven by underlying causal relationships among multiple variables. Effectively identifying such causal structures from high-dimensional time series data is essential for accurate anomaly detection and root cause analysis. However, existing methods for temporal modeling and causal discovery frequently encounter challenges such as the curse of dimensionality, limited sensitivity to anomalies, and low accuracy in root cause localization. To address these limitations, this study builds upon the score-based causal discovery model NTS-NOTEARS by introducing a hypersphere constraint in the latent space, thereby enhancing the model’s sensitivity to abnormal samples. Furthermore, a unified framework is proposed that integrates anomaly detection and root cause analysis, leveraging both the learned causal structure and prediction residuals to accurately identify the root causes of anomalies.