Improving Anomaly Detection in Industrial Time Series: The Role of Segmentation and Heterogeneous Ensemble