<![CDATA[Water pump fault detection in industrial systems is challenged by extreme class imbalance, where anomalous states account for less than 7% of operational data and labeled fault examples are scarce. This paper presents a comparative evaluation of three unsupervised anomaly detection approaches Isolation Forest (IF), Long Short-Term Memory Autoencoder (LSTM-AE), and One-Dimensional Convolutional Neural Network Autoencoder (1D-CNN-AE) applied to a real-world multivariate water pump sensor dataset comprising 220,320 one-minute observations across 51 sensors. All models are trained exclusively on normal operational data, eliminating dependence on fault labels. A sliding-window preprocessing pipeline (W = 30 timesteps) with a scaler fitted only on normal training data ensures no temporal data leakage. Anomaly thresholds are derived statistically (mean + 3σ) from reconstruction errors on the normal test partition. Using a chronological 60/40 train-test split and threshold-free metrics, Isolation Forest achieves AUROC = 0.9973 and Recall = 0.9850, while LSTM-AE yields F1 = 0.9218 and Precision = 0.8892. The results reveal a fundamental sensitivity-selectivity trade-off across unsupervised paradigms and provide actionable guidance for deploying anomaly detection in water infrastructure monitoring systems.]]>
Copyrights © 2026
