<![CDATA[Fault detection in multi-setpoint industrial process control systems is complicated by the fact that normal sensor behavior shifts substantially across operating points, making conventional threshold-based alarms and single-condition models unreliable when setpoints change frequently. Recent studies on LSTM-Autoencoder for industrial anomaly detection have demonstrated promising results, yet most are evaluated under fixed operating conditions and do not examine how feature engineering choices affect detection performance across diverse setpoints. This study aims to determine whether physics-informed derived features improve LSTM-AE fault detection performance in a real-time multi-setpoint water level control system, and whether the improvement holds under practical deployment conditions. The proposed framework augments seven raw PLC sensor readings with three derived variables: delta flow, level error, and frequency-per-flow and applies a per-setpoint windowing strategy to prevent cross-setpoint data contamination during training. An ablation study compares the eleven-feature model against a seven-feature baseline under three labeling scenarios reflecting varying preprocessing quality. The eleven-feature model achieves an AUC of 1.0000 and F1-score of 0.9993 under onset-cut evaluation, and reduces the false positive rate from 18.48% to 15.21% under corrected labeling while maintaining perfect recall. Real-time validation across thirty fault injection experiments confirms a 100% detection rate with a mean latency of 6.37 ± 2.04 seconds, 38.2% faster than the baseline. These results confirm that derived features meaningfully improve both classification quality and temporal detection performance, though adaptive thresholding at high-variability setpoints remains an open challenge for future work.]]>
