An AI-Driven Framework for Thermal Sensor Stability Assessment and Predictive Fault Diagnosis in Industrial Cooling Systems: A Comparative Study of SVM and LSTM Approaches

The stability and reliability of temperature sensors in industrial cooling systems are critical to process quality, energy efficiency, and operational safety. However, existing approaches lack systematic stability metrics and intelligent predictive capabilities. This study proposes an AI-driven framework integrating stability feature engineering with machine learning models for fault identification and early prediction of temperature sensors in power plant cooling systems. The framework introduces three physics-based stability indicators—rolling standard deviation (σ_roll), variation intensity index (VII), and short-term variation magnitude (ΔT_short)—to quantify sensor signal quality. These features, combined with operational parameters, are used to train support vector machine (SVM) and Long Short-Term Memory (LSTM) models for binary classification. The framework is validated using over 260,000 one-minute records per unit collected from three parallel steam-turbine generating units (Units 1, 2, and 3) of the same coastal thermal power plant. Each unit is served by an independent once-through seawater cooling loop instrumented with redundant Pt-100 temperature sensors at the inlet and outlet manifolds; the three units differ in their operating profile—Unit 1 operates under variable load with frequent cold-start events, Unit 2 under moderate variable load, and Unit 3 under stable high-load conditions—with data collected at 1 min intervals from January to June 2025. Under an explicitly anomaly-positive evaluation, with the full confusion matrix reported for every unit and model, classification performance is limited and strongly unit-dependent. In real-time identification, AUC-based ranking ability varies across units (SVM AUC = 0.65, 0.75, and 0.98 for Units 1–3; LSTM AUC = 0.66, 0.31, and 0.52), but under the extreme class imbalance (anomaly rate ≈ 0.07–0.13% in the test partitions), the calibrated operating-point precision and F1-scores remain low for all unit–model combinations (F1 ≤ 0.26, MCC ≤ 0.28). McNemar’s test indicates statistically significant paired differences for Units 1 and 2 but not for Unit 3. These results show that, on this dataset, neither model attains reliable anomaly classification, and that all reported metrics must be interpreted together with the disclosed confusion-matrix counts and severe class imbalance. The primary contribution of the framework is therefore methodological—physics-based stability indicators, redundant sensor cross-checking, and an operational false-alarm analysis—rather than high-accuracy prediction, and the study highlights the difficulty of learning-based prediction for rare, rule-defined thermal sensor anomalies.