Optimizing SVM training for DCS fault detection: A novel ensemble‐based data selection algorithm

Traditional fault detection in distributed control systems (DCS) relies on manual inspections or model‐based methods, which are time‐consuming and struggle with complex industrial data. While support vector machines (SVM) offer promising alternatives, their practical adoption in industrial settings is constrained by two fundamental challenges: the requirement for large labelled datasets and inherent susceptibility to overfitting. This paper introduces powered SVM (PSVM), a novel ensemble‐based algorithm that identifies the optimal training dataset while detecting the overfitting threshold. PSVM employs a three‐stage mechanism. First, three complementary classifiers, like Cubic SVM, Quadratic SVM, and K‑nearest neighbours (KNN) are trained on an initial 10% dataset. Second, samples misclassified by all three methods are added, achieving rapid convergence. Third, the algorithm identifies the overfitting threshold by monitoring accuracy changes across iterations. PSVM was implemented in a gas refinery DCS via open platform communications (OPC) protocol and also validated on the dry bean dataset (13,611 samples, 7 classes, 6.8:1 imbalance). In the gas refinery, PSVM achieved 99.8% fault detection accuracy (95% CI: [99.2%, 100%]), significantly outperforming conventional SVM (94.8%, p < 0.01). On the dry bean dataset, PSVM reached 96.2% accuracy using only 1722 samples (13% of training data), compared to 93.1% for conventional SVM—a 3.1% improvement with 87% less data. The algorithm identified the overfitting threshold at 12.6% data utilization and achieves 95% accuracy in just 3 iterations. This work introduces three innovations: quantitative overfitting threshold detection, an ensemble‐based data selection mechanism achieving rapid convergence with minimal data, and experimental validation showing optimal accuracy with only 13% of training data.