Comparative Evaluation of Optimisation Algorithms for Artificial Neural Network‐Based Classification of Partial Discharge Defects in Generator Stator Windings

Accurate classification of partial discharge (PD) defects is essential for condition assessment and reliability improvement of generator stator windings. This study presents a comparative evaluation of artificial neural network (ANN) training algorithms for PD defect classification using signals generated from an Insulation Defect Simulator (IDS). Six classes of PD‐related signals, consisting of internal delamination, internal voids, slot discharges, end‐winding discharge type 1, end‐winding discharge type 2, and noise, were used to construct a dataset of 960 samples. Statistical features extracted from the PD signals were used as ANN inputs, including peak value, phase angle, mean, variance, standard deviation, entropy, skewness, kurtosis, and correlation coefficient. To ensure a fair comparison, three optimisation algorithms—Broyden–Fletcher–Goldfarb–Shanno quasi‐Newton (BFGS), scaled conjugate gradient (SCG), and Levenberg–Marquardt (LM)—were evaluated under the same dataset, feature set, and three‐layer ANN architecture, with hidden‐layer configurations of 3, 5, 7, and 9 nodes. The models were assessed using confusion matrices, accuracy, precision, recall, F1‐score, and mean square error (MSE). The experimental results show that the LM algorithm with nine hidden nodes achieved the best overall classification performance, with an accuracy of 99.06% and an average F1‐score of 0.9906. This outperformed the best SCG configuration with five hidden nodes, which achieved 91.77% accuracy and an F1‐score of 0.9165, and the best BFGS configuration with nine hidden nodes, which achieved 90.00% accuracy and an F1‐score of 0.8993. The results confirm that the LM‐optimized ANN provides superior classification accuracy, stronger precision–recall balance, and better suitability for lightweight PD defect diagnosis in generator stator windings. The proposed framework offers an interpretable and computationally efficient approach for supporting maintenance decision‐making and reliability assessment of high‐voltage rotating machines.