Comparative Performance of Machine Learning Classifiers in Detecting Vibration Anomalies in Industrial Power Systems

This study examines methodologies for detecting abnormalities in Combined Cycle Power Plants (CCPPs) through application of vibration signal analysis and machine learning algorithms. Models’ performances were evaluated using different key metrics. The results indicated that the Random Forest classifier, particularly in combination with ECPT data, exhibited superior performance, achieving perfect scores across all metrics. It highlights the robustness of the Random Forest algorithm when applied to ECPT data, making it the most effective approach for vibration anomaly detection. The K-NN classifier demonstrated satisfactory performance when applied to AS and BTT data, attaining accuracy scores of 0.49 and 0.52, respectively; however, it exhibited limitations in handling diverse data distributions, as reflected in its lower accuracy of 0.44 with LDV data. Both GBM and SVM performed suboptimal, with GBM achieving a maximum accuracy of 0.52 with AS data, while SVM attained the highest accuracy of 0.49 with the same technique. Findings underscore the critical importance of selecting an appropriate combination of machine learning models and vibration measurement techniques to enhance the accuracy of anomaly detection. Eventually, the Random Forest algorithm is well suited for complex datasets with varied patterns, while K-NN may serve as an efficient alternative for simpler, more uniform data.