Intelligent Fault Prediction in Diesel Engines: A Comparative Study of SVM and BPNN for Condition-Based Maintenance

Abstract

This study discusses the application of Support Vector Machine (SVM) and Back Propagation Neural Network (BPNN) in predicting diesel engine health based on operational data relabeled using K-Means Clustering. Two types of SVM kernels were tested, namely Radial Basis Function (RBF) and Sigmoid, with various parameter combinations. The results indicate that SVM with Sigmoid kernel achieved an accuracy of 94.06%, but was less sensitive in detecting unhealthy engine conditions. In comparison, the BPNN method with a three-hidden-layer configuration (1-2-1 neurons) and tansig activation function showed superior performance with 97.13% accuracy, MSE of 0.03, recall of 94%, precision of 100%, and F1-score of 97%. These findings prove that BPNN outperforms SVM in capturing complex data patterns and is more accurate in detecting unhealthy engine conditions. Additionally, relabeling the dataset significantly improved predictive accuracy from 72.3% to 97.13%, highlighting the importance of balanced data in modeling. Overall, this study demonstrates that optimally configured BPNN is more effective in predicting diesel engine health than SVM, making it a more reliable approach for engine condition monitoring.