Machine fault detection through sound analysis using MFCC and machine learning

This study addresses the need for automated damage and failure detection in industrial machinery through sound analysis and machine learning. Traditional methods rely on human experts to identify faults using microphones, which can be time-consuming, stressful, and prone to errors such as limited perception, subjectivity, and inconsistency. This study leverages machine learning to create a more objective and efficient alternative. Mel-Frequency Cepstral Coefficients (MFCCs) were employed for feature extraction, capturing intricate sound patterns associated with machinery faults. Through rigorous experimentation, 11 MFCC coefficients were identified as optimal. The Support Vector Machine (SVM) emerged as the best-performing classifier compared to LightGBM and XGBoost, achieving a training accuracy of 83.12% and testing accuracy of 82.50%. The dataset was split between 80% for training and 20% for testing. The small gap between training and testing accuracy indicates an ideal model with no signs of over fitting, under fitting, or data leakage. Real-world simulations validated the model’s efficacy under various operational scenarios, demonstrating its readiness for industrial deployment. This study highlights the effectiveness of sound analysis and SVM classification in proactive maintenance, offering a reliable tool to reduce downtime and maintenance costs while enhancing operational efficiency and reliability.