Audio Based Machine Fault Diagnosis using Hybrid Feature Extraction and Ensemble Learning

Over the last decade, proliferation of mechanical machines has surged exponentially, amplifying the challenge of monitoring their operational health due to the inevitability of wear and tear. Consequently, the conventional human centric approach to tracking machine conditions has become impractical. This underscores the demand for an innovative and efficient strategy. This research involves machine generated audio signals to detect and analyze faults or irregularities. Analyzing the acoustic patterns generated during machine operation makes detecting abnormal sounds associated with specific faults possible. This paper examines various techniques and algorithms for achieving this goal and compares them to recommend the most effective method. Initially, Short Time Fourier transform was performed on the preprocessed audio data. Then, various techniques to perform feature extraction on the transformed audio, such as spectral kurtosis, spectral centroid, zero crossing rate, and Mel spectrogram. Subsequently, the aggregate feature vector was input into various algorithms, including XGBoost, Support Vector Classifier, Random Forest Classifier and Ensemble Learning. The MIMII dataset was trained on these models. The evaluation showed that XGBoost Classifier performed the best as compared to the other models, achieving a total accuracy rate of 98 percent upon incorporating features extraction Melspectrogram, Spectral Kurtosis and Spectral Centroid.