Towards enhanced acoustic fan booster damage detection: a comparative study of feature-based and machine learning approaches

Abstract

Machine failure detection frequently uses non-destructive monitoring techniques such as vibration analysis. Although vibration analysis can identify machine degradation, the apparatus is often costly and necessitates specialist knowledge. Additionally, many existing methods in audio classification rely on characteristics represented as pictures or vectors, which increases computational complexity. In contrast, this research introduces a novel method that substitutes vibration data with a singular numerical feature derived from audio signals, addressing both cost and complexity issues. Our objective is to develop a rapid and precise audio-based method for detecting machine damage. The acoustic signals from the machine apparatus were classified into three categories: normal, belt damage, and combined belt and bearing defect. The data processing technique involved lowering the sample rate and segmenting the data to improve computational efficiency and classification performance. We use the Welch method and appropriate statistical techniques to analyze Power Spectral Density (PSD). The performance of seven classifier models, KNN, LDA, SVM, NB, ANN, RF, and DT, was evaluated using accuracy, precision, sensitivity, specificity, and F-score. LDA achieved the highest accuracy at 92.83%, followed by ANN (92.75%), NB (92.74%), and DT (92.34%). These models outperformed KNN (89.90%) and RF (89.40%), with SVM recording the lowest accuracy at 85.40%. LDA was highly effective, achieving the highest accuracy with a single average PSD-type feature, showcasing its robustness in machine defect diagnosis. Compared to previous methods, this approach simplifies feature extraction, reduces computational demands, and maintains high diagnostic performance, providing notable benefits in terms of effectiveness and precision.