Dynamic feature fusion of multi-differences MFCC and 1D-CNN framework for acoustic fault diagnosis of air compressors

Abstract

Early fault diagnosis of air compressors is critical for industrial safety, yet it remains challenging due to the subtle dynamic operation characteristics and the reliance on manual feature determination. To address these issues, this study proposes a dynamic feature fusion method integrating multi-differences Mel-frequency cepstral coefficients with a lightweight one-dimensional convolutional neural network for acoustic fault diagnosis. A hybrid acoustic feature matrix is constructed by fusing static Mel-frequency cepstral coefficients with its first- and second-order differences, explicitly capturing the transient temporal dynamics in acoustic signals. Moreover, the proposed end-to-end network architecture enables automatic feature learning and classification. Experimental results demonstrate that the proposed framework achieves a classification accuracy of 99.44% on a public acoustic dataset across eight operational states. The diagnostic accuracy is improved by 2.22% with the proposed dynamic feature fusion method over static features. Compared with other diagnostic methods, the proposed method achieves an accuracy increase of 1.82% while using 14.73% fewer parameters. Meanwhile, it reduces training time per epoch by 80% and performs feature extraction 81% faster. Thus, this work provides a solution for non-contact fault diagnosis in rotating machinery, with potential applications in predictive maintenance of industrial equipment.