From human hearing to machine diagnostics: Incipient fault diagnosis of rotating machines using multi-bioscale attention entropy and acoustic signals

Fault diagnosis in rotating machinery is traditionally performed using vibration signals; however, acoustic signals captured by non-contact sensors provide a promising alternative for early fault detection. Although entropy-based methods have been widely applied for signal complexity analysis, existing multi-scale entropy approaches are not specifically designed for acoustic signals and may not fully exploit their perceptual frequency characteristics. To address this limitation, this study proposes a novel bio-inspired feature extraction method called multi-bioscale attention entropy (MBSAE). The proposed approach is inspired by the human auditory system and employs a Gammatone filterbank to decompose acoustic signals into multiple frequency sub-bands. Attention entropy is then computed for each sub-signal to construct a discriminative feature vector for fault classification. An extreme learning machine optimized using a simplified version of harris hawks optimization (HHO) algorithm is employed as the classifier. The proposed framework is validated using acoustic datasets obtained from a custom-built rotor-bearing and a gearbox experimental setup. To evaluate the effectiveness of the proposed approach within modern diagnostic frameworks, comprehensive comparisons are conducted with several state-of-the-art deep learning architectures, including convolutional neural networks and transformer-based models. In this framework, acoustic signals are first converted into time–frequency image representations, including short-time fourier transform (STFT), Morlet continuous wavelet transform, Mel-spectrogram, and Cochleagram, and then processed using well-known and advanced image-based classifiers. Experimental results demonstrate that the proposed MBSAE-HHO-ELM framework achieves superior diagnostic accuracy compared with modern deep learning models, while requiring significantly lower computational complexity. These findings indicate that the proposed MBSAE framework provides an effective and computationally efficient solution for acoustic-based fault diagnosis of rotating machinery, particularly in industrial scenarios where available training data and computational resources are limited.