Acoustic-Based Machine Condition Monitoring Using Sparse Optimized Spectrum

In practical industrial scenarios, acoustic monitoring often faces challenges related to contactless data acquisition and intense background noise. To address these issues, this study presents a novel gearbox fault diagnosis framework that combines sparse decomposition with a joint optimization strategy. First, a redundant Fourier dictionary is designed to encompass theoretical fault frequencies and their corresponding harmonics. Subsequently, an enhanced orthogonal matching pursuit (OMP) algorithm is utilized to isolate sparse components associated with mechanical faults. Classification gradients are introduced to guide atom selection, enhancing the suppression of environmental noise. Furthermore, an objective function combining reconstruction error and classification loss is established, and joint optimization is achieved through alternating updates of sparse coefficients and classifier weights, improving sensitivity and robustness to fault features of the model. Experimental validation demonstrates that the proposed method effectively extracts fault characteristics under strong noise conditions, and due to its non-contact measurement nature, it eliminates the need for reserved sensor mounting positions or structural modifications, making it suitable for condition monitoring of high-speed, heavy-duty, and complex gearboxes.