Decomposing and Modeling Acoustic Signals to Identify Machinery Defects in Industrial Soundscapes

Acoustic sound-based condition monitoring (ASCM) systems, which typically utilize machine learning algorithms on established audio features, have demonstrated effectiveness under controlled conditions. However, their application in real-world industrial environments presents significant challenges due to complex and variable soundscapes with high noise and limited fault data. The presence of random interfering sounds and variability in operating conditions can lead to lower performance and high false-positive rates. To overcome these limitations, we propose a fault detection method that leverages the underlying physical characteristics of the sound signals. By investigating the components of the acoustic signal, we found that fault-related sounds can be modeled as exponentially decaying oscillations. This insight allows for the development of a physically based signal model, setting our approach apart from purely data-driven methods. Using this model, we developed a robust detection method based on a Generalized Likelihood Ratio Test (GLRT). The effectiveness of this approach was validated using both synthetic and real-world data from a steel industry facility. Our results demonstrate that the proposed model-based approach provides superior performance compared to standard audio features, particularly in high-noise conditions. On real-world data, the GLRT-based approach outperformed all audio features, as clearly shown by the Receiver Operating Characteristic (ROC) analysis. Specifically, the Partial Area Under the Curve (pAUC) of the GLRT is more than twice that of the best-performing audio feature, demonstrating good detection at significantly lower-false-positive rates compared to audio features. Furthermore, simulations showed that our method maintains robust detection down to a Signal-to-Noise Ratio (SNR) of -13 dB, significantly outperforming audio feature-based detection, which was limited to approximately -10 dB. The physically informed nature of our model not only provides a more reliable and robust solution but also enables the method to be generalized to other industrial scenarios with similar fault properties, offering broader applicability for reliable acoustic condition monitoring.