Machinery condition monitoring through acoustic emission

Sammanfattning

Unplanned failures in rotating machinery cause significant financial losses and safety risks across industry. Condition-based maintenance systems that monitor equipment health in real time offer a proactive alternative to reactive repair, but traditional vibration-based sensing often fails to detect early-stage faults due to background noise. Acoustic emission sensing, which captures high-frequency elastic stress waves generated by phenomena such as crack initiation and material friction, provides higher sensitivity to early mechanical degradation. This project developed a diagnostic pipeline for classifying rotating machinery as normal or faulty from acoustic recordings. Raw audio signals were transformed into Melspectrograms using the Short-Time Fourier Transform and used to train a custom lightweight Convolutional Neural Network with 298,562 parameters. The model was evaluated on the Revix Engine Knock dataset, comprising 1,199 recordings split into 80% training, 10% validation, and 10% test subsets using a stratified split. A Support Vector Machine and four large pretrained deep learning models were used as baselines for comparison. The custom network achieved a macro F1-score of 0.838 on the test set, outperforming the Support Vector Machine baseline(0.807) and matching or exceeding three of the four pretrained models. Only the largest pretrained model exceeded the custom network, achieving a macroF1-score of 0.858 with 79 times more parameters. The inference time of 0.018milliseconds per sample satisfied the real-time requirement by a large margin. The results demonstrate that Mel spectrograms combined with a purpose-builtconvolutional network provide an effective and computationally efficient approach to acoustic fault detection. The primary performance requirement of a macroF1-score of at least 0.85 was not met by the custom network, primarily due to the limited size of the dataset and a low recall for the faulty class. The system is suitable as a supplementary monitoring tool and provides a foundation for further work on larger datasets and real-world deployment.