A Case Study of Acoustic-based Anomaly Detection for Industrial Predictive Maintenance

Unplanned downtime in industrial machinery is expensive, which makes catching faults early genuinely valuable. This project builds an acoustic anomaly detection framework that listens to a machine and flags irregular sounds before they turn into failures. Raw recordings are converted into Mel-spectrograms and passed through an Autoencoder trained only on normal operation, so that unfamiliar sounds surface as high reconstruction error, with Local Outlier Factor, Isolation Forest, and PCA included as comparative benchmarks. Evaluated on both synthetic and real-world datasets, the study finds that Local Outlier Factor holds up best under noisy, variable conditions, while the Autoencoder is competitive in structured, low-noise settings and cheaper to run once trained. Their complementary strengths point to hybrid approaches for reliable acoustic monitoring in real industrial environments.