AI for Industry
Anomaly Sound Detection
Overview:
In order to automatize product quality control, we develop new methods to detect malfunctioning products by detecting their anomalous operatings sounds.
Description:
The main difficulty with Anomaly Sound Detection is the lack of anomalous data, leading to unbalanced normal and anomaly datasets. We have investigated two approaches to tackle this issue: (1) the binary classification approach with Data Augmentation and Transfer Learning techniques; (2) the one-class classification approach with Bayesian Deep Learning techniques.
Contributions
Several Data Augmentation techniques able to generate realistic anomalous sounds.
A new binary classifier technique based on a Multi-Task Learning framework for Transfer Learning (Proof of Concept).
A new one-class classifier technique based on Bayesian Density Estimation which leverages the full 2D spectro-temporal information contained in the sound spectrograms. The results of this work were presented at the International Conference on Industry 4.0 and Smart Manufacturing (ISM2023) at Lisbon in November 2023.
