Enhancing Domain Generalization for Unsupervised Anomalous Sound Detection Through Domain Classification Loss Integration

Machine condition monitoring is a task of high practical relevance in various industries, as early detection of potential equipment malfunctions can reduce downtime of machinery, increasing productivity and cutting down on costly breakdowns. Anomalous sound detection of machines using deep learning has been gaining attention recently due to the low installation cost and scalability of such systems. Autoencoder neural networks have been widely adopted for this problem due to their capability of training in an unsupervised manner. Autoencoders can learn to compress and reconstruct normal machine sounds and hence can classify machine sounds where the reconstruction error exceeds some threshold as anomalies. However, the performance of autoencoders tends to decline in real-world scenarios as the operational and environmental conditions of the monitored machine are constantly changing. This exposes autoencoder models to new acoustic events that can yield high reconstruction errors due to their significant deviation from the statistical distribution of the training data and thus can be classified as anomalies even though they are normal. Therefore, we propose a new autoencoder architecture that augments traditional autoencoders with a domain classifier to enhance their domain generalization ability. The aim of the domain classifier is to make the encoder network generate domain-invariant latent representations, consequently reducing the distribution difference between the source and target domains during the training of the autoencoder. This makes the autoencoder learn the normal sound patterns of the machine regardless of the domain, effectively improving performance in real-world scenarios. To demonstrate the effectiveness of our proposed autoencoder architecture, we use the gearbox machine as a case study. The proposed autoencoder architecture obtains competitive results on the gearbox machine dataset extracted from the MIMII DG dataset.