Cross-modality domain adaptation for mechanical anomaly detection: A von mises-fisher VAE with enhanced interpretability

Abstract

Deep learning has shown great promise in mechanical fault detection, yet many existing methods struggle to adapt across varying operating conditions and often lack transparency in their decision-making processes. To overcome these limitations, we propose a novel approach that combines cross-modality domain adaptation with an interpretable deep learning framework. Our method leverages a von Mises-Fisher variational autoencoder (vMF-VAE) to learn meaningful representations of normal machine behavior, while domain adaptation techniques help ensure robust performance across different environments. To enhance interpretability, we incorporate a visualization module that highlights the critical data regions influencing the model's decisions. Extensive experiments on both vibration and acoustic datasets validate the effectiveness and adaptability of our approach. The results show that our method outperforms existing techniques and offers greater insight into model behavior, supporting more reliable and trustworthy fault detection in real-world applications.