Deep Hierarchical Knowledge Loss for Fault Intensity Diagnosis

A collaborative research team from multiple institutions has developed a novel AI framework called Deep Hierarchical Knowledge Loss (DHK) to significantly improve Fault Intensity Diagnosis (FID) in industrial settings. The core innovation addresses a critical limitation in existing methods: their failure to account for the hierarchical dependencies between different fault classes and intensities. By neglecting these relationships, conventional models struggle with practical deployment, especially in detecting subtle, early-stage faults that precede catastrophic failures.

The DHK framework introduces two specialized loss functions that work in tandem. The first is a hierarchical tree loss that creates a holistic mapping of classes sharing similar attributes, using both positive and negative hierarchical constraints derived from a tree structure. The second is a group tree triplet loss with a hierarchical dynamic margin, which incorporates group concepts and tree distance to better model the structural boundaries between different fault classes. This joint approach forces the AI model to learn a more consistent and nuanced representation of the fault hierarchy.

Extensive validation was conducted on four real-world datasets from diverse industrial domains, including three proprietary cavitation datasets from industrial manufacturer SAMSON AG and one publicly available dataset. The results, accepted for publication at the prestigious KDD 2026 conference, show the DHK framework consistently outperforming recent state-of-the-art FID methods. The system's improved recognition of subtle fault signatures allows for earlier and more accurate intervention, moving from simple fault detection to precise intensity diagnosis.