Publikation
A Domain-Agnostic Neuro-Symbolic Architecture for Multimodal Human-in-the-Loop Anomaly Detection and Complex Fault Diagnosis
Tim Bohne; Anne-Kathrin Patricia Windler; Martin Atzmueller
In: IEEE Access (IEEE), Vol. 13, Pages 210201-210236, IEEE, 12/2025.
Zusammenfassung
This paper presents a general architecture for iterative, hybrid neuro-symbolic anomaly detection and complex fault diagnosis, in which symbolic knowledge-based methods and neural machine learning methods reinforce each other. For evaluation, we introduce a neuro-symbolic diagnosis benchmark that systematically assesses the architecture using randomized, parametrized synthetic problem instances with ground truth solutions. These are derived from an abstract formalization of the general problem of diagnosing systems composed of causally interconnected components based on sensor signal evaluation. It results in a domain-agnostic diagnostic framework, where synthetic instances capture a multitude of practical domains, enabling robust, empirically grounded conclusions. Explainability and interpretability emerge naturally through the specific neural-symbolic interplay. The architecture serves as a transferable blueprint for diagnosing systems across domains involving causal structure and sensory assessment.
Projekte
- AW4.0 - Autowerkstatt4.0
- Railway-X - KI-basiertes Wissens- und Datenmanagement zur Entscheidungsunterstützung und Handlungsempfehlung in der internationalen Bahnindustrie