CMR-Mamba: Causal Mechanism Monitoring for Cross-Domain Industrial Fault Detection

A recent research article available on arXiv (2608.14666) introduces CMR-Mamba, an innovative technique for unsupervised fault detection in industrial systems, specifically focusing on coupling faults—instances where the interrelations among sensor groups fail, even though the individual sensor distributions appear normal. Conventional reconstruction-based approaches tend to overlook these hidden failures, which can jeopardize both reliability and safety. CMR-Mamba employs per-domain Mamba state-space encoders trained on healthy datasets and incorporates a causal cross-modal predictor to regulate the encoders, ensuring that the effect-channel manifold accurately represents normal cause-and-effect relationships. Anomalies are assessed using k-nearest-neighbour (kNN) distances on this manifold or by analyzing the residuals between observed and causally predicted effect embeddings. The method has been tested on electromechanical (Paderborn bearings), hydraulic (ZeMA), and cyber-physical datasets. This paper was authored by a team of researchers and recently submitted to arXiv.