BRIDGE: big data-powered large language models for real-time multi-modal decision support in industrial time series
Year: 2026
Type: article
Abstract: Industrial Internet-of-Things environments generate heterogeneous sensor streams, topology metadata, and operational logs that must support forecasting, anomaly detection, and operator-facing explanation under deployment constraints. We present BRIDGE (Big data-driven Real-time Industrial multi-modal Decision support via Graph-Enhanced modeling), a topology- and log-aware industrial time-series framework that integrates: (i) a Graph-enhanced Topology Encoder using physical, metadata-based, training-only inferred, or identity graphs; (ii) an xLSTM temporal backbone trained with Contiguous Graph Patch Masking to simulate localized sensor outages; and (iii) an optional Cross-modal LLM Alignment module that generates reports without fine-tuning the LLM backbone. Because no single benchmark contains all modalities, we distinguish numerical-only, topology-aware, and log-enhanced evidence tiers. BRIDGE is evaluated on GiftEval-ZS and Chronos-ZS forecasting benchmarks, four industrial anomaly benchmarks, and 200 log-enhanced report windows. It obtains a CRPS of 0.396 on GiftEval-ZS and an $$F_1$$ score of $$92.4\%$$ on SMAP. The numerical predictive path requires 12 ms per sample at batch size 128 on the tested hardware, excluding autoregressive LLM decoding. The results support topology-aware integration for industrial monitoring while limiting claims about complete multi-modal effectiveness to the modalities actually evaluated.
Authors: Fulin Li, Jun Nie, Zhijing Li
Institution: Guangdong University of Science and Technology
Language: English
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,990
OpenAlex ID: W7155074659
