VLT: A Vision-Language-Time Series Multimodal Foundation Model for Industrial Intelligence

arXiv cs.AI · 6 hours ago

Researchers introduced VLT, a multimodal foundation model that combines time series data, frequency spectrum visualizations, and text to improve industrial equipment monitoring and diagnostics. The model uses frequency spectrum as a visual bridge to connect continuous signals with discrete language, employs a Time-aware Mixture-of-Experts component for temporal dynamics, and applies gradient alignment to handle cross-modal conflicts. The approach demonstrates improved performance on industrial datasets across few-shot, noisy, and incomplete data scenarios compared to existing single-modality methods.