VLT Model Boosts Industrial Intelligence with Multimodal Fusion

Industrial prognostics and health management (PHM) relies heavily on time series data to ensure the reliability and safety of equipment like aero-engines. However, current approaches often model only a single data modality, limiting their effectiveness in complex industrial scenarios. While large language models offer new opportunities for multimodal learning, effectively combining continuous time-series signals with discrete textual semantics remains a significant challenge. To address this, researchers propose VLT, a novel multimodal foundation model. VLT jointly processes time-series data, visual representations derived from frequency spectrums, and textual knowledge. A key innovation is using the frequency spectrum as an intermediary visual bridge to link continuous temporal signals with discrete semantic information. VLT incorporates a Time-aware Mixture-of-Experts (Time-MoE) to capture diverse temporal dynamics and a Frequency-Text Augmented Learner for joint modeling of spectral and semantic features. A time-centric gradient alignment mechanism further mitigates cross-modal optimization conflicts. Extensive experiments across various industrial datasets demonstrate VLT's superior performance, robustness, and generalization, even in few-shot, noisy, and incomplete-modality conditions.