A novel LLM-enhanced anomaly prediction framework for IoT time-series data

Le résumé fourni par la source

The rapid development of IoT has enhanced production flexibility, but concomitantly increased the risk of system failures. Existing anomaly prediction approaches necessitate the adaptation of feature extraction or model retraining for different scenarios, which complicates practical engineering. To address it, this paper proposes an LLM-enhanced anomaly prediction framework for IoT time-series data. First, a general LLM is fine-tuned to equip it with domain prediction capabilities. Next, a lightweight TCN-LSTM model is designed for resource-constrained edges. An adversarial distillation strategy is then proposed to align the hidden vectors of TCN-LSTM with those of the LLM, facilitating knowledge transfer from the fine-tuned LLM to TCN-LSTM. Experimental results demonstrate the efficacy of the proposed design.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.