Answers are grounded in this paper and cite their sources.
Investigates long-term forecasting for servo motors in battery swapping systems, suggesting a new memory-enhanced framework.
Long-term forecasting of servo-motor signals is fundamental to predictive maintenance in battery swapping stations to prevent severe downtime. SegRNN is a relevant reference model for industrial long-horizon forecasting due to its competitive accuracy and memory costs, while LLMs offer a promising means of transforming forecasting results into actionable maintenance insights. However, SegRNN may still suffer from long-term memory degradation over extended horizons, and LLM-based reasoning is often constrained by static or insufficiently governed knowledge under evolving maintenance standards. To address these challenges, a long-term memory-enhanced forecasting framework is proposed, which combines robust numerical forecasting with expert-guided semantic reasoning. A time-frequency residual memory module is introduced to preserve multi-scale temporal patterns and alleviate long-horizon forgetting. In addition, a human-in-the-loop mechanism enables LLMs to dynamically refine the external knowledge base through expert feedback, improving explanation consistency and traceability. Experiments on a private real-world dataset—comprising over one million shuttle car servo-motor data points and 500 + fault reports—demonstrate that at a 720-step horizon, the proposed method reduces torque and speed RMSE by 0.1255 and 0.3486, respectively, compared to SegRNN. Additionally, the proposed HLME mechanism based on Qwen3 improves the ROUGE-L score of generated explanations by 0.137 over the RAG baseline.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: