ConMem: Contribution-Aware Memory for Long-Horizon Manufacturing Inspection Logs

2026-07-30

AI Summary

Existing long-cycle inspection systems for steel plant equipment struggle to effectively retain historical records with early-risk indicative value, thereby limiting degradation预警 capabilities. This work proposes ConMem, a novel framework that segments inspection logs into functional evidence units and evaluates their diagnostic contribution using a Shapley-value-inspired metric. Under constrained memory budgets, ConMem prioritizes the retention of high-value evidence, enabling cross-cycle preservation and retrieval of subtle early degradation signals. The approach integrates function-role-aware segmentation, contribution-driven memory pruning, and retrieval-augmented generation. Evaluated on real-world data, it achieves a 76.0% question-answering accuracy—significantly outperforming the strongest baseline—while reducing input tokens by 88.2% and response time by 86.6% compared to an 8K-context LLM. Notably, it successfully captured early seal wear signals across three consecutive inspections.

Abstract

Long-horizon steel-equipment inspection requires reasoning over heterogeneous records accumulated across repeated inspection cycles. Existing retrieval-augmented generation systems treat historical logs as a static corpus and retain records without estimating their diagnostic value, failing to report early risk. To this end, we propose ConMem, a contribution-aware memory framework for LLM-assisted equipment inspection, supporting a human-in-the-loop early-risk screening system. Specifically, our ConMem first segments inspection logs into functional evidence units, then estimates each memory unit's contribution to downstream diagnosis through a Shapley-style estimation, and finally retains high-value evidence under a constrained memory budget. In experiments, we evaluate ConMem on real-world dataset and ConMem achieves 76.0% QA accuracy, exceeding the strongest directly comparable baseline. Relative to the naive 8K-context LLM baselines, it reduces the average number of input tokens by 88.2% and response time by 86.6%. Ablation studies also show that the functional-role-aware segmentation and contribution-based valuation are helping prioritize weak degradation signals for targeted field inspection. Practical deployments further confirm that ConMem retains the weak early signal across three inspection cycles, providing an early-stage seal-wear alert targeted for on-site inspectors.

Problem

• long-horizon inspection
• early-risk detection
• diagnostic value
• heterogeneous records
• memory retention

Innovation

• Contribution-aware memory
• Shapley-style estimation
• Functional evidence segmentation
• Long-horizon inspection
• Retrieval-augmented generation