2717. Granularity-Driven Post-Correction Model for Digitizing Handwritten Aviation Maintenance Records
Invited abstract in session MD-28: Human-AI Collaboration and Ethics, stream Decision Support Systems.
Monday, 14:30-16:00Room: Maurice Keyworth 1.03
Authors (first author is the speaker)
| 1. | Jianwu XUE
|
| 2. | Zhaoyang Zhang
|
Abstract
This study addresses the critical challenges in digitizing handwritten civil aviation maintenance records through an innovative approach that integrates granularity-aware tokenization with post-recognition correction mechanisms. First, we construct a domain-specific corpus optimized for tokenization by combining web-crawled aviation maintenance terminology with offline collected records, analyzed using WordSmith to establish linguistic patterns, resulting in a 23.6% improvement in labeling consistency. We then develop a correction model embedding tokenization-aware mechanisms within KenLM (statistical) and Transformer (neural) architectures to address seven distinct error categories prevalent in maintenance documentation, including character mis-segmentation and domain-specific terminology errors. Evaluation using engine maintenance records demonstrated that our Transformer-f model achieved a 12.3% accuracy improvement and 18% speed enhancement over baseline approaches. Finally, we implement a system using Spring MVC architecture that integrates recognition modules with granularity-driven correction algorithms and dynamic corpus management capabilities. Field testing across multiple maintenance facilities yields a 94.7% correction accuracy and reduced manual review requirements by 40%. This research provides a practical solution for aviation maintenance digitization that effectively balances accuracy with operational efficiency while maintaining regulatory compliance.
Keywords
- Decision Support Systems
- Knowledge Engineering and Management
- Manufacturing
Status: accepted
Back to the list of papers