Edge AI system extracts structured data from maintenance work orders
ARID: A Deployable Edge AI System for Structured Information Extraction from Industrial Maintenance Work Orders
Computation and LanguageArtificial IntelligenceMachine Learning
Summary
Maintenance work orders often need to be processed on small computing devices without internet access, but the results must be clear and structured. The authors present ARID, a system that reads and organizes key information from these work orders into a fixed digital format on an embedded NVIDIA Jetson device. ARID uses several techniques to improve accuracy and efficiency, including filtering training data and controlling output formatting. It works well on tested data and runs quickly on low-power hardware, though its understanding of new task types can be limited.
What this means in practice
- •For industrial maintenance teams: Implement on-site extraction of component and failure details from work orders on embedded devices without cloud dependency.
- •For field service software developers: Integrate ARID’s fixed-schema output to automate maintenance record digitization on low-power edge hardware.
Authors
Kuanlin Chen, Chen-Wei Kuo
Abstract
Maintenance work orders must often be processed offline on embedded hardware, yet downstream software requires predictable structured output. We present ARID (Aviation-inspired Routing for Industrial Deployment), which extracts component, failure mode, symptom, and maintenance action into fixed-schema JSON on an 8 GB NVIDIA Jetson Orin NX. ARID combines conservative dual-teacher filtering, targeted noise-aware synthesis, one routing decision per work order, 4-bit inference, and grammar-constrained decoding. From 2,326 unlabeled OMIn records, it retains 716 training pairs and adds 99 topology-constrained records targeting action extraction. On 300 human-labeled records, ARID reaches 84.8% token-F1 on the reference stack and 82.9% on the deployed Jetson. Resident serving achieves 5,310/5,656 ms P50/P99 at 12.5 W. On zero-shot MaintNet transfer, semantic F1 falls to 46.4% while parser success remains at least 99.8%, showing that output validity transfers but field semantics do not.