AI-Powered Automated Exam Generation for Aviation Maintenance Training

The maintenance department of JD Aviation faces rapid staff growth and extensive training requirements, needing to cover dozens of manuals and frequent regulatory updates, which makes manual question creation labor‑intensive and ineffective.

To improve training outcomes, a "test‑driven training" approach is proposed: generate a massive question bank covering all knowledge points, continuously update it as source documents change, and ensure 100% knowledge mastery.

Tool Selection : The AutoBots platform is chosen for its configurable vector store, knowledge retrieval accuracy, and workflow orchestration capabilities, enabling quick integration of large models, knowledge bases, and document parsing.

Overall Process Design : Users upload training materials to JoySpace; AutoBots parses and stores them in a vector database. Trainers input parameters (question type, difficulty, quantity, knowledge scope) into the training system, which assembles prompts and calls the large model via AutoBots to generate questions, returning results in JSON format for preview and one‑click addition to the question bank.

Large Model Workflow includes:

Receiving a command with parameters (e.g., query, knowledge range, question type).

Knowledge recall from the vector store to retrieve relevant manual excerpts.

Prompting the large model to generate single‑choice, multiple‑choice, true/false, or short‑answer questions with strict output formats.

Example prompt constraints:

{
  "type": "singleChoice",
  "question": "...",
  "options": [{"optionName":"A","optionContent":"..."}, ...],
  "correctAnswer": "A"
}

Sample generated questions for CCAR‑396‑R3 regulations are shown, demonstrating correct answer formatting and option structure.

System Integration : Generated questions are structured for easy ingestion, allowing users to add them to the question bank with a single click, supporting batch generation and future full‑document processing.

Continuous Upgrades include scaling generation volume, handling entire documents, adding duplicate‑rate checks, and implementing automated validity assessment of existing questions via periodic knowledge recall and large‑model evaluation.

The solution aims to free training staff from low‑efficiency, high‑effort tasks, showcasing a practical AI application that can be adapted to other domains requiring automated question or questionnaire generation.