DESIGN AND IMPLEMENTATION OF A HYBRID AI ARCHITECTURE FOR MANUFACTURING PROCESS PLANNING

Abstract

Modern manufacturing demands efficient, precise, and cost-effective machining process planning, yet traditional methods rely on manual expertise, leading to inconsistencies and knowledge transfer challenges. Computer-aided process planning (CAPP) systems automate selections but lack explainability and adaptability, while general-purpose large language models (LLMs) suffer from high hallucination rates (∼ 41%) and insufficient grounding in engineering standards, risking production errors and intellectual property breaches. This paper introduces a hybrid AI architecture for manufacturing process planning, integrating rule-based database logic with LLM-enhanced natural language processing to ensure accuracy, transparency, and privacy. The system employs a two-tier framework: Tier 1 filters operations through deterministic checks on ISO tolerances (IT01–IT18), geometric compatibility, and precision/roughness boundaries; Tier 2 leverages the locally deployed Qwen3:14b LLM (Ollama) for contextual explanations and query resolution, constrained to validated data to mitigate hallucinations. Key contributions include: a domain-constrained AI agent with query classification for theoretical, practical, and computational routing; a three-layer web architecture (Angular frontend, Flask backend, MySQL database) supporting dual interfaces—a conversational educational mode and a tabular planner for multi-surface analysis; multi-criteria optimization for ranked recommendations, including economic vs. maximum precision classifications and exportable reports. Implemented on local hardware (9 GB RAM), the system processes specifications (geometry, R_a values) in under 5 seconds, aligning with Industry 5.0's paradigms. Evaluations demonstrate superior reliability over traditional CAPP and LLMs, fostering knowledge transfer and adaptive planning.