Title: Optimize Performance Through Strategic Planning and Evaluation: A DMAIC-Based Agentic Framework for Industrial Anomaly Detection

Abstract:

While Large Language Model (LLM) agents demonstrate significant potential in automating intricate data-analysis pipelines, their reliable integration into high-stakes industrial environments remains a formidable challenge. Industrial anomaly detection (IAD) is critical for ensuring manufacturing quality, safety, and operational efficiency. However, current LLM-driven IAD solutions predominantly prioritize execution, often neglecting the crucial phase of strategic formulation. This imbalance limits their ability to manage heterogeneous data modalities in a unified and cost-efficient way.

Drawing inspiration from the DMAIC quality-management framework, we introduce DMAIC-IAD (DMAIC-inspired Agentic Industrial Anomaly Detection). This "Plan First, Judge Later" multi-agent system aligns LLM agents with structured industrial problem-solving methodologies. DMAIC-IAD first converts heterogeneous reference data into standardized operating procedures (SOPs) prior to strategy generation. Furthermore, it employs a pre-trained, execution-free judge model to rank candidate strategies, thereby eliminating the need for expensive runtime trials. Extensive experiments across four distinct modalities demonstrate that DMAIC-IAD enhances average detection performance by 37.76% compared to relevant agentic baselines.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC