Skip to main content
Back

HKUST's DMAIC-IAD Beats Factory Defect Detection Baselines by 37.76%

DMAIC-IAD borrows a Six Sigma quality framework to give LLM agents a structured plan before running industrial defect detection, beating baselines by 37.76%

Paper, Full Access
  • DMAIC-IAD is a new multi-agent system for industrial anomaly detection that structures LLM agents around the Six Sigma DMAIC quality framework.
  • It improves average detection performance over agentic baselines by 37.76% across four data modalities: 2D images, 3D point clouds, multimodal, and time-series.
  • A key innovation is an execution-free judge model that ranks candidate detection strategies without running them, saving significant compute cost.
  • The system converts heterogeneous reference materials into standardized operating procedures (SOPs) before generating any strategy, enabling cross-modality generalization.
  • Five specialized agents map to DMAIC phases: TAD Consultant, Data Inspector, IAD Strategist, Judge Model, and IAD Executor.
  • No code or weights are publicly released yet; the paper is available on arXiv.

Factory floors generate a relentless stream of sensor readings, camera feeds, 3D scans, and log files. Catching defects in that data is not just a computer vision problem -- it is a multi-modal reasoning problem that existing LLM-based agents have been ill-equipped to handle. A new paper from researchers at HKUST and collaborating institutions introduces DMAIC-IAD, a multi-agent system that borrows its structure from a decades-old manufacturing quality framework and applies it to automated anomaly detection -- with a 37.76% average improvement over comparable agentic baselines.

Why factory AI is harder than it looks

Anomaly detection is crucial for maintaining the safety, reliability, and optimal performance of complex systems across diverse domains such as industrial manufacturing, cybersecurity, and autonomous systems. But the industrial setting is uniquely punishing. Tasks usually require building models that can detect fine-grained visual defects under highly imbalanced data distribution and limited annotation. Defect samples are rare by definition -- you cannot train on what you are trying to prevent.

Building effective IAD collaboration frameworks still demands substantial manual effort across the entire pipeline, from data preprocessing to model selection and training strategy design. Recent agentic frameworks have tried to automate this, but they hit a fundamental wall: existing LLM-based agents are primarily designed for general-purpose task or code generation, but lack the domain-specific capabilities needed for IAD. They jump straight to execution without a coherent plan, and when the data changes modality -- say, from RGB images to time-series sensor logs -- they have no principled way to adapt.

Borrowing from Six Sigma

The core insight of DMAIC-IAD is to stop treating anomaly detection as a single-step inference problem and start treating it as a structured engineering process. The inspiration comes from DMAIC, a structured, data-driven problem-solving method used in Lean Six Sigma to improve existing processes. The acronym stands for Define, Measure, Analyze, Improve, and Control, and it provides a clear roadmap for identifying root causes, reducing variation, and creating reliable, repeatable results across any industry. Six Sigma was originally a set of techniques meant to improve and refine Motorola's manufacturing processes, and its core was known as DMAIC.