Plan First, Judge Later, Run Better: A DMAIC-Inspired Agentic System for Industrial Anomaly Detection

Researchers propose DMAIC-IAD, an LLM-based multi-agent system for industrial anomaly detection that combines structured planning with pre-trained judgment models. The system achieves 37.76% performance improvement over existing agentic baselines by standardizing heterogeneous data inputs and evaluating strategies without costly runtime execution.

The research addresses a critical gap in deploying large language model agents for high-stakes industrial environments. While LLMs have demonstrated capability in automating complex workflows, their application to anomaly detection has been constrained by poor strategy formulation and inefficient handling of diverse data types. DMAIC-IAD resolves these limitations through a two-phase approach: first converting heterogeneous industrial references into standardized operating procedures, then employing a pre-trained judge model to evaluate candidate strategies without requiring expensive computational trials.

This advancement emerges within the broader context of AI reliability in critical infrastructure. Manufacturing quality, safety, and operational efficiency depend on rapid, accurate anomaly identification—domains where traditional statistical methods struggle with complex, multimodal data patterns. The DMAIC framework, borrowed from quality management practices, provides the structural discipline necessary to make LLM-based systems deterministic and auditable.

For industrial enterprises and AI practitioners, the implications are substantial. The 37.76% performance improvement translates directly to fewer false negatives in defect detection and reduced false-positive maintenance alerts. The execution-free judgment mechanism significantly lowers operational costs by eliminating wasteful trial runs, making enterprise AI deployment economically viable at scale. Organizations can now leverage LLM agents with confidence in safety-critical scenarios previously requiring specialized domain expertise.

The multi-modality support across four data types indicates the system handles real-world industrial complexity—sensor data, images, logs, and metadata. Going forward, the critical test involves deployment validation across diverse manufacturing environments and integration with existing quality-management systems.

• →DMAIC-IAD achieves 37.76% performance improvement by combining structured planning with pre-trained evaluation models for industrial anomaly detection.
• →The system standardizes heterogeneous data inputs into SOPs before strategy generation, enabling unified handling of multiple data modalities.
• →Pre-trained judge models eliminate costly runtime trials, significantly reducing computational expenses in LLM-based industrial applications.
• →The framework bridges the gap between LLM flexibility and industrial reliability requirements, enabling deployment in high-stakes manufacturing scenarios.
• →Multi-modality support across four data types demonstrates practical applicability to real-world manufacturing complexity.