IEEE IRAI 2026 · Peer-Reviewed Paper

An Agentic Generative AI Framework for Industrial Predictive Maintenance: Physics-Aware Anomaly Detection with Multimodal Retrieval-Augmented Generation

R. W. V. Imansha Dilshan, A. L. Supun Tharaka, P. D. D. Ransika, W. G. N. Dananjaya, M. B. Abeywardena, Dr. Akila Wijethunge

Accepted · to appear in IEEE Xplore IRAI26-000215

Abstract

This paper presents a neuro-symbolic agentic generative AI framework for industrial predictive maintenance, unifying sub-symbolic fault detection with symbolic Large Language Model (LLM) reasoning and multimodal retrieval-augmented generation (RAG). Four contributions are introduced: a multimodal RAG engine employing GPT-4o Vision captioning within a unified 1536-dimensional pgvector embedding space; a six-node LangGraph multi-agent pipeline with safety-critical terminal validation; a physics-aware hybrid confidence layer fusing dual-architecture autoencoders with manufacturer-specified operational constraints; and an institutional intelligence subsystem vectorizing resolved incidents through a five-gate quality pipeline into queryable organizational memory. Empirical evaluation at a tea manufacturing facility yields 90.15% detection precision at 3.35% False Positive Rate (FPR), 6.8-second fault-to-procedure latency representing a 300× reduction over manual baselines, cross-machine F1 > 0.71 across four asset types, and validated bidirectional institutional learning.

Agentic AI
Retrieval-Augmented Generation
Predictive Maintenance
Multi-Agent Systems
Anomaly Detection
Industrial IoT
Human-in-the-Loop
Large Language Models

Conference

1st IEEE International Conference on Responsible Artificial Intelligence

3-5 September 2026
Melbourne, Australia
La Trobe University City Campus
IEEE Industrial Electronics Society (IES)

Conference website See the product

90.15%

Precision

3.35%

False positive rate

0.8475

AUC-ROC

183×

MSE separation

6.8 s

End-to-end latency

> 0.71

Cross-machine F1

Contributions

Four Key Contributions.

Multimodal RAG

Caption-based vision-language alignment (YOLOv8-DocLayNet, GPT-4o Vision, Voronoi tessellation, Mobile SAM) that enables cross-modal retrieval in one embedding space, with no auxiliary visual index.

Six-node LangGraph DAG

Diagnostics split into specialised cognitive roles with auditable reasoning and a terminal Safety Critic that enforces LOTO and PPE compliance.

Physics-aware hybrid confidence

Dual-architecture autoencoders combined with manufacturer operating limits, which cuts false positives well below rule-based SCADA.

Institutional intelligence

A five-gate quality pipeline that turns resolved incidents into queryable organisational memory. This gives human-in-the-loop learning without retraining.

Figures

Architecture & Figures.

System architecture of the agentic predictive maintenance framework — Fig. 1: System architecture: the Sense -> Detect -> Reason -> Advise -> Learn pipeline.

Five-gate quality pipeline for organizational memory — Fig. 3: Five-gate quality pipeline for archiving resolved incidents into organisational memory.

ROC curve with AUC 0.8475 — Fig. 5: ROC curve for the primary asset (AUC-ROC = 0.8475).

Confusion matrix over 20,000 observations — Fig. 7: Confusion matrix over N = 20,000 observations.

Generalisation

Evaluation Across Four Asset Types.

Per-machine models were evaluated on four industrial asset categories with no architectural changes.

Asset	Type	Acc. %	Prec. %	F1	AUC	FPR %
Tea Pourer	Pouring Machine	89.13	90.15	0.7979	0.8475	3.35
Lathe	Lathe Machine	87.75	86.00	0.7758	0.8487	4.93
Pump	Centrifugal Pump	85.96	85.15	0.7335	0.8347	4.81
Turbine	Industrial Turbine	85.02	84.09	0.7121	0.7980	5.01

Evaluation used synthetic telemetry parameterised from field-recorded specifications. Further validation on live operational fault data is conducted prior to production deployment.

Cite this work

@inproceedings{zynaptrix2026agentic,
  title     = {An Agentic Generative AI Framework for Industrial Predictive Maintenance: Physics-Aware Anomaly Detection with Multimodal Retrieval-Augmented Generation},
  author    = {R. W. V. Imansha Dilshan and A. L. Supun Tharaka and P. D. D. Ransika and W. G. N. Dananjaya and M. B. Abeywardena and Akila Wijethunge},
  booktitle = {Proc. 1st IEEE International Conference on Responsible Artificial Intelligence (IEEE IRAI 2026)},
  address   = {Melbourne, Australia},
  year      = {2026},
  publisher = {IEEE},
  note      = {Accepted, to appear}
}

The official IEEE Xplore DOI will be updated upon proceedings release.

Pilot this on your assets

We're onboarding enterprise partners to evaluate our multi-agent diagnostic framework on live enterprise data.

Start a conversation zynaptrix.team@gmail.com