The predictive_maintenance_pipeline project demonstrates a complete agentic edge AI system for Predictive Maintenance (PdM) of critical infrastructure (e.g., pipelines, bridges, solar panels) using Intel Edge hardware and software stack.
It combines real-time inference, structured data storage, and a multi-agent reasoning layer coordinated by a Meta-Agent.
Note: The terms PACE (PdM using Agents for Critical infrastructure on the Edge) and predictive_maintenance_pipeline are used interchangeably throughout this project.
This is a Proof of Concept (PoC) and is not intended for production systems. Use at your own risk — we take no responsibility for any deployment, data usage, or results. Currently, this PoC supports only the pipeline defect detection use case.
See also: CONTRIBUTING.md · SECURITY.md
See QUICKSTART.md for setup, training, model conversion, and running the pipeline.
The architecture consists of three major units:
- Intel DLStreamer (OpenVINO) inference pipeline
- YOLO object detection on Intel iGPU or CPU
- Inline visualization via
gvawatermark - Video and image mode support
- Detections written to SQLite
- Embedded, serverless database at
out/sql_data/detections.db - Stores per-frame detections (label, confidence, bounding box)
- Supports text-to-SQL via SQLCoder model (optional)
A hub-and-spoke agentic system powered by LangGraph:
| Agent | Role |
|---|---|
| Meta-Agent | Central coordinator; orchestrates all agents |
| Policy Agent | Applies filtering and threshold rules via SQL |
| Analysis Agent | Generates summaries and confidence statistics |
| Evidence Agent | Stores justifications and decisions for traceability |
| LLM Glue Layer | Text-to-SQL, schema validation, reasoning |
All agents communicate only through the Meta-Agent — no direct agent-to-agent messaging.
LLM Modes: fallback (rule-based, no LLM) · model (local OpenVINO LLM) · server (remote LLM server)
pace/
├── src/
│ ├── agents/ # Multi-agent system (meta, policy, analysis, evidence)
│ │ └── utility/ # LLM client, DB backend, state, caching
│ └── utility/ # SQLite client, prompt loader
├── scripts/ # Download data/models, run agents, LLM server, utilities
├── setup/ # setup.sh, convert_to_openvino.py, requirements.txt
├── config/ # Use-case YAML configs
├── models/
│ ├── pt_models/ # PyTorch models
│ └── ov_models/ # OpenVINO models (YOLO + LLMs)
├── datasets/ # Dataset files and dataset.yaml
├── prompts/ # Agent prompt templates per use case
├── docs/user-guide/ # Documentation
├── web_app/ # Web application interface
├── out/ # Output (SQLite DB, agent reports, visualizations)
├── config.json # Main config (use-case-id)
├── run_complete_pipeline.py # End-to-end pipeline
├── run_inference_oep.py # DLStreamer inference
└── interactive_chat.py # Interactive agent chat
config.json — Sets the active use case:
{ "use-case-id": "pipeline_defects_detection" }config/pipeline_defects_detection.yaml — Use-case specific settings for inference, agents, LLM mode, and SQL. See the file for all options.
| Document | Description |
|---|---|
| QUICKSTART.md | Setup, training, model conversion, running the pipeline |
| AGENT_ARCHITECTURE.md | Agent system design |
| LLM_SERVER.md | LLM server architecture |
| TROUBLESHOOTING.md | Common issues and solutions |
| prompts/README.md | Agent prompt format guide |
- Intel OpenVINO Documentation
- OpenVINO GenAI
- Ultralytics YOLO Documentation
- LangGraph Documentation
- SQLite Documentation
Predictive Maintenance Pipeline is licensed under the Apache License 2.0.
The project also includes Intel binary components (openvino, openvino-genai, optimum-intel[openvino], dlstreamer) governed by the Intel Simplified Software License (Version October 2022).
See THIRD-PARTY-PROGRAMS for the full list of third-party dependencies and their license terms.
Note: This project uses Ultralytics YOLO which is licensed under AGPL-3.0. Commercial use without open-sourcing your application requires a separate commercial license from Ultralytics.
For quick start instructions, see QUICKSTART.md.