Release Notes: Agentic Predictive Maintenance#

Version 2026.2.0#

TBD

Features:

  • Initial release of the Agentic Predictive Maintenance (APM) blueprint.

  • Configuration-driven multi-agent pipeline using LangGraph. Adapt to any defect detection use case by editing four configuration files — no code changes required.

  • Four-agent reasoning pipeline: Policy Agent, Analysis Agent, Evidence Agent, and Ticketing Agent run sequentially to analyze detections and generate structured maintenance tickets. The reasoning agent (apm-agent) is not built from this repo — it is an external image (Intel EAL’s agent-quality-handler microservice) pulled via docker/compose.agents.yaml. See the agent-service integration guide for its contract.

  • Two operating modes: Large Language Model (LLM) mode for AI-generated analysis (using OpenVINO Model Server) and fallback mode for rule-based operation without an LLM service.

  • Real-time video inference via Deep Learning Streamer (DL Streamer) with YOLO-based object detection; DL Streamer publishes detection events over Message Queuing Telemetry Transport (MQTT).

  • SQLite database-backed storage service with Representational State Transfer (REST) API for querying detections and statistics.

  • Web dashboard (React) with live detection feed, run history, and ticket viewer.

  • The storage service and agent service both expose Prometheus metrics.

  • Reference use case: pipeline-defect-detection with four defect classes — Rupture, Deformation, Disconnect, and Obstacle.

  • Interactive Chat (Ask & Analyze): conversational interface in the web dashboard that lets operators query maintenance analysis, detection evidence, and pipeline data using natural language. Answers are generated from available pipeline data via the LLM service.

  • Data preparation script for downloading and building sample video from a public Kaggle dataset.

  • On-demand “Run Pipeline” trigger: one full detect-then-reason cycle per click — the DL Streamer pipeline runs once over the (finite) source video, then the agent pipeline reasons over exactly the detections that the run produced (an id-based window). Only one run may be in flight at a time; the agent-service rejects concurrent triggers with 409. Live and continuous background detection is planned for a future iteration.

Hardware Used for Validation:

  • 5th Gen Intel® Xeon® processors (CPU-only)

  • Intel® Core™ Ultra processors with Intel® Arc™ GPU (LLM mode)

Known Limitations:

  • Neural Processing Unit (NPU) inference support for the LLM service is experimental and is not validated for all model and configuration combinations.

  • Only the pipeline-defect-detection use case is provided as a reference configuration. Additional use cases require manual configuration file setup.

  • This release does not include a Helm chart for Kubernetes deployment.

On this page

This Page