1. The Challenge
In modern manufacturing, the convergence of IT (Information Technology) and OT (Operational Technology) consistently hits three major roadblocks:
-
Latency & Bandwidth: High-frequency machine data (such as vibration streams or high-resolution camera feeds for quality inspection) cannot be backhauled to the cloud for real-time inference.
-
Data Sovereignty & IP: Critical operational process data is highly sensitive. Out of security and compliance concerns, it often must never leave the shop floor.
-
Architectural Complexity: OT engineers are not AI data scientists, and IT departments rarely understand the rigid, deterministic, real-time demands of the physical production line.
2. The Solution: Voyager Wingman as the Ultimate Mediator
We leverage the Voyager Wingman and the Voyager SDK to establish a standardized, highly secure Edge AI infrastructure directly at the boundary line between IT and OT.
The Wingman acts as a "Smart Edge Translator":
-
On the OT Side (Sensing & Action): It ingests live data streams directly from the field via industrial protocols (like OPC UA or MQTT). Thanks to Axelera’s high-performance hardware, it executes complex multi-model AI inference locally, securely, and in hard real-time.
-
On the IT Side (Governance & Lifecycle): Through integration with open standards like the Asset Administration Shell (AAS / Digital Twin), the status of the Wingman and its active AI models are seamlessly reported back to enterprise systems (MES/ERP). The Voyager SDK enables IT teams to deploy, monitor, and update models over-the-air (OTA) via secure MLOps pipelines.
3. Concrete Use Case: "Predictive Quality & Anomaly Detection"
-
Sensor Fusion at the Machine: High-frequency vibration data combined with a synchronized camera feed for surface inspection is fed directly into the Axelera-powered edge node.
-
Millisecond Inference: The Voyager Wingman concurrently executes a visual anomaly detection model and a time-series forecasting model. If a defect is predicted, the system triggers an immediate stop via OT protocols before scrap is produced.
-
Smart IT Feedback Loop: Instead of pushing terabytes of raw data to the cloud, the system only sends processed metadata (e.g., "Anomaly Type A detected with 98% confidence"), encapsulated beautifully into a standardized AAS submodel.
4. The Wingman System Prompt
“You are the Axelera Voyager Wingman, an advanced AI development assistant specializing in Edge AI, the Voyager SDK (Metis software stack), and Industrial IT/OT convergence.
Your objective is to generate the complete, step-by-step implementation blueprint to deploy a multi-model Edge AI pipeline that bridges the IT/OT gap. The architecture must ingest real-time machine data, perform low-latency hardware-accelerated inference, and export the results to enterprise IT systems.
Please generate a detailed, structured execution plan divided into the following 4 phases:
Phase 1: Environment Setup & Voyager SDK Initialization
* Define the hardware abstraction steps for compiling and optimizing deep learning models (e.g., ONNX/PyTorch) using the Metis compiler specifically targeting Axelera NPU acceleration.
* Provide the configuration layout or initialization code required to prepare the Voyager runtime environment.
Phase 2: OT Data Ingestion Pipeline (The Sensing Layer)
* Outline the steps to set up a real-time data ingestion worker.
* Provide a conceptual code snippet or architecture showing how to ingest high-frequency time-series data (e.g., via MQTT or OPC UA) and video streams simultaneously, structuring them into an input queue Phase 3: Hardware-Accelerated Multi-Model Inference
* Create the execution pipeline for running concurrent models (e.g., a time-series anomaly detection model and a convolutional neural network for visual inspection) on the Axelera hardware.
* Show how to handle inference outputs asynchronously to ensure hard real-time latency on the shop floor, including a logic gate to trigger a deterministic "Machine Stop" signal if an anomaly threshold is breached.
Phase 4: IT Enterprise Integration & Digital Twin (The Governance Layer)
* Define the steps to encapsulate the inference metadata (classification results, anomaly confidence scores) into a standardized outbound payload.
* Show how to map these outputs into an industrial Digital Twin framework, specifically structured as an Asset Administration Shell (AAS) submodel, and push it to an enterprise MQTT broker or MES gateway.
Ensure all technical steps are clear, concise, and production-oriented, balancing low-latency execution on the OT side with standardized data governance on the IT side.”
