AI Suite on Industrial Edge

Download detailed architecture (PDF)

Download detailed PDF

Field level: Industrial Edge as the AI execution layer

Industrial Edge devices sit directly on the shop floor and connect to PLCs, drives, robots, cameras, and any other automation equipment using pre-configured connectors for PROFINET, S7, OPC UA, EtherNet/IP, Modbus TCP, and others. Because the connector library covers equipment from any vendor, the architecture also fits brownfield environments without requiring hardware replacement.

A set of local apps runs on the edge device alongside the connectors:

• AI Inference Server for on-device model execution, supporting vision, time-series, and batch inference use

  cases

• Vision Connector Application for connecting to GigE industrial cameras and RTSP cameras to deliver vision data for inference
• Vision Data Collector for capturing images and metadata from cameras and vision systems, along with inference results from runtime, feeding the (re)training data pipeline
• Industrial Information Hub, which maps raw PLC tags and inference results to a consistent semantic data model before data leaves the device
• LiveTwin and Virtual PLC for digital twin simulation and virtual control
• Mendix on Edge for role-based operator interfaces that span both the edge and upstream systems
• Energy Manager and Performance Insight for operational KPIs including energy consumption and OEE
• IT Connectors for connectivity to enterprise systems

Databus, based on MQTT, connects these apps to each other on the device and provides the publish-subscribe backbone for passing inference results, sensor readings, and events up to the factory level. Vision data between vision connector and inference server is transmitted using ZMQ for handling larger, high frequency payloads.

Factory level: the AI operations layer

The AI Asset Manager runs on a virtual Industrial Edge device at factory level and acts as the one-stop shop for all AI-related activities on the shop floor. It sits between the development environment above and the edge devices below, coordinating the full operational lifecycle of AI solutions.

AI Asset Manager: model distribution and operations

The AI Asset Manager's job is to receive packaged AI models from the development environment, deploy them to the correct AI Inference Server instances across the fleet, and collect metrics on model performance and inference activity. It manages AI solution versioning, monitors device-level deployment status, and provides the operational interface through which automation teams manage AI without needing to interact with development toolchains.

Use the AI Asset Manager for:

• Pulling packaged models from the IT-level development pipeline and distributing them to edge devices
• Managing model versions across a fleet of Industrial Edge devices, including rollback and staged rollout
• Collecting inference metrics and performance data from deployed models
• Providing a single operational view of AI solution status across all devices and sites

The AI Asset Manager is not a development tool. It does not train models, validate datasets, or manage development infrastructure. Those responsibilities belong to the MLOps workflow in the cloud or on-premises development environment. AI SDK packages the AI Model and delivers read-to-deploy artifacts to the Factory level architecture layer, where the AI Asset Manager's scope begins[AN1] and ends when operational metrics feed back into the development cycle.

Industrial Edge Management (Virtual, Pro, or Cloud) handles the broader device management layer: deploying apps, pushing firmware and configuration updates, monitoring device health, and managing the Industrial Edge Hub as the global app repository. It works alongside the AI Asset Manager rather than replacing it — Edge Management handles the platform; the AI Asset Manager handles the AI solutions running on that platform.

IT and enterprise level: the AI development environment

Model development takes place in cloud or on-premises environments using the Siemens AI SDK. The pipeline at this level covers the full development lifecycle before models reach the factory.

Siemens AI SDK: model development and packaging

The AI SDK provides data scientists with the tooling to package and validate their AI models in an environment of their choice. It is a python library that provides methods to define data interfaces for AI models with other systems (automation, for example), define runtime requirements and package the AI model along with the business logic into an artifact that can be executed completely offline on the shopfloor.

Use the AI SDK for:

• Packaging AI models and generate validated, deployable artifacts for the AI Asset Manager, which eventually can be executed by AI Inference Server on the shopfloor, using real time production data from diverse sources.
• Integrating with AWS, Azure, or on-premises MLOps environments to deliver packaged AI Models to the factory level

Once packaged, models are pulled by the AI Asset Manager and distributed to the fleet. Updated models trained on new production data follow the same path, closing the development-to-deployment loop.

Why the full suite is deployed together

A realistic deployment uses all three levels in combination because they handle distinct problems. Consider a visual quality inspection deployment on an electronics assembly line:

• The Vision Data Collector along with vision connector application captures images of assembled boards at each inspection station. Images and metadata flow into the data landing zone (Cloud storage, (S)FTP) for consumption by the MLOps workflow
• Data scientists use their own MLOps workflow to (re)train a defect classification AI model on that production data, validate it, and package it as a deployable artifact using the AI SDK
• The AI Asset Manager pulls the packaged model and deploys it to the AI Inference Server on the relevant Industrial Edge devices across all inspection stations
• Vision Connector application provides connectivity to the station cameras for capturing the image of the board and provides it as an input to the AI model on the Inference Server
• The AI Inference Server runs the model locally at each station, classifying boards as pass or fail in real time without a cloud dependency
• Inference results are published to the Databus and ro

  uted to quality management systems or operator dashboards

• Asset manager also collects indicative metrics from each deployment and allows the user to dashboard for easy visualization and alarming based on rules
• Defect images and classification results flow back into the data pipeline via vision data collector. The model is retrained on expanded data, repackaged, and pushed back out to the fleet

Without the AI Inference Server, inference requires cloud connectivity and introduces latency incompatible with line-speed inspection, apart from the costs that are incurred for each data transaction. Without the AI Asset Manager, deploying an updated model to fifty stations across three sites would be fifty manual operations. Without the vision data collector and a structured data pipeline, training data does not reflect real production conditions and model quality degrades over time. AI SDK enables piecing together the repeatable delivery by standardizing the delivered artifact, agnostic to the kind of AI model being deployed.