• highly robust when faced with changes in illumination and environment
• can be adapted to new product variants through retraining
• 100% test rates without manual parameter optimization
• seamless integration into control systems with TwinCAT

• Quality and surface inspection
  Detection of cracks, scratches, soiling, or shape deviations on metallic, wooden, or polymer surfaces
  (Example: end-of-line testing of metal components, wooden panels, or injection-molded parts)
• Classification and sorting
  Sorting natural products according to quality grades, color, ripeness, or damage
  (Example: classification of eggs, fruit, or other agricultural products)
• Anomaly detection in processes
  Visual monitoring of packaging, welding, or assembly processes to detect faulty procedures
  (Example: detection of defective seal seams in food packaging)
• Object recognition and localization
  AI-supported detection of components, holes, or markings for robot-assisted pick-and-place or assembly processes
  (Example: aligning a bottle using the PET logo for precise labeling)

• recognition of subtle patterns beyond classic threshold values
• combination of multiple sensor signals into holistic state models
• seamless integration into control systems with TwinCAT

• Process monitoring and quality assurance
  AI-based assessment of welding, pressing, or packaging processes based on electrical or mechanical signals
  (Example: detection of faulty seals using servo current profiles)
• Process optimization
  Dynamic adjustment of parameters such as pressing force, feed rate, or temperature for maximum energy efficiency and product quality
  (Example: adaptive feed control for milling processes based on current curves)
• Forecasting and control
  Short-term forecasts of process variables such as wind direction, pressure, or temperature for predictive control
  (Example: prediction of wind speed and wind direction for the optimum alignment of wind turbines)
• Condition monitoring and predictive maintenance
  Early detection of bearing damage, imbalances, or pump wear through analysis of vibration and current signals
  (Example: detection of compressor defects via temperature and current curves)

Each application and also each IT infrastructure places different demands on the method of collecting machine data: SQL or noSQL, file-based, local or remote, limited port releases, cloud-based data lake, and many more. For all of these scenarios there are a large number of established TwinCAT products available, such as the TwinCAT 3 Database Server TF6420, the TwinCAT 3 Scope Server TF3300, TwinCAT 3 Analytics Logger TF3500, or the TwinCAT 3 IoT Data Agent TF6720. For image data, TwinCAT Vision provides an entire product family for image acquisition, image (pre)processing, and image storage.

Machine learning needs clean, representative data sets as a foundation. Depending on the task, these are pre-processed, labeled, and then trained – either automatically or manually.

For automation and process experts

With the TE3850 TwinCAT 3 Machine Learning Creator, users without prior AI knowledge can create AI models independently – quickly, intuitively, and without AI expertise.

• automated, intuitive training via a web-based no-code interface
• optimized for accuracy and latency for Beckhoff hardware
• export in open ONNX format for maximum interoperability
• provision of a PLCopen XML file for direct integration in TwinCAT

This creates a standardized, reproducible training process that requires no data science expertise.

For AI experts

Depending on their objectives and working methods, AI experts have two possible roles in the Beckhoff ecosystem:

• The AI expert with the Machine Learning Creator
  The TwinCAT 3 Machine Learning Creator can serve as an efficiency tool to standardize, accelerate, and document the training process. Data scientists use the tool to quickly generate an initial model (“version 0”), create benchmark data, or compare different configurations. The models generated can then be imported into specialized AI frameworks via ONNX and further refined there.
• The AI expert as an independent framework user
  Alternatively, data scientists can work entirely in their familiar environment. Models trained using frameworks such as PyTorch or TensorFlow can be exported directly as ONNX files and integrated into the Beckhoff AI system. This preserves complete freedom in model design, while execution takes place seamlessly in TwinCAT.

Both approaches are fully compatible and interoperable. This allows companies to optimally combine their existing AI expertise with the real-time and automation advantages of the Beckhoff environment without lock-in effects, but with maximum integration depth.

If the trained model is available as an ONNX file, it can be loaded and executed directly on the control computer. This makes AI an integral part of the PLC application – with all the advantages that entails:

• no additional hardware or interfaces required
• standardized system for maintenance, security, and updates
• direct access to all machine data in real time
• cost-efficient, scalable, and transparent

Hardware

Depending on the model size and latency requirements, various hardware options are available, ranging from CPU-based to GPU-accelerated solutions.

CPU-based execution: For many applications, executing AI models on standard CPUs without additional specialized hardware is sufficient. The familiar Beckhoff portfolio, ranging from Intel Atom^® to Intel^® Xeon^®, covers all performance levels and enables real-time, cost-efficient integration of AI models into the TwinCAT runtime process.

GPU-accelerated execution: For larger neural networks or high latency requirements, the use of GPU-based systems is recommended. The C6043 ultra-compact Industrial PC with Intel^® Core^TM i and NVIDIA^® GPU (e.g., RTX^TM A500 or RTX^TM 2000) provides the computing power needed to run complex deep learning or vision models in near real time, directly called from the PLC code.

Software

Two architectures are available for the productive use of AI models, which differ in terms of execution type, latency behavior, and hardware use. Both variants integrate fully into the TwinCAT environment and support the open ONNX standard.

Inference in the TwinCAT runtime: The AI model is executed directly in the real-time context of the control system (XAR). Calculations are performed synchronously and deterministically on the CPU – ideal for applications with fixed cycle times or hard real-time requirements. This variant requires no additional hardware and uses the familiar Beckhoff IPC portfolio. It is particularly suitable for relatively compact deep learning or classic ML models where low latency and precise timing are crucial.

Inference server (GPU-accelerated): Alternatively, AI execution can take place in a separate server process. This architecture enables near real-time processing with significantly higher AI computing power thanks to GPU support – especially in combination with the C6043 ultra-compact Industrial PC with integrated NVIDIA^® RTX^TM GPU. It is ideal for larger deep learning models that combine high model complexity with short response times. In addition, the server approach allows multi-client connections and thus central provision of AI models for multiple controllers.

The following table shows the respective properties:

AI models are capable of improving through training with larger sets of data. Likewise, general conditions can change gradually or spontaneously when the machine is being operated. This means you can update the trained AI models during the machine’s normal runtime – i.e., without stopping the machine, without recompilation, and also completely remotely through the standard IT infrastructure. Regardless of which TwinCAT product you choose to execute the AI models, a new AI model can be transferred to the control computer and reloaded via PLC functions.

What’s more, you can also operate your training environment remotely or locally on the industrial PC within the operating system context, allowing you to retrain, exchange, and load models with TwinCAT in close proximity to the process.