TE3850 TwinCAT 3 Machine Learning Creator is a cloud-based development environment that automates the entire AI model pipeline – from uploading and annotating data to the trained ONNX model ready for use in your TwinCAT 3 PLC. The system standardizes and automates model creation, reduces development work, and speeds up the time to result by more than one order of magnitude.
It enables automation engineers to create and implement AI models independently without the need for specialized data science expertise. Hardware-compatible training with configurable runtime restrictions ensures that the models are optimized for execution at control level. Integrated analysis and validation functions ensure transparent model behavior and enable evaluation before use in industrial applications.
To use the Machine Learning Creator, log in with your myBeckhoff account:
Product roadmap and release status
Stay up to date with the latest functions and upcoming innovations of the TwinCAT 3 Machine Learning Creator. The roadmap provides an overview of available, beta, and planned functions for computer vision and signal-based AI applications.
If you are interested in beta or upcoming functions, contact your local Beckhoff representative for more information and access options.
Assigns entire images to predefined categories based on learned visual characteristics. Ideal for applications where the complete image represents a clear state or situation.
- assigns a label to each image (no object localization required)
- requires labeled training data for all relevant classes
- well-suited to quality inspections, sorting processes, and global OK/nOK decisions
- simple output, can be used directly in the control logic
Detects deviations from a learned normal state without the need for explicit error classes. Ideal for applications in which errors are rare, unknown, or generally difficult to define completely.
- recognizes deviations from a learned “normal” visual state
- can work with no or very few images of defects
- is mainly trained based on “normal” data in general
- “non-normal” data can be included in training and is extremely beneficial
- pixel-precise localization of anomalous areas is possible
- well-suited to unknown, rare, or changing defects
Recognizes, localizes, and classifies one or more predefined object classes within a single image. Ideal for applications in which certain objects need to be identified, counted, and evaluated based on their position and size.
- recognizes and classifies one or more objects within an image
- recognizes several instances of the same class at the same time
- provides position and size information for each detected object
- is particularly suitable for counting, localization, and size-based evaluation
- enables more detailed decisions than pure image classification
Classifies sequential or structured data such as sensor signals into predefined categories based on learned patterns. Ideal for applications in which machine or process states can be identified based on characteristic signal curves over time, position, or other ranges:
- assigns a classification to each signal sequence or each defined window
- works with data such as currents, vibrations, temperatures, or pressures
- supports signals over time, position, angle, or other continuous ranges
- requires classified examples for each relevant class
- is well-suited to condition monitoring and error detection
- enables automated detection of machine and process states
Predicts future values of sequential or structured data based on learned patterns. Ideal for applications in which future system behavior or trends need to be anticipated to enable optimization or control over a time frame or other continuous ranges.
- forecasts future values for a defined time frame
- processes data such as energy consumption, temperature, or process variables
- supports temporal signals and other continuous data ranges
- recognizes dependencies and recurring patterns in data
- is particularly suitable for predictive control and process optimization
- enables proactive decision-making instead of reactive solutions
Industrial AI in practice
Metal processing
Inline quality inspection
The geometric shape is a key quality feature in the production of metal workpieces. Instead of complex measurement methods, simple evaluation as OK/nOK is sufficient in many cases.
An AI model for classifying workpieces was trained based on a small, annotated image data set. Despite the fact that there are numerous different defect patterns, the model reliably detects deviations. The model was created with the TwinCAT Machine Learning Creator without expert AI knowledge and can be seamlessly integrated into the machine control system.
Read the application report here.
Wood industry
Inline quality inspection
Traces of glue are checked on wooden components in an automated joining process. Under ultraviolet illumination, these appear as bright stripes on the surface and serve as a quality feature for the process.
Due to the wide variety of defect patterns and varying conditions such as glue type, temperature, or extraneous light, conventional image analysis can only do so much.
By using an AI-based classification model, the traces of glue can be robustly evaluated. The model was created efficiently with the TwinCAT Machine Learning Creator and enables reliable inline quality inspection, even if there is high variance in the input data.
Food industry
Sorting and classifying
Automating food sorting is a particular challenge due to its natural variance. Eggs, for example, are classified into the categories OK, soiled, and broken.
An AI-based image classification model enables reliable and robust evaluation even if product properties vary widely.
The model was created efficiently with the TwinCAT Machine Learning Creator and can be seamlessly integrated into existing sorting processes.
Production automation
Sorting and classifying
Until now, seven pins had to be manually sorted and stored for the stripping dies of board die-cutting machines. This is a task that fewer and fewer workers are able to undertake – in some cases, none could do it.
The solution: the pins are fed in unsorted, separated, and then classified, measured, and automatically sorted by a camera-based AI system.
The AI model was created quickly and efficiently with the TwinCAT Machine Learning Creator – without specialist AI expertise. The model accuracy is an impressive 100%.