Machine learning as a sub-area of artificial intelligence is increasingly making inroads into automation and control technology and bringing advantages in the development of complex systems: instead of classic engineering, deep learning techniques such as machine learning or neural networks use extensive sets of process data. Models are created and trained that can recognize patterns and laws from these data. The trained model is then applied to new process data and detects anomalies that indicate incorrectly produced parts or changes in the machine.
Knowledge-based systems, such as machine learning or neural networks offer the greatest advantage when they are executed with real-time capability in machines. This also benefits demanding applications, such as wear and energy-optimized motion applications. An inference machine executes the model within the controller and links these technologies to the classic production environment.
The workflow starts with the data, to which initially several machine learning methods are applied to identify and train the optimal model. A variety of software libraries and frameworks (including Scikit-learn, PyTorch, TensorFlow, MATLAB®, MXNet) are available for this part of the engineering. The trained model is available in an exchange format such as the Open Neural Network Exchange (ONNX) for processing with additional tools and for execution in the inference machine in the controller.
Machine learning in real-time capable controllers opens up new fields of application:
- Recognition of the individual usage behavior of a machine and adaptation of maintenance models
- Optimization of the energy consumption and wear behavior of machines
- Evaluation of objects, patterns and structures in camera images for quality assurance
- Detection of anomalies in machine operation and identification of new operating states
- Control of difficult-to-control processes with adaptive control algorithms
- Improvement of overall equipment effectiveness (OEE) with autonomous edge analytics
Beckhoff combines these innovative technologies with PC-based control technology:
- The open ONNX exchange format allows a free choice of the training tool.
- The inference machine is deeply integrated in the TwinCAT runtime.
- Hard real-time requirements are met at all times.
- Manufacturers of standard machines can also deliver their machine learning applications to their customers in a closed binary data format (BML).
- Reconfigurations, such as when changing products on the machine, can be done with a direct change of the machine learning model without compiling or restarting the TwinCAT runtime.
- All process data can be read and written in TwinCAT from the I/O system or via database access.
Further information about our technologies and products for system-integrated machine learning in real-time systems can be found here: