Even if the practical application of machine learning in an industrial environment is still in its infancy, one benefit is already clear today: areas such as robotics, anomaly detection or model-predictive control would benefit directly. With TwinCAT Machine Learning, Beckhoff now offers automation engineers and machine builders the possibility of integrating inference - i.e. the execution of a trained ML model - in an industrially compatible way and in real-time.
Machine learning - particularly with neural networks - is structurally conditioned by deterministic runtime complexity and can approximately map any continuous function. This already applies to flat networks, which, in addition to deterministic runtime complexity, also have a very short runtime if implemented appropriately. This makes neural networks interesting for automation in motion control - Neural Automation.
The ML-XTS demonstrator with the TwinCAT 3 Neural Network Inference Engine, integrated in TwinCAT 3 Motion Control, shows exactly this. The core of the demonstrator consists of two XTS systems and a conveyor belt placed in between. The ten XTS Movers are each synchronised one after the other to markings on the assembly line and are driven parallel to the respective marking for a distance of one metre. They then uncouple and move back to the starting point of the assembly line for the next synchronous movement.
An XTS is controlled using a classic method, in which high wear and energy consumption values are evident due to the high dynamics, especially during decoupling and shortly before resynchronisation. In contrast, the second XTS, which is controlled by a neural network specially trained to minimize energy and wear, operates with significantly lower overall energy consumption and less wear. This is because the neural network generates "softer" profiles, i.e. it distributes the dynamics required for synchronization over a larger range of the XTS.
TwinCAT Machine Learning