Tool allows machine learning at a machine, avoiding the cloud
29 November 2024 
The German automation supplier Weidmüller has announced a machine learning (ML) tool that allows you to implement ML at a machine, without needing to connect to the cloud or the Internet. The edgeML tool can run learning algorithms at the edge on a PLC or industrial PC.
The software is also available as a Docker container, allowing it to be used on any industrial controller that supports the execution of these containers. This manufacturer-independent capability can execute various machine learning models. A no-code design makes it possible to deploy ML models on controllers without needing any knowledge of Python or data science.
According to Weidmüller, machine learning directly at a system offers numerous advantages, including the local collection, storage and processing of data. Unlike cloud-based ML systems, the machine or controller does not need to transfer the data to the cloud, ensuring that sensitive data does not leave a company.
Any discrepancies in the production process are detected at the machine. This speeds up troubleshooting, prevents prolonged downtimes, and reduces the production of rejects, Weidmüller says. At the same time, edgeML cuts costs because it eliminates the need for cloud licences, as well as fees for data transmission and storage.
Production lines with machines and systems that cannot be connected to the Internet for security reasons can also benefit from local ML capabilities.
The path to a machine learning model at the edge begins by collecting data which is imported from the system into ModelBuilder software, where users can create ML models based on the data. These models are then transferred to edgeML.
Because edgeML supports the standard ONNX format, users can create models other than ModelBuilder – using Python, for example. This allows them to implement ML in a familiar environment. edgeML also allows existing ONNX models to be re-used.
If a model is no longer performing as desired, it can be replaced easily without having to adjust any communications settings. In this way, edgeML make it easier to manage the lifecycles of ML models (MLOps).
To minimise the use of time and resources when creating ML solutions, Weidmüller plans to allow the calibration of created models in the future. This function is already available in ModelRuntime. A standard model for a machine family thus becomes a template that can be extended to other machines of the same class. The applied model continues to learn from these machines to adapt to the system, enabling the ML models to be scaled and re-used as needed.
For its next stage of development, Weidmüller plans to further improve the accessibility of edgeML. A connector will overcome the boundaries of fieldbuses and protocols, making it usable on any system.
