Machine learning processes and artificial intelligence are providing important technological impetus for the use of data in mechanical and plant engineering. Numerous companies see data-based value-added services as crucial for their future business. They are therefore looking into data analysis solutions that can optimize production processes and/or create completely new data-based business models. Weidmüller wants to enable machine builders and operators to create machine learning (ML) models quickly and easily and thus transform the collected data into concrete recommendations for action.

Domain knowledge is in demand

Until now, data scientists have analyzed the data and created ML models. This process is largely manual and exploratory. The process of modeling and creating the ML pipeline is very complex. In total, there are up to 10 to the power of 40 possible combinations to build an ML solution. The actual design of the ML pipeline is also specific to each use case. Weidmüller wants to simplify the application of machine learning to such an extent that domain experts can use their knowledge of the machine or production process to implement ML solutions independently - without data science expertise.

The Industrial AutoML software guides the user through the process of model development; Weidmüller also refers to this as 'guided analytics'. The software helps to translate and archive the domain experts' application knowledge into a reliable machine learning application by querying the existing knowledge and combining it with the ML process working in the background. Compared to an automated modeling process, the model quality improves substantially when domain knowledge is integrated.

The Industrial AutoML solution essentially consists of two modules: The 'Model Builder' and the Runtime cover the complete cycle of development, execution and optimization of ML models over their lifecycle. With the 'Model Builder', the user can create ML solutions for anomaly detection, classification and error prediction. Both supervised and unsupervised machine learning are combined, thus integrating domain knowledge into the model building process. The user's task is to mark the normal or possible misbehavior of the machines in the training data and, if necessary, to generate their own features that are particularly relevant for the use case. This creates a data set enriched with domain knowledge, on which the training, optimization and validation of alternative ML models are performed automatically.
The topic of 'Explainable AI' is of particular importance in this AutoML approach: the Industrial AutoML software enables domain experts to understand the influence of their application knowledge on the model quality and ultimately to understand why an ML model leads to a certain analysis result.

The second module of the Industrial AutoML solution is the execution environment, which is used to run the ML models in the cloud or even directly on the machine in an on-premise application. The execution environment presents the model results in an understandable way so that the user can implement specific actions, for example to avoid errors.

Cooperation between Weidmüller and Microsoft

Dipl.-Ing. Silke Lödige is a trade press officer at Weidmüller Interface.

© Weidmüller

The Industrial AutoML solution has been available worldwide via the Microsoft Azure Marketplace for several weeks now. As part of this collaboration, both companies are combining their respective strengths: Weidmüller is contributing its industrial and machine learning expertise for use cases in mechanical and plant engineering. Microsoft 's Trusted Cloud infrastructure and Azure services enable the Industrial AutoML software to be used by users.

This includes fast access via the Marketplace as well as simple deployment and scaling during operation. The cooperation between the companies goes beyond a technological exchange: Weidmüller is part of Microsoft's partner network and has achieved 'co-sell' status. This means that Microsoft and Weidmüller also operate together on the market and offer a service package. "Together, we are accelerating machine builders and operators to transform their domain knowledge into data-driven innovation and digital value-added services by enabling quick and easy adoption of ML and IIoT technologies," explains Tobias Gaukstern, Vice President Business Unit Industrial Analytics at Weidmüller.
"Weidmüller enables the democratization of ML in the industry by removing a bottleneck: the availability of expensive and only temporarily needed data scientists. At the same time, Weidmüller offers a clear process model and helps to conserve crucial domain knowledge," says Oliver Niedung, IoT specialist at Microsoft. "No company can or should build truly complex IoT solutions alone. Close collaboration in the development and creation of global marketplaces is more important than ever. Here, the engagement with Weidmüller was absolutely exemplary."

The collaboration between Weidmüller and Microsoft ranges from development to holistic operation and ongoing innovation of the solutions. This enables users to focus on their business, their competencies and their differentiation.

Accelerating machine learning applications

Tobias Gaukstern, Vice President Industrial Analytics at Weidmüller.

© Weidmüller

Tobias Gaukstern, Vice President Industrial Analytics at Weidmüller, explains in an interview how Industrial AutoML works and what benefits the software offers users.

The 'Model Builder' is used by domain experts to classify data. How exactly does this work?

The user can mark the machine data with the help of an intuitive Graphical User Interface - GUI for short - and thus mark certain time ranges as normal or abnormal machine behavior. This process can also be used to mark and label any other behavior or state in the data. The user can carry out labeling or tagging using any data series and thus link their application knowledge with the data set. The correlations to all other data series are then calculated automatically as part of the modeling process.

What exactly is supervised and unsupervised machine learning?

In supervised learning, the correct answer to each training task is available in the form of a label. In a classification task, for example, the valid classes and the assignment of states to the respective classes are specified. In the case of unsupervised learning, there are no labels for the training data. The aim is therefore to recognize structures in the data whose categories are not previously known. Clustering methods are used for this purpose. We combine both approaches in our Industrial AutoML product. This allows a model to be trained for anomaly detection without certain anomalies being known. Unsupervised learning is therefore used to learn the normal behavior. During operation, any deviation from the desired behaviour is then recognized as an anomaly. If specific anomalies are also known, the ML models can be validated using unsupervised learning.

How much training data is required for an ML model?

A general answer cannot be given, as the amount of data required depends on many factors. These include, for example, the abstractness of the data, the frequency of the anomaly or error, the number of features and the data tracks that correlate with the anomaly or error. To get started, a few megabytes of training data are usually sufficient to validate whether machine learning is suitable for the use case and to train an initial model. We are therefore talking about a small-data approach here. The special feature of our approach is that training data from the normal behavior of a machine is already sufficient to create a model for anomaly detection, which means that almost any user can embark on the ML journey without any hurdles.

How does the data classified by the domain expert differ from the data selected by a data scientist? And to what extent would differences affect the model?

The engineer can use their application knowledge to select the feature and time ranges that are relevant for the use case from the outset and simultaneously assign the corresponding labels. The data scientist cannot contribute this knowledge, as they simply do not usually have it. This also applies to features that the domain expert also creates. It makes a big difference whether the domain expert contributes their knowledge directly to the data set or whether a data scientist is involved in the project, who laboriously asks for the knowledge and then takes it into account when creating the model. This is also reflected in the model performance: models that are heavily enriched with domain knowledge are usually superior to other models - but it always depends on the individual case.

The ML models are generated automatically by the software. How can the software know which model is best suited to the respective application?

It is precisely the combination of automated machine learning with the user's domain knowledge that leads to suitable models for the respective use cases. The domain knowledge is used to set the objective and the framework conditions to be taken into account. As part of the automatic modelling process, a large number of combinations of features, ML methods and their hyperparameters are created, optimized and validated - always against the background of the specific application. Ultimately, the best model is the one with the best performance in terms of the target values set by the user.

What specific benefits can machine and plant manufacturers generate with the Industrial AutoML software?

First of all, the Industrial AutoML software accelerates the implementation of ML solutions by up to 80 percent. These are enormous time and cost benefits. Machine manufacturers can also achieve significant sales and margin benefits with new data-based services. For machine operators, manufacturing processes can be optimized, for example to improve product properties, ensure product quality or increase productivity.