Machine Learning easy to use without expert knowledge in Data Science

Weidmüller Automated Machine Learning Tool: Assisted creation of models using artificial intelligence (AI).

Companies, which would also like to be profitable in the future, must set the course for the future today. They must deal with the topic of digitalization, because data-based services determine the business success (of) tomorrow. This makes the use of artificial intelligence (AI) in industrial production plants one of the central challenges for machine and plant operators today. Artificial intelligence solutions are used in many areas from detecting anomalies, classifying and predicting wear or damage to quality control. How to get started with AI? With the Automated Machine Learning (AML)Tool, Weidmüller provides the user with the appropriate software. The AML Tool enables domain experts to independently create Machine Learning (ML) models based on their application knowledge. This way, they can contribute their highly-specialized knowledge of their machine and plant processes into the software tool. At the end of the modelling process, the expert receives the most suitable model for his application.

Complex modelling process

Today, data scientists analyse data and create ML models. This process is largely manual and explorative. This not only creates the actual model, but also a so-called ML-Pipeline, in which the data passes through many processing steps, at the end of which the model is presented and the result is outputted. The process of creating the model and the ML-Pipeline is very complex. In total, there are up to 1040 possible combinations to build an ML-Solution. The concrete design of the ML-Pipeline is in each case unique. Of course, there are some software tools for the data scientists that support the basic structure of the Pipeline and thus, simplify the work of data scientists. However, most of the parameters for the ML-Solution must be determined manually in a creative way, which is at the same time arduous work. During the modelling and construction of the Pipeline, the data scientist continuously discusses the relationships found in the data with the machine and process experts. The results are jointly interpreted, whereby the parameters for the model and the pipeline are ultimately identified and then established. Therefore, the application knowledge of the domain experts plays a decisive role in the success of a good ML-Solution.

Democratising the use of machine learning

Weidmüller’s vision is to democratise the application of machine learning, i.e., to make it accessible to every domain expert in the industry and ensure the application of ML in the industry is not limited by the number of data scientists. This make for the best use of the existing knowledge of domain experts. Consequently, the use of ML for industrial applications needs to be standardised and simplified to such an extent that domain experts can independently generate ML solutions without expert knowledge in the field of data science. This also includes the greatest possible automation of modelling as well as the generation of the ML-Pipeline to accelerate the creation of the ML-Solution. The technological approach behind this is described with the term, ‘’automated machine learning’’, although it does claim to completely automate the process of generating ML solutions. On the contrary, the domain experts should actively link their knowledge to the AML-Process to create excellent ML-Solutions.

Guided Analytics

With AML-Software, domain experts can create ML-Models. The AML-Software guides the user through the process of model development, which is why guided analytics is mentioned. The expert(s) focuses on their knowledge of machine and process behaviour and links this to the ML-Processes running in the background. This means that the software helps translate and archive of the existing and valuable application knowledge into a reliable machine-learning application by cleverly interrogating the existing knowledge and combine it with the ML-Processes running in the background.

Domain experts independently develop machine-learning solutions

The AML-Solution essentially consists of two modules for the model creation, execution and optimization, as well as for the management of the models throughout their life cycle. With the module for modelling, the domain expert(s) can create ML-Solutions based on the training data and his application knowledge for anomaly detection, classification, and failure prediction. The anomaly detection is unique in the world based only on ‘’good data’’, so-called ‘’unsupervised’’ training. An algorithm learns the typical data patterns of a normal machine behaviour based on historical data. While running, deviations from these patterns can be identified. The detected anomalies can be inefficiencies, minor faults or major errors. Thanks to this approach, the system can detect even previously completely unknown error cases as soon as they occur. The result of the modelling process is a completely configured ML-Pipeline, including the model. In addition, the model-builder module is used to optimize the ML-Models during operation. New events, such as certain operating situations, anomalies or errors that occur while a machine operates and were not included in the training data, can be added to the models with just a few clicks. This allows the models to be continuously improved over their life cycle. The second module of the AML-Solution is the execution environment, which is used to run the ML-Models in the cloud or in an on-site application. It is not dependent on a certain platform and scales automatically according to the number of models to be executed. In addition, the execution environment presents the model results in an understandable way, allowing the user to take concrete actions, e.g., to avoid errors. Since the models are enriched over their life cycle and thus new model variants are created, model management is another component of the execution environment. Among other things, model management provides functions for model versioning, model recovery, and model monitoring.

The knowledge of the application exerts is decisive

In automatic modelling, suitable ML-Procedures are first automatically selected based on the structure of the training data used to analyse the task and of the application knowledge. In doing so, up to 300 features are generated for each data track from the raw data, thus covering a relatively large solution space. Then, alternative ML models with different feature combinations are trained and their hyperparameters optimized. Finally, the models are validated and integrated into the ML-Pipeline generated in parallel. All these steps run completely automatically. Depending on the complexity, the calculation of the models can take minutes or hours. The first models are already available after a few minutes, allowing the user to get feedback on the quality of the models in a timely fashion and can then decide whether the model building process should be continued or aborted. Crucial for the success of the model-building process is the application knowledge of the domain experts with which the training data set is improved. Based on their machine and process knowledge, they can label the data, e.g., mark desired and undesired machine behaviours in the data. According to this same principle, certain process or production steps can be labelled. A typical example is the start-up behaviour of a machine. The user can also create his own features, which are not included in the raw data but still help to evaluate the manufacturing process. The data set enriched with the application knowledge is the input variable for subsequent automatic generation of the ML-Models. This results in ML-Solutions that are comparable to the solutions manually created by data scientists. At the end of the modelling process, the user selects the model best suited for his application according to certain criteria such as model quality or execution time. The favoured model can be exported and saved or integrated into the execution environment.

The focus is on the application know-how of the users

Weidmüller’s AML-Solution enables practically every application expert in the industry the ability to create and use ML-Models for a wide range of uses. It is the first solution that allows ML-Models to be created without any expert knowledge of data science and only on the basis of domain knowledge. The user along with their application knowledge is at the centre of attention, which ultimately and additionally contributes to increasing the acceptance of ML in industry. The software is an end-to-end solution for the creation, operation, and optimization of ML-Models, which so far has not been available for industrial application in its’ current form. It reduces the complexity of machine learning and accelerates the implementation of ML-Solutions considerably. This makes for a strong contribution to the industry with focus on ML applications, which is pivotal for the success in the European economy. The feedback from the first pilot users among machine manufacturers and operators demonstrate how user-friendly the Automated Machine Learning Tool is and how it is optimally tailored to the needs of users in its functionality and user guidance.

Downloads