Data is continuously generated in machines and production plants. Companies who succeed in translating this data into innovations gain decisive competitive advantages. With user-friendly software, Weidmüller is now making artificial intelligence methods available to machine builders and manufacturing companies.
For the analysis of machine and process data with Industrial Analytics, complex models are used that are capable of detecting anomalies or even predicting future machine behaviour. Artificial intelligence (AI) methods and machine learning (ML) are used to uncover previously unknown relationships between measured values using features derived from raw data.
Combined know-how required
The necessary information is available in almost every company. When developing meaningful analysis models, medium-sized companies in particular are often still dependent on the external support of data scientists. Weidmüller has developed a groundbreaking solution that enables them to act without the need for data scientist resources. In close cooperation with the end user, the data experts identify relevant correlations in the measured values and train the initial model. After successful application, the initial model is repeatedly fed with new data and further developed over the entire life cycle of the machine. This increases the quality of the information over time.
Learning Machine Learning
Many machine builders and manufacturing companies have not yet been able to use the available machine learning tools independently, as their operation has been optimised for the data-driven activities of analytics experts. Companies can either train their existing employees for a huge amount of money, or hire a data scientist themselves. This results in an inhibition threshold that is currently slowing down the spread of artificial intelligence in industry.
An alternative is to develop user-friendly software solutions that even users without any statistical training are able to understand and generate analytics models. Weidmüller‘s Industrial Analytics business unit has put this idea into practice with its automated machine learning software. The name of the application itself implies that the models are largely developed automatically.
“Similar applications are currently being used in the areas of fintech, banking and marketing. However, the existing solutions are not suitable for machine and plant engineering, because they do not support the relevant data types from the automation industry. They always require an ideal database,” explains Dr Carlos Paiz Gatica, Product Manager at the BU Industrial Analytics. “In addition, they don‘t provide the ability to integrate the user‘s domain knowledge, which is essential for industrial applications.”
For the automated machine learning software, Weidmüller’s analytics experts combine the domain expert’s data and information with algorithms to automatically generate suitable models. The following working steps describe the model generation process using anomaly detection as example:
1. Selection of training data
The domain expert decides which data sets should be used to learn the normal behavior of a machine or plant. For this purpose, an overview of the raw data is first generated, which supports the user in evaluating the information content of the data. The preparation of the measured values takes place completely automatically.
2. Feature engineering
If the raw data is not sufficient, additional information can be generated on its basis. The user can use his domain knowledge to create new features. These can, for example, describe the course of temperature change instead of only showing individual states. Using such features, the machine condition can often be better evaluated than with the raw data.
3. Labeling the machine behaviour
With a label the user marks areas in the data in which a normal (green) or undesired (red) behavior is present. This enables the user to increase the information content of the training data with his domain knowledge. Assistance systems support the labeling process by directly highlighting similar situations in the data set.
4. Model training
The labelled data sets are converted into models and trained with various ML methods. This fully-automated process results in a list of alternative models, which are provided with information on the quality of the result, execution time and training duration. The so-called Anomaly Score Plot directly represents results of the models so that the expert can directly compare the performance of the models. If the desired model performance has not yet been achieved, the user can edit the features and labels of the model again. The model can then be transferred directly into the architecture of the target system.
Extending AI applications
“With the automated machine learning software, machine builders and manufacturing companies have the opportunity to independently exploit the benefits of artificial intelligence and machine learning, without having to become data experts themselves,” says Paiz. “The universal application supports the users in both initial model generation and further development. In this way, companies are no longer dependent on the resource of the data scientists and do not have to share their process and machine knowledge with external partners.”