Analysts predict a more than rosy future for the smart factory approach, which largely monitors and controls itself. In its study 'Smart Factories @ Scale', the Capgemini Research Institute assumes that intelligent factories will bring the global economy an annual productivity gain of between 2.8% and 4.1% by 2023. Expressed in hard figures, the experts see an increase in value of between 1.21 and 1.86 billion euros.

Data forms the foundation of a smart factory, on which applications such as PLC, SCADA, MES and ERP solutions are based. This data is generated in Industrial Internet of Things (IIoT) environments via sensors that record a wide variety of values - parameters such as temperature, pressure, (flow) speed, acceleration and values relating to speed, switching status, position or even sound. The corresponding real sensors are attached to the points of the systems and machines to be monitored. However, this is not always unproblematic: the use of physical sensors may not be possible due to spatial constraints or may simply be too costly. Another crux is the susceptibility of such sensors to errors.

Another challenge is values that can only be measured in the laboratory.
in the laboratory. This applies, for example, to data on material properties that affect the quality of the product. Due to the time required to determine such data, its usefulness for current production is limited.

Virtual sensors as a solution?

The dashboard of the Pythia Virtual Sensors app - users can create their own virtual sensors via this interface.

© PerfectPattern

Virtual sensors are often seen as the solution to these challenges. Put simply, virtual sensors combine various variables measured by real sensors and then deliver the desired result. Based on the data generated by physical sensors and other available parameters that influence the process, they are able to predict dependent variables - such as real sensor values, laboratory measured values or variables that cannot be measured directly, such as quality or service life parameters. This data is determined using prediction models, which the virtual sensors use to forecast the effects of the measured variables on the desired target variable. This means, for example, that it is possible to work with continuously available, virtual live measurement values instead of sporadic and delayed laboratory measurement values.

More than just a downer, however, was the previously data-, time- and cost-intensive creation of virtual sensors. In-depth data science expertise was also required to set up the prediction models. For this reason, virtual sensors have been slow to catch on, despite all their advantages.

New possibilities thanks to AI

Comparison of 'Pythia Virtual Sensors' with conventional methods for the creation of virtual sensors.

© PerfectPattern

A new technology called 'Pythia Virtual Sensors' from PerfectPattern is now set to help virtual sensors achieve a breakthrough. Based on the 'Pythia' AI technology platform also developed by this company, the system is able to clean up data independently and unsupervised. This involves synchronizing and removing anomalies and mathematically adding missing values.

On this basis, the system also independently defines the necessary prediction models. Even very hidden patterns and dependencies are found. The system learns how to predict and control each desired variable. It also finds relationships between data that are often barely recognizable. The result is a virtual sensor that is available both as an embeddable library and as an executable program or web service. The execution time for a predicted value is in the range of milliseconds. Thanks to extremely low hardware requirements, it can be executed in the cloud, on standard platforms and on embedded systems (on edge). While the creation of a virtual sensor for complex tasks used to take days or weeks, this is now usually a matter of an hour. Even adapting to changes in the system environment only requires a new learning process for the models.

The 'Pythia' AI platform on which 'Pythia Virtual Sensors' is based enables automated, comprehensible and transparent data analysis and prediction. To do this, the technology uses innovative mathematical concepts, basically a system of stochastic differential equations.

Democratization of data science

The aim in developing 'Pythia' was to create a solution that enables people from industrial practice to quickly gain insights from their data and arrive at data-driven, transparent conclusions in order to increase the efficiency and safety of their production processes. Compared to 'conventional AI systems', Pythia requires significantly smaller amounts of data and works with unprocessed, unfiltered raw data.

The indicators determined by 'Pythia' for the cause of disruptions in production processes are important for understanding them. For example, production managers can find out which combination of shift, equipment, machine, material and more is not working as expected and why. They also recognize which critical relationships and dependencies play a role.

The system does not require any data science expertise and can therefore also be used by process engineers and domain experts.

Practical example: Cost-efficient production

A company manufactures its products using calendering. In this process, a melt of raw materials, which are often subject to natural fluctuations in quality, is passed in sheets through a system of heated rollers in order to increase the density of the sheet and improve the surface quality of the end product. Production must remain within a narrow corridor: If the quality required by the customer is not achieved, the products are not accepted - if it is exceeded, an unnecessary amount of material is consumed. The process therefore requires continuous adjustment based on hourly laboratory quality measurements. As these are time-consuming, the data required for production control is only available with a considerable delay. As a result, production has to work with a high safety margin - a cost driver.

For this reason, a virtual sensor was implemented. This enables live prediction of the quality of the end product. This continuous 'virtual' measurement enables a lower safety margin and thus increases the cost efficiency of production. When creating the virtual sensor, the company only defined the goal of 'predicting quality'. The basis consisted of 8000 signals from a time series database. 'Pythia Virtual Sensors' learned independently that only around 150 of these signals are relevant for predicting quality.