AI is not an end in itself – in other words, just because you use AI, you do not necessarily achieve added value. You should know exactly what AI can and cannot do, where potential for AI-based improvement can be found in your own production and how this can be accomplished. The following overview is supposed to help with this.
Many use cases in AI follow one of two approaches that are closely related but often used independently in practice: Retrospective analysis and live prediction using a mathematical model.
Analysis is about finding hidden correlations from historical operating data in order to answer a question and present the answer in a transparent and explainable way, for example in the form of a report. The added value lies in the fact that process engineers, technologists or even machine operators can gain new insights from the answers and derive concrete improvements which, once implemented, generate lasting added value.
For example, such an analysis could indicate that the quality always drops when the temperature exceeds a certain value. So, for example, a warning can be implemented via condition monitoring if this temperature is exceeded.
In contrast, models, be they prediction models or monitoring models, are first and foremost software modules. As such, they can usually only achieve their added value if they are integrated into the continuous flow of current operating data.
They must be continuously fed with current data in order to continuously generate a prediction via the mathematical model on which they are based. For example, failure probabilities, quality estimates and much more can be calculated. The result is made available to the users during operation.
Based on these two approaches, a wide range of applications can be realized – in the following, we present five ways in which AI can be impactful used in production:
This analysis is useful, for example, to better understand certain KPIs such as quality. Assume that quality can only be measured in the laboratory. However, one would now like to know whether there are other process parameters besides those already known that have a significant influence on quality.
The dependency analysis provides an answer to this by analyzing the historical process data and finding out which of the process parameters influence the quality with which time lead and to what extent. In this way, new and valuable information can be obtained from which, for example, refined process specifications can be derived in order to sustainably increase quality.
A key use of AI in production is to identify the root causes of process malfunctions. To this end, recurring malfunctions are analyzed to identify indicators of their root cause. These help process engineers to find and eliminate the underlying causes of the malfunction.
A malfunction can be any form of undesired behavior, be it a production stop or the process moving outside a required target corridor. The task of root cause analysis is to identify, sort and disclose the scenarios in the historical process data that directly preceded the undesired behavior.
Furthermore, it should show which signals and relationships have behaved in a characteristically different way per scenario than they usually do. With this information, process engineers can conduct targeted root cause research and, in the best case, prevent the process malfunctions in the future through targeted one-time measures.
A virtual sensor, also called a software sensor or soft sensor for short, is a software module at the core of which a mathematical model calculates an output variable from special input variables.
Virtual sensors are generally used where physical sensors are too prone to failure or too expensive, or for determining critical process parameters that are not at all accessible for sensor measurements. For example, a virtual sensor can determine the quality directly and without delay from the running process data, which conventionally can only be determined after sampling in the laboratory (predictive quality). This allows operators to react immediately to deviations instead of having to wait until the next laboratory measurement is available before they can intervene.
Classic condition monitoring is used to monitor whether all process parameters are within specifications. AI-supported component monitoring goes a step further here. It monitors the normal behavior of the component, which it has learned from a recording of the component’s signal data during smooth operation (predictive monitoring).
Often, new sensors have to be installed to measure vibration or temperature, for example. For some AI systems, however, the sensors and data already installed in the automation technology are sufficient for monitoring.
The monitoring model permanently compares key signals and signal relationships of the component for abnormal behavior. The anomaly score ranges from 0 (everything OK) to 100 (severe abnormality). It thus provides quickly and easily indications of the health of a component and allows timely scheduling of maintenance, provided that the abnormal behavior becomes visible with sufficient lead time (predictive maintenance).
Component monitoring is used primarily where components, each of which rarely fails, are present in such large quantities that failures are nevertheless reported frequently – typical examples are valves, pumps, and motors.
Malfunction monitoring picks up where the analysis of process malfunctions (see above) leaves off. This is because, in addition to malfunctions that can be permanently eliminated by targeted one-off measures, there are also those that can only be permanently prevented or at least minimized by regular, timely intervention.
Here, the first step is to examine the critical scenarios from the malfunction analysis and to develop appropriate countermeasures for each scenario in order to bring the process back to a non-critical state should it be in this scenario. Next, a monitoring model is needed that continuously monitors the relevant process parameters and notifies the operator as soon as the process is in one of these critical scenarios. By taking the countermeasures immediately, effective action can thus be taken against the process malfunction.
These five examples are common applications that are already being used successfully in a variety of industries to increase production safety, productivity and product quality.
Our trailblazing AI technology aivis is capable to cover all of them and more. And the good thing is: you don´t need to be a data scientist to exploit the power of aivis in your day-to-day production processes.
You are a process engineer in charge of a data-intensive plant? You would love to exploit AI to further improve processes, better understand problems and further increase productivity? Then all you need is your production data to get started. The easiest way is a simple CSV export form your production system. This data either already exists, or you can just record it for a certain time.
As soon as the CSV file is exported, it just needs to be converted in aivis simple CSV format and it is ready to be analysed by aivis.