Machines learn from machines
Be a Digital Enterprise. Gain productivity with digitalization. This is, what we aim to at our own factories. Read my introduction to a series of blogposts how we do this. Harnessing the power of artificial intelligence (AI), engineers at our manufacturing plant in Amberg can predict when a key component is likely to fail – up to 36 hours before the failure actually happens. This allows them to react in plenty of time to avoid a costly breakdown of the machine.
In our electronics manufacturing facility in Amberg, we have several PCB cutting machines that are deployed for a number of our SIMATIC products – including the S7-300 and ET 200.
As this machines work, aggressive milling dust builds up and, eventually, the spindles bearing jams. This generally shuts down the respective machine for 1-2 shifts, because the drive system for the spindle needs to be changed, which requires new adjustment parameters from the producer. If the shutdown happens late in the day and those parameters aren’t available for the night shift, the machine stays offline until the next day. Meanwhile, the costs quickly add up.
This is an unpleasant situation that every manufacturer will be familiar with, but there is a way to solve the problem.
Today, we already have suitable products, the process knowhow, and qualified people who can implement artificial intelligence (AI) solutions on a large scale. And the solution we found for Amberg is a combination of AI and edge computing that makes it possible to detect bearing erosion and predict when the machine is going to breakdown between 12 and 36 hours before it actually happens.
How it works
The first step was to analyze the machine data with Artificial Intelligence – specifically, a machine-learning algorithm. The aim is to detect anomalies in the behavior of the spindle that would indicate an impending failure. Of course, the key question here is what data do we need to analyze? Our team isolated two key parameters: the spindle’s rotation speed, and the current needed to power the drive system.
The data is extracted via analog outputs and then transferred to the edge device via MQTT, where it is then analyzed. The machine-learning algorithm calculates in real time, and once the anomaly score exceeds a certain threshold, we know we have to take steps to prevent a breakdown from occurring.
The relevant data is visualized in the SIMATIC Performance Insight app, on our MindSphere platform. This doesn’t mean, however, that someone needs to sit at a computer all day monitoring this. Instead, MindSphere sends a notification of the exceeded threshold directly to the service technician’s device.
MindSphere does not simply pass on the news that an anomaly has been detected in the data. After all, the learning algorithm has to be trained somewhere. So MindSphere collects and archives all the machine data, enabling the algorithm to learn what it needs to know. The more data we collect, the better the algorithm becomes at predicting – that’s why we call it “predictive learning.”
As the algorithm gets better, the new improved version is transferred from the cloud to the edge device – using the edge management system, which is part of the package we offer. The result is greater availability of the cutting machine, significantly reducing the costs caused by unexpected downtime – including potential knock-on effects all the way down the supply chain. At the same time, planned maintenance of our equipment also becomes more efficient. This puts us in the best position to help our customers realize the same benefits in their own plants.
The big advantage is that customers can gather and analyze vast amounts of data right there at the source, creating seamless interplay between automation, cloud and edge. What’s more, it’s an open system that allows customers to add their own expertise and program their own algorithm/apps in our solution. With an open IOT platform like Mindsphere this is only the first step in predictive maintenance and many other cases will follow.