Machines learning with Optiklink worldwide
When Roger Feist has visitors, one of his favourite things to do is sit them in front of his rolling mill simulator. Modelled on the original, this control desk consists of three screens, buttons and levers, a temperature display and various flashing warning lights. It was originally a trade fair exhibit made by the machinery and equipment manufacturer Achenbach Buschhütten.16 Jan 2019
“It’s time to prove yourself! The aluminium strip mustn’t tear,” jokes Feist, who has been responsible for digitizing the machinery at Achenbach for many years, and is now starting the first level. Operating the machine is complicated. The strip keeps on tearing, the lights flash and the displays give warning after warning. “You need to have a good feel for the machine,” explains Feist. The production and performance data are stored in Google Cloud and after three attempts, Feist releases his guest. “Torn three times. We can save ourselves the trouble of evaluating the data,” he laughs, immediately launching into his topic: data and machine learning.
“Our customers, but also we ourselves, wanted to understand the production steps more precisely and correlate, for example, data from the foil rolling mill with the slitting machine and feedback from the customer.” The aim is to make production more transparent. All the data from the Bachmann M1 controller is sent via OPC UA to a small single-board computer, which can then subscribe to the information and store it in a cloud memory.
The machine controller can neither be viewed from the Internet nor addressed. The machine operator therefore has sole control over which data is transferred to the cloud and which isn’t,” explains Feist. Around three gigabytes can accumulate in one day for each machine – mostly OPC UA and SQL data, which will later be of key importance for machine learning. And because there is virtually unlimited storage space in the cloud, data never has to be deleted to save space. “Neither our customers nor we can say today what questions we will need the data to answer in future. Only if specific problems occur with a particular material or if a customer is struggling with failures of a particular subsystem do we know which data is relevant to solve the problem. Problem solving would often be far more difficult if this data was not stored in advance or was deleted too soon for storage space reasons.” Machine learning would be wasted.
Achenbach has developed a business model, a product, from the data: Achenbach OPTILINK – in the first instance a cockpit or analysis tool for customers worldwide. Customers can query the current status of their machines via a web interface. Achenbach provides the customer with a basic set of analysis tools, but customers can also create and perform analyses themselves.
However, Feist and his colleagues are still not satisfied. Artificial intelligence is the name of the game, and not just with the focus on ‘deep learning’ as strongly promoted by Google. Achenbach employs ‘unsupervised machine learning’ in many solutions. The idea behind this is that the rolling mill tries to identify patterns in the data that deviate from the unstructured noise. Ideally, this will enable a recommendation for action – such as ordering a spare part from Achenbach – to be given to the operator. Linking the OPTILINK system to an electronic spare parts catalogue was one of the first enhancements to the software package. In line with this OPTILINK can trigger defined work processes in the company’s ERP system via ‘tickets’, thus connecting the work of humans and computers.
To be able to maintain a high development speed for ML algorithms, the developers at Achenbach have integrated three tool packages into their portal. One component is based on ‘Matlab’ and observes production processes, while another was created with ‘Rapidminer’ and analyses incidents (such as strip tears) on the machines. For certain forecasting models, however, neural networks based on ‘TensorFlow’ were used. The next goal is to automate the decisions that are, in most cases, currently still made by humans on the basis of this information. Many of the ML algorithms have been known for years, but the correct data filtering and selection in relation to production data is still largely untested. “We’re working on it” is all that Feist is prepared to say on the matter. He doesn’t want any AI hype.
A large amount of data is usually required for the application of unsupervised machine learning algorithms. Feist and his customers store this data in Google Cloud – “without any limits”, Feist reiterates. Without a sufficient volume of data, the algorithms are not able to cluster the data. The machine becomes smarter with each data set. But this also carries a risk: overfitting. Known data is processed well, but the machine finds it more difficult to process new data. This is something that all AI pioneers are struggling with. The other extreme is underfitting – the data is missing. “We’re training the machine and need to find the right data balance,” explains Feist.
Interested in news about exhibitors, top offers and trends in the industry?
Your web browser is outdated. Update your browser for more security, speed and optimal presentation of this page.Update Browser