AI on the line – three cases of digital pyrotechnics on the production line
“It’s like a secret recipe,” says Bernd Gross, chief technology officer at German firm Software AG, when asked about why the car industry is not crowing, yet, about its paint-shop analytics solution, developed with German robot maker Dürr, and rolled out by at least two prominent – unnamed, but not hard to guess – German car makers.
No one is telling, hardly. At the same time, demand for edge-based compute power is spiralling upwards. Revenue from the sale of artificial intelligence (AI) chipsets for edge inference and inference training will grow at 65 per cent and 137 per cent respectively between 2018 and 2023, according to ABI Research.
Shipment revenues from edge AI processing reached $1.3 billion in 2017, and will climb to $23 billion by 2023, it says. The demand for production-line AI may be real but progress is slow and challenges remain. “It’s not mainstream; some manufacturers even at the high end are only just experimenting,” says Viral Chawda, principal for innovation and enterprise solutions at KPMG.
These challenges are discussed — and this discussion is expanded – in a new report from Enterprise IoT Insights, called ‘AI on the line – how advanced analytics and artificial intelligence are transforming the production line’. The report is available to download for free, here.
But rarefied instances of production-line analytics and supply-chain integration are out there, and three of these are detailed in the narrative below, lifted directly from the report. These discuss, in turn, the new dynamism and performance gains AI has brought to the production of cars, the processing of food, and quality assurance in electronics manufacturing.
We should start with Software AG, just because it appears to have seized on the opportunity of new production-line analytics more successfully than most other (often bigger) brands. Indeed, the Darmstadt outfit ranked highest (with familiars like PTC and Hitachi, as ‘visionaries’) in Gartner’s latest review of the top industrial IoT platforms.
Its work with the ADAMOS (ADAptive Manufacturing Open Solutions) collective – founded with Dürr, as well as fellow German machine makers DMG MORI, and Zeiss, specifically to develop high-end process analytics for industrial machines – has borne fruit, notably with paint-shop robots for German auto plants.
Dürr, which makes the robots, has reworked its DXQ equipment analytics programme with Software AG to record, analyse, and eliminate faults in the painting process. The software has so far been rolled-out to 10 factories, belonging to a certain premium car marque. “It goes plant by plant. You need to implement the software and train the people. You can’t just push a button, and it’s everywhere.”
But it will be rolled out generally, reckons Gross, even as its first customer keeps stuum about its impact. “It’s a tricky question. No one is really talking. I ask the same thing,” he says of his first client’s reticence. “But it’s so compelling, It will be done by most of them.” A second German automotive icon has already picked it up, he adds.
The solution itself is an elegant example of production-line pyrotechnics, which replaces manual on-the-hoof inspections with automated real-time cut-offs as soon as the spray job goes awry. “A lot of data is coming off these robots – two million kilobytes per day. We are evaluating 230 different signals from each, at any one point in time – out of 100,000 data points we are collecting.”
It is standard procedure, he explains; 230 data points are sufficient to correlate live errors in the process. The rest are weeded out in the ‘co-creation’ phase, with the ADAMOS trinity of software provider, hardware maker, and factory operator banging their heads together to finesse the analytics. “You don’t need everything,” says Gross.
Where typically car manufacturers run manual checks on every tenth paint job, and check over or respray the previous nine to be sure, the new system from Dürr and Software AG stops the line at once, showing the painting error and flagging-up the fault in the machine.
“We stop the process right away. The other nine cars are not affected. And the operator knows exactly where the problem is, as soon as it arises – because the nozzle’s gummed-up, say, and there’s not enough liquid in the paint mix.”
The trickery comes in the algorithmic shuffle of three treatments of machine data, in streaming analytics, time-series analytics, and batch analytics. These are pooled in Software AG’s Cumulocity IoT platform – in this case, an edge-based version of the system that has turned heads at Gartner.
The first of these analytics tools, based on Software AG’s Apama product, is key: it embeds the rules (iterated through the co-creation process), analyses the data, and issues the ‘alarms’ as a live-feed off the production line. “That’s the real-time analytics part,” says Gross.
The time-series engine, based on its TrendMiner product, trawls historical data, as it arrives and as it is stored, to reveal the errors in the painting – “what happened when the job went out of sync; the patterns and the fingerprints”. These are rendered in a visual dashboard, for factory-line operators to take appropriate action, and re-spin the cycle.
The final piece, the batch analytics, organises the traditional ‘big data’ rules, as provided by any number of business intelligence (BI) programmes. Software AG’s innovation has been to create a ‘data hub’ interface that exposes the data, in this case from the auto-plant paint shop, in third-party BI engines such as Tableau and Microsoft Power BI.
“There are so many batch analytics engines around. We didn’t want to reinvent the wheel,” says Gross.
There are other powerful examples of how advanced analytics, verging on artificial intelligence (AI), is transforming the production line, whether in car-plants or food factories. Quality assurance (QA), such as with paint shops, is a primary recourse for high-end analytics, and considers the product on the line directly.
A neat instance – detailed in these pages before, but worth a second look – comes from Microsoft, via Swiss technology firm Bühler.
Bühler has developed an optical sorter to remove carcinogenic corn kernels, infected by a mould called aflatoxin, from the production line in food processing plants. In the Bühler solution, corn gets fed from a truck to a hopper, and into a chute, and falls at 3.5 metres per second as a waterfall-like corn-feed in front of an ‘AI-enabled’ camera setup.
The cameras project UV light to illuminate the grains; a telltale fluorescence shows up the aflatoxin, and high-speed air jets shoot contaminated kernels into the bin – as they fall! (The exclamation point is fully warranted.) The rest passes into shipping containers.
Microsoft is providing the analytics tools. Diego Tamburini, the company’s principal industry lead for manufacturing within its Microsoft Azure division, says: “We’re talking milliseconds; just imagine: the corn kernel is falling at speed, and the machine finds time to take a picture, process it, make a decision, and take an action.”
The Bühler solution, called LumoVision, processes 10 to 15 tonnes of maize – or an entire truckload – in an hour. It was tested with Italian agricultural cooperative Capa Cologna, and shown at Hannover Messe last year. The phone has been ringing off the hook ever since.
Another example of rarified production-line analytics, achieving major QA gains compared with old manual techniques, comes from HPE and Foxconn, which have developed a machine version of Where’s Wally? to discern the manufacturing equivalent of striped hats in discrete electronics, and raise the alarm.
It is a symbiotic arrangement, using HPE edge components to automate the QA process on a Foxconn production line carrying HPE servers at a factory in the Czech Republic. The system, developed with video analytics firm Relimetrics, is being positioned as a support line for the standing workforce, to relieve it of duties it was never really qualified for in the first place.
“Things are getting smaller, things are getting more varied, and yet we need more precision,” comments John Gallagher, operations manager of Foxconn, speaking at the opening of HPE’s new innovation lab in Geneva earlier this year. Batch sizes are small and production is swift. The production line for HPE servers jumps to a new configuration every three or four units.
Human eyes are fallible, and increasingly unsuited to checking for assembly errors in shrinking electronics. “People miss a lot of things,” he says. The new edge system catches the faults before they cause delays, and makes good on Foxconn’s quality promise, a crucial consideration for any supply business.
These two examples take big data analytics to a new level, bringing higher-order intelligence to the automation of the production line. They both employ this technique of ‘computer vision’, whereby data from rapid-fire imaging is filtered through hard-worked algorithms at such a pace that the latency of the QA process is set to almost zero (‘real time’).
Or to effectively zero, at least. The wheat is sorted from the chaff (the bad seeds and assemblages, in these cases), without disrupting the rest of the production line. These twin examples advance the Dürr/Software AG setup insofar as they introduce video. But the impact is equivalent; the latter pairing has achieved the same by interrogating existing data, without introducing data-hungry video feeds, it might be noted.
Importantly, this capability to keep the production line humming is new. The disruption of laboured manual checks is reduced significantly by these AI solutions, and productivity and profitability are raised as a result. Which explains the deathless pursuit (and paranoid secrecy) of industrial innovation, which is jumping with new digital tech.
This article is taken from a new report from Enterprise IoT Insights, called ‘AI on the line — how advanced analytics and artificial intelligence are transforming the production line’. The report is available to download for free, here.