Doomed to fail without AI?

EMO Han­nover 2019: Apply­ing an “engi­neer­ing approach” to the sys­tem­at­ic use of arti­fi­cial intelligence 

The head­line is inten­tion­al­ly provoca­tive. Accord­ing to a 2018 study by Boston Con­sult­ing, Chi­na, Japan and the USA have been upgrad­ing their arti­fi­cial intel­li­gence (AI) sys­tems for some time now, mean­ing that Europe urgent­ly needs to take action. Prof. Jörg Krüger, Head of the Indus­tri­al Automa­tion Tech­nol­o­gy Depart­ment, at the Insti­tut für Werkzeug­maschi­nen und Fab­rik­be­trieb (IWF) at the Tech­ni­cal Uni­ver­si­ty of Berlin, and Head of the Automa­tion Tech­nol­o­gy Divi­sion at Fraun­hofer IPK, is now mak­ing the nec­es­sary wake-up call. The AI pioneer’s tip for pro­duc­tion engi­neers: come to EMO Han­nover 2019 and be inspired by the WGP (Wis­senschaftliche Gesellschaft für Pro­duk­tion­stech­nik — Ger­man Aca­d­e­m­ic Soci­ety for Pro­duc­tion Engi­neer­ing) and trail­blaz­ing machine man­u­fac­tur­ers to take your first steps towards AI.

Pro­fes­sor Krüger, how and when did you did first come into con­tact with AI?

Jörg Krüger: That was in 1992 when I was explor­ing the use of arti­fi­cial neur­al net­works in the diag­no­sis of machine tool axes. Today I am fas­ci­nat­ed by the tremen­dous advances made in deep learn­ing (part of machine learn­ing based on neur­al net­works and large amounts of data, author’s note) and con­vo­lu­tion­al neur­al net­works (accord­ing to Wikipedia, a machine learn­ing con­cept inspired by bio­log­i­cal process­es, author’s note), some of which already exceed human capa­bil­i­ties in pat­tern and image recog­ni­tion in the pro­cess­ing of audio and video data.

Mas­ter­ing com­plex AI sys­tems with sim­pler structures

In Novem­ber 2018 the book Der unter­legene Men­sch: Wir uns uns digi­tisieren mit Algo­rith­men, kün­stlich­er Intel­li­genz und Robot­ern selb­st weg? was pub­lished by Riva-Ver­lag in Munich. What do you think of the pes­simistic views on AI expound­ed by the author Prof. Armin Grun­wald, who, as head of the Office of Tech­nol­o­gy Assess­ment at the Ger­man Bun­destag (TAB), has a cer­tain amount of influence?
Jörg Krüger: Some aspects are becom­ing ever more com­plex and so I under­stand his scep­ti­cism regard­ing the inscrutabil­i­ty of these sys­tems. We need to cre­ate sim­pler automa­tion struc­tures that facil­i­tate bet­ter under­stand­ing of and con­trol over the com­plex­i­ty of machine learn­ing. The issue is whether pro­duc­tion engi­neers can trust a self-learn­ing sys­tem enough for them to deploy it in pro­duc­tion. Edu­ca­tion and train­ing often don’t adapt quick­ly enough to such rapid devel­op­ments in research. This means that only with a con­sid­er­able delay are we able to acquire the skills nec­es­sary to mas­ter this com­plex­i­ty and build con­fi­dence in these new tech­nolo­gies. All this, of course, fuels the notion that humans will even­tu­al­ly be pushed aside.

Exploit­ing employ­ees’ knowl­edge of their domain

And then the pro­duc­tion engi­neers find them­selves con­front­ed with a moun­tain of Big Data: What should they do with this?
Jörg Krüger: It depends very much on what kind of data you’re talk­ing about and what the pro­duc­tion engineer’s objec­tive is. One user, for exam­ple, want­ed to use image recog­ni­tion to iden­ti­fy and man­age 50,000 to 60,000 parts in a ware­house. Pre­vi­ous­ly, a ware­house work­er would do the work of iden­ti­fy­ing and sort­ing the parts using a cat­a­logue. We only had a lim­it­ed num­ber of images tak­en of each part, not enough to map a neur­al net­work struc­ture. How­ev­er, we found that by using pre-trained net­works based on non-indus­tri­al image data in com­bi­na­tion with a lim­it­ed amount of indus­tri­al com­po­nent data, it was pos­si­ble to achieve accept­able recog­ni­tion rates for pro­vid­ing assis­tance func­tions even from an ear­ly stage. The AI sys­tem then works as a semi-auto­mat­ic assis­tant that dis­plays the five most like­ly parts to the ware­house work­er. Thanks to this assis­tance, he now works much more effi­cient­ly and accu­rate­ly. But this only works if you know the process­es very well. That’s why my mes­sage is: don’t just invest in hard­ware and soft­ware, you also need to make sys­tem­at­ic use of the “domain knowl­edge” of the pro­duc­tion staff. Peo­ple must also learn to eval­u­ate process­es in the fac­to­ry and decide which tasks AI can take over. The inclu­sion of domain knowl­edge from pro­duc­tion is cru­cial in order to iden­ti­fy new areas of val­ue cre­ation poten­tial quick­ly and systematically.

Data is dig­i­tal gold dust

Smart assis­tance is one aspect, but what else does AI have to offer?
Jörg Krüger: The data gen­er­at­ed by com­pa­nies is dig­i­tal gold dust for me. In my expe­ri­ence, many com­pa­nies are not yet aware of just what val­ue cre­ation poten­tial it holds. Machine learn­ing tools are becom­ing increas­ing­ly pow­er­ful. In pro­duc­tion, we should now be sys­tem­at­i­cal­ly com­bin­ing data with domain knowl­edge in order to refine process­es and make them more effi­cient. I’m inter­est­ed in tack­ling this issue with col­leagues from WGP. For exam­ple, we should no longer be focus­ing on increas­ing recog­ni­tion rates with the help of machine learn­ing meth­ods as we did in the past. Instead we should be analysing the poten­tial of exist­ing pro­duc­tion data for machine learn­ing more sys­tem­at­i­cal­ly than before and exploit­ing the result­ing poten­tial for rais­ing pro­duc­tion effi­cien­cy lev­els more sys­tem­at­i­cal­ly. I rec­om­mend watch­ing the       YouTube video by the Cana­di­an sci­en­tist Ajay Agraw­al or read­ing  his book Pre­dic­tion Machines: The Sim­ple Eco­nom­ics of Arti­fi­cial Intel­li­gence. Here, automa­tion or pro­duc­tion engi­neers view the oppor­tu­ni­ties of AI from an eco­nom­ic per­spec­tive in order to dis­cov­er the dig­i­tal gold dust in their own com­pa­ny.  Com­plete­ly new val­ue cre­ation mod­els and nich­es sud­den­ly emerge for small busi­ness­es, espe­cial­ly for start-ups.

What role will sen­sors play?
Jörg Krüger: “Sen­sori­sa­tion” is usu­al­ly the first step in acquir­ing data for learn­ing. The more pow­er­ful and cost-effec­tive the tools for machine learn­ing are, the more valu­able the under­ly­ing data becomes. Espe­cial­ly in the field of sen­sori­sa­tion, great progress is being made in research and devel­op­ment in rela­tion to Indus­try 4.0 — a good pre­req­ui­site for tak­ing the next step towards machine learn­ing in production.

Sys­tem­at­ic AI solu­tions based on Ger­man “Inge­nieur-Denke” (engi­neer­ing thinking) 

But what are our chances against coun­tries like Chi­na, which are invest­ing vast amounts in AI?
Jörg Krüger: The lev­els of invest­ment in AI infra­struc­ture, as we’ve seen in Chi­na in par­tic­u­lar, are indeed impres­sive: it’s dif­fi­cult to see how we could match that here in Ger­many. In terms of inter­na­tion­al com­pe­ti­tion, how­ev­er, I see a bright future for Ger­many in apply­ing struc­tured engi­neer­ing think­ing to the indus­tri­al use of AI or machine learn­ing. We should then be able to main­tain and build upon our very good glob­al posi­tion in the field of automa­tion in the future.

What are you most inter­est­ed in see­ing at EMO Han­nover 2019 – not only in your capac­i­ty as a pro­duc­tion engi­neer­ing researcher inter­est­ed in AI?

Jörg Krüger: I’ll cer­tain­ly be inter­est­ed in see­ing the exhibits being pre­sent­ed by our WGP col­leagues. For exam­ple, I heard that an insti­tute will be show­cas­ing some­thing very excit­ing in the field of pat­tern recog­ni­tion on machine tool dri­ves. That’s all I can tell you. It should also be worth­while check­ing out the stands of some machine man­u­fac­tur­ers and automa­tion companies.

Pro­fes­sor Krüger, thank you for talk­ing to us.

Trumpf: Using AI in production
“Arti­fi­cial intel­li­gence does­n’t kill jobs; it’s sim­ply a fur­ther log­i­cal step towards secur­ing Germany’s com­pet­i­tive­ness and pros­per­i­ty,” says Dr. Thomas Schnei­der, Head of Devel­op­ment at the Machine Tool Divi­sion of Trumpf GmbH & Co. KG, Ditzin­gen. “The mechan­i­cal engi­neer­ing knowl­edge we have gath­ered over decades is essen­tial for mak­ing use of arti­fi­cial intel­li­gence in indus­try. We must seize this oppor­tu­ni­ty.” In Swabia they are already deploy­ing AI: 25 Trumpf employ­ees are work­ing to ensure the trans­par­ent and group-wide coor­di­na­tion of the company’s AI activ­i­ties. Arti­fi­cial Intel­li­gence can, for exam­ple, use an auto­mat­ic laser sys­tem to analyse the removal of cut sheet met­al parts – ini­tial­ly unsuc­cess­ful­ly, but ulti­mate­ly suc­cess­ful­ly – and auto­mate the pro­ce­dure using the result­ing data. Trumpf can then apply this method to all machines of this type. AI has also proved suc­cess­ful in the company’s own pro­duc­tion: sen­sors in the machine record large amounts of data dur­ing a short test and send them via the con­troller to the cloud, where the AI solu­tion analy­ses them auto­mat­i­cal­ly. An inge­nious com­bi­na­tion of sim­u­la­tion, mea­sure­ment and analy­sis meth­ods enables the machine to be mon­i­tored in count­less oper­at­ing states. If there is a prob­lem in the data, it not only recog­nis­es the error, but also knows how to cor­rect it from its pre­vi­ous analyses.

Author: Niko­laus Fecht, spe­cial­ist jour­nal­ist from Gelsenkirchen
((Size: around 8,800 char­ac­ters includ­ing blanks))