Rise of AI Summit 2020 · Berlin · Applied AI Day | Future Space

How to Successfully Implement AI in Manufacturing

Thibault de Assi · General Manager and Global Head of Industrial Computing at Siemens Digital Industries Factory Automation

At Rise of AI Summit 2020, Thibault de Assi, General Manager and Global Head of Industrial Computing at Siemens Digital Industries Factory Automation, presents Siemens' approach to deploying AI on the factory shop floor. He opens with figures on the scale of the opportunity: 85 percent of potential industrial assets remain unconnected, 87 percent of machine learning projects never reach production, and he expects 41.6 billion industrial IoT devices within five years. He describes Siemens' stack of control layer, edge layer and cloud layer (built on MindSphere and Mendix) and walks through six client use cases spanning die-casting defect prediction, bakery visual inspection, additive manufacturing quality checks, automated guided vehicle failure prediction, beer quality optimization and PCB inspection, each with concrete before-and-after metrics such as defect-rate reductions and cost savings. He answers audience poll results showing most shop-floor customers still run only proof-of-concept AI projects, and discusses lack of trust, insufficient in-house AI expertise and keeping models updated as the top barriers to adoption, along with a typical six-to-twelve-month payback period.

Key takeaways

  • De Assi cites 87 percent of machine learning projects failing to reach production as the core problem Siemens' approach targets.
  • An edge-deployed AI model for die-casting defect prediction cut defective parts by 30 percent and reduced the defect rate by 40 percent.
  • A single stopped automated guided vehicle costs 500 euros per minute, motivating Siemens' failure-prediction use case for AGVs.
  • PCB inspection with an AI model achieved 30 percent fewer X-ray tests and roughly half a million euros in reduced capital investment.
  • De Assi estimates that around 50 percent of Siemens' customers are still only at the proof-of-concept stage for shop-floor AI.

Detailed notes

The scale of the opportunity

De Assi opens with figures framing why manufacturers need a systematic approach to AI: 85 percent of potential industrial assets remain unconnected, 87 percent of machine learning projects never make it into production, and he projects 41.6 billion industrial IoT devices within five years. He says Siemens' approach starts by understanding a customer's data flow across the value chain and how operational technology integrates with information technology, then combines an AI model, a software application with domain know-how and a human-machine interface, and industrial-grade hardware such as industrial PCs and edge devices to run it, with Siemens also managing the full lifecycle from connectivity to deployment, monitoring and retraining.

The technology stack

De Assi describes a layered architecture: a control layer of industrial PCs and PLCs controlling field devices such as power plants, production machines or elevators; an edge layer that he says is not a simple gateway but covers connectivity, openness, local data pre-processing, security and machine learning; and a cloud layer built on Siemens' MindSphere and Mendix platforms for developing and deploying industrial IoT applications. In a body-shop example, he shows automation and control (Simatic and third-party products), edge computing running AI and non-AI applications close to the machine, a centralized edge management system for mass and security updates, and an app store where Siemens and partners provide applications for deployment on the shop floor.

Six client use cases

De Assi presents six examples with results: defect prediction in die-casting, where an edge-deployed AI model cut defective parts by 30 percent and reduced the defect rate by 40 percent; visual quality inspection on a bakery line, improving counting accuracy and reducing manual inspection time; layer-by-layer defect detection in binder-jet 3D printing, saving about 2.5 hours of rework per day; failure prediction for automated guided vehicles, where a single stopped vehicle costs 500 euros per minute, avoiding unplanned stops; process optimization for beer quality, correlating raw-material and process parameters against a minimum quality grade of 6.8; and PCB inspection in an electronics plant, achieving 30 percent fewer X-ray tests, a 100 percent quality rate and roughly half a million euros in reduced capital investment.

Q&A highlights

Live poll results shown during the session indicate most attendees' companies are not yet using AI on the shop floor at all; de Assi says across Siemens' own customer base roughly 50 percent are running proof-of-concept projects, about 30 percent are actively scaling implementations, and 20 to 30 percent are just starting to explore where AI could help. On the top-cited barriers, lack of trust and insufficient AI expertise, he says Siemens addresses trust by monitoring deployed models for performance drift and retraining them with new data, and addresses the expertise gap by pairing its own machine learning experts with customers' process knowledge. He estimates a typical return on investment of six to twelve months, depending on customer and industry, contingent on jointly aligning on expectations and KPIs from the start.

Quotes

85 percent of potential assets remain unconnected, yet 87 percent of machine learning projects never make it into production.

Thibault de Assi

AI enabled us to trigger tasks that human beings are not able to do before.

Thibault de Assi

Questions answered in this talk

How does Siemens help manufacturers implement AI on the shop floor?

According to Thibault de Assi at Rise of AI Summit 2020, Siemens combines an AI model, software with domain know-how and a human-machine interface, and industrial-grade edge hardware, layered on top of a control, edge and cloud (MindSphere/Mendix) architecture, and manages the full lifecycle from connectivity to deployment and retraining.

What results has AI achieved in die-casting quality inspection?

De Assi states in this talk that an edge-deployed AI model for defect prediction in die-casting reduced defective parts by 30 percent and cut the overall defect rate by 40 percent for a Siemens customer.

What stops manufacturers from adopting AI on the shop floor?

De Assi identifies lack of trust in AI solutions and insufficient in-house AI expertise as the top two barriers cited by the audience at this talk, addressed by continuous model monitoring and retraining, and by pairing Siemens experts with customers' process knowledge.

Mentioned

industrial IoT, edge computing, die-casting defect prediction, automated guided vehicles, additive manufacturing, PCB inspection, model drift monitoring

Keep exploring

More from Rise of AI Summit 2020

More on Enterprise AI & Adoption

Transcript

Auto-generated transcript, not edited. · 4,013 words

Show the full transcript

our next speaker has an impressive academic record he has earned multiple master's degrees in engineering in computer science telecommunications engineering and image treatment and cad design he also hauls an mba from imd business school in luzan he currently leads a global department with expertise in artificial intelligence machine learning software development automation and simulation please welcome the general manager and global head of industrial computing at siemens digital industries factory automation tibor de asi hello tibo hello thank you very much good to see you we can see you we can hear you the floor is yours perfect look around you look at the assets around your mobile phone your production machine your robotic station 85 huge number 85 percent of potential assets remain unconnected yet 87 of machine learning projects never make it into production there will be 41.6 41.6 billion industrial iot devices in five years from now i cannot see your faces unfortunately i wasn't able to come to join you in berlin and i am sure you are wondering how we siemens are helping

our clients in connecting and protecting and interning data into valuable insights and to answer to that question i'm here with you today tibo do asi when we help our customers in their journey we really take the time we really take the time to understand their pain points to really understand the data flow across the whole value chain and to really and really understand how operational technology layer is integrated with information technology layer and then we bring our hardware and software integrated solution together with our industry know-how and our machine learning experts to answer so our clients challenges it all start with data there is different data sources in the shop room and from there take a look at the circle and take a look at the left side of the circle from there we're going to create a near-eye model but this is not enough take a look now at the right side of the circle from there we're gonna create a software with domain knowhow including a human machine interface helping operators to take

the right decision but that's also not enough we need to execute that software application on industrial grade hardware and for that we need an industrial pc we need an edge devices to realize customer value we at siemens we are also offering in operation to manage the ar life cycle from connectivity to deployment to monitoring and training and now take a look at the bottom of that slide that could be that could be your power plants that could be your production machine that could be the elevator that take you to the berlin conference or to your home this morning on top of it we have what we call the control layer this is where with the industrial pcs or with the plc with the controller are controlling those field devices then we have our edge layer and for us edge is not a simple gateway for us edge is about connectivity it's about openness it's about local data pre-processing it's also about data and security because security is at the heart is at the core of

our product portfolio and it's about machine learning and artificial intelligence the layer above is the cloud layer with our platform mindsphere and with mendix we are offering to our clients an efficient way to develop their own industrial iot applications and deploy those application on in the ot on the on on edge and on the edge layer this this is an example of a body shop this is the whole ecosystem at the left side of the screen the number one this is where we have the automation the control layer with our simatic products or other third-party products then we have the edge computing this is our edge devices the industrial grade hardware where we are running ai application and non-ai application close to the machine it's a cloud agnostic and then we have is an edge management system this is a centralized management system for mass updates for security updates and for distributing application across the whole manufacturing line we have also an app store enabling our our partners but also siemens to provide ai and

non-ai application and enabling to deploy those application on the shop room these are use cases where we are developing and creating impacts with ai at our customer side in different industries take a look at the second use case in the first row smart assistant siri i'm sure you have been using siri to find the nearest french restaurant uber is becoming ever today we we have some some cinematic assistant for machine it embed natural language processing and knowledge graph enabling the field maintenance engineer to communicate and to collect data from the machine and to execute its maintenance of a procedure in efficient manner let's take a look on some of those use cases that you have been implementing in the industry the first one the first one is about defect prediction for diecasting today our customers are inspecting the parts at the end of the line he's looking at the part manually or he is using automatic optical inspection he is cropping more than 10 000 part per year a different trade between two to four

percent and other customers it can go up to ten percent what we have been doing here we have been identifying we have been identifying process parameter affecting the throughput of the die-casting machine we built an ai model and we deployed that model on our edge devices to make real-time annually detection and defect prediction and feedback that result to the plc to prove to prevent further productions of defective parts it helped our customer to further reduce the process costs and defected products by 30 percent and you manage to optimize the process and decrease the defect rate by 40 percent and i am sure in your own words you also would like to have such savings let's have a look on the architecture so we are collecting data with our simatic s7 the controller what you see on the left side of the screen also additional data coming from different quality data coming from quality management system and other sources feed these data is feeded into our edge platform that run on our edge devices there we

have our trained ar model for clustering and classification purposes with the data we have also a storage and from there we have also dashboard this is where the operator can observe the data to take some decisions that would impact the die casting process in this case we have also a cloud gateway to send the data to the to the clouds for uh for data storage further further statistical analysis training purposes and later on deployments the second use case the second used case is coming from the bakery line is about visual quality inspection the baked product is ejected out of the oven on a conveyor belt down to the down to the packaging line today our customers are counting manually the baked items defects also could come from the degree of burn or the size of the of the shape of the items so what you have been doing we have been implementing an imaging system to get a near accurate count of the baked items to localize the defects that might happen after uh at

the end of the oven this data would be sent to the to the to the plc where we have our ai model running there and through the hmi we would be able to look if these big items are has good or bad quality it helped our customer to save time in inspection to improve or reduce inaccuracy in the counting area in the counting and to reduce defense the third use case is about visual inspection in additive manufacturing a bio.jets 3d printing is an additive manufacturing technology for making or for printing a product out of a digital file or out of a cad file this process could go up to 2500 layer per job it could last 30 hours to get the product printed and what you have been doing here we have been implementing efficient system to detect defects layer per layer in case of defect a trigger a stop signal for the printing machine the operator could take out the products take some measures either put it back or discard it and run another

manufacturing cycle it helps our customers to reduce time and rework around by around 2.5 hours per day the fourth use case is about failure prediction for automatic guided vehicles today at this custom agvs are being charged at the end of the manufacturing cycle one stop of a single agv results in 500 euro per minute 500 euro per minute production cost our approach was to analyze hiv alarms from the reports getting data from the charging stations and develop an air model to predict accurately when the maintenance has to be done and this helped our clients to avoid a stop indeed in the operations reduce production costs and optimize the charging processes here you have from architecture standpoint how this would work we have the agvs they have a controller we collect the data from the from the agv controllers this data will be sent to the to next devices where we have our ai model running there monitoring system that provide the information when do we have to send the data when do we have to

send the agv to the charging stations the fifth use case is about process optimization it's about improving beer quality i'm sure you all you love drinking beers i do so today there is a kind of inconsistency per batch process parameters are influencing the beer quality at the end of life and there is a kind of a beer sommelier after the packaging who is testing the beers and associating a grades at the minimum grade it has to be six point eight what you have been doing here we have been identifying key process parameter influencing the build quality if you look at the overall process we have raw material brewing cold block and fermentation packaging quality check parameters like time ph value alarms yeast type in the fermentation area dosing data and the fermentation quality energy and so on these are the parameters that have huge influence on the manufacturing processes and we developed the ai-based solution to predict parameters for achieving such optimized process to get the right quality at the end of life and we

observe that we observe that trend how we get to this 6.8 grade and we correlate all processes together to achieve that results we deploy we developed there in the model and we train that model on the on the edge as you can see here this helped our customer to improve processes to optimize his processes but also to improve the quality of the product coming from its own supplier the last six use case is coming from our electronic plants from our amber plants today to inspect pcbs we are using a very expensive machine x-ray machines what we have been doing there we have been gathering data we have been gathering data from the different process areas in the pcb manufacturing plants from the solder paste printer from the solar paste inspection from the pick and place machine from the ovens from the automatic oil inspections we developed the ai model data are being sent to our cloud platform to our mine sphere and this ai model being trained with the software application deployed and run on

the edge and there we would be able to predict the quality of our pcs it helped us to achieve 30 fewer x-ray tests it helped us to achieve 100 quality rates and reduce our capital investments by half a million we have our software and we have our industrial grade hardware to help our customers to deploy and run those use cases from the ultra compact from the ultra compact defenders device what you see on the left side of the screen this is where we have been running our additive manufacturing use cases he embedded this is where we have been running our agvs and behind when it comes to a real time quality inspection for high-end images or for training data on premise this is where you would require such powerful devices ai is one of the major breakthrough technology in our center ai enabled us to to to trigger tasks that the human being are not able to do before ai is providing the right information the right data to the operators helping him to take

the right decision how do we imagine ai in 2030 i do see smart assistance i do see smart assistants those smart assistants getting data from different sources from the operational technologies but also from the informational technology providing that information to the operators to perform uh efficiently and take the right decision on the shop floor i do see agvs auc agvs optimizing the roots by themself within the factories i to see more and more factories going into autonomous factories i obviously robots with more and more vision vision embedded system with ai technology being able to handle different parts different geometries so ai will bring a lot of a lot of productivity efficiency developing our own people to put us at the next level in terms of manufacturing and in terms of industry and now before going to q a and when i say go and not until i say go i want you all to go to the chatroom i want you all to the chatroom to go into the chatroom and write one word and

one word only so now look around you and imagine an object imagine an object that can be your smartphone your kitchen appliance that was produced using artificial intelligence ready go to the chat and type it thank you very much thank you thanks a lot thank you very much for your energy and your passion and uh i'm not sure if our audience has seen it but you have also prepared a couple of questions in the poll section of the swap card app so i would ask you all to look at these polls and please give us your answers so we can talk a little about that um if i maybe maybe maybe i want to build on some of the colleagues that are posting posting some words in the chat so uh so for example dr nico erdmann he mentioned car and absolutely here in the in the welding in the body shop in the welding process we are we are applying artificial intelligence to improve the quality of the spot welding thank you thank you

dr edmund for our for your contribution right we will i think we will we will see how the polls change over the next couple of minutes uh i just want to ask you something you gave us such a broad portfolio of things that are possible that can be done with your technology or technology maybe not from siemens but preferably so i have the feeling that there is something in there for every manufacturer do i understand that right would you say that it is a fool who manufactures anything and is not looking at the possibilities today yeah and in fact we uh we from uh the digital industry we are addressing uh broad range broad range of verticals so we are in in automotive we are in food and beverage with electronics batteries also in the within the process industry in the pharmaceutical in the chemical industry so uh i would say we have the right industrial industrial portfolio to help our clients to take advantage out of ai right okay thank you um let us

look maybe at the answers of the polls now and see if anything surprises you for instance i'm nobody voted for batteries being benefited by artificial intelligence most people settle on automotive and logistics um most of the people are not at all using ai solutions on the shop floor yet i'm not sure if this is like um a problem of our audience or is this something you encounter very often that people say very interesting but we don't do anything at all and we probably don't have any particular plans at the moment yeah i mean absolutely i mean actually if if we look to over all customers that that we are that you are helping them today i would say we have around 50 50 50 of those customer that are running proof of concept on the shop flow third i would say 30 percent they are really forging ahead uh in terms in terms of implementing and taking advantages and twenty to thirty percent twenty percent to thirty percent the early they have no idea said

you know what we know a little bit about ai here and there how you see ms could help us to really take take take advantage out of artificial intelligence and from a customer perspective we do see i would say different different buckets one who has its own experts and on ai and they are looking for industrial grade solution software and hardware the ones who want to partner with us and the ones who are asking us for a machine learning key expert in combination with our siemens hardware and software portfolio elements right um apart from the industries that you mentioned now do you see the possibility that in some completely different fields of industry suddenly there will be a lot of benefits and gains and we are not thinking about it yet because maybe technology hasn't evolved enough yeah i mean absolutely i mean i was i was i was for example uh at that one customer um he was doing he's an ice cream manufacturer and and and guess what uh he wants to apply

ai to improve the quality of the images that is on the packaging line so what happened today is that it's a manual activity right is it where we have the operator carrying big boxes of big samples of of those ice cream packaging of ice cream and they are inspecting manually so we do see in in in a lot of fields right ai could really uh could really bring value in there in the appliance for the kitchen appliance right we also see a lot of in the consumer world a lot of ai in there right thank you when we look at the question what keeps companies away from implementing ai on the shop floor there are top two maybe yeah the third one is also very close like the top two reasons is lack of trust regarding ai solutions and insufficient ai expertise for creating and deploying ai models how do you think these first of all do you think these concerns are merited and how could you work against them absolutely and first of all

the first word regarding trust i think my colleague olivatinger he will have a session around around uh trustworthy ai and he's gonna and he's gonna talk about this this this fema very very deep uh in the uh in in this uh in this event uh what do we what we do see today is two things first is lack of trust means uh how my a i am i gonna be sure that my ai model would be performing as expected in the shop floor and here come i would say our expertise our industry know-how right uh together with with our partners with our clients because they know their machine they know the process better than us and bringing our experts together with them we would be able to monitor accurately the behavior of the ai model and in case of drifting right of the performance then we would retrain the model with more and more data to make sure that the model is performing accurately the first point second point is yes absolutely right there is

a clear lack of expertise i was hearing also the colleagues regarding quantum computing probably it's not at that scale but actually there is there is a lack of expertise specifically in the industry when it comes to machine learning and and ar if you look to companies like google and facebook you can have a lot of these high high potential people in there but in the industry is still a little bit uh there is still there is still a lot of luck right thank you the third biggest hindrance the audience sees is keeping ai solutions updated according to production changes uh that frankly that surprises me i i would not think that it is so hard what what is your take on that do you have experience is it hard to keep the uh the thing updated or or is it easier than people fear uh abs absolutely and also i totally i totally share the feedback from the audience i mean assets assets are the behavior of the assets right are is are is being

changed and here comes here comes technologies and software and hardware being able let's say to to look at the assets behavior that would impact at the end or the whole process right and and in case and here our solution in terms of monitoring the ai model right it's crucial to be able to to get closer to to that our customers and provide additional services to improve uh to improve the ai model and bring it back on track right thank you maybe as a yes sorry i want to build back to the example of the beer right you have beer are being produced per batches so quality the raw material quality would have massive impact on the overall quality of the beer at the end of life right the raw material quality the meal gap the the milk batch and so on so this is also an important element that contributes to the mis to the mistrust of the performance of the air modem right thank you uh tibo maybe just uh just a rule of

thumb from your experience when when companies invest now in solutions like you propose them uh when does it pay off right so when do they have a return on invest um how how long does it take for them to clap themselves on the shoulder and say well i've done the right thing i would say i would say it depends it depends from customer to customer industry industry to industry so for some for some of the customer that that that i mentioned or some of the use cases that i highlighted before we do see we do see if we pay off after a period between six to six six six to 12 months and here it all depends for us our approach is really working together as a partner with our clients so here we need to both together align on expectation align on kpi bringing knowledge from both parties in order to achieve such return right thank you very much so let me pat you on the shoulder tibu thank you very much for your

talk and i hope you have a lot of success uh convincing people but i think with that much energy that you have it won't be a problem at all thank you very much thank you so much see you hopefully next year on stage thanks a lot have a good day bye

Rise of AI 2027 · Berlin · May 2027

Be in the room for the next one.

Ten years of talks started as conversations in one curated room in Berlin. The next edition brings 300 selected leaders together in May 2027.