video_OCP2026_英業達_NVIDIA_AI工廠代理_20261015

All right. I think it's time to get started. Uh welcome everybody. I'll be talking alongside JK from Inventec about bringing autonomous agents down to the factory floor. Um so I'll kind of tee it up and then we'll see how Inventec has applied this technology. So in a lot of today's factories, there's a lot of the inherently legacy challenges, right? Everything's kind of rules-based, is very rigid, siloed systems leading to unplanned downtime, quality issues, and such. So all of this yields slow reactive time. And especially in terms of decision-making and when you're trying to, you know, get things out the door, um it is very important to try to compact that time to a decision-making. Whether it be how do I plan my line changeover, how do I bring up a a a machine that's going back down, how do I do a particular SOP, how do I make my my product line more efficient. So through the evolution of AI in the factory, which started off with pretty much computer vision-based quality inspection, um we have moved on to generative AI, which has brought that data hurdle down, and then agents came about that can then correlate multimodalities, whether it be video against your PLC signals, your logs, your ERP base systems. We have proven the value of AI in the factory, right? To actually have implications. So now it really becomes how quickly and how easily can I integrate these AI agents into my existing brownfield and legacy systems, right? On the OT side. And how do I orchestrate this and get these agents uh working together? And that's really where the factory operations blueprint comes into play. It's taking that same paradigm of what tool calling and skills and agent hardnesses has done for software development and bringing it down to factory operations. So now you can have an agent or a multiple of agents that are secure, that have governance, right? With any kind of agent harness being either run through frontier models or local models within the factory, and have it be able to work um discover skills, discover other AI systems, talk to um uh normalized OT data to understand what's going on in your factory operations for either root cause analytics, for maintenance or for planning. So you can have an agent and you say, "Hey, how many did I ship yesterday? How many did I uh how many had an issue? What was the history of my defects? How do I bring up a particular machine? How do I handle this error?" And it did not have to be planned out that way ahead of time. Uh we almost say that the dashboard in in manufacturing is dead because the agents now can dynamically create that dashboard and that intelligence for you along the way. So in terms of the benefits of what factory or bringing uh autonomous AI agents down to the factory, of course is integration, right? We know AI agents provide value. It's just how quickly do I have those agents working with my machines, with my legacy ERP base systems, all of those working together in order to correlate it. And that's going to allow me to build quicker and then also to operate a lot faster and get to a lot more uh efficiency. So with that, I am going to hand it over to JK that will show you how Advantech has applied this technology in their own manufacturing. >> So Hello. Hello. Mike. Hello. Sorry, the uh Hey. Uh thanks Alvin for the uh introduction of NVIDIA uh factory operation blueprint, FAX. So uh we Inventec has been adopting AI technologies to facilitate our server manufacturing for a long time. Uh recently by adopting uh NVIDIA FAX, we're making the AI solution much more agentic and autonomous. So referring to the uh the diagram here on the left, we use NVIDIA frameworks to do data collection, do uh processing of data, and to to synthesize data whenever necessary. Then we use uh frameworks such as uh Cosmos and uh Nemo uh to train the image recognition model and uh large language model we use for the factory operation. And over in the middle, uh the um with the trained model, we build agents which can understand factory contexts. We use such agents to to do things like anomaly detection, uh cycle time counting, and monitoring as SOP uh compliance, and also do root root cause analysis. So with all the uh factory operation data we collected and also the uh anomaly or maybe the uh incompliance we have found, we can generate reports and turn factory data into business insights, which are valuable to both us and our customers. And finally, we can uh the agents can find inconsistencies or insufficiencies in our uh AI models and take such inconsistencies as example, uh loop back and train our uh AI model again. So, this form a closed-loop AI solution in our factory operation. So, in the rest of the talk, I will uh share with you three use cases in our factory. The first one is uh laptop surface defect detection. Uh in the past, we had we need to have human operators to uh uh look at the uh sur- the laptop to find defects on the surface. Uh this process is quite inefficient. Uh also due to like lighting condition difference, uh viewing angle, or maybe human operator fatigue inconsistency, which all result in in in unsatisfac- factory uh results. So, uh a couple years ago, we introduced image recognition models to uh uh look for such defects. And this turned out to be pretty good, but then we we were hit by another problem. Uh every time we introduce a new laptop model, due to the uh difference in color or texture of the laptop, we need to recollect data and retrain our model to be able to work on the new new laptop model. That's when we uh learned uh NVIDIA's anomaly gen solution. Let me show you with a video here. The uh NVIDIA anomaly gen solution can when we give give some defects example to the uh anomaly gen, uh it automatically segment out the anomalies, create masks. And with the mask uh defects, it use generative AI models to produce more examples of the same or similar distribution. Meaning the uh defects in different sizes and different shapes, but with similar uh look. So, now whenever we introduce a new laptop model, we only need to collect a few samples and use anomaly gen to to turn that into hundreds of or thousands of examples to train our AI model. And that solves our uh problem. And the second example I'll share with you is the cycle time detection. Uh with uh in in the past, we need to have line manager or uh production engineer using a stopwatch at the station to to count cycle time, which is very limiting and and also inefficient. Uh with trained AI model, uh this can be automated. And uh right now we have real-time data of cycle time, and we have full log of all the stations with cycle time logs, and we can use that to optimize our operation and find inconsistency in our cycle time. Look at the footage video footage and do improvement right away. The next example is um the use of uh uh Cosmos model. Uh in this example, we we feed the model with our uh operation SOP. Uh we don't write any additional rules. The model understands uh for example, this this this is a CPU socket assembly on a motherboard. There are six screws that needs to be applied applied to each socket, and the screws need to be applied in in a specific order. Uh in this video, the the model understand the SOP and by looking at the video, it finds inconsistency or incompliance of the SOP right away. Not only that, the uh the model also try to analyze uh what the root cause might be and suggest solutions to for improvements. So, uh in this talk, I shared three uh use cases with you. Um we use uh perception AI and generative AI solution to solve the laptop surface defect detection problem. And we also use a genetic AI to solve the cycle time counting. The SOP compliance and also do root cause analysis. In addition to that, we also try to apply more and more robotics to our factory operation. Yep. So, with all of that, I think we are only halfway through our journey with the um uh factory factory AI introduction. Um yeah. I think as outlined by Alvin earlier, we think in the future we will build factory AI agents that are more um that can do more reasoning and analysis and also task orchestration among all the different agents. And with that, we hope to improve our uh the cost, the efficiency, and also yield rate of factory continuously. Yep. Uh thanks. And by the way, Innodisk has a exhibit exhibition booth at Hall number two. If you want to learn more about the talk and also maybe about Innodisk product and service, feel free to come over to have a chat. Thank you.