If you are curious how a Data Science leader thinks about applying Artificial Intelligence to physical dynamic systems in the areas of Industrial IoT, then this is the episode for you!
QingHui Yuan, a Director of Modeling and Data Science at Donaldson joins us to talk about all of the topics above and much, much more! Prior to joining Donaldson, he was a manager at Eaton in their Hydraulics Advanced Technology Group. QingHui has Master's Degrees in Electrical and Computer Engineering and a Ph.D. in Mechanical engineering, both from the University of Minnesota.
I love QingHui's definition of Artificial Intelligence. He states, "Learn from the past, predict the future." Spot on and conscience!
If you are interested in learning about how AI is being applied across multiple industries, be sure to join us at a future AppliedAI Monthly meetup and help support us so we can make future Emerging Technologies North non-profit events!
Resources and Topics Mentioned in this Episode
A lot of mathematic foundation established a long time ago, I feel maybes adoption, or the attraction of AI really happening in recent years because otherwise mental infrastructure, when I say infrastructure is a highly aware software so that a lot of things which we think impossible is impossible to do now is it become very practical exercise.
AI Announcer 0:29
Welcome to the conversations on applied AI podcast where Justin Grammens and the team at emerging technologies North talk with experts in the fields of artificial intelligence and deep learning. In each episode, we cut through the hype and dive into how these technologies are being applied to real world problems today. We hope that you find this episode educational and applicable to your industry and connect with us to learn more about our organization at applied ai.mn. Enjoy.
Justin Grammens 1:00
Welcome everyone to the conversations on applied AI podcast. Today we have Cheng Wei, he won, a director of modeling and data science at Donaldson. Prior to joining Donaldson, he was a manager at Eaton in their hydraulics Advanced Technology Group. jingwei has master's degrees in Electrical and Computer Engineering and a PhD in mechanical engineering, both from the University of Minnesota. Thanks, Janeway for being a part of the program today.
Thanks for having me.
Justin Grammens 1:24
Awesome, awesome, cool. Well, you know, I gave a little bit of an intro with regards to where you're at today. And how you kind of went from Eton to being a Donaldson, maybe you could put a little more color around it, I guess, how did how did you get into the field? And, and, you know, how have you kind of progressed in your career?
Since for the question, actually, it's a pretty unique experience. For me, my background is control engineering, as what I gather my PhD in control, dynamical system control you mechanical engineering, control is a pretty interesting field, right? It's dealing with all kinds of dynamical system behavior and design, you're trying to bring the behavior system to the desired state. So I think, in Eton days, I was having opportunity to dealing with a lot of complex problem. So I think the one example like I'm not sure you're familiar with like electrohydraulic above it is a control device to control the hydraulic fluid entering and leaving the hydraulic actuator. So all the bigger construction equipment, agriculture equipment, majority are controlled by the this type of fluid control device is actually highly nonlinear, if you think about the manufacturing a millions of different control device, you know, you're going to have a tolerance, different tolerance, and used to be a very, very tedious manual tuning process to make sure each bottle working for each machine, it is unbelievable, tedious. So, I had an opportunity to develop a some kind of electric hydraulic control system, then the one thing I say, How can we overcome this kind of variation from one system to another are we able to introduce some of the intelligence to help us understand like a mechanical data band or different system response, so that they can learn and learn the system behavior and adjust certain parameter so that for different emission different system, they can still maintain optimal or semi optimal performance? So that part of work actually reflect what is a control domain trying to do right you, you already have a concept of the training your control algorithm to achieve the optimal, you already have, like a monitoring kind of mechanism. But so so far, the embedded system is a single system. So I think I personally develop the interest the one look at a faster growing into me on the data science, you know, in an ISA is a similar it's a similar problems is in a bigger scale, right? You have a more data, you have a different dynamical system you're trying to predict and hopefully control. So I say, yeah, you know, what I learned in in the different scale can be applied to a much, much larger scale, and does that have an impact for the different ecosystem? So I started with the RV interested from control to eventually data science, and I feel fundamentally a lot of connection and a common ground. Yeah, just a little bit of background how I switched from one domain to another.
Justin Grammens 4:44
Yeah, for sure. Well, I mentioned during the entry of a PhD in mechanical engineering, do you sort of see that being a unique skill set to because like you say, you're talking about these physical systems and all the mechanics going on inside of it, but yet, you have all this experience in electrical and computer and engine
I feel the multi discipline a certainly a help, because, you know, in a data science, eventually you solve a lot of different problems. So I think it's very hard to be successful, just stick to input output and a focus on algorithm, you have to pull back to look at what is a problem, why there's a problem important, you have to formulate something, right. And then after you're gone for going through the algorithm exercise in apply your artificial intelligence and gather some performance, you have to look at it was as a means, right? What is the impact that was added means? So I feel like my training on mechanical engineering, electrical engineering and computer science and in particular control mindset, I feel give a me a lot of help to understand this kind of multi discipline element, you know, perfect. Yeah, I mean, anytime you're dealing with a physical object in the field, and you're you're trying to get data from it, there's all sorts of things that are out of your control. And so being able to apply, like you said, this multi discipline approach certainly helps, because it's not as straightforward as people would would think when you're talking about these control systems, especially the very complex ones that you're working on. One of the people that I like to ask, one of the questions is, how do you define AI? Ai is such a broad concept. But you know, do you have a short elevator pitch? Either, sort of around how you conceive AI, when people ask you, what do you do in your job? Do you have sort of a way that you'd like to explain it? plosive definition AI these days. So I think, from my perspective, I came with a very simple, and I call it learn from past predict the future. So So basically, if you think about you, comission learning AI is a learning, right? intelligence, like a human intelligence, the reason we be able to hold our bar that high is because we have continuous learning experience from past from childhood from school from experience, that gave us the capability to formulate a thought and predict I think, before artificial there, I feel is is the same thing. It's a very simple, you get you give them opportunity to learn. And it gives them opportunity to predict for sure, for sure. Yeah, yeah. Yeah, this concept of AI, I think, has been around for probably since the early days of computing, right? It's everyone's been wanting to make their computers more intelligent. What do you think is different, I guess, today than maybe what we have seen in the past eight, you know, say the past 10 years versus the past 50 years? Yeah, I think if you I mean, even from the past a book called control theory, and machine learning a lot of mathematics foundation established a long time ago, a long time ago, I feel maybes adoption, or the attraction of AI really happening in recent years, because otherwise ment of infrastructure. So I want to say infrastructure is a highly aware software. So so that a lot of things which we think impossible is impossible to do now is it become very practical exercise. So So that basically opened the door right for many, many, many opportunity to apply as our ism developed many, many years ago to growing application. So I think that the enabling piece I feel is really key to drive adoption and actually going to be continue being adopt in the coming year. It just becomes a toolset. Right? For for many, many applications.
Justin Grammens 8:40
Sure. Sure. Yeah. So like, what's an example you think of with regards to things that you thought maybe we thought were impossible, you know, 10 years ago or so. And all of a sudden, I guess I wouldn't say easily solvable, but they're actually solvable today. Do you have an example you could think of?
I mean, one example, I feel probably autonomous drive, right? If you think about you, you're dealing with so much information, and you have to deal with in real time. And even for 20 years ago, you think about that kind of intensity of computation is never achieved when you embedded a system. And well, now it's it's, you know, so many companies are doing that. It's a common practice now.
Justin Grammens 9:21
Yeah, for sure. For sure. Great example, well, what are some of the problems that you and I guess your organization are trying to solve today that are pretty difficult for you, but maybe in the next when we talk five years from now, you'll have them soft, but I'm just kind of curious. Maybe you could shine some light on that.
Yeah, I think for enterprise AI, I call business AI sometime. We're not necessarily trying to bring the most challenging technology in house so many times is trying to understand what are the opportunities, just the understand that what opportunities we can apply this kind of new tech. Knowledge is a tool converted from many process to certain automation and a certain intelligence to make the job impossible possible. I have, you know, I think a few example I think, in Donaldson, we actually explore several swimming lines in AI or machine learning domain. I'll give a few examples. Because a pretty broad, it's not like we're just doing this, or we're just doing that it is continuous expansion of our data driven culture in the organization. And the way for us to do that is always find the most promising opportunity and apply to certain technologies, see the success and use success story to educate prior organization and then build attraction gradually, as a couple interested in seeing what we're what I'm thinking the AI help us is the first why I will say IoT IoT space. Donaldson is a filtration company where leading industrial filtration with a fundamental core to providing clean air to all the application right. So, we used to make manufacturing companies is a create media put into element sell two different application and one airflow contaminant being captured, then you get cleaner, but there is no real understanding. Like before this filter element, what is really happening for the dirty air is a filter really reached the end of life or is a filter is a still in the beginning of usable life, that there is a owning schedule based maintenance practice in the field. So every three months, I mean, talking about different application in the truck every six months is is changing the filter. in industrial manufacturing application every two or three months, you just change a filter. But But now we need to actually help a customer under understand the value story. If you change You're too late, or did you change too soon, because there are consequences on each of the sides, right? I can't tell you for industrial field, there is a measure pretty expensive. It's another like two or $3 is a very expensive, we'll talk about hundreds of dollar, if you just randomly change your base on week or monster schedule, you may actually waste a lot of material, waste a lot of materials or fuel or maybe only 10% on life, you didn't attack your factory in 100% of capacity, you still just the change of the field or as euro as a waste, right? Yes, yeah, for sure. Another side, you know, if you say you know, I really want to push everything to the boundary, I really want to push everything to the extreme, then you run the risk of the filter actually reach end or life lose a capability to filter, the contaminant, then contaminant getting into your equipment, I think we have a very wide application from aerospace, transportation, or for highway industrial manufacturing, but in and you know, your equipment are going to be vulnerable for the performance and the life, you know, some Third of May maybe kill your engine. So we basically think, you know, introduce sensing as a starting point into our filter. And then when you receive like a pressure signal or temperature signal, you can start to learn, you can start to learn what's going on, you can start to learn what is filters real life based on the sensing data, that's one thing. And you also can also learn what is operating procedure and the duty cycle for different applications. So that'll go into that training and the learning, going to give a tremendous insight for us to understand what is the right maintenance schedule, what is a recommended the insider to the customer in a different field. And as I eventually created the fundamentally created value for them, right, create a value for them. So this is one example I think, is a very practical and you deploy the sensing element to the field. And you collect that data you use AI to study to understand the pattern, understand the trend, create the insight, and the customer make a change on their action, you know, so there's a control system you think about is a control system, they change their behavior, they change their supply chain system, so you don't need to put a pile of the element in your field as a human. So because all the information flowing you can streamline the supply chain and once you reach the end of life three weeks ahead of time you can inform ordering system. To kick off, then you've got a parse right on your door, right? So that that kind of thing is pretty amazing on the create tremendous opportunity.
Justin Grammens 15:09
You're right, very, very practical example, because it's not a one size fits all, every organization is going to be using their filters differently. And you're right, if it's just a standard time come out and change it too early, too late. And there's consequences on both sides of that. And then I think what's really cool too, is that you, you're talking about a percolating through the entire system, right? So you're right, there's the entire supply chain side of it, where now you don't need to store all these things. It's just in time ordering. So it makes it makes a ton of sense for an organization to sort of employ, or I say, I would guess would say, to deploy some of these technologies within their company. What's some of the challenges? Like I mean, obviously, that's okay. So we have this solution we're looking at, there's everything from the sensors to the connectivity that need to be set up to the data ingestion to I mean, just to filtering out all the noise that maybe is going on within that, what are some of those challenges, then if we, if we stay on that sort of, like, really awesome business use case? What are some of the challenges that you're seeing as you actually go to build out this technology? In general, I guess, you know,
our practice to solve the challenge problem, that Bell Labs, could it be a lot of technology review meeting, it's a conversation with a supplier or technology provider partner, and also have a some kind of logistical piece, you have to consider investment budget, you know, hey, we want to have this a big investment. That's he or for this technology, we're how we figured out investment the framework, because it's too late, you want to do tomorrow, and you talk about the money, right? So right does that sound sound is always a part. And hiring, hiring and recruiting in data science to me, I feel is very challenged, actually, I'm growing the team on the data science side, I witness of some of the challenge of how you find right talent in this domain. And you have to be able to spend more time to understand market dynamic, and be flexible and able to get the right people into the job. And personally, I think I want to reserve everyday reserve some time to absorb development in this area, you know, this is a fast growing area, both data engineering and data science, you know, almost every domain, you know, 10,000 company pop up, right in this domain every year. Sometime you probably feel 204. But you know, have to maintain that kind of learning agility every day to see what's going on. And make sure you stay on top of the field. But you just feel like Jesus so exciting every day and time is is a is a biggest enemy. So yeah, for sure.
Justin Grammens 17:54
No, I know what you mean about trying to stay engaged and learning new stuff. That was kind of the reason behind sort of this podcast. And the applied AI group is just getting more and more people together to talk and learn from each other. I'd like to dig a little bit on the hiring and recruiting stuff like what would you suggest people do? Are there any Are there any books or classes or conferences or stuff like that any places that you have found other podcasts or places that you found sources of inspiration that maybe you might suggest to people
was a lot of perspective, maybe I will talk line by line? Just my perspective, from a cost perspective, right now, there's just so many resources available, I feel like you know, you can just grab any of the reputated University Institute, there are a lot of high quality courses available, I think a message action is a maybe on one focus on some fundamental really understand what are the operating principle underneath some key hours, because there are a lot of open package open source packages available, a lot of solution available. And it's a permanent getting less important for implementation to let you write the 1000 line code for one particular algorithm it is a utilization and choose the right one to solve a rider problem. So build a fundamental understanding of the different piece how they play together working well on different condition is more important than you know the implementation of individual algorithm. The number two on the cost side I think I feel like right now the things are moving from PC and on premise setting to cloud so I feel for data science practitioner, maybe find a wall to I don't care is AWS or Microsoft, you know, you choose why, you know, just get a little bit flavor and a certificate on how the data science platform working together with the ecosystem right database, you know, cloud aware. House stream ally, a mechanism pass, you know, visualization, this kind of little bit of flavor of, Okay, I eventually going to deploy that not on my PC right behind on this internet. So what are the other basic element I need to be aware? And to connect with the algorithm and AI ps? I feel like is that going to be provide a little bit of boost to the career?
Justin Grammens 20:27
For sure, for sure. Yeah, no, it's, it's, it's a big difference between like you said, I was kind of laughing the guards like, you know, running something in a Jupyter Notebook on your computer to bring it into a full sort of ml ops pipeline where, you know, models can be adjusted on the fly, and all that sort of stuff out in the field. So great, great perspective. Well, we're getting close to the end here. I'll put some liner notes. So I'll put a link off here to whatever sites you want to have people check out both across the board here. But yeah, how do people reach out to you?
Yeah, LinkedIn is a perfect choice. And I actually spend a lot of time on LinkedIn for the continual learning. So I pretty proactive over there. So you know, contact me, feel free to contact me over there, send a message, you know, and be willing to share whatever I have.
Justin Grammens 21:14
Awesome. Cool. Was there anything else you wanted to talk about? Any, any subjects you wanted to touch on around artificial intelligence?
Just interesting fact. Right? I I just recently learned how much the impact of machine learning and AI to our daily life I wish I didn't pay attention previously, because we think, you know, data itself is a bigger problem. Connecticut data Canada algorithm deployed, I think that occupied majority of my site. Recently, I heard the energy consumption of training for a while with a study from IBM showing one sophisticated algorithm, the algorithm, just the for training, the electricity for the training is equivalent to carbon footprint of five car in a lifetime. So it is a think about this effect. You know, it's a fascinating eventually, if you think about, okay, it's not only about the digital impact, because the digital are going to have consequences on electricity, and on the physical world. So if we deploy more and more, and this kind of energy consumption and other things that will be eventually linked to our daily life. So I know that data center already consumed like a 10% that you know, TriCity. So I just consider moving forward and more model being deployed, and Zana probably going to be a bigger factor in future. So I didn't feel like it's interesting to share that observation. And we'll see how that goes, you know, people will starting to pay attention to energy efficient training, energy efficient computation, or maybe more some training from cloud to edge, you know, all kinds of things are going to develop. And, and it is a very interesting angle. I just lie. I feel like maybe we'll share with the audience.
Justin Grammens 23:13
Yeah, that's great. That is great. Yeah, I mean, training models aren't free, like you said, from an electrical standpoint. So just you think you're going to pay for it down the road. Because you could potentially deploy this model out somewhere and an air conditioning unit that would be smarter and would only turn on when it needed to, and all that type of stuff. But the impact actually train that model cost to do something. So it's sort of a, it needs to be a net positive at the end of the day. And you're right, I guess I hadn't really even thought a lot about that most people are so busy just training models for the sake of training models. And they don't really know that there is actually a cost to that. That's great perspective. Excellent, excellent. Anyway, I appreciate you taking the time to talk with us today and sharing your knowledge. And I look forward to hearing more from Donaldson and yourself, obviously, as you move throughout your career. And all of us are kind of learning and sort of building on each other in this new and exciting space. But I want to say thanks again for being on the show. And hopefully we'll have you back in the future again to talk about what's what's changed.
Justin thanks for the invitation is a great pleasure to be with you today. Yeah, anytime in future.
Justin Grammens 24:20
Great. Take care. Bye.
AI Announcer 24:24
You've listened to another episode of the conversations on applied AI podcast. We hope you're eager to learn more about applying artificial intelligence and deep learning within your organization. You can visit us at applied ai.mn to keep up to date on our events and connect with our amazing community. Please don't hesitate to reach out to Justin at applied ai.mn if you're interested in participating in a future episode. Thank you for listening
Transcribed by https://otter.ai