• Benjamin Moses

AMT Tech Trends: Better Research, Better Podcasts

Updated: Apr 28

Release date: 24 April 2020


Episode 24: Ben finally got his home automation running seamlessly… for now! Steve caved and replaced his overheating Raspberry Pi 4 with a liquid-cooled gaming computer that he promises will be used for some work. Stephen then opens with some cool ML work MIT has done implementing a vision system to monitor social distancing. Ben shares a cool video highlighting the usability of Boston Dynamic’s Spot industrial robot, a dog-like quadrupedal drone. Stephen talks about the metrology used to perform alignments on military vehicles. Ben closes with a CIRP paper on “Machine learning-based image processing for on-line defect recognition in additive manufacturing.”



Benjamin’s Linked In www.linkedin.com/in/benjamin-moses-b13b44a2/ Amateur Machinist Blog swarfysteve.blogspot.com/ Music provided by www.freestockmusic.com



Transcript:

Benjamin Moses: Welcome to the Tech Trans podcast where we discuss the latest manufacturing technology research and news. I'm Benjamin Moses the director of manufacturing technology and I'm here with...


Stephen LaMarca: Stephen LaMarca technology analyst.


Benjamin Moses: Steve-


Stephen LaMarca: Ew, dude, I just did that creepy twitch thing with my head that the guy from the Papa John's commercials does. Better research, better podcasts, AMT tech trends.


Benjamin Moses: Oh man, welcome to day 1000 of the coronavirus.


Stephen LaMarca: It's only day 1000.


Benjamin Moses: We had a lot going on. One good thing though, weather's kind of warming up. I got the sprinklers back up and running. Felt good. As one of the couple of automation things I've got going on home and I'm really happy with the way things have been automated. Sprinklers I've had since after we recorded the last episode and they've been running great. The evidence I have that they've been working well is the grass is getting more green.


Stephen LaMarca: Nice. That's good data.


Benjamin Moses: That's good data. Yeah, all the brown patches, all the hibernation and the recovery from the winter has been going great. Also the robot vacuum cleaner has actually been fixed now. So I had issues where it would get stuck every night and just wouldn't vacuum most of the floor. I moved it to a different location. I don't know if that fixed it, but for me it's working more and it's not getting stuck as often. So that's been working out great.


Stephen LaMarca: Interesting.


Benjamin Moses: Yeah, it's weird.


Stephen LaMarca: Well that's awesome. The home automation is finally a past the teething process it sounds.


Benjamin Moses: Yeah. I am a little wary though. So the sprinkler system working great and the people around me have permanent installs put in, which annoys me. So every season I've got the above ground systems. So I take it all apart for the winter time and take it back out. I'm not going to invest in ground system just yet. I'm going to hold out, but I have a external timers. So they're connected right to the spigot outside.


Benjamin Moses: You're not supposed to get to below 32 degrees according to the sticker on the timer. I got to 33 degrees last night, which is weird for springtime, I think. So it was a little cold. Everything was wet this morning so I think it worked or it could be leaking. [inaudible 00:02:22] I don't know, but it got a little chilly last night.


Stephen LaMarca: It did. It's actually been kind of chilly.


Benjamin Moses: Yeah. Yeah.


Stephen LaMarca: But I mean, it's been really bright. Of course we've got like one of the most beautiful springs at least in of any year in my recent memory of this area and were locked inside.


Benjamin Moses: It was like, I think it had a high of 47 degrees today. I don't know.


Stephen LaMarca: Oh Really?


Benjamin Moses: Yeah. [inaudible 00:02:49].


Stephen LaMarca: It's pretty, but it's cold.


Benjamin Moses: Yeah.


Stephen LaMarca: But speaking of temperatures with respect to my home test bed, which really is just the Raspberry Pi that I've been blabbing about for the past few episodes, as you know, it's been overheating. It's had an overheating issue, and it's a known issue with all Raspberry Pi 4s. It's not just me. I'm not abusing it.


Benjamin Moses: Which you should. I recommend abusing any electronics.


Stephen LaMarca: I mean, yeah. For a half hour at a time you should abuse it.


Benjamin Moses: You should abuse it.


Stephen LaMarca: But so it's been overheating. I tried, as you know from the last podcast, I tried adding some heat sinks and it didn't really quite work.


Benjamin Moses: These are passive heat sinks right? So there's no fans.


Stephen LaMarca: Theses are just passive heat sinks. They have this thermal compound tape, which I'm not even sure is actual thermal compound. I think it's just-


Benjamin Moses: Just tape.


Stephen LaMarca: I think it's just double ... it's probably not even brand name 3M double sided tape. It's probably like some off-brand stuff.


Benjamin Moses: 5M double sided tape.


Stephen LaMarca: Yes. 5M from China. It was double sided tape and he stick the heat sink on to the chips that you want cooled. And I put a heat sink on the CPU, the ram and some other chip that I probably should know what it is, but I've no idea.


Benjamin Moses: Sure.


Stephen LaMarca: It's not like it has a GPU. So I have really no idea what it is. Didn't work. Most recently, I ordered some pretty high end ... I've been watching a lot of PC videos on YouTube. Ordered some pretty high end thermal compound, Thermal Grizzly thermal compound, which is the stuff all the computer nerds use. That's the good stuff.


Benjamin Moses: Yeah, they use that in between the CPU and the CPU cooler.


Stephen LaMarca: Exactly. Yeah. So what I did was I pulled the heat sinks back off. I pried them off, I cleaned up the chips and the heat syncs with 91% isopropyl alcohol to remove all of the adhesives or move the thermal compounds [crosstalk 00:05:05]chips.


Benjamin Moses: You know the proper way to use 91% alcohol right?


Stephen LaMarca: How do you use it?


Benjamin Moses: You take a sip and then you use it.


Stephen LaMarca: Okay. isopropyl doesn't work like that.


Benjamin Moses: Oh no.


Stephen LaMarca: Please, everybody listening, do not drink isopropyl alcohol. It actually turns into, I think methanol in your body.


Benjamin Moses: Yes. Yeah. It's very, very awful.


Stephen LaMarca: It does something really nasty to you. Not fun.


Benjamin Moses: So disclaimer, that was a joke. Okay.


Stephen LaMarca: I got you. Got you. But I don't know who's listening man. You can't let them ... Don't let people try that.


Benjamin Moses: If you get in trouble fax us a letter.


Stephen LaMarca: Write your complaint on the back of a hundred dollar bill. But so I cleaned up the board with in chips with the isopropyl alcohol and I cleaned the heat sinks with it. Then I applied a little bit of the thermal compound and stuck the heat sinks back on. And it seemed to help at first and then a half hour later I get the overheat warnings, your CPU's being throttled. And so I gave up and not immediately, but I didn't realize that okay I was seeing at first these things working and then they didn't.


Stephen LaMarca: And I realized that while I wasn't preventing, I wasn't able to prevent the PI 4 from overheating, I was slowing down the change in temperature.


Benjamin Moses: Oh, okay.


Stephen LaMarca: That's what military weapons companies use with like machine guns. On a machine gun, you use a heavier barrel, a thicker wall to barrel on a machine gun, then you would a standard issue rifle. And the purpose of that is not only for weight. It's typically a machine gun mounted to a vehicle or something like that. But with sustained fire, which rapidly gets the temperature up on the gun barrel, throwing more material at it, which is effectively what you're doing with the heat sink slows how fast the temperature changes which can have a negative effect. Because that also slows how quickly the barrel cools too. But it's essentially the same science used for the cooling of a computer chip with a heat sink that is. We're not talking about using fans, but essentially-


Benjamin Moses: So you increase basically the thermal mass of it that it has to accumulate.


Stephen LaMarca: Yeah, and it was a fail on the poor little Raspberry PI just because the Raspberry PI 4 is awesome.


Stephen LaMarca: It's the most, it's the most powerful PI that that company has made. And it's a very impressive computer for the size. And it's simple, which is awesome. But you ruin all of that simplicity when you realize that the only way to effectively keep this thing cool and powerful is by putting some sort of fan cooling system on it. So all of the simplicity that you engineered into it is out the door-


Benjamin Moses: So what are you going to do next?


Stephen LaMarca: ... which is a shame. What I do next is ... So I replace it with a very expensive gaming PC that's liquid cooled, and it's very nice.


Benjamin Moses: I like it.


Stephen LaMarca: This is really for me, and it's also part of being cooped up at home during this pandemic it gives me something else to do.


Benjamin Moses: Yeah, maybe I'll take your Pi and use it in the future automation.


Stephen LaMarca: Yeah, no we can definitely do the brotherhood of the traveling Pi.


Benjamin Moses: All right. Into that.


Stephen LaMarca: And send it on. It's awesome. It's got a great enclosure for it. It's got the heat sinks on it, but had to be replaced. By speaking of which this new gaming rig hasn't all been done for play. I've actually ran a Autodesk's Fusion 360 to look at and tweak with the AMT test bed model that I'm doing and man does a powerful computer mean a whole world of difference when you're doing like CAD design.


Benjamin Moses: Yeah. So for background, you are using it on a basic business laptop that's-


Stephen LaMarca: Yeah, that has onboard graphic.


Benjamin Moses: Onboard graphics, medium grade CPU, low ram. Now you've gone to a full desktop computer that's got his own discrete video card, a high end processor and tons of ram. I mean those must be night and day difference. Plus you're on a big screen. You're using your TV as a monitor, right? So it must be awesome to move the model around and you probably can manipulate the entire cell if you want it to.


Stephen LaMarca: Right. And everything I've done on this computer so far with either CAD or video games not only has the temperature not even crept into the 60s it's been in the low fifties the whole time. But the thing has been silent.


Benjamin Moses: Oh nice.


Stephen LaMarca: It hasn't made a peep and there's no lag or any sort of like imaging fractals that you'll see on the screen or when I'm playing in Fusion 360. I can't find enough stuff to throw at this thing.


Benjamin Moses: I mean, so I'm assuming you're enjoying the water cooling then.


Stephen LaMarca: Yeah, I really am.


Benjamin Moses: that's the way to go.


Stephen LaMarca: The silence is nice.


Benjamin Moses: Yeah. So tell me about your first article. You mentioned you got something on some machine learning and social distancing.


Stephen LaMarca: Yeah, so somebody sent me over the weekend, I think Saturday night somebody actually reached out and sent me this really cool MIT article about ... The title is, Machine Learning Could Check if You're Social Distancing Properly At Work. And I figured this would be a great article to bring up for handful of reasons. Number one is by NIT. When they speak people listen. And secondly, we're talking about social distancing and the pandemic and just being safe to flatten the curve and whatnot. And then thirdly we want to talk about it because they're using machine learning and you don't even need to read the article. I behoove you and all of our listeners to check out this article. I'll link it in the description just to click the link because the first thing that you'll see right below the title is like a GIF. Or that's either a GIF or a short video clip of essentially people walking up and down this city street somewhere.


Stephen LaMarca: I don't know where it is. I think it's in Europe. But people walking up and down the city street and the computer that is monitoring this video feed is putting a box around every person, even a little baby that's being pushed in the stroller has a box around it, which is kind of creepy, but everybody's got a box around them. And if the box around a person is green, that indicates that that person is six or more feet away from everybody else. Any other identified person on this street. If people are too close, it puts a red box around them. If they're under six feet from each other and puts a red box around them and a line connecting to try to visualize who that person is too close to that many people. It's wild. I mean yeah this technology has been around for awhile, but it's good to see it being used in like a non creepy big brother esque platform instead of just like, "Hey, hey, hey you guys six feet."


Benjamin Moses: And I do like-


Stephen LaMarca: Instead it's just being like a pool lifeguard.


Benjamin Moses: Yeah, that's a good point. I do like an article. So the third paragraph talking about the under the hood and one thing that I've always wondered about. So they have a use case, how do I use that to use case in my own scenario? And it gets into how do I calibrate it to other scenarios. They get into other scenarios they could use it for, like a really large factor you want to implement, like an Amazon warehouse or something like that.


Stephen LaMarca: Yeah.


Benjamin Moses: But you want to implement ... ensure our social distancing.


Stephen LaMarca: Potential collisions, like if you could measure-


Stephen LaMarca: If a machine learning system like that can measure the distance between known moving objects and it can probably also calculate their speed and then you can have a differential equation implemented into the vision system in that if this object is going this speed, it has to have a larger distance between other objects so it can maneuver safely at that speed. Or if something's getting too close, it needs to slow down regardless of whether or not that other object is getting closer to being in its path or not.


Benjamin Moses: Yeah, and it talks about what's required for calibrating to the local environment. Basically taking known measurements out in that space and putting that back into the algorithm. So they're using a neural network. So that's a solid implantation. We have a lot to learn from what's being done in autonomous vehicles and automated ground vehicles to other applications. I mean, this is the similar object image recognition that you would see like in any other autonomous or supportive vehicle. So cool. Awesome.


Stephen LaMarca: Remember when a Tesla, I think it was Tesla, had that hiccup with their autonomous vehicle programming?


Benjamin Moses: Mm-hmm (affirmative).


Stephen LaMarca: That if the car needed to do a U-turn, it would do a, you turn to the capabilities that the vehicle do a U-turn.


Benjamin Moses: It hauled so fast.


Stephen LaMarca: And so the FBI style J turns and then people realize ... well, then they realize, "Oh Whoa, we need to slow that. Not make it to the vehicles capabilities, not to the vehicles all limits. It should be to the limits of human comfort."


Benjamin Moses: That's a good point.


Stephen LaMarca: So you learn a little things like that too in that kind of programming.


Benjamin Moses: Speaking of automated vehicles, the next article I have ... actually, it's a video from YouTube. It's Adam Savages Tested channel. So Adam Savage of course is originally from MythBusters. He has a YouTube channel. So it goes around doing Adam Savage type things, getting some cool technology and talking about it. I saw another video that he did where he talked about AV equipment for the play Hamilton. I was really nerding out about the AV equipment that they had, but as a different podcast we can talk about. And he was taking the Boston dynamics robot named Spot around.


Stephen LaMarca: Oh, no way.


Benjamin Moses: I don't know if he has it or they lent it to him, but he had one that he could use and he was there by himself in a open area and he was showing how to use it. And I think it's a really good video that shows that it's not a mysterious thing that's working as a sentient object. It's basically a significantly more advanced drone. So if you've seen any of the four blade drones, you've got a tablet or a phone and a bunch of controls. In the video that Adam shows, he's got a controller with a tablet built into it and he's moving around based on that tablet.


Stephen LaMarca: No way.


Benjamin Moses: Now of course you can have maybe program different locations, but the way he was using it was, "Hey, robot, go in this direction, go at this speed, go at this gallop." And well, at a very basic level that sounds really immature. But when you look at the controls and some of the interesting thing that's he points out in the video, there's some really cool collision detection going on and it's probably coupled with the machine learning type object recognition that you talked about near article. I have a couple of scenarios that he points out.


Benjamin Moses: So one where he's moving the robot away machine ... I'm calling it a machine and I'm not calling it a dog, I'm calling it a machine or he's moving the machine away and it gets closer to like a lamppost and he gets near the lamppost but then it walks next to it. It doesn't actually collide, doesn't do anything. But one thing they point out in the video where the leg is actually really, really close and if you use the standard gate, the leg woulda hit the post but on its own controls it move the leg and did as kind of jog with that one leg to avoid touching the lamppost.


Stephen LaMarca: No way. So instead of like moving the leg in -


Benjamin Moses: A front back motion.


Stephen LaMarca: ... The most direct path, it did it in sort of like a curved path to go-


Benjamin Moses: Yeah, it jogged around it.


Stephen LaMarca: Wow.


Benjamin Moses: Yep. So there was scenario where he's asking the robot a machine to go forward and it's close to a building. So the front of it's about to collide into the wall. And of course it's smart enough not to collide into it, but it's still receiving command to go forward. So what it does, it starts moving perpendicular, I'm sorry, parallel to the wall, but his body is still perpendicular. So it was looking for a way to go around that object and still continued to go forward. So I thought that was really fascinating where it didn't just stop like an emergency stop or didn't know what to do. It's still receiving commands from the operator to go forward and still trying to figure out how to go forward. So I thought that was fascinating.


Stephen LaMarca: Wow.


Benjamin Moses: The last scenario that he points out was it tries to go ... it goes up a flight of stairs that has a landing, so it's got to go back onto itself. So it's smart enough to kind of place each foot on a step, comfortably articulate around the landing, which is fairly tight and then continued up the stairs. I think Adam was forcing or controlling it to go around the landing. It wasn't command to say go upstairs he was driving it around as, as you would a drone. But the sensors on the robot ... And actually it's interesting that you in seeing the video where it has a warmup phase. So you actually put the machine down upside down and it has a warmup phase where it rights itself up and then stretches around and looks around. You can see the different cameras all around the robot. So it is monitoring its own feet. So it's got not only the forward looking vision system, but it's got some under the belly and some on the rear. So it can see almost-


Stephen LaMarca: Wow.


Benjamin Moses: ... 360 degrees and can see self-aware because it could see his own legs.


Stephen LaMarca: So when you deploy this, and we won't call it a dog, we'll call it a quadrupedal. When you deploy this drone, unlike an aerial drone where you like you need to launch it from like a flat surface and you put it upright, this has to start upside down.


Benjamin Moses: It looked like it. Yeah. So you carry it's got handles. So you mentioned that basically the entire thing is an entire pitch point. So unless you use the handles, your hand will get pinched at some point.


Stephen LaMarca: Got you.


Benjamin Moses: Because it's got your, the legs are so close to the robot.


Stephen LaMarca: Interesting.


Benjamin Moses: It was really interesting. It's a really good look at, "Hey this thing isn't mysterious. You control the like you would anything else except it's smart enough to be more self aware so it wouldn't call it collide and clash crash into things." If you think about the scenarios that he walked us through in that video is indicative of what the implications of how to use that in real life. So it's not a warehouse robot where you know you build a flat factory and it picks up a bunch of things. It's a say like a post off delivery thing where you can have a go open downstairs or you can issue commands and it'll figure out certain things within a set of constraints. So it was a pretty good video. I recommend it.


Stephen LaMarca: That is cool. I will them watch that. I will definitely watch that after this. You know what would be really cool? And it's kind of silly for me to say this, but I think it would be a really cool project to do for science. I think somebody, whether it's Boston dynamics or somebody who just buys a handful of these Spots, should take these two the [inaudible 00:21:03] see how they handled themselves against a bunch of Huskies.


Benjamin Moses: There was a manufacturing paper that was presented five years ago on swarm robotics. So the idea was let's construct something but use super simple robots that have super simple sensors, maybe like one or two but all working together somehow to effectively construct something. Either they would use themselves as a construction device, so they're building themselves or go and fetch objects. So the idea of swarm robotics is growing, not at the pace you would expect, but the idea of five or six things giving a singular task and they can work together to figure it out, it's moving along.


Stephen LaMarca: Wow.


Benjamin Moses: A little scary at the same time.


Stephen LaMarca: But cool.


Benjamin Moses: But cool. Yeah. Tell me about your paper on metrology.


Stephen LaMarca: So my last one, my last article that I have, I'm just Scanning Metrology News, I love that website. But I came across an article released two days ago. It doesn't matter. Optical Metrology System Improves Alignment Of Military Vehicle Performance. And you and I both being gear heads we get our cars aligned and the technology just to align a civilian automobile, it has come a long way and a ridiculously long way. Not only do our alignments now performed, not so much by, they're still performed, executed by hand, but all of the measurement is done with lasers and stuff like that. And those alignment racks that they put your car on at a service center, they have come a long way. And you know, you look at, "Well why the hell is a an alignment, a hundred $150 to get it done?" And it's just like, not only does it take a lot of time with the manual labor, it's relatively ... it seems like relatively easy work, especially when you have a lot of the advanced computer systems, but they're still expensive because you have to pay for all that high end metrology to do that.


Stephen LaMarca: What I'm trying to get as the technologies come along the way. When I dropped my car off at the service center in Tyson's, I actually pull into this Bay where I just drop off my car and I get out. But as I'm pulling in, I noticed that I'm driving over these sensors and cameras and I asked somebody there what they were doing. It's cameras, vision systems and lasers measuring not only the tread depth of my vehicles, tires-


Benjamin Moses: That's cool.


Stephen LaMarca: ...to measure how much life is left expected in the tires that I'm running, but it actually measures the alignment of the treads of one tire to all four of the other tires to verify the total alignment of the vehicle's suspension, which is wild.


Benjamin Moses: That's wild.


Stephen LaMarca: So they can actually tell you whether or not you're wasting your time asking for an alignment or whether you might need to or should you need to replace your tires due to uneven tire wear, which may be caused by an alignment. So there's a lot of that stuff. And that's just for civilian vehicles. this article gets into the metrology system and the means of aligning a heavy military vehicle like a tank or a Bradley fighting vehicle or a striker APC, which you know, I'm sure most mechanics, most skilled mechanics can probably tell you, "Oh, it's all the same thing. They're just bigger and heavier." But there's probably different ... We probably can't put an APC on the same lift that you put a Camry on.


Benjamin Moses: You could, it won't end well.


Stephen LaMarca: It won't end well. But it's a really fun article to read. And also metrology just doesn't get enough of the limelight.


Benjamin Moses: I agree. Optical metrology is underplayed quite a bit.


Stephen LaMarca: Oh yeah, and it's only getting better.


Benjamin Moses: Only getting better. Absolutely. The last paper I've got is a research paper. Let me read the title here. It's Machine Learning Based Image Processing For Inline Detection Recognition in Additive Manufacturing. So there's an international paper. Most of the work was done from Italy and the US and the underlying idea is to do semi real time metrology or flood detection and selective laser melting machines. The idea would be they're using machine learning to one, recognize the image, and then also understand if it's a flaw detection. So two applications of machine learning. So what they're doing is not real time as in when the laser's passing. What they're doing is waiting for after the laser passes on the bed and the bed has solidified, they're taking super high resolution pictures of the bed once it's solidified before the next layers printed.


Benjamin Moses: So it's a really interesting application. And the reason why I found it super useful is that additive has growing quite a bit and getting into production. And when you look at like aerospace applications, because with my background in aerospace, I'm really interested in those applications. 100% of the parts are either CT scan or some level of subsurface inspection, 100% of the parts. And that's really, really I'll say time consuming and tedious and maybe not the best thing for large volume production. So the idea would be how do you move forward from there?


Benjamin Moses: And I think one thing that we always forget is what are the parallels back to castings? So if we look at castings have been around for a hundred some years, castings, aerospace, have been around for a really long time. What have they been doing? And they'd been doing similar type processing where if I pour a casting, I do x-ray inspection for all my parts. I do 20, 30 different angles. But those have been semi automated where I can digitize those also. And so I'm doing digital x-ray and I can see those slaw details. I can come back excavate and re-weld and process and if I PR reprocesses within the mass spec limits to save the part. But this method looks at them as the parts are being processed, collecting enough information to determine whether or not I have flawless subsurface and I thought that was really valuable.


Benjamin Moses: But in the end, the main technique that they're using for all the machine learning data scientist nerds out there, they're doing deep convolutional neural networks. So they're both doing that for image collection, image recognition and classifying of the flaws and jumped at the conclusion here they achieved a 99.4% accuracy and defective conditions. So I thought that was really positive news for the industry to help us further along additive manufacturing into production world where you want high volume, you need fast processing. You need these unique designs, but you don't want to spend time looking at parts and the human looking at these parts all day also. So I thought that was really good.


Stephen LaMarca: Yeah.


Benjamin Moses: Steve that was an amazing podcast. Gave a really good update on the news.


Stephen LaMarca: It was.


Benjamin Moses: Really hoping for some really, really hot and humid Virginia weather sound. I don't know about you.


Stephen LaMarca: Not too hot, not too humid.


Benjamin Moses: You're not a fan of the hot weather warmer.


Stephen LaMarca: A little bit warmer.


Benjamin Moses: A little bit warmer. Yeah, I agree.


Stephen LaMarca: The car really likes the cold air.


Benjamin Moses: That's true.


Stephen LaMarca: And so do I.


Benjamin Moses: Yeah. Well, tell us how they can find more on our info on the podcast today.


Stephen LaMarca: You can find everything in the description below. That's including the

articles that we're going to, as well as Ben's LinkedIn account.


Benjamin Moses: And you can find us-


Stephen LaMarca: Any other features today?


Benjamin Moses: And you can find and you can find us on AMT news.org.


Stephen LaMarca: That's right.


Benjamin Moses: That's where we're posting all the transcripts also.


Stephen LaMarca: That's right.


Benjamin Moses: All right. Bye everybody.


Stephen LaMarca: Bye everybody.

Recent Posts

See All