YT YouTube2Text

YouTube2Text → Transcripts → Day 6 | AIO Automated Optical Inspiration | X-Ray Defect Detection | Manufacturing Optimization

Day 6 | AIO Automated Optical Inspiration | X-Ray Defect Detection | Manufacturing Optimization — Transcript

by STR2E · 6,853 words · 1,053 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00not evolve that was not progressive. So
0:02every time something new came again to
0:05plan new maybe old rules became obsolute
0:07or not no more functional because
0:09technology has changed and it used to go
0:12on going so it used to take forever
0:14that's why there is very little evidence
0:16that companies have implemented on a
0:19very large scale in a very complex
0:21environments like aerospace or medical.
0:23There are some companies that has that
0:24have done that but because the cost and
0:27the complexity involved and chances
0:30is very very less very big companies
0:32have only done that till late now once
0:36after this transformers paper came from
0:39Google and chat GPT and all this this
0:41progress the computing limits have been
0:44broken. So AI has become a reality. The
0:47deep learning the deep learning
0:49technologies where the machine will
0:52learn itself or machine learning
0:55whatever is the way we show them some
0:57examples
0:59examples which are characterized. Okay,
1:00this is good, this is bad, this is not
1:01acceptable, this is acceptable, this is
1:03partly acceptable so that it learns the
1:05patterns on it once it learns the
1:07patterns so it it keeps improving itself
1:11making it smarter like a regular way. So
1:15this data model is you feed all that
1:19photos and then get the output. This is
1:22one of the ways in which artificial
1:26intelligence can be used from the
1:29optical perspective in inspection. The
1:31next thing is where optical fields. So
1:34we go for X-ray defect X-ray. So when if
1:38you have seen a typical X-ray what it
1:39shows it shows your bones. Why? Because
1:42your bones are mostly metallic.
1:44So and electronics also involves what?
1:47Copper, aluminium,
1:49uh gold,
1:51right? So these are all metals. If you
1:53and alloys of the shoulder alloys are
1:55also metals. If you see them through a
1:59x-ray, we will find that other things
2:01are either translucent or they are not
2:04at all visible in the x-ray. So we are
2:06able to focus on our interest points.
2:08Like here if you can see some images
2:10here the x-ray has been taken of a
2:12device which is wire bonded and here you
2:14can see there's an open opening here.
2:17Similarly here a device has been
2:20soldered and it is below the PCB and we
2:23cannot see cannot see through eyes but
2:25X-ray because X-ray can pass through it.
2:28We have seen that there's an extra
2:29solder
2:30uh there's insufficient reflow there's a
2:34short there's a void there's a
2:35misalignment. So what it basically how
2:38it basically helps is where optical or
2:40our eyes cannot because we need a line
2:42of sight X-ray helps not to a 100%
2:46extent so that if we say we can say that
2:48optical inspection has covered this much
2:49and if we go by X then we are 100% sure
2:52that circuit is fine but okay to an
2:54acceptable level so these two techniques
2:57optical instruction X-ray have been used
2:59extensively till date and they are still
3:01being used and in medical industry uh
3:04what I have found
3:06AI based tools have come. It can analyze
3:08the patient's X-ray and tell the doctor
3:12can take decisions. He can refuse or
3:14[clears throat] take that is done
3:16faster. Same way we can al also do for
3:18electronics.
3:19Next is failure prediction.
3:23uh we keep generating data as I told you
3:26there are thousands of joints nearly one
3:29lakh joints in a satellite of all the
3:33circuit of satellite
3:35ever since the design starts till the
3:38spacecraft is launched until the
3:40lifetime is completed keep generating a
3:42lot of data lot of data but till now we
3:44not have any way to make sense of the
3:47data okay data analysis tools were there
3:49but they were still dependent on the
3:51skill of the person who is using that
3:53data. There was no concurrency. There
3:56was a consistency between person to
3:58person and it and we had to dedicate so
4:02many resources to find the relationship
4:05between a certain factor to a certain
4:08output and it was not every time very
4:11linear or very uh single polar it was
4:15very multivaried. So if you have seen a
4:17fishbone diagram you can find that for
4:19everything there is a there are pins
4:21that are associated similarly
4:24here what will happen is let me take
4:27this example let's say there is a
4:28satellite that is 2 years in orbit now 2
4:32years in orbit means that light would
4:33have much been designed before 2 years
4:36then it would have taken some
4:37realization time then all these data
4:41mean it it would have gone through all
4:42the processes and steps all these data
4:44has been captured
4:45So what will what happens is there's a
4:48model that is specifically trained for
4:49this. We give all this we keep giving
4:52all this data. So let's take an example
4:54of a transistor. A transistor is going
4:56on a PCB. That PCB is going in a
4:58subsystem. The sub then going in a
5:00satellite and the satellite is being
5:01launched. So we track the transistor.
5:04This we have seen the history of the
5:05transistor that okay if it experiences
5:08thousands of temperature cycles of this
5:10temperature sinks for these many hours
5:12at this functional level. What we found
5:14that the functionality of the transistor
5:16has gone down. Let's say degradation has
5:17started. So or let's say the temperature
5:20is not this much. If the temperature
5:23range is less, if the timing is less, if
5:25the functionality percent is less, the
5:27life is more. So this way we know that
5:30our predicted lifetime of a thing
5:33component is this under the this
5:35conditions. So previously what used to
5:37do we used to draw a simple model. wish
5:39to try or design a simple model model
5:42for each and every component which are
5:43not necessarily true which are most of
5:45the time it was not true. So each model
5:49each component can be modeled in various
5:51subm models if you can see those subm
5:54models can be analyzed and based on
5:57analysis of those subm models we can
5:58arrive at a figure which is more or less
6:01accurate based on the historical data
6:03what we have to model. So it will
6:06analyze the temperature cycles that
6:08would undergo the vibrations that will
6:10happen the satellite operational hours
6:12the all the radiation activity and
6:14everything whatever sensor is sent to
6:16the earth station. Now the AI says
6:19there's a warning because you have
6:22subjected this particular module or
6:25particular circuit to this this many
6:27extreme harsh conditions. It will fail
6:28in 2 weeks 3 weeks 1 month. So next step
6:32into size of what you do if you want to
6:35if you want this component to work more
6:37you reduce the temperature limits you
6:39turn on something so that the
6:41temperature goes out or what you do you
6:44reduce the working hours you take the
6:45load from that particular component to
6:47the other component because in satellite
6:49follow the concept of mean and redundant
6:50there is no there's nothing a satellite
6:52that goes usually typically that is the
6:55single point failure we always give a
6:57backup we call it main and redundant so
7:00it will ask if
7:01So that mean get some stress sleep or if
7:05you have backup for if you have some
7:07other backup. So this way not only it
7:10helps us predict how what will the
7:12failure what can be the failure it will
7:14also tell what is the remaining useful
7:15life. This way we can even extend the
7:18life from the design. If in maybe the
7:21data can help us to find it okay we have
7:24planned the design life as per these
7:26this condition but if we switch this
7:28condition that AI insights will tell
7:30that okay we can increase the right and
7:33the next one is very important and it's
7:36a very cool one a digital
7:39so what is a digital twin first so till
7:42now we all know that we have been
7:44designing electrical circuits we have
7:46transistors diodes and everything and we
7:48have equations so if we see what was
7:51those equations actually there just
7:52there are nothing but just models of the
7:54actual
7:56physical hardware correct we have V is
7:59equals to I for a resistor so this
8:02equation is what it represents a
8:03relationship it's a model similarly
8:05there are in network theory we have Z
8:07parameters H parameters and everything
8:09so what is actually do help us model it
8:12right so this modeling if we take to a
8:15very large level from the very component
8:17level to the very macro level from chip
8:20we have models and chip to package model
8:23package to PCB PCB to subsystem and
8:26subsystem to satellite so if you combine
8:29all these models we have a completely
8:31functional satellite that is working
8:33inside a computer so how what so it's a
8:36virtual copy there's a satellite that is
8:39actually in the orbit and there's a
8:41virtual copy that it's exactly same as a
8:43satellite which is in orbit and it is
8:45completely mirroring how it mirrors is
8:47real satellite temperature or some
8:49sensor data
8:51it it sends to the earth station that
8:53sensor data is fed to the digital twin
8:57the digital twin because it knows that
8:59if these parameters come to the
9:01satellite this happens so the digital
9:04twin also shows similar impacts so the
9:06manifestation of those environmental
9:09conditions or those factors will be seen
9:11in the digital twin now a will compare
9:14both the versions and the changes will
9:16not be a single variable a here and
9:18there the changes will be from the micro
9:21level to the macro level to track those
9:23changes through hard code or any other
9:26way is invisible. This is where the AI
9:28and because it has big data crunching
9:30capability it comes it will find each
9:32and every factor isolate each and every
9:34factor where the difference is now it
9:36will improve the digital twin. So after
9:39some time what happens is the base
9:40satellite functions and the way digital
9:42twin functions will be nearly identical
9:46and we will be able to guess what is
9:49going to happen because we can subject
9:51the digital twin to whatever factors we
9:53want because it is a virtual copy. So we
9:56can do anomaly detection. Now we we
10:00subject the digital twin to extreme
10:02ranges to find where anomaly can happen
10:04because we know that let's say after one
10:06week the satellite is going to a zone
10:08where the radiation will be too hard. So
10:11one week prior we'll have the data we'll
10:13have the information okay this plane has
10:15grown like this. So these are the things
10:17that will get suffered will either turn
10:19off their voltages current power
10:22whatever it is life prediction crystal
10:25is complete copy of that. So here only
10:28it will show that okay because you are
10:29subjecting like subjecting this
10:32particular subsystem to these these
10:34things this will go back life will or
10:37life or failure prediction they are just
10:38inverses of each other and mission
10:40planning. Although when mission planning
10:42goes there are things that we have to
10:45decide from earth based on the sensor
10:47that we receive and then we do
10:49calculations and then we send to a
10:50satellite to make its course correction
10:52change the orbit uh change orientation
10:55alignment and all the things.
10:57But here what will happen is we will
11:00simulate all the situations with the
11:02digital
11:03whatever is possible in the next 1 week
11:051 month 2 months 1 year after that and
11:08we have all the necessary steps that
11:10will be required and we can feed in
11:11advance to the satellite okay do this
11:13when this comes satellite has very
11:15limited things
11:17this is a very useful thing they are
11:20being used in industry but not in
11:22aerospace till now because of the
11:24complexities involved and there are some
11:26other reli ability concerns. But okay,
11:27this offers a very useful application
11:31and it offers a very huge uh what to say
11:35huge computational benefit you will get
11:37because once this comes into picture and
11:39people are working on this it's it seems
11:41that NASA has already done it and 2010
11:45they have been actively using it but
11:47there's a very little data on how much
11:49accurate and what is the reliability of
11:51that of those spins but yes people are
11:53doing it and we also will do
11:55in the future. Next thing is
11:59let's go back to the shoulder joint
12:01example. One leg joints.
12:04So those joints they have done not all
12:06the soldering in batches for machine.
12:09The machine works on certain parameters
12:11like speed, process, temperature,
12:13pressure, ranges and all those things.
12:16Then there's a yield. We let's say we
12:18pass 100 PCBs in an hour. So we receive
12:2299 PCBs after soldering which are
12:25acceptable. So that's called yield
12:27inspection time. So because inspection
12:30is also part of manufacturing because
12:31until unless that's inspected and
12:33cleared we don't say it is completely
12:34manufactured for the next process. We
12:37have we could have all these parameters
12:39we could optimize. We can learn full
12:41patterns because till now once a pattern
12:43is certified or qualified we go through
12:45the same and we usually we usually don't
12:48try to deviate from that but once this
12:50data is there we'll be able to
12:52understand what happens if we deate from
12:54this is one example that is very
12:55actively being used in industry and
12:57other other places also. Now if AI is so
13:01good applications are there like
13:03automatic optical inspection is already
13:05in they in industry your smartphones if
13:08you if I give you the very best example
13:10your smartphones like many of the
13:12manufacturing companies your laptops
13:14especially in China where they are made
13:15in the uh third third party contractor's
13:18facility in the same facility you will
13:21find that one line goes for the HP other
13:23line goes for the Lenovo they have this
13:25automatic in optical inspection systems
13:27where PCBs are inspected in hundreds
13:29per hour if not thousands. Accident
13:32effect detection detection systems are
13:33also working very fine. They're very
13:34good. Especially in medical industry,
13:35they are doing like anything in normal
13:38consumer electronics also they doing
13:40like anything. This uh the next thing is
13:43manufacturing optimizations have
13:45happened at scale we have never seen
13:46earlier because previously it used to
13:48take years to realize okay this is the
13:49pattern but now we doing in minutes and
13:51hours are better the useful life
13:53everything is going but still why is
13:55that AI is so great and why is
13:58everywhere so there are some challenges
14:02that is stopping us to put it everywhere
14:05we only know chat GPT Gemini they can
14:09check and generate code they can
14:10generate text, they can generate images.
14:12Fine, because this is what we have seen.
14:16So there are some challenges. So number
14:17one is training data. So when it comes
14:21to satellites or aerospace in general,
14:24they have very limited tearing data. As
14:26I give you the example of smarter
14:27manufacturer, iPhone or Apple itself
14:30ships nearly more than a million iPhones
14:33and combined with Samsung if you go I
14:35think it would be 3 to 4 million per
14:37year and each smartphone contains at
14:40least five sub five PCBs inside it
14:43minimum. So you can see the how much
14:46quantum of data they have to train the
14:48model. But in satellites we don't have
14:51that luxury.
14:52If we can realize 8 10 maximum 12
14:55missions we don't neither have the
14:57manpower nor that much capability nor
14:59the infrastructure and it is very
15:00difficult to scale that will not be
15:02justifiable to scale it to that level
15:03that okay we want to generate data for
15:0550 satellites 100 satellites in a year
15:06so it's not scalable as of now how do
15:09you train that if you have no data so
15:11the way is synthetic data so instead of
15:14waiting for real failures we create fake
15:17ones so how we create so there's
15:19something called physics simulations
15:20somebody would have used console or
15:23tools like
15:25and uh this
15:28guidance.
15:29There are some tools we that that helps
15:32you see what happens that if you put
15:35certain heat on that particular
15:36component, you put certain force on that
15:39component, if you give some radiation to
15:41that component, if some ionic activity
15:44happens. So these are lot of physics
15:45based models that are you that there are
15:48certain tools for that you simulate the
15:50data and we see okay this happens so
15:52this happen so this gives us some data
15:55of the failure so we try to simulate
15:57what are the failure points that are
16:00fairly accurate because those models
16:02have been working for years in the
16:04simulation tools and they have given
16:06kind of good real life what to say uh
16:10not with very much deviations so we run
16:14computer models to guess and see what
16:16happens in stress digital to as I
16:18explained earlier but that also requires
16:20data image augmentation because here our
16:24focus is on inspection of soldering
16:26joints images let's say a PCB
16:30is slightly tilted when it comes to the
16:32camera so what will the model do it
16:35won't be able to find because it has not
16:37been trained on that so all those
16:39conditions will make different images
16:41for the same image rotate
16:43[clears throat] it, flip it, modify the
16:45photos, change the background, change
16:47the transparency and everything. So that
16:49for one single image we are generate
16:51able to generate n number of
16:52[clears throat] images which are useful
16:54for us. Then there are CAD models. Now
16:56CAD models are also very good because
16:59each CAD design can give you n number of
17:02perspectives the way you want. You can
17:04rotate it in the way you can go inside
17:06and all those things. So if we capture
17:08each and every perspective whatever is
17:10important from us from the design file
17:12or the CAD model then we have the data
17:16instead of subjecting the actual model.
17:18So this is the problem number one and
17:20this is one of the many solutions that
17:22we can use in our mean like satellite
17:25domain. It's not that it's everywhere
17:28people would are using but in satellite
17:29it can be useful. The next problem that
17:32is
17:34the Y problem or the explanability and
17:36trust. Few years back Anthropic released
17:39a paper saying that we don't know how it
17:42is working. Okay, the conclusion or the
17:45summary of that paper was they don't
17:47know how their AI tool that CLA is
17:49working. So they are feeding some prompt
17:51they're getting some output and they
17:53don't know how that input is related to
17:55the output. There's no there there's a
17:57complete black box missions like
18:00satellites or even in medical I'll tell
18:04you let's say for example quality says
18:06AI have rejected this joint and somebody
18:08asks why but you say this AI says it is
18:1199% defect but we need to know why why
18:14AI says it is there what are the
18:16criterias on which it says how how will
18:18you explain that certain certain thing
18:20is acceptable certain things is rejected
18:22because that possibility is very
18:23important once something happens or even
18:25something goes slightly off the course.
18:28There is the concept of explainable AI
18:29that has
18:31gained some trend in last few years once
18:34this blackboxing came into picture and
18:36people were able to see that okay this
18:38is not good because we don't know how it
18:40is doing and if you don't know the
18:41process how it is doing then predicting
18:43the output of that process is very
18:45difficult and if that's the case then it
18:48cannot be given for some missionical
18:52so how explainable a health is it reads
18:55on why it has done the way it has
18:57If you write some code in thinking model
19:00of anything it is said it is thinking
19:01like this. If somebody would have used
19:03perplexity says thinking the way first
19:05step is the second but still that is not
19:07convincing. So what we want actually is
19:11if if something says something is accept
19:13or reject it should say why it is
19:15acceptable what are the parameter that
19:17it has met how it has done. Similarly if
19:20it says it has rejected let's say it has
19:22rejected because of a crack. It should
19:24highlight the crack with red box. Let's
19:26say there's a suspicious solder region
19:28that okay whether the solder is done or
19:29not. It should point because it cannot
19:31take a decision. It doesn't have a data
19:33to analyze on a suspicious case. If
19:36there's a void, it should measure the
19:38void because 100% void cannot be taken
19:42out. There will be some void after a
19:44solder. If some abnormal actually
19:46features are there that is beyond its
19:48that is beyond its trained data, it
19:50should not hallucinate. It should not
19:52just generate some conclusion just for
19:54the sake of conclusion. that that's what
19:56happening with the current tools right
19:58so this is the second problem and this
20:00explainable AI may give some solutions
20:03in the future that's all third one is
20:06power and computing
20:08if you are uh in touch with news you
20:12might have heard that Facebook is trying
20:14to revive or old nuclear power power
20:17plant to power its data center its
20:21computing center basically for its uh AI
20:24AI projects right.
20:27>> So each
20:29AI center has terabytes of RAM memory,
20:34hundreds if not thousands of GPU
20:37clusters,
20:38associated heat machinery,
20:41infrastructure and everything. So the
20:44scale is in megawatt. It's not even
20:47kilowatt. It's megawatt and sometime now
20:49it is going up to gaw. But satellites
20:52satellites have only few kilowatt 2 3 5
20:5410 max that's it memory
20:58let's say even if the best of what we
21:00have we can maximum we can pick up let's
21:02say just assume 10 some one or two TV at
21:06the max not more than that heat here on
21:10data centers we have dedicated heat
21:13cooling system that keep running
21:14continuously 24x7 at very high
21:16efficiency take out they're all
21:19conduction convection based we don't
21:20have luxury there. Now the example is a
21:23highend GPU like H00 or H 200 of that I
21:27don't know what's that called
21:30Nvidia has us 300 to 500 W maybe
21:33sometimes more we don't know satellites
21:36have limited also these it becomes
21:38infusible to put to have to assume that
21:41satellites will have that computing
21:42capability so the answer is H AI
21:46AI why it came into picture is we don't
21:49want everything to happen at a single
21:52dedicated central location every
21:54decision and everything to happen and
21:55there were concerns regarding data and
21:57everything and uh also letting the uh
22:01thing that every data has to be
22:02transmitted and to take its time. So why
22:05not do take the decision at the place of
22:08actuation.
22:10So very small models not billions of
22:13parameters very few let's say a million
22:15parameters which are fairly accurate
22:18okay and uh which have been trained on a
22:21sufficiently good data and they have
22:23been fed to the satellites
22:26the satellites computing power per watt
22:29is important per MB is important
22:32all those things need to take into
22:34consideration we should find the model
22:36we should tune the model so that it
22:37works there for the life of a satellite
22:40Maybe it's not that it should work every
22:42time but should work let's say for a
22:44satellite duration of 15 years it works
22:45half a time let's say every half an hour
22:48it works or every 15 minutes it works so
22:50that it saves some time it saves some
22:52power it saves some memory so the simple
22:55is that those data centers are like
22:57smart supercomputers
22:59but what we have in our hands is a
23:01smartphone it meets our requirement it's
23:04very limited it it it does not uh meet
23:06those computing requirements or the
23:08performance requirement It meets or
23:10similar in the case of
23:12another thing is next that is that
23:15defines this domain is qualification
23:18certification. Everything that goes into
23:20a space has to go all has to pass all
23:24the reliability parameters
23:27not only on the product itself but in
23:30the process also and that qualification
23:33certification requires time money effort
23:37and it has to pass to many standards.
23:39Now component goes to vibration testing,
23:42thermal radiation
23:45qualifications everything for these
23:47things qualification criteria have been
23:49already defined in the aerospace
23:51standards and they are working very fine
23:53till date and there is no doubt on those
23:56things. We cannot say that okay this
23:57standard is lagging in this thing. Okay
23:59some minor improvements happen here and
24:01there with every revision but let's
24:04overall broadly it is still the same.
24:06Those qual qualification parameters are
24:08still working very good throughout the
24:11world. But how to certify AI? AI is just
24:14a model trained on some data. Okay. And
24:17it gives some data and we also don't
24:19know how that how it gives some data.
24:20How is that model work? That thing is
24:23happening. AI
24:26this models have to be changed that the
24:28system can work have to be retrained
24:31have to be configured. What happens if
24:33you do like that? Because if there's a
24:35very simple there's the concept of men
24:38machine material method 4M if any of
24:41these changes in space we go for
24:43qualific certification
24:45men means if some person who was doing
24:47the job is qualified he changes or he
24:49happens something qualification material
24:52let's say you're using soldier A that
24:54was qualified we want to move to solder
24:56B qualification
24:58method we were using the reflow process
25:01now you want to go for advanc something
25:04again qualification
25:06this material sorry men method material
25:09machine let's say you're using a machine
25:11X now you want to use the machine Y
25:13because some issues again you go for
25:15qualification but in AI how will you do
25:17like that because all these things these
25:19four M's happen very rarely but AI
25:22changes its parameters daily if you
25:24train it right so the solution is AI
25:27assurance frameworks so we need to
25:29define a framework standard such
25:33How AI should behave is how it should be
25:35first trained. What type of data should
25:37be captured and after telling what are
25:40the minimum expected criteria that it
25:43should perform such that it becomes
25:45spacew worthy or tritew worthy
25:48we need to prove that whatever AI is
25:50doing it's correct. You have to validate
25:52it. It's not only verify we have to
25:54validate it. Like we have ISO we should
25:58have certification standards. till now
26:00they are not there in the space as of
26:02now there's no industry that has any
26:04certification or qualification standard
26:06for AI audit it's a human error
26:09definitely because human made error same
26:11error is being learned by AI
26:14operator bias let's say something
26:17appears shining to me and something
26:19appears less shiny to somebody else so
26:22for me will be good for him it will not
26:24be good the same data labeled by two
26:26different people will be given different
26:30category because we had some bias. These
26:33are the problems that are the data
26:36quality AI systems optical or even X-ray
26:39requires cameras, hardware computing,
26:41software,
26:42field engineers, testing, it requires
26:46some millions of dollars to do all these
26:48things. But whether it is justifiable if
26:51you see the long-term savings that you
26:53will get by preventing one satellite
26:56failure or a delay for everything. One
26:59space mission cost nearly half a billion
27:01dollars. So here we are spending just 1
27:04billion. So if I if I go by the
27:07calculation it's not even.1 or.5%.
27:12And if you just uh spread that
27:14calculation for let's say 10 years, 15
27:16years, 20 years becomes even less. So
27:19although it is costly there is the
27:21initial implementation cost setup cost
27:24and everything but if we see the spread
27:27the way the way it is done in the
27:29balance sheet of any company like that
27:31depreciation for any equipment similarly
27:33if you spread that cost is capital cost
27:35then we'll be able to realize how
27:36important it is and how much cost
27:38effective it is if you're a if you're
27:40able to do in the correct way so it will
27:43reduce infection time man is used it is
27:46used manufacturing defects definitely
27:48The earlier you catch the much cheaper
27:50it is. Rework cost because rework also
27:53requires some money. Mission risk most
27:55dangerous of all that mission risk will
27:58be minimized.
28:00Till now we are minimizing it through
28:02manual impact. Now there will be the
28:04system that will continuously working
28:06the chances will be less and yes because
28:08it will be reduced. The chances that
28:10that our missions will fail or not meet
28:11our objectives will be very very less.
28:13So that money will be justified. So the
28:15cost spread has to be taken into
28:17consideration.
28:19Next is human expertise. Now even after
28:22everything is there AI is very good on
28:24repetitive and consistent things. But
28:27what what happens next? Something
28:29unusual comes then there AI will either
28:31hallucinate which it should not do case
28:34it should simply say oh okay I don't
28:35know let's bring a human into picture.
28:38Now there are some failures that have
28:40never occurred before. Now it is
28:42happening. Yeah, you should say okay I
28:44don't know please come. Similarly
28:46mission specific requirements which are
28:48not trained to the digital twin now the
28:51digital twin needs that so human has to
28:53come into picture and now there will be
28:55trade-off between computing computing
28:56goes like okay the system says that
28:59computation requirement is it should
29:01work continuously for 10 hours on
29:03onboard satellite but there is a but
29:05there's a sensor that is giving an alarm
29:07that if you work for more than 6 hours
29:09that it will fail. So how how you should
29:11do because you need 10 hours there has
29:13to be some trade-off that decision
29:14cannot be taken by AI human still has to
29:18be there that's still a problem cannot
29:19it's not critical so I don't count it as
29:21a problem actually I I count it as a
29:23opportunity AI will not replace
29:26experienced people the people who have
29:27knowledge of that domain
29:30if we see it as our competitor then we
29:33will not be able to take the full as
29:35other people are selling okay it should
29:38be AI plus engineers that should
29:40such that not only AI progresses but we
29:44also progress like previously
29:46uh I will give a very simple thing
29:48previously you wanted to code the only
29:51way you had is you have to learn coding
29:53but now
29:55even you need to learn code to write a
29:57very good to make a design and develop a
29:59very good software but to get hands on
30:01it to make a general thing you don't
30:03need to learn code you can write promps
30:04and I will help you it's not it has
30:07replaced developers or coders So it has
30:10been most extensively used in industry
30:12in that way. That's why I'm giving with
30:13this example. Similarly, if you are
30:16working let's say if you're electronics
30:17engineer, compress engineer, electrical
30:19engineer, it's not that it will replace
30:20it will help you. You justly if you are
30:22taking any number of let's say 10 hours
30:25to design a circuit maybe it will help
30:27you doing in 1 hour. So the scale of
30:29work will only increase the output will
30:31only increase. So here it's not a
30:35problem. It's an opportunity for all of
30:37us to go ahead such that if we take it
30:40as a competitor then let us face it AI
30:42is going to stay and please be ready for
30:45this. This with this I'd like to end.
30:48Thank you. Over to you ma'am. Moving on
30:51into the lab. Uh we just need to uh get
30:55into a motivational background. Why
30:59emerging tech?
31:02Are we witnessing some revolution?
31:04So
31:06already
31:08we have been honored to hear from
31:10professor Jam Kumar Sar from IIT Kpur a
31:13very elaborative and insightful session
31:16on industry 4.0 and industry 5.0.
31:19So just an overview.
31:23See technology
31:26is something
31:29uh which
31:31by the virtue of application of
31:33scientific principles.
31:35It makes problems easier. It finds
31:39solution to problems
31:41and the application of science and
31:45scientific principles
31:48comes out as technological solution. We
31:51are using mobile phone without bothering
31:54to learn about embedded electronics
31:57without bothering to learn about uh the
32:01chips integrated circuits and VSSI
32:04but mobile phone is inevitable. we
32:07cannot uh um you know survive or we
32:11cannot
32:13live uh comfortably without using
32:16smartphones and similarly with internet
32:18and similarly today AI is also becoming
32:22a part of our
32:24uh you know lifestyle. It is becoming a
32:27part of our lifestyle. It is um so these
32:32uh the technological interventions the
32:34technological advancements which we are
32:36witnessing today
32:38has been uh
32:41an outcome of an evolution. This
32:43evolution is displayed over here uh uh
32:46on the slide. So in 1780s we experienced
32:52uh huge transition in industry by the
32:56advent of power of steam.
33:00the steam was produced by boiling large
33:05tons of water, tons of water uh through
33:08using the energy as a coal energy from
33:11combustion of coal and this power of a
33:14steam
33:16led to this is the scientific principle
33:19steam and [clears throat] Pascal's law
33:21and pressure thermodynamics this is
33:23science and the technology is steam
33:25engine the steam engines resulted in two
33:28major mechanization
33:31and the industry
33:33spec specifically textile mean. See in
33:35ages and in different eras we have
33:38witnessed certain specific domain of
33:41industry undergoing some revolution. For
33:43example, today we are witnessing an uh
33:46revolution in information and
33:47communication technologies. ICT is
33:50leading the world. So in 1780s textile
33:53mills and clothes making industries were
33:56on peak. So textile mills experienced
33:59mechanization
34:01by the advent of the steam power.
34:03Similarly 100 years forward we
34:07experienced
34:09a revolution in industry because of
34:12advent of electricity.
34:14So see we are using light bulbs we are
34:17using electrical appliances. Everyone,
34:19every literate or illiterate person has
34:21to uh we are all familiar about uh
34:26electricity with electrical appliances
34:29and know but we only a few of us are
34:32well versed with
34:34the electrostatic principles the current
34:37electricity so all the science which we
34:40deal with. So this is known as
34:42revolution that
34:44the technological solution becomes a
34:46part of our life. We um cannot uh avoid
34:51the using those technological uh
34:54applications without bothering into
34:56entering the details of the scientific
34:59principles which back which is at the
35:01back end. In 1970s
35:04see electronics and computers were
35:08discovered invented and
35:11large applications of electronics
35:13know semiconductor industry was
35:16emerging. So the energy source was
35:20electricity and the transformation was
35:22directly reflected in automation. So we
35:26had CNC machines, industrial robots and
35:31these uh revolutions came into picture.
35:35Modern manu the traditional
35:37manufacturing was converted into
35:39transformed into semi-automated
35:41manufacturing by the advent of
35:42industrial robots and CNC machines. So
35:451970s experience revolution and 2011
35:49onwards 2011 onwards we are witnessing a
35:53humongous uh application of a bunch of
35:57emerging technologies. These
35:59technologies can be named as sensor
36:02technology, communication uh
36:05technologies like 4G and 5G, uh internet
36:08of things, data revolution, we can say
36:11big data, data analytics, artificial
36:14intelligence, robotics,
36:16so uh green energy, cloud computing,
36:20blockchains, cyber physical systems,
36:24Wi-Fi, broadbands. So these set of
36:27technologies
36:29in many applications
36:32used collectively
36:34there are several applications we can
36:37think of where in Ive named seven eight
36:39technologies or 10 technologies we are
36:42using five to six technologies at the
36:44same time. For example if you go in a
36:47modern diagnostic lab uh which is a
36:52facility of some hospitals. So if you go
36:55for a modern diagnostic lab for a test
36:58like MRI and CT scan, we can see cloud
37:01computing. We can see uh robotics.
37:05Okay. Robotic endectors are used to uh
37:08carry the plates and capture the images
37:11from the patient's uh
37:14uh organs or parts which have to be
37:16scanned. So
37:19robotics is there, internet is there,
37:21cloud computing is there, data has been
37:23data received in the test, data captured
37:26in in those test is saved on some cloud
37:30um from which from where uh that
37:32platform uh doctors or different
37:35stakeholders of the medical
37:37infrastructure may retrieve and process
37:39those datas. We are see many hospitals
37:42like as even uh the medical colleges,
37:46state medical colleges, Tata memorial
37:47research center we are all working on e-
37:51medicine electronic medical reports are
37:53there. So doctor on several clicks can
37:56retrieve all the files all the details
37:58of the patients which have been
38:00previously
38:02uh evaluated. So this is all a result of
38:06uh collection of these collective
38:07technologies.
38:09So this revolution is known as fourth
38:11industrial revolution or industry 4.0.
38:15One of major applications an industry
38:18which has transformed a lot through
38:21fourth industrial revolution is
38:23manufacturing sector and one of these
38:26technologies is additive manufacturing
38:29or 3D printing. This 3D printing so this
38:32is known as people planet. We are all
38:36very much aware about the sustainability
38:39approaches
38:41uh considering climate, environment,
38:44conservation practices, use of natural
38:46resources, judicious rule, use of
38:48natural resources. This is all the
38:50planet vertical and profit means
38:53uh uh adaptation and implementation of
38:56these uh technologies should
39:00>> know what the part will be created what
39:02shall be the shape what would be the
39:04features
39:06uh it has all to be decided in computer
39:10by using software technologies known as
39:12computerated design CAD. So CAD is a
39:17platform where there are different tools
39:19like Solid Works, Fusion 360, Katya
39:23wherein uh these 3D models are designed
39:26and based upon those 3D models uh a
39:30digital file is created. Digital file
39:32carries the information on the path and
39:35how uh the material has to be deposited
39:38in what pattern.
39:42So here is a uh very simple
39:44representation of traditional
39:45manufacturing versus active
39:46manufacturing. Traditional manufacturing
39:48previously we were using metal cutting
39:51and even with technology we used CNC
39:54machines that is also doing metal
39:57cutting operations. So see we have a
40:00block and there is a finished part. So
40:02uh we have to create certain holes and
40:04remove material from the block. Then
40:06only this part can be created. This will
40:08result into waste generation. Why? If
40:12the same part is to be produced by
40:13additive manufacturing,
40:15the extruder will will deposit materials
40:18only where it is desired. If there is a
40:20hole, material will not be deposited in
40:22that hole. We can optimize
40:25this path of design, this path of this
40:28this uh entire formula of production, we
40:30can optimize and we can generate in
40:32computer softwares. While slicing the
40:35files, while slicing the models, we can
40:37um optimize [clears throat] all these
40:39parameters so that zero or very less
40:43waste is generated because there is no
40:46scope of waste generation because the
40:48material is deposited only where uh it
40:51is desired as per the design.
40:55Several examples and applications
40:57turbine blades, aerospace brackets,
41:01dental aligners, medical [snorts]
41:02implants. we'll see certain these uh
41:05prototypes printed parts in our lab. Uh
41:08meanwhile I'll request our uh
41:12coordinators and Dr. Sanjiv Sharmas to
41:15kindly in for dental aligners. We need
41:18to take X-ray of the dental you know jaw
41:21align and dental sockets
41:24um you know features of the jaw and the
41:28gums. So if there is any deviation
41:32that may result into mismatch and if the
41:36patient has to compromise with a
41:38mismatched wrongly developed dental
41:41aligners it may result into several
41:45um long-term diseases or chronic
41:48diseases problems because of you know
41:51this deviation because of unhygienic
41:57conditions in the mouth which will
41:58develop over time it is also a result of
42:01oral cancers.
42:03So precision here again is very
42:06important and we cannot compromise with
42:08the precise designs. In this case in
42:11these examples
42:133D printing is the only solution. Why?
42:16Because we can by reverse engineering we
42:20can create the exact CAD models of the
42:23need. We can create the designs the
42:27computer design models and based upon
42:30computer design models our printer will
42:32deposit material exactly following the
42:35design. If there is any error in the
42:37design that will be replicated in the
42:38product. If there is no error in the
42:40design. So we can design things we can
42:43simulate. We can model we can test even
42:46the before manufacturing the products.
42:48This is the beauty of this technology.
42:51Before manufacturing the product,
42:52product does not exist and we have
42:54tested,
42:56we have tested the design, we have
42:58tested the prototypes, we have by
43:00techniques like simulation. Think about
43:02the stents, the medical implants which
43:06are placed by massive surgeries and know
43:09which are placed artificial parts and
43:11implants which are placed in the body
43:14like [clears throat] graft wall and
43:16cardiac surgeries. How

About this transcript

This page contains the full transcript of Day 6 | AIO Automated Optical Inspiration | X-Ray Defect Detection | Manufacturing Optimization by STR2E, generated from the public captions YouTube serves with the video. The transcript has 6,853 words across 1,053 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.