Anyone else using the assembled PCB neural network Optical Inspection service?

[Calder]

It has been three weeks since I got an user account to use the neural network powered AOI service (from a company which posted here, in the forum. Not going to name it)
Used with more than 2500+ assembled PCBs

The thing is... it is working very, very well for us. But, where is the catch? Every inspection is cheap, the results: accurate and fast. Are there any other companies offering the same? Why I can only find just one? Pioneers?

We already have a KohYoung AOI system here but the solution offered by this new company seems to be the future (or present) for small/ medium assembly PCB companies. We have saved a LOT OF TIME (+money, of course) just using our camera (Panasonic lumix gx880) and good lighting to inspect our boards. The solution is fast, very affordable and...just works. Check images attached: they shows the defect/S detected (in red) on some of our PCBAs.

Are we missing another popular inspection service company as the one mentioned above? Does anyone know any alternative?
This sounds so good that makes me doubt about everything/it.
Thanks community

C.

[fcb]

Sounds like an interesting service - any reason why you won't/can't name them?

[olkipukki]

Are we missing another popular inspection service company as the one mentioned above? Does anyone know any alternative?
This sounds so good that makes me doubt about everything/it.

If everything works fine, why do you need 'alternative'? 

[mikeselectricstuff]

here is the original thread :
https://www.eevblog.com/forum/eda/neural-network-powered-assembled-pcb-inspection-tool-(aoi)/msg3331110/#msg3331110

[Calder]

It is www.Agnospcb.com
You can check the integrity of your PCBAs just from a photography.

[Calder]

Are we missing another popular inspection service company as the one mentioned above? Does anyone know any alternative?
This sounds so good that makes me doubt about everything/it.

If everything works fine, why do you need 'alternative'? 

It was strange to find a solution like this by "itself": I was expecting big companies to offer a similar solutions (using neural networks to find faults), but I guess Agnospcb has been the one capable to do it.

[Calder]

here is the original thread :
https://www.eevblog.com/forum/eda/neural-network-powered-assembled-pcb-inspection-tool-(aoi)/msg3331110/#msg3331110

Yes, this is the one. Simon, from Agnospcb created a user account for me with free "inspection credits". It took spending about 10 of them before starting to get everything together. Now, we are inspecting a board every 5 seconds. Just with one click. The software they have released recently (here https://www.agnospcb.com/inspection-software-user-manual) can be used to check the boards with "comparison" magnification areas. Extremely useful.
This video sums everything up: https://video.wixstatic.com/video/87e1b6_3677138671c24f88b9830a340a942b00/1080p/mp4/file.mp4

[48X24X48X]

Would it be able to detect very mild tombstone (joints that look like reflow from the top but only from side view you can see the 1 side of the pad is slightly lifted)?

[Calder]

Would it be able to detect very mild tombstone (joints that look like reflow from the top but only from side view you can see the 1 side of the pad is slightly lifted)?

I would suggest you to try it by yourself. I do not have any example of mild tombstone here, but the ones we have had, have been detected. The neural network results are "conservative", so, In the case of doubt, we have noticed that this tool, tags the element as faulty.

[BravoV]

Would it be able to detect very mild tombstone (joints that look like reflow from the top but only from side view you can see the 1 side of the pad is slightly lifted)?

I guess, if a human eyes, even looking at it thoroughly and still can not tell the different or not sure from that view angle, then how to expect it to perform better ?

[olkipukki]

It was strange to find a solution like this by "itself": I was expecting big companies to offer a similar solutions (using neural networks to find faults), but I guess Agnospcb has been the one capable to do it.

Are you in the same category?

There're a plenty solutions I saw at last Productronica, but eye-watering prices does not fit into the same category that Agnospcb is targeting....

[olkipukki]

Would it be able to detect very mild tombstone (joints that look like reflow from the top but only from side view you can see the 1 side of the pad is slightly lifted)?

I guess, if a human eyes, even looking at it thoroughly and still can not tell the different or not sure from that view angle, then how to expect it to perform better ?

Well, everything is possible 
as example (taking into account 5 years old!) 

[tszaboo]

here is the original thread :
https://www.eevblog.com/forum/eda/neural-network-powered-assembled-pcb-inspection-tool-(aoi)/msg3331110/#msg3331110

Yes, this is the one. Simon, from Agnospcb created a user account for me with free "inspection credits". It took spending about 10 of them before starting to get everything together. Now, we are inspecting a board every 5 seconds. Just with one click. The software they have released recently (here https://www.agnospcb.com/inspection-software-user-manual) can be used to check the boards with "comparison" magnification areas. Extremely useful.
This video sums everything up: https://video.wixstatic.com/video/87e1b6_3677138671c24f88b9830a340a942b00/1080p/mp4/file.mp4

Two questions:
1) Are you a shill?
2) Who is the target of this product? I expect full AOI from the assembly house, with high res cameras and possibly multiple angles, excellent lighting conditions. And X-Ray if necessary.

[Calder]

Would it be able to detect very mild tombstone (joints that look like reflow from the top but only from side view you can see the 1 side of the pad is slightly lifted)?

I guess, if a human eyes, even looking at it thoroughly and still can not tell the different or not sure from that view angle, then how to expect it to perform better ?

Well, everything is possible 
as example (taking into account 5 years old!) 

Our AOI system (the KohYoung) can do similar things with the cameras, but at the end of the road, it is not practical for low/ medium batches. You need to upload the gerber files, set tolerances...and takes about a minute to check a single PCBA. I have contacted several AOIs companies for a similar solutions. Lets see if they come up with something like the agnospcb tool. The thing is: I do not know how the neural network really works. Among different PCBAs there are always (obviously) differences like slightly rotated components, different labels, element’s geometry, but they are not marked as "faults". It is like if I was there, checking the boards by myself. Curious how the AI works...

[Calder]

here is the original thread :
https://www.eevblog.com/forum/eda/neural-network-powered-assembled-pcb-inspection-tool-(aoi)/msg3331110/#msg3331110

Yes, this is the one. Simon, from Agnospcb created a user account for me with free "inspection credits". It took spending about 10 of them before starting to get everything together. Now, we are inspecting a board every 5 seconds. Just with one click. The software they have released recently (here https://www.agnospcb.com/inspection-software-user-manual) can be used to check the boards with "comparison" magnification areas. Extremely useful.
This video sums everything up: https://video.wixstatic.com/video/87e1b6_3677138671c24f88b9830a340a942b00/1080p/mp4/file.mp4

Two questions:
1) Are you a shill?
2) Who is the target of this product? I expect full AOI from the assembly house, with high res cameras and possibly multiple angles, excellent lighting conditions. And X-Ray if necessary.

I am not related to the company. That is one of the reasons I did not want to name it here... 
I do not really know their target, but for our company, it is a solution to a problem we had: Low/ medium batches fast supervision. For large batches, we are still using the expensive AOI systems (for now)

[olkipukki]

I have contacted several AOIs companies for a similar solutions. Lets see if they come up with something like the agnospcb tool. The thing is: I do not know how the neural network really works. 

In-house or similar cloud services?
If first - what is your budget? 

Curious how the AI works...

Well, not hard to find these days, as example, that's basic explanation

https://www.mvtec.com/technologies/deep-learning/how-to-start

[olkipukki]

You need to upload the gerber files, set tolerances...and takes about a minute to check a single PCBA.

No wonder if they're using own software solutions (rather than rely on some software-specialized vendors) , always will be behind a current trend - develop machines and software that's a LOT of work required...

[SMTech]

OK if we ignore the "neural" bit which presumably helps it learn as it goes on more than anything else. What this offers is a low cost entry into some level of inspection for small companies who could be OEM or CEM elements of it are unique and others are probably not. I'm sorry @NANDblog but there PLENTY of small CEM companies who don't have AOI and even more who don't have X-ray, those are expensive tools to acquire and often don't suit the batch sizes many companies operate in. AOI needs extensive data, training and often golden samples as well to build up their dataset. Interestingly its only very recent AOI machines that make any mention of machine learning/cloud etc being used to improve the function of the product.

I've looked at the product in question briefly, but I haven't had the time to extricate myself from assembly to play with this in QA some more, its clearly going to have the same issues as any top down inspection method that doesn't use angled cameras or polarised/angled light and lasers to get a 2.5/3D image and carry out measurements. Of course there are plenty of things you can spot without those tools and that saves buckets of time over trying to do it manually which in many cases is basically impossible. I'm totally sold on it being a useful tool, just not the only tool you might need, all depends on your product, client, volume etc.

Its rival products, given its pure 2D nature to my mind are other FAI/"light AOI" machines, normally based around an adapted flatbed scanner, these are not that expensive @ around £10k and most of the people who make one of those will probably sort you out a payment plan to help. They are slower than true AOI machines from what I have seen taking about 90seconds to scan and analyse a board, much slower than Agnos. The Koh Young time quoted above seems quite high compared to AOI demos I have seen but they are very full featured top end machines and maybe that is an older example or one with a complex board and lots of checks... (I saw an Aleader at a show inspect a 1600 component board in 16 seconds). The other rival I have seen on here somewhere is a German piece of software that costs around €5k that works with an off the shelf scanner IIRC.

The Agnos pricing of $5/50 scan seems pretty reasonable, the hardware setup they will sort you also seems reasonable, the two higher tier high capacity plans https://www.agnospcb.com/pricing @ $3k / month?? I'm less sure, it would be intriguing to put them head to head against a low end AOI/FAI machine to see how they perform relative to each other. The inspection area is notably small even with the setup they will sell you using a 24mp camera, you obviously get a much higher effective resolution using  a roving camera or scanner. I think it is surprising this is the first commercial offering of its type given the tech behind it has been out there and used extensively elsewhere for a good 5 years or so. I would bet there are multiple in house implementations in the wild, I know a company we work with was working on something in this area and I don't think they were planning on sharing it.

[mairo]

I wonder if AgnosPCB are based/using Amazon Rekognition. I listen recently that one of the target markets for Amazon Rekognition will be infarct manufacturing and assembly floors and the presenter specifically  said that one of the larger targets would be the electronics assembly market.


Few issues I see with images being process in the cloud particularly by a third party provider (an there will be probably few layers of providers), if lets say the account is being hacked:
- being able to easily obtain high res images of a PCB - this is not that big of a deal, but could be in some instances;
- being able to estimate the manufacturing volumes of a specific product

[jpedroc]

I wonder if AgnosPCB are based/using Amazon Rekognition. I listen recently that one of the target markets for Amazon Rekognition will be infarct manufacturing and assembly floors and the presenter specifically  said that one of the larger targets would be the electronics assembly market.


Few issues I see with images being process in the cloud particularly by a third party provider (and there will be probably few layers of providers), if lets say the account is being hacked:
- being able to easily obtain high res images of a PCB - this is not that big of a deal, but could be in some instances;
- being able to estimate the manufacturing volumes of a specific product

Hi,
 I just found this thread. My name is John, technician in Agnospcb.com ;-)
 We do not use Amazon Rekognition, that kind of service has been created for... lets say, "standard procedures" like tagging persons on a photo or a dog in a video.

Agnospcb has been created from scratch, tailored to understand what a PBCA is. Let me explain: the neural network knows what an electronic component is (among other entities like labels, shadows, solder paste, holes, bridges, pads...). It does recognize the element on the PCB surface and can check if it has been placed incorrectly, rotated, shifted, if there are short-circuits between pads, pins...etc.
 
We have used more than 150K "defect patterns" to teach it what a fault really is (and how to indicate where that fault is).

 It does not matter the minor differences you can find among different PCBAs (PCB color / manufacture variations, silk layer, tiny soldering variations...), the neural network will only spot the real faults. And (this is important): you do not need to re-program the FAI/AOI system every time you need to check the integrity of a batch.

Check the image attached. It shows: a good PCBA (left), the inspected board and the faults detected. On the right image, inside orange boxes, the entities that have been discarded as defects.: connector serial number, silk layer "label" differences, minor manufacturing discrepancies....

NOTE: we remove all the images from our server daily. For security reasons and because we can not store the amount of data our clients are continuously uploading 


If anyone wants to give it a try, just let me know. I will create a user account with free inspection credits for you.
 We have the webapp tool: https://app.agnospcb.com/ So you do not even need to install anything to get a glimpse of the Agnospcb potential, just drag and drop photos of your PCBA to the webpage.

[ThomasRR]

I only have good words about agnospcb's service. If you try it, avoid their Webapp tool: it is clumsy and intended (I guess) to get a glimpse of the inspection service capabilities. Use their inspection tool software instead from the very beginning. https://www.agnospcb.com/inspection-software-user-manual
We have been playing a lot with this service the last month using it to detect defects just after the "pick and place" process (before the reflow oven): it is really capable of detecting misplaced/rotated/missing elements.
Screen-capture from their inspection tool below.
We are not using it for large batches as we already have regular AOI equipment here (3 units), but it is extremely useful for low/medium runs, checking the "repaired" PCBAs and now (so far) for pre-oven units.