Machine Vision

Hello all

Beginning new R&D. Need to learn all I can about machine vision. So far Seen just about every video on youtube, every webinar, and tutorial over the last four days.

Problem: recognize part, use engineering files provided by customer to identify, exercise quality control and clean.

Items are small, inconsistently space, of irregular design, and a low probability that any two would ever be alike. On the order of 53,000 pcs in a 22 day month.

Goal: to recognize, identify, qualify (does the item possess the dimensions it was designed to ) , and clean.

Only thing I'm real Curious about is the "Universal Gripper" Developed in 2010 at Cornell University ( Ground Coffee, Party ball, and a vacuum ). Neat

https://www.youtube.com/watch?v=0d4f8fEysf8

In all the years of MRO, I have worked with all manner of automated systems, just never robotics, or Machine vision. (shamefully it was a minor in college)

In the 32 hours invested in research, I have found that Keyence, Cognex, and Sick are not capable of sorting the irregular, Most are ok with inconsistent and small.

Next stop will be the robotic manufactures for their developed systems. ABB Yumi
and Yaskawa SDA 5 and 10 series, and one offering from Kuka.

Your Input on any of these systems would be greatly appreciated. Also any systems, ideas, theories I might be overlooking would be greatly appreciated.

Thank you all for your valuable time and attention to this.

Check out Codian Robotics NA for the robots too! They have a HUGE selection of delta style robots, unlike most of the other companies out there.

http://www.codian-robotics.com/en/

The company is out of the Netherlands, but they have a state-side group owned and operated in Michigan.

https://www.facebook.com/Codian-Robotics-of-the-Americas-LLC-246338062073567/

What I am reading is this: the vision system needs to first identify a part from a "library" of known parts, and if a match is found, it will be inspected against customer-supplied specifications in 2(?) dimensions. Furthermore, the parts will be moving down some type of conveyor with unknown orientation, spacing, and potential overlap. It is inferred that the parts are small (millimeters across?), and within a reasonable size range. Inspection rate is on the order of 2 pieces per minute, assuming a single line and vision system for a 24-hour operation.

I'm mostly not sure about the "clean" requirement. Is that a physical cleaning?

Were you actually told that there is a engineered vision system out there which could recognize, identify and qualify irregular items? (I am not sure what 'clean' means within the above context......Remove from line?...Washed?)

Of course there are, at least in the manufacturers' advertisement propaganda and the 'infected' (by said propaganda) business management type who doesn't have any engineering clues whatsoever but obviously believes in 'artificial intelligence'.

I am truly sorry to disappoint the believers of above paragraph scenario but the reality is a far cry from the claimed functionality.

Any 'automation vision' system is based on a given, distinct visual pattern; the vision system will compare what it 'sees' within its limited range of visibility to any of the stored patterns and then make a 'decision' based on user defined 'qualifiers'...That's it...The more patterns are to be monitored the longer will take the vision system to arrive to a 'decision' and a lesser accurate 'decision' that will be.

Similarly, the more and 'complicated' patterns are to be monitored the less of the production line speed will be required.

Moreover, ambient lighting consistency is an absolute must; the vision system 'training' of the to be monitored patterns lighting will need to be carried to production time as 100% identical down to no windows at all and same manufacturer, wattage, light 'color' and inspection area coverage.

I could go on and on for a while......Vision 'works' when a small amount of qualifiers are being used...It is not 100% accurate(I'd say some 92-93%) yet but moving in the right direction...However, again, the vision system needs to know exactly what to look for...There is only 'white or black' (or a defined % of 'grey') in industrial automation vision...

Mispeld said:

What I am reading is this: the vision system needs to first identify a part from a "library" of known parts, and if a match is found, it will be inspected against customer-supplied specifications in 2(?) dimensions. Furthermore, the parts will be moving down some type of conveyor with unknown orientation, spacing, and potential overlap. It is inferred that the parts are small (millimeters across?), and within a reasonable size range. Inspection rate is on the order of 2 pieces per minute, assuming a single line and vision system for a 24-hour operation.

I'm mostly not sure about the "clean" requirement. Is that a physical cleaning?

Click to expand...

You have the relative gist of it, Cleaning would be the physical removal of the leftover material either my compressed air or a liquid rinse of sorts.

dmargineau said:

Were you actually told that there is a engineered vision system out there which could recognize, identify and qualify irregular items? (I am not sure what 'clean' means within the above context......Remove from line?...Washed?)

Of course there are, at least in the manufacturers' advertisement propaganda and the 'infected' (by said propaganda) business management type who doesn't have any engineering clues whatsoever but obviously believes in 'artificial intelligence'.

I am truly sorry to disappoint the believers of above paragraph scenario but the reality is a far cry from the claimed functionality.

Any 'automation vision' system is based on a given, distinct visual pattern; the vision system will compare what it 'sees' within its limited range of visibility to any of the stored patterns and then make a 'decision' based on user defined 'qualifiers'...That's it...The more patterns are to be monitored the longer will take the vision system to arrive to a 'decision' and a lesser accurate 'decision' that will be.

Similarly, the more and 'complicated' patterns are to be monitored the less of the production line speed will be required.

Moreover, ambient lighting consistency is an absolute must; the vision system 'training' of the to be monitored patterns lighting will need to be carried to production time as 100% identical down to no windows at all and same manufacturer, wattage, light 'color' and inspection area coverage.

I could go on and on for a while......Vision 'works' when a small amount of qualifiers are being used...It is not 100% accurate(I'd say some 92-93%) yet but moving in the right direction...However, again, the vision system needs to know exactly what to look for...There is only 'white or black' (or a defined % of 'grey') in industrial automation vision...

Click to expand...

Thank you for your time and information. No sir, I was not told there was a system that could do exactly this. Most of the engineered system that are available work on a consistant pattern recognition. Even the ones that can bin pick scattered pieces work on a consistant pattern recognition then apply orientation optics ( mostly IR).

The theory of the engineered library is a batch process. The twenty/thirty pieces would be process then disregarded for the next batch. So I believe the scan time would remain consistent, thought I do not know what the duration is yet.

I was working on a process definition by where a two camera system would work like a rough tune and fine tune. The robotic system would identify individual pieces probably on a patterned surface like gradient square or small circle pattern for location recognition, any disruption in the pattern would indicate location. Not accounting for overlap or touching parts ( probably over come with a small vibratory ). Once that hurdle is overcome the part is moved to another location for qualification with another vision system that would judge against an identifiable pattern such as CAD or other three "D" engineering engine. ( This is were Keyence and Cognex got off )

Using a combination of IR ( Fine lines ) and 3D vision ( over all) try and match patterns. The trick here is to either move ( rotate) the part ( Disregarding the end effector) or move the camera, in which case the sixth side would be an assumption. I dislike assumptions. But your 92-93% is more than likely accurate.

Ni has a vision system that is customizable but pricy for a one off, and it will work with all manufactured visioning equipment. I am discussing with the engineers the possibility irregular identification in batch processing. ( Giving this to their R&D is like giving a kid a new toy, I think they're bored).

I was just wondering IF any other businesses where entertaining this type of idea.

Thank you Both for your input and feedback.

Sounds like a somewhat challenging project. In reading the additional description, it appears you have the opportunity to influence the physical design of this inspection/cleaning operation. My advice is to leverage the system design as much as possible to make the automation task less complicated and expensive. There are many opportunities, such as what was mentioned (separating the parts), as well as more basic aspects like cleaning before imaging and inspection.

Mispeld said:

Sounds like a somewhat challenging project. In reading the additional description, it appears you have the opportunity to influence the physical design of this inspection/cleaning operation. My advice is to leverage the system design as much as possible to make the automation task less complicated and expensive. There are many opportunities, such as what was mentioned (separating the parts), as well as more basic aspects like cleaning before imaging and inspection.

Click to expand...

Sage advice.

Thank you Mispeld