印刷电路板市场中的机器视觉

HUNNISH 注：这篇文章对机器视觉在电路板检测上面的应用给出了全面的阐述。本文对有志于从事这方面领域的人来说，不失为一篇很好的参考文章。

===================================================

Machine Vision in the Assembled Printed Circuit Board Market
by Nello Zuech, Vision Systems International

The part 1

The Past

Much has happened since the first machine vision systems were introduced to inspect populated printed circuit boards in the early 1980s. Today these systems are no longer called machine vision but rather “automated optical inspection” or AOI systems. Originally the term AOI was associated with machine vision systems addressing bareboard inspection applications. Today the term is widely used for all automated visual inspection systems used in electronic manufacturing. It is likely the term AOI has been favored because the performance of the early machine vision systems addressing assembled board applications was only marginally acceptable. Regardless, the underlying technology is still machine vision.

Since the early 1980s the fundamental manufacturing process for electronic assembly has changed. At that time the principal approach involved lead-through-hole mounting. The machine vision application included top-side presence and orientation but also bottom-side lead presence and clinch angle verification. Today the principal manufacturing approach involves surface mount technology. Now the application includes both presence and placement validation. Significantly, there are still many boards that combine the two assembly approaches. Board designs that use component direct attach methods are gaining in popularity.

All these assembly approaches yield a unique set of appearance variables. Variables are the “gotchas” of machine vision systems. Today’s machine vision systems are far better at handling these variables than ever before and they are only getting better with the advances in the underlying compute power which have made it possible to handle the resolution required at the throughputs required, all while reducing the false calls and false accepts and reducing training time for new board designs. The technology is finally living up to its much-heralded promises of the past.

Adoption Drivers

Driving demand for these AOI systems are factors such as ever-smaller components, ever finer pitch density, ever more components per square inch, higher production rates – all of which make it increasingly more difficult for people to inspect. Bob Ries of CyberOptics specifically cited “Primary drivers are PCB complexity and component size.” And “The need to improve first pass ICT yields.”

The subjectivity of people leads to even greater quality judgment uncertainty as boards become more complex. Similarly these factors make it ever more difficult for conventional in-circuit testing. In this country using people to perform these inspections requires companies provide them with ergonomically correct inspection workstations, which are expensive. Increasingly electronic manufacturing service companies or contract manufacturers are doing board assembly. This is a business that is very competitive price-wise. Consequently, there are economic drivers to substitute capital for people. In the case of AOI systems additional benefits include improved quality, improved yield, reduction of rework, improved customer satisfaction – all results that go to the bottom-line.

The cost of failure in electronic manufacturing is appreciably greater with each value-adding step, with the cost of the failure of a product reaching the consumer the greatest cost. The ultimate price of many consumer electronic products is relatively low, to the point where many are not worth repairing. In the case of cellular phones many are actually given away as a promotion to provide the phone service itself. However, if the phone does not work one has a very unhappy customer and one likely not to use the service promoted or re-enlist in the service. Hence, there are significant economic drivers among both contract manufacturers and OEM manufacturers to avoid failure.

Chuck Gamble of Leica Microsystems made the following observation: “AOI is usually associated with three things, PAIN, they usually have so many problems producing a product, that they turn to AOI in a move of desperation. SALES, their customer tells them they won’t get an order unless they have AOI in place for their products. And KNOWLEDGE, the PCB assembler has learned through experience that they can ultimately produce a superior product, have faster start-up times, raise yields and lower overall production costs with the proper implementation of AOI.”

The North American electronic industry is apparently the earliest champion for in-line AOI systems. Bob Ries attributes this to “the relative complexity and high cost of PCBs assembled here.” Chuck Gamble suggests that <place w:st="on">Europe</place> is following closely but that the “Pacific Rim tends to be strictly an inspection (good or bad) market.”

Where to Use AOI?

The biggest challenge is not whether to use AOI systems, but rather where to use them. To avoid shipping a reject product suggests the use of AOI systems at the end of the production line. The challenge is that at the end of a production line AOI systems are not likely to provide a comprehensive solution. Rather they have to be considered a complement to X-Ray-based machine vision systems, which have the capacity to assure the quality of a solder joint as well as verify the joint’s presence. X-Ray-based systems used in the electronic industry will be the subject of another article.

Using AOI systems to sort rejects at the end of the production line does not avoid rework, which can be reduced by using AOI systems to monitor the results of each of the value adding steps along a board assembly line. In other words, a line designed with prevention in mind might include the following operations:
<!--[if !supportLineBreakNewLine]-->
<!--[endif]-->

SIDE 1

Screen print solder paste or adhesive dispensing solder dots

AOI

After chip placement

AOI

After fine-pitch placement

AOI

After reflow

AOI

Through-hole assembly

SIDE 2

Screen print solder paste or adhesive dispensing solder dots

AOI

After chip placement

AOI

After fine-pitch placement

AOI

After reflow

AOI

Wave solder

AOI

In other words, based on the philosophy of process monitoring or prevention of rejects, up to nine AOI systems could be required. Given an average price of as little as $100K suggests that it would cost at least $<chmetcnv unitname="m" sourcevalue="1" hasspace="False" negative="False" numbertype="1" tcsc="0" w:st="on">1M</chmetcnv> to install AOI systems as part of a comprehensive process control strategy and even more if one were to include X-Ray-based systems as part of the strategy. This is a highly unlikely scenario! Using an AOI system after fine-pitch placement would detect conditions associated with chip placement machines, however, it would not provide immediate feedback as to the performance of the chip shooter. The key is to provide data immediately after a process step so an indication of pending failure can be flagged and corrective action take place before failure is actually experienced.

If the boards were rerouted through the same component assembly machines for side 2, only five systems would be required and if no wave soldering took place, only four systems would be required. While more manageable these would still be expensive scenarios. It is also noted that today even for lead-through-hole components solder paste can be applied in the holes in the same manner as for surface mount devices. This practice eliminates the need for wave solder.

Where is the Most Value?

The challenge is where to get the most “bang-for-the-buck?” One factor will be the nature of the assembly line itself. Is it one geared for high volume/low model mix, medium volume/medium model mix or low volume/high model mix production? In the high volume/low model mix scenario one stands to gain the most from a comprehensive strategy to avoid adding value. Nevertheless, four to nine systems on a production line will be unlikely.

Since assembly starts with solder paste application and any reject condition at this stage will manifest itself later on it makes sense to perform a comprehensive inspection after application so reject boards can be cleaned and reused. At this stage the system should inspect for solder registration with respect to the pad, sufficient/insufficient solder on the pad and solder bridging between pads. One issue at this stage whether 2-D or 3-D measurements are required. As smaller components and components with finer pitch densities are adopted, volume measurements will become more critical.

Some, however, suggest that inspection before reflow not only finds placement machine errors but also can find defects resulting from solder paste deposition. One issue at this stage is whether the system should be quantitative or qualitative or both. Systems that offer qualitative solutions can tell if components are present and correct, oriented properly and relatively aligned or positioned properly. Quantitative systems can actually measure component offsets and can be used to monitor the placement machine’s performance and flag conditions trending out of spec so corrective action can take place before actual defective conditions occur. Conditions detected before reflow can be corrected more easily than after reflow, where rework requires unsoldering which can further damage neighboring circuitry and result in ultimately throwing out the board.

Using an AOI system after reflow will likely detect most of the errors caused by solder paste deposition, placement and the reflow process itself. However, at this stage one can sort rejects and provide immediate feedback for the reflow process but not for the screen printing or chip placement processes. It is also more difficult to rework a reject condition after reflow than it would be after the previous steps in the assembly process. In other words, a good deal of the value of AOI is lost!

Understanding Requirements

If a qualitative approach is satisfactory, it is important to understand the principles of detection associated with machine vision systems. Most today would agree that in order to detect an artifact in an image that artifact should cover a minimum of 3 X 3 pixels. Sub-pixel detection schemes do not apply when simple presence is the application. If one then suggests that the smallest component on the board (say a component that is <chmetcnv unitname="”" sourcevalue="0.1" hasspace="False" negative="False" numbertype="1" tcsc="0" w:st="on">0.1”</chmetcnv> X <chmetcnv unitname="”" sourcevalue="0.1" hasspace="False" negative="False" numbertype="1" tcsc="0" w:st="on">0.1”</chmetcnv>) will be covered by 3 X 3 pixels, each pixel will be <chmetcnv unitname="”" sourcevalue="0.03" hasspace="False" negative="False" numbertype="1" tcsc="0" w:st="on">0.03”</chmetcnv> X <chmetcnv unitname="”" sourcevalue="0.03" hasspace="False" negative="False" numbertype="1" tcsc="0" w:st="on">0.03”</chmetcnv>. If the typical system based on an area array camera has nominally 500 X 500 pixels, the maximum field-of-view of that camera should be <chmetcnv unitname="”" sourcevalue="15" hasspace="False" negative="False" numbertype="1" tcsc="0" w:st="on">15”</chmetcnv> x <chmetcnv unitname="”" sourcevalue="15" hasspace="False" negative="False" numbertype="1" tcsc="0" w:st="on">15”</chmetcnv>. If, however, one is trying to detect bridging between neighboring pads or lines that are only 2 mils a part, then the 3 X 3 pixels have to cover an area of <chmetcnv unitname="”" sourcevalue="0.002" hasspace="False" negative="False" numbertype="1" tcsc="0" w:st="on">0.002”</chmetcnv> X <chmetcnv unitname="”" sourcevalue="0.002" hasspace="False" negative="False" numbertype="1" tcsc="0" w:st="on">0.002”</chmetcnv>. This then suggests the maximum field-of-view should be <chmetcnv unitname="”" sourcevalue="1" hasspace="False" negative="False" numbertype="1" tcsc="0" w:st="on">1”</chmetcnv> x <chmetcnv unitname="”" sourcevalue="1" hasspace="False" negative="False" numbertype="1" tcsc="0" w:st="on">1”</chmetcnv>. As board densities have increased the AOI industry has responded by offering systems with cameras that have nominal resolutions of 1000 X 1000 or even greater in some instances.

If a quantitative approach is required, then the rules of metrology apply. That is that the measurement instrument should have a repeatability that is 10 times the total tolerance band and ideally an accuracy that is 20 times the tolerance band. In gauging applications, since measurements are typically made to boundaries or objects and boundaries are defined by their edges, the concept of subpixel processing does apply. What all this suggests, therefore, is if one is measuring features with tolerances on the order of +/- <chmetcnv unitname="”" sourcevalue="0.001" hasspace="False" negative="False" numbertype="1" tcsc="0" w:st="on">0.001”</chmetcnv> so the tolerance band is <chmetcnv unitname="”" sourcevalue="0.002" hasspace="False" negative="False" numbertype="1" tcsc="0" w:st="on">0.002”</chmetcnv>, the repeatability of the measurement instrument or AOI system in this case should be <chmetcnv unitname="”" sourcevalue="0.0002" hasspace="False" negative="False" numbertype="1" tcsc="0" w:st="on">0.0002”</chmetcnv>. This means the AOI system must be able to discriminate to <chmetcnv unitname="”" sourcevalue="0.0001" hasspace="False" negative="False" numbertype="1" tcsc="0" w:st="on">0.0001”</chmetcnv> increments, as there is inevitably a +/- one discrimination unit of uncertainty in the measurement.

All this suggests that the AOI system should have the ability to find an edge such that the subpixel is <chmetcnv unitname="”" sourcevalue="0.0001" hasspace="False" negative="False" numbertype="1" tcsc="0" w:st="on">0.0001”</chmetcnv>. Given that a typical system has an ability to perform subpixel processing to one-tenth of a pixel suggests that a pixel unit will be <chmetcnv unitname="”" sourcevalue="0.001" hasspace="False" negative="False" numbertype="1" tcsc="0" w:st="on">0.001”</chmetcnv>. This then suggests that a camera with a nominal 500 X 500 array can only have a field-of-view of <chmetcnv unitname="”" sourcevalue="0.5" hasspace="False" negative="False" numbertype="1" tcsc="0" w:st="on">0.5”</chmetcnv> x <chmetcnv unitname="”" sourcevalue="0.5" hasspace="False" negative="False" numbertype="1" tcsc="0" w:st="on">0.5”</chmetcnv>. One could relax these rules somewhat but for reliable measurement performance it is not advisable to relax them by any more than a factor of two. In other words one might expect reasonable performance with such a camera covering a <chmetcnv unitname="”" sourcevalue="1" hasspace="False" negative="False" numbertype="1" tcsc="0" w:st="on">1”</chmetcnv> X <chmetcnv unitname="”" sourcevalue="1" hasspace="False" negative="False" numbertype="1" tcsc="0" w:st="on">1”</chmetcnv> field-of-view.

Recognizing these “rules-of-thumb,” so to speak, one can see why systems today have responded by being designed around higher resolution area imagers or higher resolution line scan imagers. Fortunately, the price/performance of the compute power required to handle the extra data at sufficient speeds to process the required imager processing and analysis algorithms and keep up with line speeds continues to improve.

Online vs. Offline

It is recognized that an online system may not always be required or even be consistent with the quality philosophy of a particular plant. Some subscribe to the sampling philosophy to monitor processes. In the case of even high volume operations conditions such as chip a sampling process along with first piece inspection to assure proper set-up could monitor placement accuracy. Companies who subscribe to this thinking are using more conventional machine vision-based off-line gauging workstations for set-up verification and equipment performance monitoring on a sample basis.

First piece and sample inspection may also be a more appropriate philosophy for lines that produce medium size volumes and handle a fair number of board designs as well as for low volume/high model mix lines.

AOI Challenges

One of the challenges of AOI systems is the time it takes to train on a new board design. While many systems are compatible with one or another of the CAD files that are associated with board designs and some even have developed “golden files” of libraries of many of the components conventionally used, training on a new board still requires manual intervention. The result is that training takes time – anywhere from four hours to a day or two. While the system is being trained on a new board design it is not being used for production inspection and this is a problem. Often training involves tweaking to handle both board and component appearance variables. The same component from two different sources may pose problems. While the outline may be the same, subtle color variations and certainly marking variations will exist.

Another challenge could be board warpage. Using conventional optics, board warpage essentially changes the magnification or size of a pixel in the image space. If quantitative results are sought, this will add to errors. Today systems avoid this issue by using telecentric optics which avoid warpage as well as vibration induced magnification errors.

Another challenge is the incidence of false rejects and false escapes. A false reject rate is measured in terms of the number of false defects divided by the opportunities for false defect detection, which is normally the number of components/joints. Given a board with 1000 components and assuming straightforward presence/absence vision decisions. A false reject rate of 0.1% suggests that on average one will experience one false reject per board. In other words, virtually every board will be rejected! Clearly an equivalent escape rate, that is the rate at which a system falsely accepts a bad condition as a good or acceptable condition, may be tolerable because reject rates experienced in assembly operations are typically less than 1000 PPM. In other words, only one out of every thousand boards has a reject condition. So an escape rate or false accept rate of 0.1% suggests that out of one million boards produced, 100 rejects could escape detection.

Data

The biggest gain from using AOI systems can be had by recognizing that the system is a data collector. In general these systems come with the ability to provide valuable insights into assembly processes by examining the data. For example, for pre-reflow inspection installations, a system could provide a display of a scatter plot of component offsets. For post reflow applications, the system might display a Pareto chart that summarizes the defects by their classification. Such a chart would immediately flag the most common defect being experienced. When corrected for, the incidence of rejects will decline.

Collecting data is not enough. These systems merely flag the results of machine conditions. It requires the line operator or process engineer search for the defect’s root causes and take corrective action accordingly.

Part 2
by Nello Zuech, President, Vision Systems International

This is the second part of a two-part article on machine vision in the assembled printed circuit board market. Part 1 covered a brief review of the history of machine vision-based AOI systems in applications in the electronic industry, as well a discussion of why today there is more urgency and more compelling reasons to adopt the technology. It also discusses where on the production line such systems can be deployed and where it actually makes the most sense. Some of the application issues, concerns, and requirements are also described. Part 1 concludes with the exhortation that it's the DATA that has the value but only if it will be used to make the line adjustments flagged to optimize yield.

What follows - Part 2 - depicts an example of a return-on-investment analysis that suggests in a typical assembly shop the ROI is in months and not in years. Tables are also developed that depict defect type by specific line location and vendors by product offerings. A questionnaire is developed designed to guide a prospective buyer in evaluating product offerings vis-à-vis his specific requirements. Some tips are also provided to assure a mutually rewarding experience between a vendor and buyer of an AOI system.

Return on Investment
Table 1 depicts one attempt to develop a justification for an AOI system. In this example the application involves post reflow. Many assumptions have been made: production is 20,000 boards/month; labor rates with benefits around $20.25; percent solder joint rejects 0.2%; etc. The analysis also takes into account warranty costs, indirect labor costs associated with inspection functions, inspector training costs, value of inventory in rework, etc. The net result is that under this scenario the cost of an AOI system can be recovered in less than 5 months. Note the costs of the AOI system include annual service costs and the labor costs to operate the system.

Applications
Table 2 depicts the defect types associated with the respective applications. An analysis should be conducted to determine the frequency the respective defects are actually experienced within your own board assembly operations. This should lead to conclusions about which defects a system must be able to detect reliably as distinguished from those that it would be nice if the system detects as well. In other words, a list of 'needs' versus 'wishes' will result from this exercise. Following the table is a checklist that reviews the questions that a prospective vendor should review. In ultimately determining the most qualified vendor, the answers to these questions should match your operation's specific requirements.

Questions to Ask Prospective Vendors of AOI Systems
As you examine the products from different vendors you will find most make the same claims. It is clearly important to get them to put their claims in writing. The following questions are meant to provide the framework for a systematic analysis of the competitive landscape. The answers given should be consistent with the application requirements anticipated. This list is not meant to be complete. Because of different quality management philosophies within the board assembly industry, the set of questions used should be consistent with your own espoused quality strategy.

1. Post solder paste:

1. Is your system an on-line or offline system?

2. Does it perform 100% inspection or sample inspection? In either case, please provide some measure of board density versus throughput? E.g. For board with 4000 components on an 8' board, system can handle 2 sq. in/second or whatever.

3. Is your system 2D or 3D?

4. If 2D, what does the system do and what are the specs? Accuracy, repeatability of measurements, etc.? Do you have a recommended calibration procedure to demonstrate accuracy of the system and, if so, what is it?

5. What is the finest pad pitch that can be handled?

6. If 3D what does the system do and what are the specs? Accuracy, repeatability of measurements, etc.? Do you have a recommended calibration procedure to demonstrate accuracy of the system and, if so, what is it?

7. If 3D, how is the height of the specific solder paste pad measured - based on a local reference plane? A global reference plane? How does the system handle board warpage issues?

8. If 2D is system based on area camera or line scan camera or laser scanner?

9. If 3D can you describe the fundamental underlying principles for capturing 3D data?

10. Does your system have difficulty handling a range of solder paste secularity?

11. How is the system trained to handle a new board design? CAD compatibility? Gerber file compatibility? Train-by-showing? Other? Combination?

12. How long does it take to train on a new board?

13. What is the changeover time where boards have been previously trained?

14. What is the throughput at what specific pixel size?

15. What is your false reject rate? Escape rate? How have these been demonstrated?

16. Is there an action that takes place if there are 'x' number of consecutive rejects at the same location? Or 'Y' over the entire board?

17. Is the system design based on your own proprietary hardware or commercially available products such as frame grabbers or vision processors or is it a host-based processing system?

18. Do you offer an upgrade patch for future generation products?

19. How many cameras does your system have and why?

20. Can you comment on the strength of your lighting arrangement and relevance to the application?

21. Can you comment on the strength of your optics and relevance to the application?

22. Does your system have the ability to adapt field-of-view/resolution as a function of the board design for a specific board design?

23. Does your system have Internet trouble-shooting compatibility?

2. Pre-reflow:

1. Is your system an on-line or offline system?

2. Does it perform 100% inspection or sample inspection? In either case, please provide some measure of board density versus throughput? E.g. For board with 2000 components, system can handle 2 sq. in/second or whatever.

3. Is your system 2D or 3D?

4. Is your system color-based? What advantages does color offer?

5. What is the finest pitch component that can be handled?

6. What does the system do:

1. Component presence

2. Component missing

3. Correct component

4. Polarity

5. Orientation

6. Misplaced/offset

7. Skewed

7. If it makes measurements, what are the specs? Accuracy, repeatability of measurements, etc.? Do you have a recommended calibration procedure to demonstrate accuracy of the system and, if so, what is it?

8. If 2D is system based on area camera or line scan camera or laser scanner?

9. If 3D can you describe the fundamental underlying principles for capturing 3D data?

10. Does your system have difficulty handling a range of solder paste and board secularity?

11. How is the system trained to handle a new board design? CAD compatibility? Gerber file compatibility? Train-by-showing? Other? Combination? Do you include part libraries? Do you have the capability for offline programming to avoid system downtime during training on new board design?

12. How long does it take to train on a new board?

13. What is the changeover time where boards have been previously trained?

14. What is the maximum height of a component that the system can handle?

15. What is the throughput at what specific pixel size?

16. What is your false reject rate? Escape rate? How have these been demonstrated?

17. Does board warpage affect results or what amount of board warpage can the system handle without degrading results?

18. Is there an action that takes place if there are 'x' number of consecutive rejects at the same location? Or 'Y' over the entire board?

19. Is the system design based on your own proprietary hardware or commercially available products such as frame grabbers or vision processors or is it a host-based processing system?

20. Do you offer an upgrade patch for future generation products?

21. How many cameras does your system have and why?

22. Can you comment on the strength of your lighting arrangement and relevance to the application?

23. Can you comment on the strength of your optics and relevance to the application?

24. Does your system have the ability to adapt field-of-view/resolution as a function of the board design for a specific board design?

25. Can you comment on your fundamental inspection approach and indicate strengths: template matching (discuss basis of template), normalized gray scale correlation, edge segmented-based pattern recognition, vector-based modeling, neural net, other, etc.? Does approach have capacity for continuous learning to adapt to ongoing board/component appearance variables?

26. Can you comment on your system's suitability for: low mix/high volume operations, medium mix/medium volume and high mix/low volume?

27. Does the system interface to a rework station? Do you offer a rework station?

28. Does your system have Internet trouble-shooting compatibility?

29. What is the price range of your systems?

30. What options, if any, are offered for your system?

3. Post reflow and post wave solder, use the same questions as 2 with the following modified version of (f):

1. What does the system do:

1. Component presence

2. Component missing

3. Correct component

4. Polarity

5. Orientation

6. Misplaced/offset

7. Skewed

8. Tombstones

9. Solder presence

10. Insufficient solder

11. Solder bridges both between leads and between components

12. Solder wick

13. Cold solder joint

14. dewetting

15. Solder voids

16. Bent leads

17. Lifted leads/chips

18. Solder balls

2. Can your systems handle lead-thru-hole components and, if so, pre-wave or post-wave or both? If there are different answers to the above performance related questions for the versions of the systems that you offer for pre or post wave applications, please comment accordingly.

Vendors
The table depicts some of the vendors known to be selling systems into the North American electronic market. It is understood that there are over 25 companies aggressively pursuing these AOI applications in the worldwide market. An attempt has been made to identify the specific applications the respective companies suggest their products address. This data was largely obtained from information on their website. An attempt was made to verify this information from the vendors themselves but very few cooperated.

What is Required to Succeed
Bob Ries of cyberoptics offers the following counsel. The vendors must 'start with setting the right expectations about what AOI will and will not do. Then the user and vendor must deploy enough resources (engineering and production) to make it work. These systems do not today (and will never) run themselves. Programming and day to day maintenance take ongoing resources.'

Chuck Gamble at Leica Microsystems amplifies on the vendor's responsibility by suggesting 'The vendor has SIX AREAS where he has to provide a proper level for the AOI implementation to be successful. 1. Repeatable data, both 2D and 3D, 2. Speed, the system must be able to keep up with pulse rates, 3. Programming and ease of use. The system must be able to be programmed in less than 30 minutes and must be capable of being run and programmed by line operators. 4. Reliable and Robust machine performance in a 24 X 7 environment, 5. Support Services, such as training, spares, documentation, global support network, rapid response, etc. and 6. Price - All of this at a reasonable price, …'