AOI Program Tuning: Cutting False Calls Without Escapes

An AOI program is a set of decisions about what an acceptable joint looks like, written as tolerances on measurements the machine can make. The machine does not see a good joint; it sees the area of a fillet, the position of a centroid, the contrast of an edge and the colour of a surface, and compares each against a window. Tuning the program is therefore a matter of setting those windows so that the machine rejects what an inspector would reject and passes what an inspector would pass. The difficulty is that both halves of that requirement pull in opposite directions.

What the Program Is Actually Deciding

Most AOI algorithms for solder joints fall into three families. Area-based algorithms measure the proportion of a window covered by solder and compare it against limits; they detect missing and excessive solder well and are insensitive to the shape of the fillet. Shape-based algorithms fit an expected geometry, such as a fillet with a concave profile, and detect lift and poor wetting. Colour or intensity-based algorithms use the difference between the specular surface of molten solder and the matte surface of a cold or oxidised joint.

Knowing which family a library entry uses tells the engineer what it can and cannot find. An area-based library will not flag a joint that has the right amount of solder in the wrong place, and a shape-based library will not flag a joint whose fillet is correct but whose alloy is wrong. Program tuning starts with matching the algorithm to the defect that matters for that component, rather than with adjusting tolerance values on whatever algorithm happened to be supplied.

Libraries, Tolerances and the Definition of a Defect

A library is a template of the component, its pads and the expected joint appearance, taken from a known-good board. The tolerances around that template are the program. Setting them from the supplier’s defaults produces a program that is either too loose, and misses defects, or too tight, and stops the line for cosmetic variation. The correct starting point is the acceptance standard the product is built to, translated into the geometric terms the library uses.

The translation is where most programs go wrong. A standard that says a fillet must be visible and concave becomes, in the library, a minimum and maximum coverage percentage plus a shape score. If the coverage limits are set from a single golden board, they will be narrower than the process can hold, because one board does not describe the distribution. Taking measurements from thirty boards and setting the limits at the edges of the natural spread gives a program that follows the process rather than fighting it.

False Calls and Their Real Cost

A false call costs the time to review the image, the time to look at the board, and the loss of confidence that follows when the operator finds nothing wrong. When false calls are frequent the operator’s attention degrades, and the call that matters is dismissed along with the rest. The cost is not the review time alone; it is the escape that follows from a program nobody believes.

False calls cluster. They come from a particular component, a particular lighting condition or a particular board position, and grouping the call log by library entry and by position shows which. A component whose joints vary because the pad is connected to a large copper area will call repeatedly for insufficient solder at one end, and the answer is either a library entry specific to that orientation or a change to the inspection standard for that joint, rather than a global loosening of the tolerance.

Escape Rate and Inspection Coverage

The escape rate is the proportion of real defects that pass the machine, and it can only be measured by comparing the machine against a slower, more thorough inspection. Auditing a sample of boards that the machine passed, with a trained inspector and a microscope, gives a figure that is meaningful for that product and that program. Without such an audit the escape rate is unknown, and the program can only be discussed in terms of how often it stops the line.

Coverage is the other side of the same question. A machine that inspects only one side of the board, or that skips a component because it is inside a shield can, has a coverage gap that must be closed by another method. Mapping the coverage of every inspection station on the line, and checking that each defect type identified in the process FMEA is covered by at least one of them, is more valuable than improving the resolution of any single station.

Golden Boards and Library Qualification

A golden board is a reference unit whose joints have been verified by an inspector and, where needed, by cross-section. It is used to qualify a library after a change and to check the machine before a production run. A single golden board is not enough for qualification, because it cannot show the tolerance the process needs. A set of boards that spans the acceptable range, including a marginal one at each limit, gives the program a chance to be tested rather than assumed.

Golden boards age. A board that has been probed and handled repeatedly develops scratches and contamination that the machine may read as defects, and a board that has been reflowed more than once has a different surface finish. The set should be kept in a case, used only for qualification, and replaced on a defined interval or whenever its own inspection result changes. Recording the date and the result of each qualification run turns the golden board from a prop into evidence, and pairs it with the visual inspection checklist that the same joints were judged against.

Lighting and Camera Settings by Joint Type

Lighting is the parameter that most often decides whether a library works. A joint on a bright pad with a large copper area reflects differently from a joint on an isolated pad, and a program tuned under one condition will misfire under the other. Multi-angle illumination, with separate colour channels, allows the library to use the angle at which the fillet is brightest, which is more stable than the absolute intensity of a single channel.

Camera settings should be fixed once the program is qualified, and any change should be treated as a program change requiring requalification. Exposure and gain in particular shift the measured areas directly. Where the machine supports it, storing the settings with the program prevents them from drifting when another product is set up on the same machine, which is a common cause of a program that worked last month and does not work now.

Debug, Verification and the Operator Loop

The operator’s decision on each call is data. A simple record of whether the call was confirmed as a defect or marked as a false call, kept per library entry, shows which entries are being argued about. An entry with a high false-call rate and no real defects needs its tolerance widened or its algorithm changed; an entry that never calls and yet produces escapes in the field needs its role in the process reviewed, because it may be inspecting the wrong feature.

Verification after a repair is part of the loop. A reworked joint should be re-inspected with the same program, and where the program rejects a joint that an inspector accepts, the library should be examined rather than the inspector overruled. Accumulating the accepted rework images across a few months gives a better picture of the marginal joint population than any controlled study, because it comes from the process as it actually runs, and it shows where the workmanship training and the machine are applying different criteria to the same joint.

Change Control for Inspection Programs

An inspection program is a controlled document. A change to a tolerance, a library or a camera setting alters what the line accepts, and it should be recorded with the reason and the evidence, exactly as a change to the stencil or the profile would be. The common failure is a quiet adjustment made during a shift to get the line running, which is then never reflected in the master program and reappears as an unexplained escape.

Where a change is justified, the qualification should include a run of boards that covers the tolerance range, plus an audit of a sample against a manual inspection. The result of that audit belongs with the change record, so the next engineer can see what the program was shown to detect, not merely that someone approved it. This is the same discipline that the first pass yield data depends on, because a yield figure is only comparable when the measurement that produced it has not moved.

FAQ

Should the program be tuned to the operator’s judgement? To the standard, not to the person. Where an operator consistently disagrees with the machine, the disagreement is evidence that a library entry needs review, but the resolution should be a written criterion rather than a verbal understanding, since the person will change and the criterion will not.

How often should a program be re-tuned? After any change to the paste, the stencil, the profile or the component supplier, and whenever an audit shows the escape rate has moved. A program that is left alone for years is a program whose relationship to the current process is unknown, which is a risk at the next recall.

Is more resolution always better? No. Higher resolution slows the machine, increases the data volume and can make the program more sensitive to cosmetic variation. Resolution should follow the smallest feature that has to be measured, which is set by the component pitch and the defect types that matter.