Automated Inspection in SMT: What It Finds and What It Misses
Automated inspection has changed what an assembly line can promise, and it has also produced a set of expectations that it cannot meet. A machine that examines every joint on every board at production speed is genuinely transformative, because a human performing the same task would be slower, less consistent and more expensive. The same machine also produces false alarms, misses defects that its optics cannot resolve and cannot reason about what it sees. Knowing which is which decides whether the inspection is an asset or a nuisance.
What AOI Does Well
Optical inspection is strongest on presence, position and orientation. A component that is missing, displaced beyond a tolerance, rotated, reversed or the wrong package is detected reliably, and the detection is repeatable in a way that a visual inspection is not. It is also strong on the well defined geometric features: the fillet around a gull wing lead, the alignment of a connector, the presence of a solder fillet on a chip component and the separation between adjacent pins.
Its second strength is the record. Every board examined produces an image and a result set, which means that a defect found at the end of a run can be compared with the same position earlier in the run. That comparison is what turns a single defect into a trend, and a trend is what identifies the operation that is drifting.
The third strength is coverage at speed. A machine examines a whole panel without fatigue, which matters most in the middle of a long run where the defect rate is low and the attention of an operator would naturally drop.
What It Cannot Do
Optical inspection cannot see beneath a component. A ball grid array, a quad flat no-lead package and the underside of a shielded module present joints that no camera can reach, and the fact that the visible periphery is perfect says nothing about the connections underneath. That is a limitation of physics rather than of the equipment, and no improvement in resolution will remove it.
It also struggles with the appearance of a joint where the appearance is ambiguous. A joint that is present but weakly bonded, a barrel that is filled but porous and an interface between alloy and pad that has not properly wetted are all conditions whose visual signature is subtle and whose consequences are serious. A machine that has been tuned to catch every marginal joint produces a stream of calls that a person has to review, and a machine that has been tuned to reduce the calls lets some of them through.
The False Call Rate and Why It Matters
The false call rate is the practical measure of whether an inspection system is helping. A system that reports a defect on every board is technically sensitive and operationally useless, because the review that follows is the inspection, performed manually, with the machine providing nothing except a suggestion.
A high false call rate has a second and less obvious cost. When a review routinely confirms that the machine is wrong, the review itself becomes a formality, and the habit of dismissing a call develops alongside it. The one genuine defect that the machine finds is then dismissed with the others, and the system has produced a worse outcome than no system at all.
The tuning of the thresholds is therefore a real engineering activity rather than a setup step. The limits are set from the acceptance criteria of the product, the review is performed by a person with the authority to make the decision, and the results are tracked so that a rising false call rate is treated as a signal that something in the process has changed.
Defect Classification and the Data
The value of automated inspection increases sharply when the results are classified rather than counted. A count says that forty defects were found. A classification says that twenty eight were solder related and concentrated at one component position, seven were placement related and spread across the panel, and five were false calls on a package whose body confuses the vision system.
That classification is what allows a response. The solder defects point to the printing or the profile; the placement defects point to the programme or the feeder; the false calls point to the inspection programme itself. A system that produces only a number leaves the investigation to begin from nothing, and the investigation usually ends at the rework station rather than at the cause.
Machine learning is increasingly used for this classification, and its contribution is genuine but bounded. A model trained on the images the machine produces can distinguish a marginal joint from a false call more consistently than a fixed threshold, and it improves as it sees more production. What it cannot do is identify a defect mechanism that has not appeared in its training data, and it cannot be relied upon to detect a new failure mode that no one has seen before. It is a filter that improves the ratio of useful calls, not a replacement for the engineering that interprets them.
Where the Human Still Belongs
The human role changes rather than disappears. The person who used to examine every joint now examines the calls the machine raises, decides whether they are genuine, and classifies the genuine ones so that the process can respond. That is a more valuable activity than the original inspection, and it requires more knowledge rather than less.
There is also a class of defect that only a person will notice. A board with a scratch, an unusual discolouration, a component that has been placed correctly but looks different from its neighbours, or an assembly that simply does not match the drawing in some way that the programmed checks do not cover. Those are found by looking, and an operation that has replaced looking entirely has given up a capability along with the labour. Our inspection stations run the optical and X-ray checks under quality management, the assemblies are built on the SMT assembly lines, the bottom terminated devices are examined through the process described on process, the fixtures and the functional checks belong to PCBA testing and the prototypes are proved through rapid PCBA prototyping.
Using Inspection Data to Improve
The improvement loop is where the investment in inspection is actually repaid. A defect found at a specific position on many boards is a message about the stencil, the pad geometry or the placement programme. A defect that appears only in one region of a panel is a message about the support beneath the board or about the way the panel heats. A defect that begins part way through a run is a message about a consumable: paste ageing on the stencil, a cleaning interval that is too long or a nozzle that has begun to clog.
None of those conclusions is available from a pass or fail count. They are available from the position, the timing and the classification of the calls, which is why the data is retained against the batch rather than discarded when the boards have shipped. An operation that keeps the information and compares one batch with the next is running a controlled process; one that keeps only the yield figure is running a process whose behaviour is unknown.
The same data supports the customer conversation. When a question arises about a specific unit, the images and the results for that board can be retrieved, and the discussion moves from an exchange of recollections to a review of what was recorded. That is the difference the record makes, and it is produced as a by-product of the inspection the line was performing anyway.
FAQ
Can optical inspection replace X-ray? No. It cannot see beneath a package, which is exactly where the joints of a bottom terminated device are located.
Why is a high false call rate a problem? Because it makes the review a formality, and a review that is routinely ignored will eventually dismiss a genuine defect.
What makes inspection data useful? Classification by mechanism and position, which identifies the operation to investigate rather than only reporting a count.
