Introduction: CAD-free AOI programming lets an SMT line build an inspection recipe from a single golden board, cutting the changeover time that CAD imports used to consume.
Most SMT lines still treat a CAD file as the starting point for every new AOI recipe, which means inspection readiness waits on the design team, the file format, and the coordinate data before a single board is measured. In high-mix production that dependency turns into idle conveyor time at every product change, and the wait rarely lands on a convenient day. this guide explains why CAD became the default, how one known-good board can teach a working recipe instead, and which board types still need an engineer's hand on the settings.
The CAD import habit grew out of how early AOI engines worked. Older inspection software had no way to find a component on its own; it needed a list of coordinates and a library entry for every part number before it could look at anything. A design file already contained all of that, so importing it was the shortest path from a new product to a running program. On a stable, low-mix line that arrangement worked well for years, which is exactly why it became the assumed starting point rather than a choice anyone re-examined. Four things inside a CAD file made it so useful:
Formats for that data have been standardized for years, and industry efforts such as the IPC-2581 Consortium exist mainly to make the design-to-manufacturing handoff cleaner. The weak point is timing. In contract manufacturing the file can arrive late, in an unexpected format, or after the panel has already been re-panelized for production, and the AOI program sits waiting while the line does not.
The golden-board method flips the order of operations. Instead of describing the board to the machine in advance, the machine looks at the board and describes it to itself. An operator loads one known-good, fully assembled board onto the conveyor, and the system images it across the full surface at the machine's stated 5 μm pixel resolution. From those images it reads geometry directly: where the pads sit, where each part is placed, which direction it faces, and what a properly formed joint looks like at each location. Calibration comes from the board's own fiducials, so the coordinate frame is derived from the physical product rather than from a text file. This is how an online AI AOI optical inspection machine can start judging a product it has never seen a drawing of. A guided wizard keeps the sequence short and predictable. The operator confirms board orientation, sets the inspection area, excludes any region that should not be judged, runs the learning pass, reviews the result on screen, and saves the recipe. IRIS ANT states about five minutes for this standard workflow on the T7 using one good board, and roughly 30 minutes of training before an operator can run the sequence comfortably. What the process produces is not a rule list but a reference model of normal appearance, which the deep learning engine then uses to judge every subsequent board coming down the line. The reason a single board can carry that much information is that general defect knowledge travels well. A model that has already learned what a solder bridge, a tombstone, or a missing part looks like does not need to relearn those shapes for each new product. The same principle appears in domain adaptation research, where a system trained on one data domain is adjusted for a new domain with a modest amount of targeted input instead of a full retraining campaign. In production, the first runs do the final tuning: corrections an operator makes on screen feed back into the model, so call accuracy settles over a few cycles rather than requiring a library built by hand. Factory-side deployment keeps board images and recipe data on the local network as well.
Standard boards move through the wizard cleanly. That means typical density, reasonable contrast between pads, solder, and substrate, familiar package families, and a layout that gives the camera a clear look at each joint. Dense work is where the picture changes: fine-pitch parts packed shoulder to shoulder, dark substrates with weak contrast, tall components casting shadow over their neighbours, or a panel carrying a large number of unique package types. Those boards still learn from a golden board, but the learned model usually needs segmentation, region limits, or per-area thresholds before its calls are dependable. Engineering time goes into shaping the model rather than rebuilding a program from nothing. There is also the question of what one board represents. A single golden board captures one example of a passing assembly, while a real process drifts across placement tolerance and solder volume from lot to lot. Engineers who want the model to absorb that spread often add samples from other production lots so normal variation sits inside the reference set. That is ordinary process engineering, and it is the main reason a dense, tightly toleranced product takes longer to bring online than a straightforward one. It helps to be clear about where each approach earns its place. Lines with clean CAD data, stable products, and infrequent changeovers can still find import convenient, because the file is already sitting there and nobody is waiting on it. CAD-free setup earns its keep when changeovers are frequent, when files arrive late or not at all, or when the people running the machine are not programming specialists. The published five-minute figure describes the standard workflow; complex high-density boards can still require engineering adjustment, and that is the realistic way to plan a changeover budget.
CAD dependency was a reasonable answer to a real problem: early inspection engines could not find components without being told where those components were. A golden-board workflow removes that dependency by letting the machine read the product itself, then keeping the recipe alive through feedback from normal production. The practical gain shows up at changeover, where the difference between waiting on a file and running a five-minute learning pass is measured in conveyor time. For standard boards the workflow stays largely operator-driven, while dense boards deserve scheduled engineering time to tune the model. Readers who want the exact standard-workflow figures can review the T7 documentation directly.
A：It means the inspection recipe is built from a physical board rather than from design data. The machine images one known-good assembly, derives pad and component positions from those images, builds a reference model of normal appearance, and saves that as the recipe. The line then inspects real production boards against that reference, with operator corrections feeding back into the model over the first few runs.
A：Because a modern inspection engine reads the board directly. Imaging at fine pixel resolution delivers pad geometry, component position, and polarity, while fiducials on the board itself establish the coordinate frame that CAD used to supply. The deep learning side already knows general defect patterns, so one good example gives it enough to start judging, and later production feedback sharpens the result.
A：Not always at the same speed as simple ones. Dense fine-pitch layouts, low-contrast surfaces, and shadowing from tall parts can push the learned model toward segmentation and threshold adjustments. Those boards still skip CAD import, but engineering time belongs in the plan. The five-minute figure describes standard workflows, while complex high-density boards may need adjustment before the recipe is production-ready.
Text and Synthetic Data for Domain Adaptation in End-to-End Speech Recognition
