Three fiducials are already on the board, yet the pick-and-place machine may still hunt for marks, place off-target, or ask for manual confirmation. The problem is often not "one mark short." It is that parts, solder mask, process rails, and panel relationships have destroyed recognizability. A fiducial must satisfy both optical contrast and geometric datum conditions.
When a pilot line reports "unstable fiducial recognition," the easiest action is to add another circle. There are then more points, and recognition can still be intermittent, because the machine is not looking for any copper circle. It is looking for a set of repeatable targets that can establish a coordinate system.
The engineering judgment is this: first ask whether a single mark can be seen stably, then ask whether multiple marks can define a stable board-level coordinate system. Count is only the last condition. Opening, contrast, keep-out around the mark, and the spread of three points decide whether the vision system can compute position and rotation accurately.
A mark counts only if the camera can see it
Local board views in the source material point to a common issue: if pads, traces, silkscreen, or a part edge sit near the fiducial, the vision window no longer contains only one clean circular outline. A small change in camera threshold can make the recognized center drift between adjacent high-contrast shapes.
Copper, surface finish, and solder-mask color also change reflection. The mark graphic, solder-mask opening, and surrounding keep-out should be managed as one footprint rule set. During review, do not only check whether the copper layer has a circle. Open solder mask, silkscreen, and assembly layers and inspect the full observation window.
The positions of three points decide whether rotation error is amplified
Board-level fiducials convert design coordinates into machine coordinates. Two points far apart can determine translation and rotation. The third point adds check and stability. If the three points crowd into one corner or are nearly collinear, a small center-recognition error is amplified at the far end of the board.
The full-board example in the source material spreads the points along the board edge. That is more reasonable than dropping three marks wherever space happens to exist. Exact count and location still depend on the machine, board size, and process requirements, but the geometric rule does not change: the wider the baseline, the less sensitive the angle estimate.
Three points that are well separated are more valuable than three points that look neatly grouped.
On a panel, separate global datums from single-board datums
On a panel, global fiducials on the process rail serve whole-panel location. Local fiducials on each single board serve fine-pitch parts or local compensation. Placing marks only on each small board without a defined panel datum can still leave the machine without a stable coordinate system at the large-panel level.
Conversely, process-rail global fiducials alone cannot cover every local distortion. Large, irregular, or high-density panels should keep local marks based on process capability. Design review must state which level each fiducial belongs to so production does not mix marks that serve different purposes.
Turn recognition problems into four measurable actions
Inspection can start on the physical board: confirm that marks are not covered by solder paste, flux, labels, or fixtures; measure opening diameter and surrounding keep-out; check three-point span and distance to the board edge; then have the line save an actual vision window and recognized center.
If failures are intermittent in the same lot, also compare surface-finish reflection, solder-mask shift, and board warp. Consistency in the design file does not mean the camera sees the same grayscale edge. Placing vision screenshots of failed and passing boards side by side often finds the main cause faster than changing circle size again.
Before design freeze, also check manufacturing output: whether copper and solder-mask openings in Gerber match the PCB database, whether the panel shop added its own process rails and datums, and whether the placement program calls panel coordinates or single-board coordinates. When file version or coordinate hierarchy is wrong, the graphics on the board can be fully compliant and the machine still searches in the wrong place. Bind the final panel file, coordinate file, and vision-program version together so later line changes or reorders do not require guessing the datum relationship again.
Local fiducials near fine-pitch BGA, QFN, or camera modules need a separate check. A local mark compensates stretch or location error in one part region. It cannot replace a global fiducial, and it should not share a vision window with a test point, via, or tooling hole. If a shared local mark falls under a nozzle, stencil opening, or fixture after the part is rotated, the program still loses its reference.
Production sampling can record recognition score, center deviation, and manual-intervention count, not only "pass." A recognition score that falls slowly across lots often indicates solder-mask shift, surface oxidation, or contamination eating into vision margin. Pair those data with fabricator lots so design, fabrication, and placement can judge whether the issue is a rule problem or incoming variation.
If the line changes camera, lighting, or program template, run first-article confirmation again. Vision-parameter changes alter the threshold margin of the same fiducial. Do not reuse the pass conclusion from the old line.
First-article records should be archived with the production version.
- Graphic: circular target boundary is clear and the solder-mask opening is complete.
- Geometry: global points have adequate span and are not crowded into one region.
- Line: fixtures, contamination, and reflection do not block the vision window.
Conclusion
Three fiducials are only a count. What the pick-and-place machine actually recognizes is a clear, stable board-level coordinate system.
Whether there are "enough" fiducials cannot be counted only in a BOM or a rule table. The machine sees an optical graphic and computes geometric coordinates. Distortion in either can turn "already three" into "still not recognized."
On the next board revision, put fiducial graphics, openings, keep-out, hierarchy, and line vision screenshots into the DFM check together. The problem then closes before placement.
Does the fiducial issue you see look more like a mark that cannot be found, or a mark that is found while placement still has a whole-board offset?
