Statistical Process Control on SMT Assembly Lines
An SMT line produces thousands of joints an hour, and inspecting them all is neither possible nor economic. Statistical process control takes the opposite approach: measure a small sample often, compare it with what the process normally does, and act when the pattern says something has changed. Done properly it catches drift before defects appear. Done badly it becomes a chart nobody reads. This guide covers how gopcb builds and uses statistical process control on its assembly lines.
Why SMT Lines Need Statistical Control
An SMT line has many inputs that move independently: paste viscosity, stencil wear, placement nozzle condition, oven temperature and ambient humidity. Individually these changes are small, and any one of them may stay inside its own specification while the combination pushes joints out of tolerance. Statistical process control detects that movement before it produces scrap.
The alternative is detection by inspection, which is always late. By the time AOI or functional test reports a defect, the line has already built a batch of suspect boards. Control charts shift the response from sorting bad product to correcting the process that made it, which is both cheaper and faster.
<img src="https://www.gopcba.com/wp-content/uploads/2021/05/landing_slide.jpg" alt="Statistical process control chart displayed on an SMT line monitor” />
Choosing the Right Control Chart
The chart type follows the data. Variables data, such as paste volume or placement offset, suits an X-bar and R chart or an individuals chart with a moving range. Attribute data, such as the number of defective boards per lot, suits a p or np chart. Using an attribute chart for continuous measurements throws away information.
Subgroup size also matters. Small subgroups are cheap but insensitive to small shifts, while large subgroups detect small shifts quickly but cost inspection time and may hide within-batch variation. A common compromise is a subgroup of three to five consecutive boards, measured often enough to see a shift within a shift.
Setting Control Limits from Data
Control limits are not specification limits, and confusing the two is the most common error in statistical process control. Specification limits come from the customer or the design. Control limits come from the process itself, calculated from the natural variation of recent data, and they describe what the process is actually doing.
A stable process produces points that fall inside the limits with a random pattern. When a point falls outside, or when a run of points sits on one side of the centre line, the cause is usually an assignable change rather than chance. Calculating limits from unstable data produces wide limits that hide the very shifts the chart exists to find.
<img src="https://www.gopcba.com/wp-content/uploads/2024/09/AdobeStock_540919248-2.png" alt="SMT line operator reviewing control charts and process capability data” />
Sampling Plans and Subgroup Size
Sampling has to be frequent enough to catch a shift before it fills a batch, and random enough to represent the whole run. Measuring five boards at the start of a shift and none afterwards tells you almost nothing about the other seven hours. Spreading samples across the shift, and across every machine that contributes, is what makes the chart meaningful.
The measurement itself must be capable. A gauge that cannot resolve a tenth of the tolerance will produce noisy data, and noise on a chart triggers false alarms. Gauge repeatability and reproducibility studies should be run before a characteristic is placed on a chart, and repeated when the measuring equipment changes.
Process Capability and Cpk Targets
Process capability compares the width of the process distribution with the width of the specification. The result, commonly reported as Cpk, expresses how well the process fits inside its limits and how far the mean sits from the nearest limit. A high Cpk means the process is centred and narrow; a low value means it is either off centre or too variable.
Typical targets are 1.33 for a mature process and 1.67 or higher where the consequence of failure is severe. Reaching them usually requires reducing variation rather than shifting the mean. Where the process capability is low, the correct response is an engineering project on the process, not tighter inspection of the output.
Reaction Rules When a Chart Signals
A chart only creates value if somebody acts. Written reaction rules should say who is notified, what is checked first and when production stops. A single point outside the control limit normally triggers an immediate check of the machine and the last few boards, while a trend or a run of seven points on one side triggers a broader investigation.
The investigation should follow the process sequence backwards: paste, stencil, placement, reflow. Our overview of solder defects and board failures lists the usual causes in that order, which makes a useful checklist when a chart signals. Recording the finding, and whether the correction worked, closes the loop.
Applying SPC to Paste, Placement and Reflow
Paste volume is the most useful characteristic to chart on an SMT line, because it is measured automatically by the printer or by an SPI system on every board. Tracking volume, height and area with control charts turns the printer into a monitored process rather than a machine that is only examined when something looks wrong.
Placement offset is the second candidate. Modern machines report their own pick and place data, so an X and Y offset chart for a critical component can be produced from machine logs at no extra cost. Reflow is charted from profile data, with peak temperature and time above liquidus as the key characteristics.
Common Mistakes and Data Quality
The classic mistakes are easy to list and surprisingly common. Using specification limits as control limits, recalculating limits after every excursion, excluding inconvenient points, and charting a characteristic that nobody can influence are all ways to make statistical process control decorative rather than useful.
Data quality deserves equal attention. Charts built from manual entry contain transcription errors, and charts built from automatic systems contain configuration errors. Someone should own the data path, verify the measurement source and confirm that the numbers on the chart match what the machine actually produced.
Linking SPC to Continuous Improvement
Charts are the input to improvement, not the result. Once a process is stable, the remaining variation can be attacked systematically, and every reduction in variation shows up as a higher process capability and a lower defect rate. That sequence is what turns a control chart into a reduction in cost.
The discipline also supports audits and customer reviews. Records that show a stable, capable process reviewed over months are far more persuasive than a certificate, and they align with the evidence expected when judging PCB quality. The information then feeds the same records used across the production process flow.
FAQ
What is the difference between SPC and inspection? Inspection sorts good product from bad after the fact, while statistical process control watches the process that creates the product and signals when it changes. SPC prevents defects by keeping the process stable, whereas inspection only finds them once they exist.
How many samples per shift are enough? Enough to detect a shift before it produces a full batch of scrap, which usually means small subgroups taken at regular intervals across the whole shift. The exact number depends on how fast the process drifts and how expensive a scrap batch would be.
What Cpk should an SMT line target? Most lines aim for 1.33 on critical characteristics and higher on safety related ones. The number should be agreed with the customer and reviewed periodically, and it should reflect the real consequence of a failure rather than a default value.
