How to Feed Inspection Data into SPC

A vision station that reports only pass or fail is throwing away the most useful number it produced. To decide that a part passes, the station measured something — a distance, an offset, an angle, a count, a grey-level difference between two regions. That measurement carries information the verdict does not, because it begins to drift long before it crosses the tolerance and therefore gives warning while there is still time to act. The verdict tells you what happened; the value tells you what is happening.

From verdict back to measurement

The first task is to recover the measurement and give it a defensible definition. A station that judges a gap should output the gap in millimetres, with the method written down: which edges were detected, over which span, with which sub-pixel rule. That definition is what makes the number comparable between shifts, between stations and after any software change. A number whose provenance cannot be stated cannot be put on a control chart, because nobody can tell whether a shift in the chart came from the process or from a change in how the number was computed.

The second task is to name it. Each value needs a characteristic to belong to, and that name is the interface to the SPC system: the measurement will be filed under a characteristic identifier, against a part or a batch reference, with a timestamp and the station that produced it. The convention for how those identifiers are formed and how the reference is carried is a design decision in its own right, and it is the same one that governs how values are written back into an upper system, described in writeback field design. Getting it consistent at the start is far cheaper than reconciling two naming schemes after the first month of production data has accumulated.

Every part, or some parts

A vision station can measure every part, which is its principal advantage over sampling inspection. A control chart built on every part is not the same kind of object as a chart built on five parts an hour: with every part measured, the point-to-point variation includes the high-frequency process noise that sampling averages away, and control limits derived from a sampled process will produce constant false signals. This is not an argument against measuring every part. It is an argument for deciding deliberately which population the chart represents and setting the limits accordingly.

Separate the measurement from the process first

A control chart is only meaningful if the variation it displays comes from the process. A vision measurement carries its own variation: lighting drift, camera noise, calibration drift, edge-detection repeatability on a rough surface. If the measurement variation is a large fraction of the process variation, the chart is mostly displaying the camera, and reacting to it means adjusting a process that was not changing. The standard way to establish the ratio is a gauge study, in which the same parts are imaged repeatedly while the process is held still; the spread of those repeated measurements is the measurement variation, and it can be compared directly against the process spread.

The usual convention is that measurement variation should be a small fraction of the tolerance — a tenth is the classic target, a fifth is often accepted where the process is comfortably capable. This is the same principle that governs the calibration drift limit, since a drifting calibration is one of the contributors to measurement variation, and it is discussed in calibration and drift. Where the ratio cannot be brought down far enough, the honest alternative is to chart the verdict counts rather than the values, or to chart a moving average that suppresses the measurement noise, and to say so in the documentation.