SPC software: getting measurements into charts that somebody actually reacts to

Capture decides whether the effort survives its first month

Data that has to be typed will be typed late, in batches, sometimes from memory, and occasionally invented to fill a gap. Gauges that can output readings directly should do so, whether through a wired interface, a wireless collection hub or a station terminal, and the aim is that recording a measurement costs an operator nothing beyond making it. Where entry must be manual, reduce the field count ruthlessly and validate at the point of entry so an obviously impossible reading is challenged while the part is still in hand. A programme that starts with enthusiastic manual recording and quietly decays is the standard pattern.

Configuration is where statistics becomes somebody's job

The software will ask questions the plant has not settled: what constitutes a subgroup, how often to sample, which chart suits a characteristic that is bounded at zero, when limits are calculated and when they are recalculated, and who may change them. These are engineering decisions with consequences, since limits set during an unrepresentative period will either alarm constantly or never. Assign them to a named engineer, document what was chosen for each characteristic and why, and review them when the process genuinely changes rather than when the chart becomes inconvenient. Limits edited to stop an alarm are the clearest sign the system has been captured.

An alarm with no named recipient is only noise

A signal is worth having only if it reaches somebody who can act within the time it takes to produce defective product. That means routing rules that know the shift pattern, escalation when nobody acknowledges, and a defined reaction attached to the characteristic itself: what to check, whether to stop, whether to quarantine what has been made since the last good check. Writing that reaction plan is unglamorous and it is what separates a plant that uses charts from one that files them. Where the answer is that nothing would be done differently, that characteristic probably should not be charted at all.

Charting everything is the classic way to achieve nothing

It is tempting to instrument every dimension on the drawing, and the result is hundreds of charts nobody reviews and an alarm rate that guarantees indifference. Selection should follow consequence: characteristics where variation causes functional failure, customer complaints, downstream assembly problems or regulatory exposure. That list is usually short, and it changes as the process matures. A periodic review that removes characteristics which have shown stable behaviour for a long period, and adds ones implicated in recent failures, keeps the effort proportionate and keeps operators paying attention to what remains.

What the accumulated data is worth afterwards

Beyond the immediate signal, a populated history answers questions that otherwise require new studies: whether a process was capable before a customer complaint, whether a change in material coincided with a shift in the mean, what capability to quote when tendering for tighter work. Customers in demanding sectors ask for evidence of this kind directly. Keeping the data usable means recording context alongside the measurement — machine, tool, operator, material lot — because a chart without that context shows that something moved without offering any route to what caused it.

Frequently asked questions

Do we need dedicated software or will a spreadsheet do?

A spreadsheet can plot a chart perfectly well, and for a single critical characteristic on one machine it may be enough. It struggles once you need automatic capture from gauges, alarms that reach a person on shift, control of who may change limits, and retrieval of history across products and periods. It also struggles with credibility, since a spreadsheet chart can be edited without trace and a customer auditor will notice. Scale and scrutiny, rather than statistical need, usually decide the point of change.

Our operators ignore the alarms. What is the fix?

Look at the alarm rate before the training programme. If signals appear many times per shift and nothing bad follows, indifference is a rational response and the limits or the sampling scheme are wrong. If the rate is low but responses are still absent, the problem is the reaction plan: nobody has said what to do, or doing it requires permission from someone unavailable at night. Fixing either of those changes behaviour faster than briefing people about the importance of charts.

Should measurement data live in the quality system or the production system?

The practical answer is that it needs to be linked to both, and the argument about ownership is less important than the identifiers. A measurement is only useful if it carries the order, batch, machine and time, so whichever system stores it must receive those from production. Many plants keep the raw stream in a dedicated store and push exceptions into the quality case system, which keeps the case load manageable while preserving the full record for later analysis.