Statistical Process Control in Waterproof Bag Manufacturing

Inspection sorts defects; control prevents them

Every factory inspects, and most do it reasonably well. But statistical process control asks a different question from AQL sampling: not "is this shipment acceptable?" but "is the process still where it was yesterday, and where is it heading?" In waterproof bag manufacturing that difference is not academic — it decides whether a leaking construction is caught as one bad weld or as eleven thousand units already boxed and shipped.

A sampling plan accepts some known level of defects by design. That is a rational, statistically defensible decision, not laziness: perfect inspection is impossible, and for many attributes inspection is destructive or simply slower than production. But acceptance sampling is a filter, and a filter always lets something through. Tightening the acceptance number makes the filter finer; it does not stop the defects being produced in the first place, and it raises the cost of every unit in order to catch problems that a process signal would have flagged for free.

The practical framing a buyer should carry is this: inspection protects you from the bad batch you have already bought, while control protects the vendor and you from ever making one. The factories that supply consistently good waterproof product are not the ones with the most inspectors. They are the ones that know, at ten in the morning, whether the machine that made nine o’clock product is still behaving the way it was when the run started.

Two kinds of variation, and two kinds of action

Make two bags from the same roll, on the same machine, with the same operator, in the same hour, and measure the peel strength of the same weld. The two numbers will not be identical. That is not carelessness; it is variation, and it is unavoidable in every process that exists. The useful insight from Walter Shewhart and W. Edwards Deming is that not all variation behaves the same way, and treating the two kinds identically makes a process worse rather than better.

Common cause variation is inherent to the process as currently designed. The machine’s temperature controller has an inherent band, the film has thickness tolerance, the compressed air supply drifts slightly across a shift. All of it is acting all of the time, it is predictable in aggregate even though individual outcomes are not, and crucially it cannot be removed by anyone working at the machine. Reducing common cause variation is a management decision: different equipment, a tighter material specification, a changed process route, training that changes how the job is done.

Special cause variation is attributable to a specific, identifiable event: a new material lot, a heater that failed partway through, a tooling change, a replacement operator, a shift in ambient temperature that pushed a cooling window. It is not always present, it is usually visible if you are looking in the right way, and it is fixable by people close to the work. The whole purpose of a control chart is to separate these two situations reliably enough that nobody confuses them.

Confusion has a cost in both directions. Treating a special cause as normal means living with a defect mode that will return every time the same condition returns. Treating common cause as special means adjusting a machine that was doing fine, which Deming demonstrated with his funnel experiment: intervening in a stable process always increases its variation. The operator who nudges the dial every time a measurement looks slightly off is adding noise, not removing it, and no amount of diligence fixes that — only the rule does.

Control limits are not specification limits

This is the single most common misunderstanding in applied quality work, and it causes expensive decisions. Specification limits come from the design or the customer: the minimum acceptable peel strength, the tolerance band on a weld width. Control limits come from the process itself, normally calculated as the process mean plus and minus three standard deviations. They are unrelated quantities that happen to be plotted on charts of similar shape, and a competent reading requires keeping them apart.

The third row is where buyers get hurt, because on a conventional inspection report it looks perfect. Every unit tested within tolerance, acceptance passed, nothing to discuss. Yet the chart shows a steady downward drift in weld strength across the shift, or a run of nine consecutive measurements on one side of the mean, which almost always means something changed: a material lot, a heater, a hand. The shipment passes. The next one may not.

This is also the clean answer to a request that comes up in supplier conversations: proving every unit met spec does not prove the process was in control. Those are different claims, requiring different evidence, and only one of them tells you anything about what happens after you place the reorder.

Choosing the few characteristics worth charting

A factory that tries to chart everything charts nothing well. Control charts cost labour every shift, and any list long enough to look impressive is long enough to be neglected by month three. What earns a chart is being critical to quality — meaning its failure genuinely harms the customer — combined with having variation large enough to be worth watching. A dimension whose natural spread is a fraction of its tolerance is not interesting, no matter how central it feels.

For welded waterproof product the honest list is short. There is rarely a reason to chart more than four to six characteristics per construction, and the same short list tends to show up across most products.

Notice what is not on it. Print placement, minor cosmetic variation within an agreed reference, stitch count on a non-structural run — these may matter commercially, but if the failure does not reach the customer or the natural variation is trivial compared with the tolerance, charting them consumes attention that belongs somewhere else. A well-chosen CTQ list is a strategic document, and reviewing it once a year is part of the discipline, because products change and what was critical last season may now be comfortably stable.

Why a measurement beats a pass tick

Quality data comes in two flavours, and the difference in information content is not subtle. Attribute data is counting: pass or fail, defect or no defect, thirteen rejects in two hundred. Variables data is measuring: 4.2 seconds to leak, 38 newtons of peel resistance, 164 degrees Celsius. Both have their place, but where a real number is obtainable it should be preferred, for two concrete reasons.

First, sensitivity. A chart based on counts needs very large samples to detect a meaningful shift in defect rate; by the time the fraction-defective chart shows a convincing signal, thousands of units have been affected. A variables chart typically stabilises its limits after roughly twenty to twenty-five subgroups of four or five pieces, and will show a trend long before any single unit crosses its specification. In weld terms, the pressure decay number creeping from 4.2 toward 3.6 is visible weeks before the first return.

Second, waterproof testing has an awkward property: the most meaningful tests are slow, destructive, or both. You cannot submerge every unit for an hour, and peel testing destroys the part. The standard resolution is to chart an easily measured process parameter as a proxy — welding temperature, dwell time, pressure — and to periodically validate with destructive testing that the proxy still predicts the outcome. The proxy gives you frequency; the destructive test gives you truth; together they give you control, and neither alone is enough.

This reframes the whole activity. Testing a thousand finished bags tells you about a fact that is already over. Plotting the parameter that caused it, several times per shift, tells you about drift that is happening now. Both cost money. Only one lets you act before the cartons close.

Validate the measurement system before trusting any chart

Every measurement contains noise from the gauge, from the fixture, from the method, and from the person doing it. If that noise is large relative to the variation you are trying to observe, your chart is plotting your ruler rather than your process. Chasing a signal that originates in the measurement system produces adjustments that damage a healthy process, and missing a real signal because it is buried in measurement noise is worse still.

The standard assessment is a gauge repeatability and reproducibility study, usually shortened to gauge R&R. Repeatability is variation from the same operator measuring the same part repeatedly with the same device; reproducibility is variation between different operators doing the same job. Industry practice from bodies such as the American Society for Quality generally treats under ten per cent of total variation as acceptable, ten to thirty per cent as conditionally usable depending on cost and risk, and above thirty per cent as requiring correction before the data is used for anything.

A workable design needs no statistics department: two operators, ten parts spanning the expected range, each measured twice in randomised order, analysed with any spreadsheet implementation of the standard method. In practice this takes an afternoon and it is the difference between a control programme that works and one that generates arguments. It should be repeated when the gauge changes, when operators turn over significantly, and whenever the chart starts behaving strangely for reasons that cannot be found in the process.

It also has to sit alongside routine calibration, which is a different thing entirely. Calibration confirms the instrument reads correctly against a traceable standard; R&R confirms the whole measurement event, including the human and the fixture, produces usable data. A perfectly calibrated tensile tester operated inconsistently by three different technicians still produces unusable numbers, which is why both activities are recorded in any serious quality system such as the one described in our overview of ISO 9001 for waterproof bag manufacturing.

Building the chart: subgroups, frequency and baseline

The design decision that most often ruins a chart is how the subgroup is formed, and it fails in one specific way. A rational subgroup consists of consecutive pieces produced close together in time, because then within-subgroup variation represents genuine short-term noise while variation between subgroups represents process movement. When instead people collect pieces scattered across a whole shift and treat them as one subgroup, within-subgroup variation absorbs all the drift, the calculated limits become absurdly wide, and the chart loses the sensitivity it existed to provide.

Typical practice for this class of product is a subgroup size of four or five consecutive units, one subgroup per hour or per defined quantity, whichever is more meaningful for the line. For the initial limits, collect around twenty-five subgroups, meaning roughly one hundred readings, then calculate centre line and control limits from that baseline. Beforehand you will not know whether the process is stable, which is exactly why you should not publish limits too early — a limit calculated from an unstable period is a limit nobody should trust.

Once established, the limits get frozen. They are recalculated only when the process genuinely changes for the better: new tooling, a material upgrade, a trained-and-sustained method change. Recalculating because the latest points look uncomfortable is the same error as moving the goalposts to suit the score, and it is the reason suppliers sometimes present charts where every point sits comfortably inside limits that nobody can explain the origin of. Ask when the limits were last set and what prompted it; the answer is diagnostic.

Frequency deserves honesty rather than ambition. Measuring five welds per shift every shift generates a usable record. Measuring twenty-five per shift for the first week and nothing in week two generates nothing. Whatever interval is chosen has to survive a busy season, a rush order and a public holiday, because a chart with gaps is not evidence of control — it is evidence of a chart.

Reading the chart beyond the obvious signal

Most people who encounter a control chart learn exactly one rule: if a point goes outside the limits, investigate. That rule matters, but it is one of eight conventional tests and arguably the least sensitive, because by the time a single mean crosses three sigma the change is already substantial. The remaining rules detect patterns that precede the excursion, and it is those patterns that make the chart worth keeping on the wall.

Running every rule at once generates false alarms, because each additional test adds its own background rate of coincidental patterns. A sensible starting configuration is rules one, two and three, which together catch large shifts, sustained mean changes and slow drift, and which account for the great majority of what happens on a welding line. Add the others when the process is stable enough that finer detection becomes useful.

The reference most practitioners use for these tests and the underlying mathematics is the freely available NIST/SEMATECH engineering statistics handbook, which is worth reading once properly rather than relying on whichever blog summarises it most confidently.

Process capability: can the process hold the tolerance

Capability indices compress a question into a number: given how this process varies and where it sits, how much of its output will fall outside the tolerance? Cp compares spread only, ignoring whether the process is centred. Cpk also accounts for centring, which is why it is the figure normally quoted. Both assume the process is in statistical control; if it is not, the indices describe a process that does not consistently exist, which makes them decorative.

Two cautions before quoting any number. First, capability from subgroup variation describes short-term potential, while capability from overall variation describes what customers actually receive; reporting one as the other is a common and consequential sleight of hand. Second, these formulas assume a roughly normal distribution. Many waterproof characteristics are one-sided — there is no upper limit on weld strength, only a minimum — and applying a two-sided index to a one-sided requirement is not legitimate without a transformation or a method designed for it.

A capability report that gives a number without stating the characteristic, the sample, the time window, the data type and whether the process was in control should be treated as marketing. Those five details take thirty seconds to add, and their absence is itself information.

Every chart needs a written reaction plan

The most common way these programmes fail has nothing to do with statistics. A chart is produced, printed, and pinned near the machine. For a few weeks people glance at it. Then a signal appears and nobody is quite certain who is supposed to do what, so the conversation happens informally, nothing is recorded, and the next signal produces the same conversation. By month three the chart is wallpaper, and the programme is reported as implemented.

The antidote is a reaction plan with three columns and no ambiguity: the signal, the immediate action, the named role. When a point breaches the control limit, stop and isolate all product made since the last known-good check, then notify the technician. When nine points sit on one side, check material lot changeover, shift change and parameter records in that order, then log what was found even when nothing is found. When a trend of six appears, verify tool condition and heater output before touching the machine settings.

The plan must also state what not to do, which is where most untrained interventions go wrong: do not adjust the machine unless a special cause has been identified, because adjusting a stable process increases variation. That sentence belongs in the work instruction, not in a training slide, and it is the one most often omitted.

Finally, name who can stop the line. If the answer requires three approvals, signals will not be reported. A control programme without a clear stop authority is a reporting exercise, and the people operating the line will correctly read it as one.

Correcting a process versus improving its capability

Once the chart separates the two kinds of variation, the appropriate response depends entirely on which one you have. A special cause calls for restoration: identify it, remove it, bring the process back to where it was, and record what happened. This work is done by people close to the process and usually within one shift.

Insufficient capability is the other situation and behaves nothing like it. When a process is stable but its natural spread cannot fit inside the tolerance, no amount of diligence at the machine changes the arithmetic. Options narrow to four: tighter incoming material specification, different or better-controlled equipment, a tolerance justified by the actual functional requirement rather than inherited from a drawing, or a different process route altogether. In this industry that last one often means comparing welding methods, such as the trade-offs set out in our comparison of RF and hot air welding in production, or reconsidering construction type in favour of a welded rather than stitched approach as discussed in the guide to stitched, welded and bonded construction.

This leads to a sharp piece of buyer intuition. A request to raise Cpk is effectively a request for capital expenditure or a design concession. If a supplier agrees to improve capability without mentioning equipment, material standards or tolerance review, they are agreeing to a number rather than a change, and the numbers will not move. Conversely, a supplier who responds by explaining what would physically have to change is telling you something true, even when that something is expensive.

Neither pathway needs a separate quality department to begin, but both need the categories kept distinct. Teams that treat everything as a special cause spend their lives chasing symptoms. Teams that treat everything as common cause accept defects as fate. The chart exists so the distinction is made from evidence rather than from opinion.

What a buyer can reasonably ask for

Most buyers never see any of this because they never ask, on the assumption that it is internal. It is internal, but it is also the most legitimate thing to ask about, because it predicts your next shipment better than any certificate on file. Six questions cover the ground, and none require specialist vocabulary to ask.

• Which characteristics do you chart? A specific answer naming two to six things beats a general assurance that quality is monitored.
• Can I see the most recent chart for one of them, with dates? Real charts have irregular handwriting and unexplained notes; templates do not.
• What signals do you react to, and what does each one trigger? Look for named actions and named roles.
• When were those control limits last recalculated, and why? "When we bought the new welder" is a good answer; "monthly" is not.
• Has the measurement been validated — gauge R&R, and calibration dates? Especially for anything involving a person reading a scale.
• What is the Cpk, and was the process in control when it was calculated? A supplier who answers the second part before you ask is usually reliable on the first.

Warning signs are consistent across factories. Certificates with no corresponding charts. Charts with no year on them. Nobody able to say who owns a given chart. "We do one hundred per cent inspection so we do not need this" — which is a statement about sorting, not about control, and usually indicates the organisation has not separated the two ideas. Our separate review of what to look for when qualifying wholesale suppliers covers these signals in the wider context of due diligence.

None of this needs to be adversarial. Asking to see the chart is simply asking how the factory knows what it knows, and suppliers with functioning systems tend to be glad someone noticed.

Implementing this without a quality department

A small or mid-sized operation should not attempt to chart every process in the first quarter. The realistic sequence is deliberately narrow, because width destroys adoption while depth builds it. Month one: choose a single CTQ that genuinely predicts warranty complaints — weld strength, in most cases — get one gauge in working order, and run a short measurement study so the numbers mean something. That is the whole deliverable for the first month, and it is more than most programmes achieve in a year.

Months two and three: measure four or five consecutive units per shift, plot them, and resist drawing conclusions. This phase establishes what normal looks like for that specific machine, material and crew, and it is frequently surprising. Once twenty-five subgroups exist, compute limits and freeze them. Month four onward: add a second CTQ, write the reaction plan so it names a role and an action, train everyone touching that machine, and start weekly reviews of whether signals were noticed and acted upon.

After roughly six months the payback usually shifts from detection to design. When you can see where the process naturally sits and how widely it varies, you can set internal parameter limits tighter than what the machine tolerates, because you know how much room exists before output fails. That is the point at which control stops being a cost and becomes a specification advantage: the parameters you can hold, others cannot.

For the buyer, that distinction is worth more than any clause in a purchase agreement. Production consistency is what turns a waterproof specification into a product customers can rely on, and it is built by people who watch a process shift by shift rather than inspecting its output afterwards. If you want that discipline applied to your own styles, Tell us the specification and we will show you where we would draw the chart. Minimum order quantity is 500 pieces per style, samples in 6–10 working days, bulk production in 35–50 days, quoted FOB Xiamen.

Frequently Asked Questions

Q1. What is the difference between statistical process control and inspection?

Inspection evaluates product that has already been made and sorts acceptable from unacceptable. SPC monitors the process while it runs, so drift is visible before it produces rejects. Both are needed, and neither replaces the other.

Q2. Why is my shipment passing inspection but failing in the field?

Because every unit tested can sit inside tolerance while the process is drifting. A run of declining weld strength shows up in the averages long before any single piece fails, and acceptance sampling is blind to that pattern.

Q3. Are control limits the same as specification limits?

No. Specification limits come from the design or customer requirement. Control limits are calculated from the actual process data, normally at the mean plus and minus three standard deviations. Keeping them separate is the foundation of reading a chart correctly.

Q4. How many characteristics should be charted?

Usually four to six per construction. Charting more spreads attention too thin and the programme quietly stops. Choose characteristics that fail expensively and vary meaningfully.

Q5. Why is measuring a value better than counting pass or fail?

Variables data detects real change with far smaller samples and shows trends early. Attribute data needs enormous volumes to reveal the same shift, by which point the affected quantity is much larger.

Q6. What is a rational subgroup?

A set of consecutive units produced close together in time. This keeps short-term noise within the subgroup and puts process movement between subgroups, which is what makes the chart sensitive.

Q7. How many subgroups before setting control limits?

Around twenty-five subgroups of four or five pieces, roughly one hundred readings. Fewer than that and the limits do not describe a stable process, because you have not yet established whether one exists.

Q8. Should control limits be recalculated regularly?

No. Freeze them and recalculate only when the process genuinely changes — new equipment, upgraded material, a sustained method change. Recalculating because points look uncomfortable hides the signal the chart exists to show.

Q9. What is gauge R&R and why do it first?

A study separating variation from the device (repeatability) and from different operators (reproducibility). If measurement noise rival the process variation, the chart plots the ruler instead of the process and every conclusion drawn from it is unreliable.

Q10. What Cpk should I expect from a supplier?

A minimum of 1.33 is the conventional expectation in demanding supply chains, with 1.67 comfortable. Always ask whether the process was in control when it was calculated, because capability computed from an unstable process is meaningless.

Q11. What is the difference between correcting and improving?

Correcting removes a special cause and restores the existing process; the line can do it today. Improving capability changes the system through material, equipment, tolerance or process route, and requires management authority and usually money.

Q12. Can a supplier really raise Cpk without new equipment?

Rarely. Capability is set by the spread the system produces. If the answer involves no equipment, no material change and no tolerance review, the requirement has been agreed to verbally rather than engineered.

Q13. Do I need a full quality department to start?

No. One characteristic, one validated gauge and a few units per shift is enough to begin. Programmes fail from being too broad at the start, not from being too narrow.

Q14. What should a reaction plan contain?

The signal, the immediate action, the named role, and an explicit instruction not to adjust the machine unless a special cause has been identified. Without the last line, well-meaning operators add variation to a healthy process.

Q15. Can this be applied to small production runs?

Yes, though it requires care. Short runs can be charted against a standardised target using deviation-from-nominal plots, so several different parts share one chart. The logic is identical; only the scaling changes.

Q16. Is ISO 9001 the same as having statistical process control?

No. ISO 9001 requires a functioning, effective quality system and leaves the methods open. SPC is one of the most economical ways to demonstrate process control within that system, but certification alone does not prove any process is under control.

Q17. How often should the CTQ list be reviewed?

At least annually, and whenever a product changes. Characteristics that were critical last season can become stable, and new failure modes appear when materials or constructions change.

People Also Ask

What is statistical process control in simple terms?

Watching measurements of the process over time so that changes are detected while production is running, rather than discovering them when finished goods are inspected or returned.

What is the difference between control limits and specification limits?

Control limits are calculated from how the process actually behaves; specification limits come from what the design or customer requires. A process can be within specification and still out of control.

What is a good Cpk value?

A minimum of 1.33 is standard expectation, with 1.67 comfortable. Below 1.00 means some output is already outside tolerance regardless of what the inspection reports say.

What is gauge R&R used for?

To confirm that variation in the measurement system is small compared with the variation being studied. Above roughly thirty per cent of total variation, the data is not fit for charting.

How do you start statistical process control in a small factory?

Pick one critical characteristic, validate the gauge with a short repeatability study, then plot four or five consecutive units per shift until twenty-five subgroups exist before drawing any conclusions.

Why is adjusting a stable process harmful?

Because common cause variation cannot be reduced by tuning, and intervening adds new variation. This is the classic over-adjustment problem, and it makes a capable measurably worse process.