Starter guide · Technical perspective · Published: October 2, 2026
Table of Contents
Key Takeaways
- A control chart answers one question — is this process stable? — and it answers with limits computed from the process’s own history, not from a specification sheet [1].
- The center line is the long-run average from a stable baseline (NIST’s metrology guidance suggests at least 100 historical check values to establish it), and control limits sit at ±3σ of the plotted statistic [1][2].
- On a stable, normally distributed process, about 99.73% of points fall inside the 3σ limits, so roughly three points in a thousand will breach a limit by pure chance. That is why one breach triggers investigation, not panic [5].
- For single water-quality readings, the individuals-and-moving-range (I-MR) chart is the standard tool: individual limits are X̄ ± 2.66 × MR̄, and the moving-range upper limit is 3.27 × MR̄ [4].
- Common-cause noise calls for system changes; special causes call for local fixes. Reacting to noise by turning knobs constantly is what Deming called “tempering the process,” and it amplifies variation [2].
- Water plants already record flow, raw water quality, chemical feed, turbidity, and finished water quality through SCADA or manual logs. Run rules such as Nelson’s six-point trend rule turn that archive into an early-warning system for drift [3][6].
What a Control Chart Actually Answers
Every water quality trend line looks exciting on a dashboard. A control chart is what turns that line into a decision. Walter Shewhart developed the chart in the 1920s; it plots a measurement against time between two statistically derived limits, and its purpose is narrow: separate the random variation built into a stable process from signals that something has actually changed [1].
That distinction matters because water treatment data is noisy by nature. Raw water quality shifts with weather, filters mature and backwash, dosing pumps cycle, analyzers recalibrate. React to every wiggle and you add your own variation to the process — chasing noise, over-dosing, resetting baselines so often that no signal survives. A control chart draws the line between “the process is talking” and “the process is shouting,” using the process’s own statistics to do it [1][2].
The Anatomy: Center Line and 3σ Limits
A Shewhart control chart has three reference features. The center line is the process average, established from a baseline period of in-control historical data; NIST’s Engineering Statistics Handbook, in its metrology applications, recommends a minimum of 100 historical check values to establish that accepted baseline [1]. Above and below the center line sit the upper and lower control limits (UCL and LCL), computed at three standard deviations of the plotted statistic — calculated from the process standard deviation, not chosen by hand and not copied from a regulation [1].
Why 3σ? Because it balances two errors. On a stable, normally distributed process, about 99.73% of points fall inside the ±3σ band, leaving roughly 0.27% — about three points in a thousand — that will breach a limit by chance alone when nothing is wrong [5]. Tighter limits flood operators with false alarms; wider limits hide real shifts until they get expensive. Three sigma is the working compromise refined over a century of practice [5].
One property of those limits deserves emphasis: they come from the process itself. That makes them descriptive, not prescriptive. They tell you what the process has been doing, and therefore what it will keep doing if nothing changes [1].
Two Charts to Start With: X̄-R and I-MR
Water quality data arrives in two shapes. Where operators pull composite or replicated samples — several readings grouped per shift — the X̄-R (averages and ranges) family applies. Where the plant has one reading per interval, which is the common case for online analyzers reporting a single value every few minutes, the right tool is the I-MR chart: a pair of charts tracking individual observations and their moving ranges [4].
The I-MR construction is simple enough for a spreadsheet. Take the absolute difference between each consecutive pair of values; average those to get MR̄. The individuals chart then uses a center line at X̄ with limits at X̄ ± 2.66 × MR̄, and the moving-range chart has its center line at MR̄ with an upper limit at 3.27 × MR̄ and a lower limit of zero [4]. Those constants embed the same 3σ logic — 2.66 converts the average two-point moving range into a three-sigma width — so the chart inherits all the interpretation rules below [4].
Common Cause vs Special Cause
Shewhart’s partition, later popularized by Deming, sorts every source of variation into two buckets [2]. Common causes are inherent to the process: the day-to-day scatter produced by its design, materials, and environment. They are stable and predictable, and the only way to reduce them is to change the system itself [2]. Special causes are assignable, local, intermittent — a failed probe, an air bubble in the flow cell, a spent chemical drum, a valve stuck mid-cycle — and they announce themselves as points or patterns the common-cause system would not produce [2].
The management rule follows directly. Special causes belong to operators: find the assignable event and correct it. Common causes belong to engineers and managers: improve the system, or accept its noise. Confusing the two is expensive in both directions. Ignore a special cause and a real failure keeps running; “fix” every common-cause fluctuation and you are tempering the process, an intervention pattern that adds variation rather than removing it [2].
Run Rules: Reading Patterns, Not Just Points
A single point beyond 3σ is the obvious signal, but processes usually announce drift gradually. Western Electric run rules and their 1984 refinement by Lloyd S. Nelson define additional patterns, each with roughly equal false-alarm probability, so that interpretation stays uniform across the control room [3]. Four rules cover most water quality work:
- Rule 1: one point more than 3σ from the center line — the classic outlier [3].
- Rule 2: nine or more points in a row on the same side of the center line — a mean shift, as when a filter’s effluent settles onto a new, slightly worse plateau [3].
- Rule 3: six or more points in a row steadily increasing or decreasing — a trend, the signature of fouling, depletion, or drift [3].
- Rule 4: fourteen or more points in a row alternating up and down — a bimodal or cycling behavior, worth checking against dosing or backwash cycles [3].
Applied together, these rules catch slow departures long before any limit is breached, which is the entire operational value of the method [3].
A Turbidity Drift Example, Step by Step
Take filter effluent turbidity, the parameter most plants already log continuously. Treatment facilities routinely record flow, raw water quality, chemical feed, turbidity, and finished water quality through SCADA or manual logs, so the raw material for control charts already sits in most historians and paper archives [6].
The starter workflow has four steps. First, pick a baseline window in which operations consider the filter to have run normally, and compute X̄ and MR̄ for the effluent turbidity signal [1]. Second, build the I-MR limits — X̄ ± 2.66 × MR̄ and 3.27 × MR̄ — and plot the baseline to confirm no rule violations, which validates the limits themselves [4]. Third, plot each new reading as it arrives and let the four rules speak: a Rule 3 run of steadily rising points on a filter that backwashed normally last week is a fouling or underdrain question to open today, not after the limit breach [3]. Fourth, act in the correct bucket — a special-cause signal sends the operator to that filter, its valves, and its analyzer; a chart that shows stable operation with a mean you dislike is an engineering project, not an emergency [2].
This is also where modern instrumentation changes the economics of SPC. Sensors that feed your AI water model, not just your dashboard — digital-output analyzers streaming Modbus data into a historian — make the baseline computation and run-rule checks automatic, so the chart reviews itself every minute while the team reviews it once a week.
Three Misreadings to Avoid
The first is treating trends as noise, waiting for a hard limit breach before reacting. The run rules exist precisely because a six-point trend is a statistically unusual event that arrives weeks before the excursion does. Ignore Rule 3 and you forfeit most of the chart’s early-warning value [3][5].
The second is confusing control limits with compliance limits. Specification limits are the voice of the customer or the regulator — where the water must perform — while control limits are the voice of the process, describing what it currently does. They come from different sources and answer different questions [5]. A plant can be perfectly in control and out of specification (a stable process that is wrong), or inside specification and out of control (a drifting process that has not yet failed). Both diagnoses demand different actions, and the chart is what tells them apart [5].
The third is rebuilding limits from live data. Control limits belong to the stable baseline period; recomputing them every time a point misbehaves launders real signals into the baseline. Set limits from a vetted historical window, review them on a defined schedule, and let the chart accumulate evidence between reviews [1].
| Chart element | What it tells you | Misuse warning |
|---|---|---|
| Center line (CL) | Long-run average from a vetted baseline period [1] | It is not a target or a spec; do not move it to make points fit |
| UCL / LCL at ±3σ | Band of pure common-cause variation, computed from process data [1][5] | Not compliance limits — specifications live in permits, not in statistics [5] |
| Point beyond a limit | Likely special cause; investigate locally [2] | Chasing common-cause noise point by point (tempering) [2] |
| Nine points one side (Nelson 2) | Sustained mean shift [3] | Normalizing the shift by accepting the new plateau silently |
| Six points rising or falling (Nelson 3) | Drift or fouling trend in progress [3] | Waiting for the breach before opening a work order |
| Moving-range chart | Short-term variation between consecutive readings [4] | Reading only the individuals chart and missing a variance change |
Scaling Up Without Drowning in Alarms
Once the first two charts prove themselves, the temptation is to chart everything and switch on every rule everywhere. Resist it. Each added rule raises the chart’s false-alarm rate, and a control room facing dozens of daily red flags will — rationally — start ignoring all of them [3]. The scaling discipline that works comes down to three things: one chart per parameter with its own baseline, never mixing statistics of different parameters into a single chart; a deliberately short list of enabled rules per parameter, matched to how that parameter fails in practice; and every alarm bound to a written response, so a flag is an instruction rather than a mood. Review the rule set on the same schedule as the limits, and retire rules that cry wolf more often than they catch [1][2][3].
Getting Started This Quarter
Pick two parameters — typically filtered-water turbidity and a distribution or process conductivity signal — and give each a 90-day baseline, an I-MR chart, and the four run rules [3][4][6]. Assign clear ownership: operators respond to special-cause signals, engineering owns common-cause improvement, and nobody adjusts limits without a documented review [1][2]. Within a quarter, the plant should be able to state, in numbers, how much of its historical “water quality drama” was noise — and act earlier and calmer on the rest.
References
- NIST/SEMATECH e-Handbook of Statistical Methods, section 2.2.2.1 — Shewhart control chart; baseline from historical data (minimum 100 check values), limits computed from process standard deviation — https://itl.nist.gov/div898/handbook/mpc/section2/mpc221.htm
- Quality Gurus — Special Cause vs. Common Cause Variation; definitions and Deming’s “tempering the process” warning — https://www.qualitygurus.com/special-cause-vs-common-cause-variation/
- Lean Six Sigma Definition — Nelson Rules; Rule 1 (3σ point), Rule 2 (nine points one side), Rule 3 (six-point trend), Rule 4 (fourteen alternating) — https://www.leansixsigmadefinition.com/glossary/nelson-rules/
- Lean 6 Sigma Hub — How to Create and Interpret an I-MR Chart; individuals limits X̄ ± 2.66 × MR̄ and moving-range limit 3.27 × MR̄ — https://lean6sigmahub.com/how-to-create-and-interpret-an-i-mr-chart-a-complete-guide-for-quality-control/
- Ucourse Academy — Why SPC control charts use ±3 sigma; the 99.73% / 0.27% false-alarm balance and control limits vs specification limits — https://ucourse.org/p/11772
- WaterWorld — Statistical Process Control Can Help Treatment Plants; SCADA-recorded flow, water quality, chemical feed, and turbidity data as SPC input — https://www.waterworld.com/home/article/16192665/statistical-process-control-can-help-treatment-plants
About the author: Written by the Shanghai ChiMay Technical Editorial Team — specialists in online water quality analyzers, multi-parameter sensing, and digital integration for industrial and municipal water treatment operations.
