High-Traffic Format · Technical Perspective · Published: October 27, 2026
Table of Contents
Key Takeaways
- In classical control-chart practice, limits sit three standard deviations either side of the mean, and only about 0.3% of results — roughly 3 in 1,000 — would land outside by chance. That is why such a point deserves an investigation rather than a shrug [1].
- Trend reading turns hindsight into foresight. Ammonia creeping from 2 mg/L to 6 mg/L over a year may never trip an alarm against a 10 mg/L permit limit, and yet the violation is visible months out [3].
- Sensor drift is confirmed operational reality. Major utilities note that pH and residual chlorine sensors drift over time, and that slow drift is difficult for operators to catch in time [4].
- Rule-based pattern checks — 8 consecutive points on one side of the center line, or 6 points moving monotonically — catch instability long before a single point breaches a hard limit [2].
- Telemetered values can appear on a public dashboard before anyone has checked them, which is why quality codes and status flags are part of professional reading, not decoration [5].
- Reacting to ordinary variation as if it were a real change (tampering) demonstrably increases output variation. Deming’s funnel experiment puts the penalty at about a factor of √2 versus leaving a stable process alone [2].
A water quality dashboard is easy to glance at and hard to read. Glancing means checking whether anything is red. Reading means extracting what the process is doing, whether it is changing, and what to do next — before the alarm fires. The difference matters, because most process damage happens inside green numbers quietly trending the wrong way. Below is the method experienced process engineers use, one rule at a time, ending with a field guide you can pin next to the HMI.
Start With the Trend, Not the Number
Look at trajectory before magnitude. A single result tells you what happened at one instant. A trend tells you where things are going. That is the whole value of trend analysis: it shows whether concentrations are rising, falling, or flat, and it flags emerging compliance risk while the response is still cheap [3].
The arithmetic is simple enough to do at the console. Say effluent ammonia sits at 2 mg/L and creeps up by about a third of a milligram per liter each month against a 10 mg/L permit limit. At that rate twelve months gets you to roughly 6 mg/L — still compliant, still climbing. No single sample looks alarming. The trend is unambiguous anyway.
Which is why daily reports built on one snapshot mislead. When you open the dashboard, give the 30-day view ten seconds before you give the live view three.
Tell a Step Change From a Drift
Once you see movement, classify it. The two families of change call for different responses.
A step change is sudden — a jump between consecutive readings. Typical causes: a pump switching over, a dose pump losing prime, a probe being cleaned or replaced, or a genuine raw-water event. If the jump lines up with a maintenance event, it is probably instrument, not water.
A drift is slow and directional. The baseline itself is moving. Drift deserves respect because it is both dangerous and hard to see. Utilities note that sensors measuring crucial parameters such as pH and total residual chlorine drift over time, and that the slow rate is exactly what makes it hard to detect in a timely way [4]. A drifting sensor does not alarm. It lies smoothly.
So the slow downward creep of chlorine residual, or the slow upward creep of a turbidity baseline, should trigger a calibration check while every value still reads comfortably in range.
The dashboard habit: whenever a parameter moves, ask two questions — how fast, and since when. Fast and coincident with an event usually means instrument or operating state. Slow and persistent usually means fouling, aging, or a real process shift.
Read Against Operating Context: Flow, Dosing, Cleaning
No parameter means anything in isolation. Experienced readers pull up the context underneath the quality trace: plant flow, chemical dosing rate, backwash and cleaning events, filter run times, temperature.
Three context checks catch most misreadings:
- Flow correlation. Conductivity, turbidity, and chlorine residual all respond to hydraulic loading. A turbidity bump that starts exactly when flow doubles is hydraulics, not filter failure.
- Dosing correlation. A pH rise starting minutes after a caustic dose step-up is control action, not a raw water change. Overlay the dosing trace before panicking about the pH trace.
- Event correlation. Cleaning cycles, probe swaps, and calibration routines leave fingerprints: a clean gap, a step, a flat line. Multi-parameter instruments reporting pH, conductivity, ORP, and temperature on one shared timestamp make this cross-reading much easier, because the traces align by construction.
The point of context reading is not to explain every wiggle. It is to stop explaining real signals away as “probably just the flow,” and to stop panicking over artifacts that actually are the flow.
Know Your Limits: Control, Alert, Action
Most dashboards show several horizontal lines. Confusing them is the classic amateur error. Professional reading separates three tiers.
Statistical control limits describe the process. Classical practice places them three standard deviations either side of the mean; assuming a normal distribution, only about 0.3% of results — 3 in 1,000 — would fall outside by chance, so such a point justifies concluding the process is no longer in statistical control [1]. These limits come from your own data, typically after at least 20 replicate results, with warning lines at two standard deviations sitting inside them [1].
Alert (or warning) limits are internal early-warning thresholds. Values here are not violations. They are the cue to look harder: more frequent checks, a calibration verification, a glance upstream.
Action (or permit) limits are the legal or contractual boundaries where a defined response is mandatory — investigation, reporting, corrective action. Environmental monitoring manuals formalize this as action-and-limit levels tied to an event-and-action plan.
The reading rule is to know which line your eye is crossing. A point outside statistical limits with no permit line crossed is a measurement-quality question. A point inside statistical limits but above your alert line is a trend question. Treating every red pixel as identical wastes operator attention on the wrong tier.
Pattern rules extend this. Before any single point breaches anything, control-chart practice watches sequences: 8 consecutive points on one side of the center line, or 6 points trending monotonically up or down, are standard signals that a special cause has arrived [2]. Train your eye on runs and drifts, not just breaches.
Trust but Verify: Data Quality Flags
A number on a dashboard is a claim, not a fact, until you know how it got there. Professional readers check data quality status as part of every review, and monitoring programs formalize it with quality codes.
Victoria (Australia)’s Water Measurement Information System is a good model: every data point carries a quality code indicating reliability, and telemetered values that appear before they have been checked or edited are flagged explicitly [5]. If a public monitoring network needs to label data “provisional,” a plant dashboard needs it more.
The flag set to expect — and to configure — on any serious dashboard:
| Flag | Meaning | How to read it |
|---|---|---|
| Offline / no data | Sensor not communicating | Check whether the gap is telemetry or sensor power before trusting neighbours |
| Maintenance / calibration | Value generated during service or adjustment | Exclude from statistics; a “clean” value here is an artifact |
| Out of range | Reading beyond physical or configured validity | Suspect the sensor first; a physically impossible value is an instrument event, not a water event |
| Flat-line | Identical value repeated abnormally long | Often a stuck sensor or a frozen communication loop — real water never holds still |
Flats and flags matter because statistics silently absorb bad data. A monthly average that includes maintenance-mode zeros or calibration artifacts is not an average of your plant. It is an average of your instrument events.
This is also the deeper case for modern instrumentation: a sensor that feeds an AI water model, not just a dashboard, has to carry its own status metadata, or the models inherit the lies.
Zoom Out: Daily, Weekly, Monthly Views
The same data answers different questions at different zoom levels. Professional reading cycles through all three on purpose.
The live and daily view answers “is the process stable right now?” Watch excursions, alarm states, and leading indicators — flow, dose, differential pressure.
The weekly view answers “is the baseline moving?” Overlay week-on-week profiles, check whether nightly minimums are drifting, compare weekends to weekdays. Slow fouling and chemical feed degradation show up here first.
The monthly view answers “are we drifting toward a limit, and what did last month cost?” Compliance summaries, chemical consumption per volume, percent of time within each band. Automated reporting from a historian turns this into an export instead of a compilation chore, but the discipline stays the same: compare like with like, and confirm every summary against the underlying quality flags.
A Field Guide to Dashboard Patterns
| View | What to look at | Abnormal pattern | Recommended action |
|---|---|---|---|
| Live trace | Current value vs. band | Step jump between consecutive readings | Check for pump switch, dosing event, or probe service before assuming water change |
| Live trace | Rate of change | Value sweeping steadily, not oscillating | Treat as directional event; verify with secondary measurement |
| Daily overlay | Day-vs-day shape | Today’s curve shifted vs. weekly norm | Correlate with flow and dosing logs; inspect upstream operations |
| 7–30 day trend | Baseline level | Slow monotonic drift, still inside limits | Schedule calibration check; model time-to-limit [3][4] |
| 7–30 day trend | Variability | Noise amplitude growing | Check sensor fouling/agitation; review control loop health [2] |
| Alarm panel | Run-based signals | 8 points one side of center; 6 monotonic points | Investigate special cause even if no limit breached [2] |
| Any view | Data quality flags | Offline, maintenance, flat-line markers | Exclude flagged data from statistics; verify instrument state [5] |
The Over-Adjustment Trap
One warning closes the method. The most expensive reading error is not missing a trend. It is reacting to ordinary variation as though it were a real change. Deming’s funnel experiment showed that tampering with a stable process increases output variation by roughly a factor of √2 compared with leaving it alone [2].
On a dashboard, that looks like chasing every wiggle with setpoint changes until the process oscillates worse than it did before. The professional stance is the reverse of the amateur’s: act on special causes — steps, runs, drifts — leave common-cause noise alone, and spend the attention you save on calibration and data quality.
Read in this order — trend, context, limits, flags, zoom — and a dashboard stops being a wall of numbers. It becomes what it was supposed to be: a running conversation with your process.
References
- Maharashtra Pollution Control Board — Guidelines for Water Quality Monitoring — regulatory guidance on Shewhart control charts: control limits at three standard deviations (about 0.3% of results expected outside by chance), warning lines at two standard deviations, and charts built from 20 or more replicate results. https://www.mpcb.gov.in/sites/default/files/water-quality/standards-protocols/GuidelinesforWQMonitoring%5B1%5D.pdf
- Quality Assurance Authority — Statistical Process Control Standards — reference for run-based special-cause signals (8 consecutive points on one side of the center line, 6 monotonically trending points) and the funnel-experiment finding that tampering increases variation by a factor of about √2. https://qualityassuranceauthority.com/quality-assurance-statistical-process-control/
- Ecesis — Water Quality Trend Analysis Software — explanation of why trends outperform single results, with the ammonia example rising from 2 mg/L to 6 mg/L over twelve months against a 10 mg/L permit limit. https://www.ecesis.net/Water-Quality-Software/Water-Quality-Trend-Analysis-Software.aspx
- PUB, Singapore’s National Water Agency — Sensor Data Integrity Monitoring — challenge statement documenting that pH and total residual chlorine sensors drift over time and that slow drift is difficult for operators to detect in a timely manner. https://www.pub.gov.sg/Industry/Events%20and%20Programmes/Global%20Innovation%20Challenge/Challenges/2020/Sensor%20Data%20Integrity%20Monitoring
- State of Victoria (Australia) — How reliable is the data? (Water Measurement Information System) — description of quality codes assigned to every data point and of telemetered values flagged as unverified before checking. https://data.water.vic.gov.au/how-reliable-data
- Shanghai ChiMay Tech ltd. — About us — company overview of Shanghai ChiMay’s online water quality analyzer product line referenced in this article. https://www.chimaytech.net/it?p=520
About the author: Written by the Shanghai ChiMay Technical Editorial Team — Shanghai ChiMay’s editorial group for its online water quality analyzer line, drawing on daily work with multi-parameter sensing, Modbus-based data integration, and plant-side monitoring workflows.
