Skip to Content
DocsUse CasesTracking Wash Cycle Quality

Use case · manufacturing

Tracking wash-cycle quality

A 90 bar pump drives a high-pressure cleaning cell. Precision-machined metal parts ride through it in cycles — each cycle either holds spec or does not. The raw outlet pressure is noisy with sensor dropouts and idle periods between washes, so the question is not whether the pump is running. It is whether each wash delivered consistent quality.

A winkComposer flow cleans the pressure signal, detects wash cycles, scores each cycle against specification limits (Cpk), and tracks sensor health as dropout frequency increases over the run. Drag the slider to move the reading point across the trace — clean wash cycles scored by Cpk, sensor dropouts caught and cleaned, and the health verdict escalating as the sensor degrades.

Loading wash cycle quality...

What You’re Seeing

The pressure trace shows two lines. The faded grey line is raw outlet pressure — look for the sharp drops to near zero during wash cycles in the second half of the run. Those are sensor dropouts where the data acquisition system briefly loses the signal. The cyan line is the cleaned signal after spike rejection: each dropout is replaced with the median of its neighbours, restoring the true wash pressure.

The green bands mark detected wash cycles. The first eight cycles (samples 0–830) arrive at high sampling rate — roughly 70 samples per cycle, each lasting about 3.7 minutes. After sample 830, the sampling rate drops to 60-second intervals. Cycles become thinner (3–5 samples each) and sensor dropouts appear.

The zoomed spec chart below the main chart clips the Y-axis to 82–100 bar so the 85–98 bar spec band dominates the frame. Each plateau is one wash cycle; idle periods are dimmed to keep the temporal rhythm visible without competing for attention. The dashed orange lines are the upper and lower spec limits; the dotted line is the 90 bar target. A plateau that rides tight to target has high Cpk; one that skews toward a limit — or dips below the frame — is marginal or incapable. Cpk is computed once per cycle, at the moment the wash ends: hover a cycle-end sample for its Cpk, hover mid-cycle to see the previous cycle’s.

The Cpk panel shows per-cycle process capability scores as they appear. Cpk measures how tightly the wash pressure stays within specification limits (85–98 bar). A score above 1.33 means capable; below 1.0 means incapable. Notice Cycle 8 — its Cpk drops below 1.0 because a startup transient pulls pressure down to 79 bar. Scores are confidence-coded: full opacity for reliable cycles (30+ samples), faded for marginal (10–29), dim for unreliable (fewer than 10).

The assessment card fuses four signals into one health verdict. The sensor-quality driver highlights with each dropout spike. The Cpk driver highlights while per-cycle Cpk drops below 1.33. The sample-confidence driver highlights while the cycle sample count drops below 30. The duration driver highlights on abnormally short washes (below 3 minutes). In the first half (clean signal, 70-sample capable cycles, normal duration) all four signals stay within their normal ranges and the verdict sits at Healthy. As dropouts accumulate, cycles thin out, and durations shorten in the second half, combined evidence climbs and the verdict escalates through Monitor toward Degraded.

How It Works

One winkComposer flow, eight building blocks plus a health assessment:

validatePressurevalidates gauge range (0–120 bar)→ flags out-of-range readingsspikeFilterdetects single-sample sensor dropouts→ cleanPressure: median-filtered signaldetectWashidentifies wash cycles on cleaned signal→ above 60 bar = washing, with hysteresiswashTimertracks each cycle’s duration and duty cycle→ duration in ms at each falling edgeinvertWashsemantic correction at cycle end→ wasWashing = true when wash just endedanalyticsGateflushes per-cycle statistics at wash end→ triggers flush on washDur ≠ nullwashPressureTWper-cycle mean, stdev, N, CV, min, max→ flushed by controller, not by windowpressurePIper-cycle Cpk against spec limits→ USL 98, LSL 85 barwashQualityfuses 4 signals into one health verdict→ spikes × Cpk × samples × duration

Range validation catches out-of-gauge readings before any analysis begins. Pressures outside 0–120 bar are flagged — a first line of defence against sensor faults.

Spike rejection detects single-sample sensor dropouts. The pressure drops from 90 bar to near zero for one reading, then recovers. A three-sample median filter identifies these events: if a sample differs from both its neighbours by more than 30 bar, it is a spike. The filter replaces it with the median, producing a clean pressure signal that downstream nodes can trust.

Wash detection uses the cleaned signal with a threshold at 60 bar and 5 bar of hysteresis. Above 60 is washing; below 55 is idle. The hysteresis band prevents chatter at transitions — the pump does not flicker between states when pressure is near the boundary.

Dwell-time tracking measures each cycle’s duration in milliseconds and the pump’s duty cycle. Duration appears at each falling edge (wash ends); duty cycle appears after two complete transitions.

Semantic correction inverts the wash flag at cycle end. The dwell-time tracker reports duration when the state changes from washing to idle — but at that moment isWashing is false. The inversion produces wasWashing = true, the correct signal for “a wash cycle just ended.”

Per-cycle flush — a controller monitors the duration field and fires a flush command to the statistics accumulator at every wash end. This is the production recipe: the accumulator runs continuously, and the controller determines when to emit a complete cycle’s statistics.

Per-cycle statistics accumulate the cleaned pressure signal into mean, standard deviation, sample count, coefficient of variation, min, and max. The window is intentionally large (100,000) so it never auto-completes — only the controller’s flush triggers emission. Wash samples dominate each accumulation window.

Per-cycle Cpk is computed from the accumulated mean and standard deviation, scored against specification limits (USL 98 bar, LSL 85 bar). Cpk confidence depends on sample count — cycles with 30 or more samples produce reliable scores; fewer than 10 is statistically unreliable. The process capability index is described in Montgomery (2009).

Health assessment fuses four signals into one verdict:

  • Sensor quality — spike magnitude; each dropout (~88 bar) builds evidence for the sensor-quality signal
  • Process capability — evidence builds while per-cycle Cpk drops below the capable threshold of 1.33
  • Sample confidence — evidence builds while per-cycle sample count drops below 30; a cycle with only 3 samples cannot produce a reliable Cpk
  • Cycle duration — evidence builds on abnormally short washes (below 3 minutes); sensor-dropout fragments can trigger false cycle-ends lasting seconds

Evidence fades between events with a half-life of 50 samples. The node self-calibrates against the first stretch of clean washes, so a pump that normally runs a little noisier than this one would hit the same verdict scale on its own baseline. Combined evidence across signals determines the verdict — Healthy, Monitor, Degraded, or Critical — and the recommended action. Each signal’s intensity and persistence appear alongside.

References

  • D. C. Montgomery, Introduction to Statistical Quality Control, 6th ed., Wiley, 2009. The standard reference for process capability indices (Cpk, Cp, Pp).
  • Dataset: 1,500 readings from an industrial washing pump, 3-second and 60-second sampling intervals, anonymised. Single outlet pressure channel. Real plant floor data from a production installation.

Next Steps

Last updated on
Classifying Driving ConditionsOverview