Tiegel home Talk to a metallurgist

CUSUM charts for metallurgical plants: early detection of small changes in recovery

Constructing a CUSUM chart to detect small, sustained changes in plant recovery or grade earlier than daily charts, with a worked example and alarm rule.

· 6 minute read · By

A gold plant's recovery may average 92% while varying between 90% and 93% from day to day because of ore changes and sampling and assay error. A single day at 91% is not significant, since such results occur every week. If the average genuinely falls from 92% to 91%, however, each individual day still appears normal, and the change cannot be seen on a daily chart.

A CUSUM chart instead accumulates the daily deviations from the target. While the process is on target, positive and negative deviations cancel and the cumulative sum remains near zero. Once the average shifts, the sum moves a little further almost every day and the line develops a steady slope.

CUSUM (cumulative sum) charts were developed for industrial quality control and are well suited to metallurgical plants, where the most costly changes are often small and persistent. This article describes construction of a CUSUM chart in a spreadsheet, its interpretation and the setting of an alarm rule.

Construction of a CUSUM chart

St=∑i=1t(xi−T)=St−1+(xt−T)

where

St
cumulative sum on day t (same units as x)
xi
daily result, for example recovery (%)
T
target or baseline mean (%)
  1. Select the variable: recovery, concentrate grade, tailings grade, reagent use per tonne, or any daily or shift figure.
  2. Select the target T: the budget, the design value, or the mean of a recent stable period. The CUSUM shows departures from whichever target is chosen.
  3. In a spreadsheet, enter the daily values in one column and x − T in the next. In a third column, calculate the running total: the first cell equals the first difference, and each subsequent cell adds the new difference to the cell above.
  4. Plot the running total against date.

Worked example

A gold plant's target recovery is 92.0%. The daily results over 30 days are:

Daily gold recoveryDaily recovery scatters between about 90% and 93% with no obvious change to the eye.88899091929394151015202530DayRecovery (%)target 92%Daily recovery: 1 → 92Daily recovery: 2 → 92Daily recovery: 3 → 92Daily recovery: 4 → 90Daily recovery: 5 → 93Daily recovery: 6 → 93Daily recovery: 7 → 92Daily recovery: 8 → 93Daily recovery: 9 → 92Daily recovery: 10 → 92Daily recovery: 11 → 93Daily recovery: 12 → 92Daily recovery: 13 → 92Daily recovery: 14 → 91Daily recovery: 15 → 92Daily recovery: 16 → 91Daily recovery: 17 → 91Daily recovery: 18 → 90Daily recovery: 19 → 91Daily recovery: 20 → 90Daily recovery: 21 → 92Daily recovery: 22 → 91Daily recovery: 23 → 91Daily recovery: 24 → 91Daily recovery: 25 → 90Daily recovery: 26 → 92Daily recovery: 27 → 93Daily recovery: 28 → 90Daily recovery: 29 → 90Daily recovery: 30 → 90Daily
The daily chart. The drop after day 15 is hidden in the day-to-day scatter. Illustrative figures.

The daily chart shows scatter between about 90% and 93.5%. Day 4 was the lowest of the first two weeks and proved to be an isolated result, and no change is evident. The CUSUM of the same data is:

Cumulative sum of (recovery − 92%)The cumulative sum wanders around zero for 15 days, then falls steadily from day 16, a slope of about −1.1 percentage points per day.-15-10-50151015202530DayCumulative sum (percentage points)change startsalarm, day 22CUSUM: 1 → 0CUSUM: 2 → -0CUSUM: 3 → -0CUSUM: 4 → -2CUSUM: 5 → -1CUSUM: 6 → -0CUSUM: 7 → -0CUSUM: 8 → 0CUSUM: 9 → 0CUSUM: 10 → -0CUSUM: 11 → 1CUSUM: 12 → 0CUSUM: 13 → 0CUSUM: 14 → -0CUSUM: 15 → 0CUSUM: 16 → -1CUSUM: 17 → -2CUSUM: 18 → -3CUSUM: 19 → -4CUSUM: 20 → -6CUSUM: 21 → -6CUSUM: 22 → -7CUSUM: 23 → -8CUSUM: 24 → -8CUSUM: 25 → -10CUSUM: 26 → -10CUSUM: 27 → -10CUSUM: 28 → -12CUSUM: 29 → -15CUSUM: 30 → -17CUSUM
The same data as a CUSUM. A steady downward slope means the mean has fallen; the slope gives its size (about 1.1%). The tabular CUSUM alarm (k = 0.4, h = 4) fires on day 22. Illustrative figures.

For the first 15 days the CUSUM varies within about ±2.5 points of zero, indicating that the process is on target. From day 16 the line falls steadily, reaching −16.8 by day 30. A steady slope indicates a sustained shift in the mean.

The shift began at the last point before the steady slope, day 15, and the plant log for days 15 and 16 should be checked for a new ore source, a reagent change, an instrument recalibration or a sampler change. The magnitude of the shift is given by the slope: −16.8 points over 15 days is −1.1 percentage points per day, so recovery has averaged about 92.0 − 1.1 = 90.9% since day 16. The average of days 16 to 30 is indeed 90.9%.

The slope, not the height, carries the information. A CUSUM that stands at −5 but is flat indicates a past dip that has ended; a CUSUM at −1 that has begun to fall steadily is the one requiring action.

The tabular CUSUM and alarm rule

Interpreting a chart by eye requires judgement. The tabular CUSUM provides a formal rule:

Ct−=max[0,Ct−1−+(T−k)−xt]Ct+=max[0,Ct−1++xt−(T+k)]

where

Ct−
lower CUSUM: builds up when results run below target
Ct+
upper CUSUM: builds up when results run above target
k
allowance or slack, usually half the shift worth detecting (often 0.5σ)
h
decision interval: signal when either sum exceeds h (often 4σ to 5σ)
σ
standard deviation of the daily result when the process is steady

The allowance k prevents ordinary noise from accumulating, and the decision interval h sets the amount of evidence required before an alarm. A common starting choice is k = 0.5σ and h = 4σ to 5σ, where σ is the standard deviation of the daily values during a stable period. These settings detect a sustained shift of about one standard deviation within one to two weeks while giving few false alarms.

In the example, σ = 0.8 percentage points over days 1 to 15, so k = 0.4 and h = 4.0.

Day1516171819202122
Recovery (%)92.490.991.490.591.190.391.791.2
Lower CUSUM C⁻0.00.70.92.02.53.83.74.1

For day 16, C⁻ = max[0, 0.0 + (92.0 − 0.4) − 90.9] = 0.7. The lower CUSUM exceeds h = 4.0 on day 22, giving an alarm one week after the change from daily results that never fell below 90.3%. The single poor result on day 4 raised C⁻ only to 1.6 before it returned to zero the following day.

Plant applications

CUSUMs of recovery and tailings grade detect slow losses caused by wear, ore changes or reagent problems. In metal accounting, a CUSUM of the difference between two measurements of the same quantity, such as plant feed by weightometer and by survey, or assayed and calculated head grade, exposes drift and bias in instruments and sampling.

In reconciliation, a CUSUM of monthly F-factors shows whether a model bias is long-standing or new; see gold reconciliation. In plant trials, a CUSUM of the difference between a test line and a control line shows the effect of the trial early, before the formal statistical analysis is complete.

Limitations

The choice of target matters. A CUSUM plotted against an unrealistic budget slopes indefinitely, so a recent stable period should be used, or T reset when a change is accepted as the new normal.

Daily results are not independent, because stockpiles and gold in circuit carry one day's effect into the next. Strong carry-over causes CUSUM alarms to occur early; weekly averages or a corrected σ reduce the effect.

A CUSUM signal identifies when a change occurred and its size, but not its cause, which requires an investigation. That step is covered in the metallurgical investigation method.

Glossary

CUSUM
Cumulative sum: the running total of each result's difference from a target.
Target, T
The reference value the CUSUM measures departures from.
Shift
A lasting change in the average of a process.
Standard deviation, σ
A measure of day-to-day scatter when the process is steady.
Reference value, k
The allowance subtracted each day so ordinary noise does not build up in a tabular CUSUM.
Decision interval, h
The threshold the tabular CUSUM must cross to signal an alarm.
Tabular CUSUM
A form of CUSUM that resets to zero and gives a numerical alarm rule.
False alarm
An alarm when nothing has actually changed.

Frequently asked questions

What is a CUSUM chart used for?

Detecting small, sustained shifts in a process average faster than an ordinary control chart, and estimating when the shift began and its size.

How is a CUSUM chart interpreted?

From the slope. A flat or wandering line indicates the process is on target; a steady downward or upward slope indicates the mean has moved. The slope per day equals the size of the shift.

What values of k and h should be used?

A common starting point is k equal to half the standard deviation and h of four to five standard deviations, adjusted for the size of shift to be detected and the acceptable rate of false alarms.

Can a CUSUM be built in Excel?

Yes. Subtract the target from each value in one column, calculate a running total in the next, and chart it. For the tabular form, use =MAX(0, previous + (T − k) − x) for the lower sum.

Discuss your plant with a metallurgist

Tiegel is a Perth metallurgy company. We help with testwork, plant optimisation, metal accounting and investigations, and we build the Tiegel Suite software.

Talk to a metallurgist

References and further reading

  1. Page, E.S. (1954). Continuous inspection schemes. Biometrika, 41(1/2), 100–115.
  2. NIST/SEMATECH e-Handbook of Statistical Methods. Section 6.3.2.3, CUSUM control charts. Link
  3. Napier-Munn, T.J. (2014). Statistical Methods for Mineral Engineers: How to Design Experiments and Analyse Data. JKMRC Monograph Series, University of Queensland.

Worked examples and charts marked illustrative use invented figures, not data from any operation.