24/7 emergencies 814 777 6921812 230 6079WhatsApp 811 786 0768
Electrónica MonterreyElectrónica MonterreyLaboratorios Creativeit Talk to an engineer
Menu

How is the uncertainty of an indirect measurement evaluated?

In an indirect measurement, uncertainty is evaluated by combining the uncertainties of the input quantities weighted by their sensitivity coefficients, which are the partial derivatives of the function that defines the measurand.

Approach

Many quantities are not measured directly: power is obtained from voltage and current, a resistance from voltage and current, a density from mass and volume. A model y = f(x1, x2, …) is defined that relates the measurand to the input quantities, and the uncertainty of each input is evaluated.

Steps

  1. Write the complete mathematical model, including corrections (temperature, buoyancy, drift) that also contribute uncertainty.
  2. Obtain the standard uncertainty of each input, either statistically (Type A) or from other information (Type B).
  3. Calculate the sensitivity coefficients: the partial derivative of f with respect to each input, evaluated at the estimated values.
  4. Combine: the combined uncertainty is the square root of the sum of the squares of each coefficient-by-uncertainty product, plus covariance terms if the inputs are correlated.
  5. Expand by multiplying by a coverage factor consistent with the desired confidence level.

An example with power

If power is the product of voltage and current, the coefficient of the voltage is the current and the coefficient of the current is the voltage. For products and quotients it is practical to work with relative uncertainties: if the inputs are uncorrelated, the relative uncertainties add in quadrature. An exponent in the model multiplies the corresponding relative uncertainty.

Correlations

If two inputs are measured with the same instrument or share a standard, their errors are not independent. Ignoring correlation can overestimate or underestimate the uncertainty, depending on the sign of the coefficients. For example, when two readings from the same instrument are subtracted, the common error partly cancels.

When the linear method is not enough

Analytical propagation assumes that the model is approximately linear around the input values and that the output distribution is close to normal. If the model is highly non-linear or a component with a non-normal distribution dominates, propagation of distributions by Monte Carlo simulation is used.

What an engineer sees in practice

The result is presented as a budget table: each row is an input with its uncertainty, distribution, sensitivity coefficient and contribution. That table reveals which input dominates and, therefore, where it pays to invest in better standards or procedures.

We do real tests, on video! Watch them on our YouTube channel →

Do you have a problem nobody has been able to measure?

Talk directly to the engineer in charge, with no middlemen.

Talk to an engineer