How To Calibrate Sensors

Caution

This page is in notes format, and may not be of the same quality as other pages on this site.

More often than not, when designing embedded systems, you’ll come across the need to calibrate a sensor. Calibration is required when the sensor output does not directly give you the required measurement (e.g. it gives you a voltage, but you need to convert it to a temperature), or when accuracy is required above and beyond the base (i.e. non calibrated) accuracy that the sensor manufacturer provides. This page covers the different ways to calibrate sensors.

Caution

There is a well known adage within the embedded/electrical engineering community: All sensors are temperature sensors, some measure other things as well.

This is a good reminder to consider other variables when calibrating sensors, especially temperature! Temperature has a rather annoying property of causing many other measurements to change or sensors to respond differently.

Linear One Point Calibration

Linear one point calibration is the simplest form of calibration. It makes the assumption that the sensor is linear and goes through the origin. You are fitting a straight line in the form $y = ax$ (note there is no $+ b$ term, which means the line goes through the origin).

Linear Two Point Calibration

Linear two point calibration is the next step up from linear one point calibration. It is simple and only requires two points of data. It makes the assumption that the sensor is linear, but does not assume it must go through the origin. We fit a straight line in the form $y = ax + b$ using two points of data, which is just enough to uniquely determine the line.

Linear Best Fit

Linear best fit is the method of fitting the data the best you can with a straight line. The straight line is commonly written as $y = ax + b$, where $a$ is the slope and $b$ is the y-intercept. The line is usually fit using the least squares method. This is where we find the line which minimizes the sum of the squared residuals. Minimizes the sum of the squares is not the only way, but is one of the most popular techniques. One reason is that typically a residual that is twice as large as a another residual is more than twice as bad¹. Squaring the residuals penalizes larger residuals more than smaller residuals and gives them more weight.

Requires a source of truth, a reference sensor which you can consider to be the “true” value.

Whilst Excel and other spreadsheet programs can do a linear best fit, I almost always prefer to write up a simple Python script (that creates graphs) as it gives me more control and flexibility. I normally type (or automate) the data into a CSV file during calibration and read that in (Python can also read spreadsheet files if you want to use that instead).

The $R^2$ value is a measure of how well the data fits the line. The closer the $R^2$ value is to 1, the better the fit.

Residuals are the left over variation in the data after accounting for model fit.

$$\text{residuals} = \text{data} - \text{model}$$

One purpose of residual plots is to identify patterns still apparent in the data even after fitting a model. This may give you some useful insight into how you can further improve the model.

EXAMPLE: Flow Sensor Calibration

Let’s calibrate a water flow sensor using a linear best fit. This is a basic hall-effect and impeller based flow sensor. It outputs a digital signal which transitions at a rate which is dependent on the flow rate. We measure this frequency with a microcontroller and output the data to serial. We also have a accurate flow meter that we trust as our “source of truth”.

The sensor gave us the following data which we have saved to a CSV file:

data.csv

actual_flow_rate_lpm,measured_frequency_hz
0,0
0.46,0
0.82,0
0.92,0
1.05,8.66
1.23,10
1.43,15
1.71,23.00
2.07,24
2.39,32.33
2.86,41.33
3.49,52.33
3.90,60.66
4.74,77.99
5.44,93.99
6.10,108.66
7.10,127
7.83,142.65
3.16,48
1.69,18
1.00,8.33
0.80,5.33
0.70,0

If we plot the data (I used Python), we get the graph shown in This is a placeholder for the reference: fig-flow-sensor-data. $Q$ is the actual flow rate in L/min and $f$ is the measured frequency in Hz.

A plot of the raw data we obtained from the flow sensor. We will use this to calibrate the sensor by performing a linear best fit.

Let’s fit a linear line to the data (a.k.a linear least squares regression). I used the NumPy np.polyfit() function to do this, but almost any programming language and/or statistical software package can do this (see the page Linear Curve Fitting for more information on how the line of best fit is calculated).

Flow sensor calibration: Frequency vs Flow Rate.

This gives us the equation $Q = 0.0511f + 0.657$ with an $R^2 = 0.993$.

One thing you’ll note is that the flow sensor does not produce any output for very low flow rates. Rather than include this in our linear fit, it is better to exclude them and build logic to assume the flow rate is 0 when the frequency is 0. This would give use a piece wise function:

$$\begin{align} Q = \begin{cases} 0 & \text{if}\ F = 0 \\ ax + b & \text{if}\ F > 0 \end{cases} \end{align}$$

If we ignore all data points where the frequency is 0, we get the linear best fit shown in This is a placeholder for the reference: fig-flow-sensor-calibration-nonzero.

A linear best fit to the data, excluding all data points where the frequency is 0.

Now we get a slightly different equation with $Q = 0.0505f + 0.709$ and an improved $R^2 = 0.998$. Done! If we wanting to now use this sensor in some embedded firmware, we might write a function like this to convert our frequency to a flow rate:

double calcFlowRateFromFrequency(double frequency) {
    if (frequency == 0) {
        return 0;
    }
    return 0.0505 * frequency + 0.709;
}

To make this more robust, you might want to assert() that the frequency is not negative, and also perhaps if it is too large (i.e. far outside the bounds of the data used for calibration, where extrapolation is likely to be inaccurate).

Polynomial Best Fit

Polynomial best fit is the method of fitting the data the best you can with a polynomial curve. The polynomial curve is commonly written as $y = a_0 + a_1 x + a_2 x^2 + ... + a_n x^n$, where $a_0, a_1, ..., a_n$ are the coefficients of the polynomial.

Non-Linear Best Fit

Further Reading

See the Linear Curve Fitting page for more information on linear least squares regression.

Footnotes

1. David Diez, Christopher Barr, & Mine Çetinkaya-Rundel. 7.3: Fitting a Line by Least Squares Regression. LibreTexts Statistics. Retrieved 2025-10-10, from https://stats.libretexts.org/Bookshelves/Introductory_Statistics/OpenIntro_Statistics_(Diez_et_al)./07%3A_Introduction_to_Linear_Regression/7.03%3A_Fitting_a_Line_by_Least_Squares_Regression. ↩