Skip to main content Skip to search Skip to main navigation

Knowledge · Tools

Allan deviation calculator: read ARW and bias instability from your own data

Free online toolYour data stays in the browserNovember 2026

−½+½0.664 · Bτ (log)σ(τ) (log)

QUICK ANSWER

The Allan deviation σ(τ) shows how the noise of a gyro or accelerometer averages out over time. On a log-log plot, white noise falls with slope −½ and gives the angle random walk at τ = 1 s. The flat minimum divided by 0.664 is the bias instability. Load a log of your sensor at rest, or simulate one from the data sheet values of an Epson IMU, and compare both curves.

01

How do I read an Allan deviation plot?

  1. 01Slope −½: white noise. Read σ at τ = 1 s. With σ in °/h, the angle random walk is σ(1 s) / 60 in °/√h.
  2. 02Flat part: bias instability. Flicker noise sets a floor. The bias instability is B = σmin / 0.664.
  3. 03Slope +½: rate random walk. For long averaging times the curve rises again. Temperature drift often shows up here as well.

The table shows what the data sheet values of four Epson IMUs mean on the plot. The last column is the averaging time at which the model curve is within 5 % of its floor.

IMUARW (°/√h)σ at 1 s (°/h)Bias instability (°/h)Floor 0.664 · B (°/h)Floor reached at τ
M-G570PR200.042.40.50.338.5 min
M-G370PDG00.063.60.80.537.5 min
M-G366PDG00.084.81.20.805.9 min
M-G330PDG00.16.031.991.5 min

Own calculation from typical data sheet values, gyro axes, without rate random walk.

02

How do I record usable data?

  • Put the sensor at rest on a stable surface without vibration.
  • Wait for warm-up and keep the temperature constant.
  • Record the raw output at a fixed sample rate, without additional filtering.
  • Record long enough: the tool evaluates τ up to one ninth of the record, so that every point rests on at least 9 clusters. For the bias instability of an M-G370PDG0 that means a little more than one hour, several hours are better.
  • Export one axis as one column. Enter sample rate and unit, then load the file.

The tool removes the mean value, integrates the rate and computes the overlapping Allan deviation. It then estimates the white-noise coefficient from the region with slope −½ and the bias instability from the smoothed minimum, with the 1σ uncertainty that follows from the number of clusters.

03

What is the math behind it?

σ²(τ) = N²/τ + (2·ln 2/π)·B² + K²·τ/3

N is the white-noise coefficient (angle or velocity random walk), B the bias instability and K the rate random walk. √(2·ln 2/π) is the factor 0.664. Quantization noise and rate ramp are part of the standard as well, but left out here. The overlapping estimator uses all 2m-spaced differences of the integrated signal θ: σ²(τ) = Σ(θk+2m − 2θk+m + θk)² / (2τ²(M − 2m)), with τ = m/fs. With n independent clusters the relative 1σ error of a point is about 1/√(2(n − 1)).

The simulated logs are generated from the same model: white noise, flicker noise as a sum of first-order Gauss-Markov processes spaced half a decade apart, and an optional random walk. If you run a simulation, the orange points have to follow the model curve within this statistical scatter.

04

Why does my curve differ from the data sheet?

  • Temperature changes and warm-up lift the right-hand side of the curve.
  • Vibration and mounting add noise that the sensor does not produce itself.
  • A short record gives few clusters at long τ, so the minimum scatters by 20 to 30 %.
  • Internal filters and decimation change the left-hand side of the curve.
  • Data sheet values are typical values. Epson states the bias instability as an average.

How the different bias lines of a data sheet relate to each other is explained in How to read an IMU datasheet. For the effect on navigation, try the IMU drift simulator.

FAQ

Frequently asked questions

Your question is not listed?

Ask our engineers →

QWhat is the difference between Allan variance and Allan deviation?

The deviation is the square root of the variance. Data sheets and this tool use the deviation, because it has the same unit as the signal.

QHow long do I have to record?

Until the curve becomes flat. For an M-G370PDG0 the model reaches its floor at an averaging time of about 7.5 minutes. With at least 9 clusters per point this needs a little more than one hour of data.

QIs my data uploaded to a server?

No. The file is read and processed locally in your browser. Nothing is transmitted.

QWhich value do I compare with the data sheet?

Compare the ARW with the coefficient from the −½ region and the bias instability with σ_min / 0.664. Some manufacturers state σ_min directly, without the factor 0.664, so check the footnote.

EVALUATION KIT

Measure the Allan deviation of an Epson IMU yourself

For qualified projects we provide evaluation kits as a free loan. You record the data under your own conditions and compare it with the data sheet here.