1. Why Bias Stability Matters in MEMS Accelerometers
MEMS (Micro‑Electro‑Mechanical Systems) accelerometers are ubiquitous in everything from smartphones to autonomous vehicles. Their ability to detect tiny accelerations hinges on the stable behavior of the sensor’s bias—a constant offset that can drift over time due to temperature, aging, or power cycling. Even a few micro‑g (µg) of bias drift can corrupt navigation solutions, degrade inertial measurement unit (IMU) performance, or trigger false alarms in safety‑critical systems.
Bias stability is defined as the maximum deviation of the sensor’s bias over a specified period. It is often expressed in units of µg or, for more granular analysis, in terms of Allan deviation (σᴀ(τ)) versus averaging time (τ). Allan variance provides a rigorous statistical framework to separate different noise processes (white noise, flicker noise, random walk) and to pinpoint the averaging time at which bias drift dominates.
Actionable Insight
In mission‑critical applications, always verify the MEMS accelerometer’s bias stability during procurement and after integration. A sensor that meets specifications in the lab may exhibit drift once exposed to real‑world temperature cycles and mechanical shocks.
2. Sensor Physics and Substrate Fundamentals
2.1. The Core of a MEMS Accelerometer
A typical capacitive MEMS accelerometer consists of a proof mass suspended by micro‑scale springs. When the device experiences acceleration, the mass deflects, altering the capacitance between the mass and fixed plates. The resulting voltage change is converted to an acceleration reading.
The key parameters influencing bias stability are:
- Spring constant (k) – determines sensitivity to acceleration.
- Mass (m) – influences dynamic range and noise floor.
- Capacitance change (ΔC) – directly linked to measurable voltage shift.
- Substrate material – silicon, glass, or composite; each has distinct thermal expansion, piezoelectric, and aging characteristics.
2.2. Substrate Thermal Effects
Silicon’s coefficient of thermal expansion (CTE) is about 2.6 × 10⁻⁶ /°C, while many MEMS packages use glass or polymer encapsulants with higher CTEs. Mismatched CTE can induce stress during temperature swings, causing micro‑fractures or changes in the spring constant, thereby shifting the bias.
Real‑World Example
The Bosch BMA280 uses a silicon‑on‑insulator (SOI) substrate with a low‑temperature‑coefficient (LTC) design that mitigates thermal drift, achieving a bias stability of 0.02 mg over ±70 °C.
2.3. Aging and Mechanical Stress
Mechanical fatigue or creep in the suspension springs leads to a gradual change in k over months or years. Packaging stresses, such as those introduced during hermetic sealing or solder reflow, can also embed permanent strain that manifests as bias drift.
3. Allan Variance Fundamentals
3.1. What is Allan Variance?
Allan variance (σ²ᴀ) is a time‑domain measure of frequency stability originally developed for oscillators. For accelerometers, it quantifies the statistical variance of the bias over different averaging times (τ). The Allan deviation (σᴀ) is simply the square root of Allan variance.
Mathematically:
\[
\sigma^2_{\text{A}}(\tau) = \frac{1}{2(N-1)} \sum_{i=1}^{N-1} \left( \bar{y}_{i+1} - \bar{y}_{i} \right)^2
\]
where \(\bar{y}_i\) is the average bias over the i‑th interval of duration τ, and N is the number of such intervals.
3.2. Noise Types and Their Signatures
| Noise Process | Allan Deviation Slope (σᴀ vs τ) | Physical Origin |
|---|---|---|
| White noise | τ⁻¹ᐟ² | Thermally driven electronic noise |
| Flicker noise | τ⁰ | 1/f noise in electronics or material defects |
| Random walk | τ¹ᐟ² | Bias drift, temperature fluctuations |
| Bias instability | τ⁰ | Dominant at the optimum averaging time |
Plotting σᴀ(τ) on a log‑log scale reveals these slopes, enabling engineers to isolate the noise source affecting bias stability.
3.3. Calculating Allan Variance: Step‑by‑Step
- Collect Time‑Series Data – Record a long, continuous bias trace (e.g., 10 000 samples at 1 kHz).
- Choose Averaging Times (τ) – Commonly 1 s, 2 s, 5 s, 10 s, 30 s, 60 s, etc.
- Segment the Data – Divide the trace into non‑overlapping blocks of duration τ.
- Compute Block Averages – For each block, calculate the mean bias.
5