Sensors

Accelerometer — MEMS sensing, axes, and g-force

An accelerometer measures proper acceleration — not velocity, and not quite what intuition suggests either. A sensor sitting motionless on a desk reads about 1g straight up, not zero, because it's measuring the force the desk applies to hold it against gravity, not gravity itself.

Proper acceleration, not velocity

"Proper acceleration" is what an accelerometer physically measures: the acceleration relative to free-fall, felt as a mechanical force on the sensing element. A sensor in true free-fall (no air resistance) reads zero on every axis, even though it's clearly accelerating downward from an outside observer's point of view — and a sensor sitting still on a table reads +1g, because the table is actively pushing back against gravity. This distinction — measuring the force felt, not the "actual" acceleration in an inertial frame — is the single most common source of confusion for anyone using one for the first time.

How MEMS sensing actually works

Inside the chip is a microscopic proof mass suspended by flexible silicon beams (springs) between fixed capacitor plates. When the sensor accelerates, the mass's inertia makes it lag relative to the frame around it — the mass shifts slightly, changing the gap (and therefore the capacitance) between it and the fixed plates on each side. An on-chip ASIC measures that tiny capacitance change and converts it into a digital value.

Axes, sensitivity, and full-scale range

Most accelerometers report three axes (X, Y, Z) simultaneously, each with its own independent proof-mass/plate structure (or a shared mass sensed along multiple directions, depending on the design). Two configuration values matter for interpreting the output:

Full-scale range

The maximum acceleration the sensor can represent — commonly selectable, e.g. ±2g, ±4g, ±8g, ±16g. A wider range tolerates bigger shocks but loses resolution for small movements.

Sensitivity

How many raw LSBs correspond to 1g at the selected range — given directly in the datasheet (e.g. "16384 LSB/g at ±2g" for a 16-bit output).

From raw registers to g

The raw value read from the sensor's data registers is a signed integer (two's complement) that needs converting using the sensitivity for whatever full-scale range is currently configured:

int16_t raw_x = (int16_t)((high_byte << 8) | low_byte);
float g_x = raw_x / sensitivity_lsb_per_g;   // e.g. / 16384.0f at ±2g
float ms2_x = g_x * 9.80665f;                 // if you want SI units

Getting the byte order and the signed-vs-unsigned cast wrong here is an extremely common first bug — a value that should read near zero while sitting flat instead reads a huge number, or flips sign at exactly the wrong moment.

Reading it: I2C or SPI

Most accelerometers support both I2C and SPI, selectable via a dedicated pin — I2C for fewer wires, SPI when higher output data rates are needed. Either way, reading a sample is the same shape: write the address of the first data register, then read six consecutive bytes (X low/high, Y low/high, Z low/high) in one transaction so all three axes represent the same instant, rather than three separate reads that could straddle a value changing mid-read.

Interrupt-driven detection

Constantly polling for events wastes power and CPU time, so most accelerometers can generate a hardware interrupt on a dedicated pin (serviced the way GPIO and interrupts describe) for conditions detected entirely on-chip:

• Tap / double-tap — a brief high-frequency spike on an axis, filtered from the slower gravity component.
• Freefall — magnitude across all three axes drops near zero, since a free-falling sensor (briefly) feels no proper acceleration at all.
• Activity / inactivity — magnitude staying above or below a threshold for a configurable duration — the basis of most "wake on motion" designs.
• Orientation change — which axis is currently closest to +1g, used for simple portrait/landscape detection.

Real-world applications

• Phone screen rotation — the classic example: comparing the gravity vector against each axis decides portrait vs. landscape.
• Fall detection — wearables and medical alert devices watch for the freefall-then-shock signature described above.
• Vibration/condition monitoring — industrial accelerometers mounted on motors and bearings detect abnormal vibration patterns that indicate developing mechanical faults.
• Pedometers and activity trackers — step counting is fundamentally peak-detection on the acceleration magnitude signal.
• Drone and robot stabilization — combined with a gyroscope (see Motion & IMU) as the core sensor feeding any flight or balance controller.

Gotchas