Reader · receive Chapter 20 of 21
Phase Extraction
arctan2(Q, I) — turning two baseband signals into the angle that carries the sensor data
What it is physically
Phase extraction is done in digital signal processing — either in the reader's FPGA or in its host software. The ADCs sample I(t) and Q(t) at ~2–5 MSPS. A CORDIC algorithm or lookup table computes arctan2(Q, I) per sample. The reader then averages the phase over each tag response (the ~100 µs EPC packet) to get one phase value per read.
In Impinj readers (widely used in RFID research), the firmware reports the phase as a 12-bit value (0–4095 mapping to 0°–360°) along with the EPC and RSSI for every tag inventory event. The sllurp library reads this via the LLRP protocol.
What it does — the computation
From the IQ mixer outputs:
- I = (A/2)·cos(φ) — the in-phase component
- Q = (A/2)·sin(φ) — the quadrature component
- φ = arctan2(Q, I) — the full 360° phase angle
- A = 2·√(I² + Q²) — the amplitude (RSSI)
arctan2 is the two-argument arctangent that resolves all four quadrants. It returns the angle of the complex number I + jQ on the unit circle, measured counterclockwise from the positive I axis.
What's in the measured phase
The reported phase φ contains EVERYTHING in the signal path:
- Propagation delay: 2×distance × 2π/λ (round trip). At 915 MHz, λ = 32.7 cm, so 360° per 16.4 cm of distance change.
- Reader TX/RX phase offset: fixed for a given reader and frequency channel.
- Antenna phase: the antenna's own phase response.
- Tag reflection phase: depends on the load impedance → THIS is what carries the sensor data.
- Multipath: reflections from walls, floor, body add random phase errors.
- Frequency hopping: each channel has a different propagation phase.
A differential design cancels all of these EXCEPT the sensor by differencing two tag ICs (A and B) that share the same antenna, same path, same reader offset, same multipath.
Why differential phase matters
- A single tag's phase is nearly useless for sensing — multipath alone can swing it by tens of degrees as people or objects move
- Δφ = φ_A − φ_B cancels all common-mode terms, leaving only the sensor-dependent part of φ_A
- The sensor's impedance change shifts IC A's reflection phase; the size of that shift is the signal
- The ratio of that shift to the residual phase noise sets the sensing resolution
- Averaging many reads per tag brings the noise down further, at the cost of latency
Phase on the Smith chart
The tag's reflection coefficient Γ is a complex number. On the Smith chart (normalized impedance plane), Γ's angle is directly related to the measured phase. When a capacitive sensor across the tag's port changes value:
- The load impedance Z₁ moves along a constant-conductance circle on the Smith chart
- Γ₁ rotates — its angle changes
- The reader sees this as a phase shift in the structural return
- The relationship is approximately linear over the small range of capacitance change
Typical values and precision
- Phase resolution: 12-bit (0.088° per LSB) in Impinj readers
- Single-read phase noise: ~2° RMS at 1m with no motion
- After 60s averaging (2000 reads): ~0.5° (theoretical √N improvement)
- Phase ambiguity: wraps at 180° (the reader can't resolve 180° ambiguity from a single channel). Continuity across reads resolves this.
- Frequency dependence: each 500 kHz channel gives a slightly different phase (used for RFID localization; a differential Δφ largely cancels it)
Signal flow
Input: Digitized I(t) and Q(t) baseband samples from the ADCs (12-16 bit, 2-5 MSPS).
Processing: CORDIC or arctan2 lookup per sample. Average over the tag's response packet. Per-channel calibration offset subtracted. Report alongside EPC and RSSI.
Output: One phase value (0°–360°) per tag per inventory round. For a differential sensor, φ_A and φ_B are reported separately (different EPCs); host software computes Δφ = φ_A − φ_B and applies a calibration curve to recover the sensed quantity.