Skip to content

accuracy.md

Latest commit

 

History

History
92 lines (75 loc) · 4.25 KB

accuracy.md

File metadata and controls

92 lines (75 loc) · 4.25 KB

Accuracy: prediction vs. real-world survey

Predictive RF planning is only as honest as its error bars. This page documents what Plexus's model does and does not capture, and a repeatable method for validating predictions against measurements in a real building — using nothing but the app's built-in survey import and a free WiFi analyzer.

Status: methodology is final; the reference dataset is being collected. If you run the procedure below in your own building, a PR adding your results table (anonymised floor plan welcome) would be a great contribution.

What the model computes

With a floor scale set (SCALE, metres per 100 px), predicted RSSI at a point is:

RSSI = EIRP − PL(d) − Σ wall losses
PL(d) = 20·log₁₀(f_MHz) − 27.55 + 10·n·log₁₀(d_m)
  • EIRP = per-AP Tx power + antenna gain − cable loss (defaults 20 dBm + 0 − 0).
  • f = a representative carrier per band: 2 437 MHz (2.4 GHz), 5 500 MHz (5 GHz), 6 525 MHz (6 GHz).
  • n = the propagation model's exponent: log-distance 2.2, ITU-R P.1238 office 3.0, COST-231 multi-wall 2.0 (with walls accounted explicitly).
  • Wall losses = per-material dB (drywall 3, wood 5, glass 6, brick 10, concrete 15, at 5 GHz) × a band factor (0.6 / 1.0 / 1.3 for 2.4 / 5 / 6 GHz) for every wall segment crossed on the straight line AP → sample point.

Known sources of error

Be suspicious of any tool that doesn't list these:

  • No multipath / reflections. Rays are straight lines; corridors and metal surfaces that duct or mirror signal are not modelled. Expect the model to under-predict down long corridors and over-predict behind large metal obstacles (racks, elevators, fridges).
  • Single representative frequency per band — no per-channel variation.
  • Wall materials are your guess. A "drywall" wall with a steel stud every 40 cm is lossier than 3 dB. Calibrate materials, not just positions.
  • 2D only. Mount height and downtilt feed EIRP, but propagation is planar; floor-to-floor leakage uses a single uniform slab attenuation.
  • Client asymmetry. Predictions are AP-side EIRP; a phone's measured RSSI also depends on its own antenna. Expect a constant offset of a few dB per device model.

Validation procedure

  1. Plan first. Load your floor plan, set SCALE accurately (use ⌖ Calibrate with a known distance — a door is ~0.9 m), trace the walls you can see, and place APs where they really are. Set each AP's real model and band.

  2. Walk the building with any WiFi analyzer that logs RSSI (e.g. a phone app or airport -s / nmcli dev wifi scripted on a laptop). At each stop, note your position on the floor plan and the RSSI of your SSID. 20–40 points spread across rooms and corridors is plenty; include a few worst-case spots (behind the server room, far corners).

  3. Build the CSV. One row per stop:

    x,y,floor,ssid,bssid,rssi,channel
0.32,0.41,Ground Floor,MyOffice,aa:bb:cc:dd:ee:ff,-58,36
0.71,0.18,Ground Floor,MyOffice,,-66,

    x/y are fractional (0–1 across the floor-plan image) or raw pixels — both are auto-detected. Only x, y, rssi matter for validation.

  4. Import via ↺ Survey CSV. Each sample renders as a dot coloured by its measured RSSI; hover for Measured / Predicted / Δ. Dots with |Δ| > 8 dB get a warning ring — those are your model errors (or your wall-material guesses being wrong).

  5. Iterate. Fix the biggest |Δ| first: usually a wrong material, a missing wall, or the wrong propagation model. In furnished offices ITU-R P.1238 (n = 3.0) usually beats the free-space-like models; in open warehouses, log-distance does.

Reporting results

When publishing a validation run, include at minimum:

Metric Value
Building / construction type e.g. 1990s office, drywall + concrete core
Sample count n
Mean error (bias) x.x dB
Mean absolute error x.x dB
90th-percentile abs. error x.x dB
Propagation model used logd / itu-indoor / multi-wall

For context: commercial predictive tools are typically considered usable at ±5–8 dB mean absolute error against a calibrated survey; anything tighter than ±5 dB on an uncalibrated first pass is luck, not skill.