uv-k1-k5v3-firmware-custom/App/apps/aprsrx/README.md

APRS RX: on-radio APRS receiver (work in progress)

Goal: an overlay app that receives APRS (AX.25 UI frames, Bell 202 AFSK 1200 bauds) on the UV-K1 / UV-K5 v3 and shows the last frame, then an APRS TX app.

Status:

Step	State
Hardware demodulation by the BK4829 FSK block	Dead end (2026-09-25, FT3D on 144.800: the chip has no Bell 202 demodulator)
Audio path an app can sample	Done by EPIRB 406: RX audio on PA4, ADC channel 4 at 9.6 kHz
Integer demodulator, modelled on synthetic audio (test/model_rx.py)	Done (see below)
Flipper Zero test transmitter (test/flipper_aprs.py)	Done, files in test/flipper/
Radio app (aprsrx_app.c)	v0.1 receives the Flipper frames on the radio; v0.3 decodes Mic-E (on air: F5RAV via F1PRY-14); v0.4 decodes continuously with 3 slicers (on air: F1PRY-14 via F5KTR-3); v0.5 shows standard positions; v0.6 adds the speaker key; v0.7 adds the corrected 20x20 APRS symbol bitmaps (not built here)
Bench test with the Flipper on 433.650 MHz	Done (2026-09-30, v0.1: _long 10/10 in STD; _badfcs not shown)
Real station: FT3D beacon on 144.800 MHz	Done (2026-09-30, v0.1: Mic-E frame >TXUPX9 received, FCS good, = 48°50.89' N 2°16.25' E)
Mic-E decoding (test/mice.py: spec encoder vs the app's decoder)	Done in the model: 6 cases + the FT3D frame

Using the app

1. Set the VFO to the APRS frequency, FM: 144.800 MHz, or 433.650 MHz for the Flipper test files.
2. Launch APRS RX.
3. The speaker is off (v0.6): the decoder does not need it, only the BK4829 AF output (tested on the radio: PA4 joins the audio before the amplifier). Key 1 turns it on to listen to the channel, as FoxHunt's audio key.

Since v0.4 the app decodes continuously, as a TNC does: no RSSI trigger, the FCS and a UI-frame check (control 0x03, PID 0xF0) sort frames from noise. Up to v0.3, captures started 10 dB above a tracked noise floor; on a busy 144.800 that floor crept up (−132 → −104 dBm), so weak stations were never sampled. Several frames in one transmission are now received too.

Keys and the screen cost sample time (the key scan alone is ~400 µs), so they are served every 50 ms only while no slicer is inside a preamble (two back-to-back flags) or a frame whose first bytes look like a callsign (busy ~4 % of the time on noise, in the model), and after 3 s of busy at the latest. The screen is redrawn after a new frame, a key, or every 5 s.

Screen	Content
Status bar	APRS RX title, then a WAIT capsule until the first frame, the speaker icon while the speaker is on, then in the scroll view ▲ while rows are hidden above and ▼ while rows are hidden below (x = 72-76 as in APRS TX, the speaker icon at x = 59-68; both drawn in one asset read)
Lines 0-3	The frame in the small font (18 characters a row, non-ASCII shown as .), 4 rows at a time, scrolled pixel by pixel with UP/DOWN (v0.7): source call and frame number in bold (F4HWN-7 #3), >DEST,DIGI*,... (as many path entries as fit two rows), then the info field, whole. For Mic-E (most Yaesu/Kenwood beacons, v0.3): latitude 48 50.89N, longitude 002 16.25E, speed km/h, course and symbol, the message type (Off Duty, En Route...), then altitude and comment without the device markers; for an uncompressed position (! = / @, v0.5): latitude, longitude, then the timestamp if any, the symbol (table + code) and the comment. Compressed positions stay raw text. A new frame scrolls back to the top. Key * switches to the compact view (below)
Lines 4-6	Under a dotted separator (y=31), the frequency drawn as the main screen draws it (big digits up to the kHz, the last two in the small font), aligned left. A matching 20x20 bitmap from the 48-symbol Yaesu set is shown at x=108 on the right for decoded Mic-E and uncompressed positions; the area stays blank for an unknown or unavailable symbol
Bottom row y=49	ok 12 -89dBm: frames received, RSSI at the end of the last frame. (v0.4-v0.5 also showed frames per slicer, sl a/b/c: on air the outer slicers found frames too; dropped in v0.6 to fit the speaker key)

Keys (UV-K5 and UV-K1):

Key	Action
UP/DOWN (held)	Scroll the frame 1 px per 50 ms slot (UV-K1: LEFT/RIGHT, as nav_dir), from the source row down to the last row
*	Scroll view / compact view, saved (the flash is written only after a change). Compact: the source in bold, then the path and 3 info rows in the tiny 3x5 font, 32 characters each (position 48 50.89N 002 16.25E on one row, speed, course, symbol and message type on the next, then the comment), as up to v0.6; no scroll
1	Speaker on/off (off at launch); FoxHunt's speaker icon in the status bar while on
MENU	Clear the last frame and the counters
EXIT	Quit

The receive path is the firmware's own (STD: 300 Hz high-pass, de-emphasis, 3 kHz low-pass). Up to v0.3, key 1 switched to RAW (filters and AFC off, as EPIRB 406); every on-air decode was made in STD and the model shows no gain for RAW, so v0.4 dropped it to fit the 4 KiB overlay.

The last good frame is kept on app_main's stack and the demodulator state on capture's: the 4 KiB overlay also holds .bss. The screen texts, symbol bitmaps and the two cos tables are assets.

Demodulator (test/model_rx.py, class Demod; the C is a transcription)

Per ADC sample (9.6 kHz, 8 samples per bit):

1. DC removal (1-pole tracker, 64 samples).
2. Band-pass biquad centred on √(1200·2200) = 1625 Hz, Q 0.9 (Q14): equal gain on both tones, removes the out-of-band noise the box correlators let through.
3. Four sliding correlators over one bit: I/Q at 1200 Hz (8-entry cos table) and 2200 Hz (48 entries = 11 cycles). The sine is the cos table read 3/4 period ahead (+6, +12). Products >> 8, so the decision below never overflows int32.
4. Magnitudes max + 3/8 min, per-tone peak trackers (attack 1/32, decay 1/1024): each tone normalised by its own peak, so the twist (pre-emphasis or not, de-emphasis or not) does not bias the decision much.
5. 3 slicers (since v0.4), decision sign(wa·Mm·Ps − wb·Ms·Pm) with wa:wb = 2:3, 1:1, 3:2, each with its own DPLL, NRZI and HDLC below (Direwolf's multi-slicer idea): the AGC does not cancel a strong twist in noise. A frame from any slicer is accepted; a copy from another slicer within 1 s is dropped.
6. DPLL: 65536 per bit, a bit at each wrap; each transition pulls the phase 1/4 of the way toward mid-bit + half a sample (the transition is only seen at the next sample: without that half-sample, a +1 % clock failed long frames).
7. NRZI, HDLC (flag, destuffing, abort after 7 ones), CRC-16/X.25 per byte, frame accepted on the residue 0xF0B8 and as a UI frame: noise brings ~0.7 candidates per second to the FCS check, so without the UI check a false frame would pass about once a day.

Slicers (test/variants.py, 4 seeds, 24 hard cases: noise 3000-4000 Hz, ±1 % clock, twist −9 to +9 dB at noise 2500 Hz, RAW and STD)

Variant	Frames /96
1 slicer (v0.1-v0.3)	57
3 slicers, weights ×2	65
3 slicers, weights ×1.5 (v0.4)	71
5 slicers (×1.5 and ×2)	71

The gain is on strong twist: sine AFSK at +6 / +9 dB in RAW goes from 0/4 to 4/4. Noise only (30 s, 2 seeds, 50 and 400 LSB rms): no frame accepted.

Model results (test/model_rx.py, 3 seeds per case, 1 slicer; 3 slicers: all ≥ these)

Channel: discriminator output + white noise (rms over 48 kHz) → radio audio path (RAW: 5 kHz low-pass; STD: 300 Hz high-pass, 750 µs de-emphasis, 3 kHz low-pass) → 12-bit ADC at 9.6 kHz, 0.065 LSB/Hz (EPIRB 406 measurement), optional clock error. 150 ms of no-carrier noise before and after each burst.

Case	RAW	STD
Flipper, clean: pos, long (233 bytes)	3/3, 3/3	3/3, 3/3
Flipper badfcs (must be rejected)	0 frames	0 frames
Flipper, noise 1500 / 3000 / 4500 Hz	3/3, 2/3, 0/3	3/3, 2/3, 0/3
Flipper, ±4 kHz carrier offset	3/3	3/3
Flipper long, ±1 % sample clock	3/3	3/3 (−1 %), 3/3 (+1 %)
Sine AFSK (a real station), twist −6 / 0 / +6 dB, noise 1500	3/3, 3/3, 2/3	2/3, 3/3, 3/3

Noise 4500 Hz is Eb/N0 ≈ 9.6 dB: the theoretical frame success of non-coherent FSK on a 520-bit frame is then ~6 %, so the failures there are expected. (At ±0.5 % clock, 6/6 in every variant tried.)

Flipper Zero test transmitter (test/flipper_aprs.py)

The Flipper's CC1101 cannot put an audio tone on FM, only switch between two frequencies (preset 2FSKDev238Async, ±2.38 kHz). Each AFSK tone is therefore sent as a square wave: the carrier toggles at twice the tone frequency, phase-continuous across bits, with edge times rounded from exact values (no drift). The receiver's discriminator outputs that square wave and its audio low-pass leaves mostly the fundamental. The Flipper cannot reach 2 m anyway; the script refuses 144-146 MHz.

test/flipper_aprs.py test/flipper [--call F4HWN] [--freq 433650000]

writes aprs_pos.sub, _digi, _msg, _status, _long (233 bytes, 2 s) and _badfcs (one info bit flipped: must not be shown). 100 ms of carrier, 40 flags (267 ms), the frame, 3 flags, 10 ms of carrier. Copy them to subghz/ on the Flipper, set the radio to 433.650 MHz FM, launch the app, then Send one file at a time, about 1 s apart. (Up to v0.4, UP/DOWN tuned ±5 kHz for a Flipper whose crystal puts the carrier off the channel; never needed on the bench, dropped in v0.5 to fit 4 KiB.)

test/ax25.py builds the frames (shared by the generator and the model); test/mice.py checks the Mic-E display decoder against a spec encoder, and the standard position display against real frames; test/variants.py compares demodulator variants on the hard cases (dev tool).

Bench checklist

• Level reaching the ADC (pp, shown up to v0.4, dropped in v0.5 to make room): estimated ~300 LSB peak-to-peak for the Flipper (±2.38 kHz at 0.065 LSB/Hz), measured 734-778 with the Flipper and 916 with a real station (F1PRY-14): ~20 % of full scale, plenty of headroom, and the per-tone AGC makes the absolute level irrelevant.