Three stages

1. Obtain physical-layer signal measurements (I/Q samples) on SoC
2. Implement lightweight feature extraction pipeline on SoC
3. Deploy ML classifier on SoC

System on Chip: nRF52833

BLE Direction Finding

• high-accuracy indoor location services
• two modes for determining signal direction
  • Angle of Arrival (AoA)
  • Angle of Departure (AoD)
• leverage angular phase-shift information from antenna array
• Constant Tone Extension (CTE): unwhitened sequence of modulated ones
• I/Q sampling during CTE

Different Protocol Options

• BLE: 1Mbit/s with GFSK, 2Mbit/s with DQPSK
• nRF proprietary: 1Mbit/s, 2Mbit/s (unspecified modulation)
• 802.15.4: data rate 250kbit/s, DSSS chip rate 2Mchip/s with O-QPSK

System Parameters

• DFECTRL1.DFEINEXTENSION allows to start antenna switching/IQ sampling in payload
• DFEMODE set to AoD disables antenna switching at receiver, samples continuously
• DFECTRL1.TSAMPLESPACING allows for maximum sampling rate of 8MHz

System Constraints

• DFEPACKET.MAXCNT restricts buffer size to 2¹⁴ − 1 bytes
  • at 8MHz sampling rate corresponds to 16383B / 32us/B ⋅ 1/8us = 63B payload
• DFECTRL1.NUMBEROF8US restricts IQ sampling time to (2⁶−1) ⋅ 8us = 504us
  • at 250kbit/s which is 4us/bit = 32us/B corresponds to 15.75B payload

Fixed Setup

• PHY layer: 802.15.4 @ 2.4GHz
• PPDU without SHR: 1+12+2B PHR + PHY payload + MAC CRC
• timing: roughly 32us/B ⋅ 15B = 500us for payload
• sampling at 8MHz for 500us results in 4000 IQ samples

Analysis of Obtained Data

Large-Scale Data Acquisition

• Anechoic chamber
• Rx: 2 nRF52833
• Tx: 30 nRF52840 = 15 DK + 15 Dongle
• 1000 frames per device pair @ 50 frames/s
  • need for fast data retrieval from SoC
• Tx power: 0dBm
• Distance 1..5 m depending on SNR

Transition to Stage 2

• Look into time recovery and fine-frequency recovery
• Offline feature extraction with Matlab/Python
• Translate to C for on-board feature extraction, compare accurancy to offline approach