Improvements of Shepherd v2

Overview

About Shepherd v1.x (from git)

Shepherd is a testbed for the batteryless Internet of Things, allowing to record harvesting conditions at multiple points in space over time. The recorded data can be replayed to attached wireless sensor nodes, examining their behavior under the constraints of spatio-temporal energy availability.

About Shepherd v2.x

Previous description is still correct, in detail:

embedded linux board with analog frontend serves as a shepherd node
harvesting-circuit can record an energy trace from an energy-source like solar, thermoelectric, …
emulation-circuit replays the energy-trace for a real target (user programmable uController) and records the drawn energy

Version 1 is build around a BQ25504-IC (plus capacitor) as harvester for recording and emulating various environments. This is a valid approach, but limited by design:

BQ-MPPT-controller is slow to react on changing environmental conditions by design (check every 16s with 64 ms downtime)
(VOC-)mppt-controller is fairly simple with pre-set setpoint (e.g. 76 % of open-circuit voltage for solar cells)
recorded harvest-environment is specific to harvest-IC + configuration
setup / configuration (especially) with storage cap is fairly static (combination of MPPT-Controller, Boost-Converter, Cap)

Version 2 improves the project by generalizing the basic idea and therefore virtualizing the harvester, storage capacitors and voltage converters. By switching to a software defined / virtual harvester-system and power-source the disadvantages from v1 vanish, with further advantages.

Virtual Source

schematic_vsource

full customization per config-parameters (e.g. boost-voltage, capacitor size, …)
- currently 10 parameters for the harvester, and
- 29 parameters for the converter-stage
harvesting-recorder has already several strategies implemented
- constant voltage (CV)
- MPPT, based on open circuit voltage or perturb & observe
harvesting can be postponed by sampling
- iv-curves from the energy-source
- model-parameters like voc and isc for solar (higher rate than iv-curves)
emulation handles harvested energy-traces or harvests directly from the iv-curves and simulates a customizable converter-stage
power-stage has already predefined templates for several setups
- direct throughput of traces
- diode + capacitor (converterless)
- buck-boost-converter (e.g. BQ25570) including the power-good-signal and efficiencies of underlying converters
- buck-converter (BQ25504)
the design enables users to automate testing of harvesting-setups e.g. by sweeping through parameters like storage-cap-size
the emulator and harvester consist of a purpose-fully chosen combination of low-noise and high-speed DACs, ADCs and Instrumentation-Amplifiers
- both circuits can handle 0 - 4.2 V (hrv even 5.0V) with up to 50 mA current
- resulting resolution 18 bit, step size is ~ 200 nA and 20 uV for voltage and current
the software-implementation updates in real time with 100 kHz

General improvements for shepherd

synchronization of sampling-trigger for node network optimized from ~ 2.4 us to under 500 ns
jitter of sampling trigger improved from about +- 600 ns to under 100 ns
two target-ports available, selectable via setup-parameter
two parallel power rails available for the targets (one with current measurement, switchable)
one default target with a nrf52-module
9 GPIO-Channels between target and system (pru-monitored)
- partly bidirectional, switchable
- relatively stable 4.03 MHz, min: 602 kHz, max: 4.55 MHz
- previously ~ 160 - 1500 kHz, with 6 GPIO
watchdog to handle hangups during unsupervised operation (and wakeups for saving power)
screw-in power-socket or type-c connector

Currently under development

software programmer (in pru) for SWD & SpyBiWire
advanced target with msp430 (with FRAM) and nrf52 for various configurations: only 1 uC active, msp as processor and nrf as radio, nrf as processor+radio and msp as low-energie storage
combined network of shepherd nodes as iot-testbed with web-based control
update doc
advanced PRU-Firmware without virtual harvester & source, but higher min rate for gpio-sampling (goal: > 5 MHz)
GPIO-Actuation

Bigger challenges in near future

hardware is getting quite old, but luckily software is improving (getting faster)
kernel 4.19 currently works (but has flaws) -> latest updates to 5.10 break kernel-module & pru
chip-shortage is limiting our supply - 2 ICs only 15 left, current restock date mid ‘23

Features in Detail

disclaimer: graphs and stats are partly outdated
scraped from the planning-git

Hardware

in Hardware

Shepherd Cape v2.4

chip shortage forced us to a ultra fine pitch (350 um) - risky and hard to hand-solder
360 components, 55 unique - > tedious, but ok to hand-solder
switch to 6 layer due to higher gpio-count
1 Free Pin left on BBone

shpcape24b

shpcape24

nRF52 FRAM Target v1.0

Target with MSP430 (FRAM) & nRF52 on one PCB

nrf52msp430target1

Harvester Circuit

Schematic

schematic_hrv

Diode selection

datasheets only promise < 40 nA

hrv_diodes

Fine tuning Filters

hrv_untuned

hrv_tuned

Resulting Performance

hrv_profiled

Virtual Harvester

either harvest right away (MPPT, constant voltage) -> iv-stream
or defer the harvesting by sampling ivcurves (or isc & voc)
configurable by 10 parameters

Implemented harvest-algorithms (& parameters)

Algorithm	Parameters
ivcurve	window size, v_min, v_max, wait-cycles, direction
isc & voc	wait-cycles
v-const	voltage
mppt-voc	setpoint, t_interval & t_duration (voc-measurement)
mppt-po	v_min, v_max, v_step, t_interval

Emulator Circuit

lowest resolution, set by software / resistor
- DAC 19.53 uV
- ADC 190 nA
- voltage set in < 8 us
switches and traces get compensated on PCB (Feedback is coming from target-header pin)

schematic_emu

Performance

at 50 mA around 3.8 V are usable without large error

emu_profiled

Virtual Source

General Features

integrated into PRU, calculated and updated at 100 kHz
fully customizable per yaml-parameter-set (29+ parameters)
- predefined sets by name ie. “virtsource: BQ25504s” for the BQ-Regulator with additional schmitt-trigger for pwr-good
- neutral parameter-set is base -> direct throughput
- inherit from existing parameter-sets -> only add altered parameters in new set
emulator can either record output or intermediate voltage (storage cap)
naming: source = harvester + converter
design enables users to automate testing of harvesting-setup e.g. by sweeping through parameters like storage-cap-size

schematic_vsource2

Examples for predefined parameter-sets

direct throughput of traces
simple diode + capacitor
buck-boost-converter (e.g. BQ25570) including the power-good-signal and efficiencies of underlying converters
buck-converter (BQ25504)

Input

oneway, imagine a perfect diode at the start so no current can flow back
diode voltage-drop can be configured from 0 to x Volt
maxima for input voltage and current (power limit)

Boost-Converter, optional

enable minimum threshold voltage for input
disable maximum threshold for boost-output (intermediate voltage)
efficiency factor with 2D-LUT (12×12),
- depending on input voltage & current
- thresholds are configurable in 2^n steps
- voltage divisions are linear, depending on lowest threshold
- current-divisions are log2, also depending on lowest threshold
- example: voltage threshold n=7 is setting first array boundary to 2^7 = 128 uV, so lut[0] is for V < 128 uV, lut[1] is for 128 to 256 uV

Capacitor, optional

capacitance from 1 nF to 1 F
initial voltage
leakage current
switchable output, hysteresis with checks at defined intervals
power-good-signal with hysteresis either in intervals or immediate (schmitt-trigger)

Buck-Converter, optional

fixed output voltage
ldo-drop-voltage, alternatively working like a diode when buck is off or intermediate voltage is below output-voltage + drop-voltage
efficiency factor with 1D-LUT
- depending on output-current
- threshold is configurable in 2^n-steps
- current-divisions are log2, depending on lowest threshold
- example: current threshold n=5 is setting first array boundary to 2^5 = 32 nA, so lut[0] is for I < 32 nA, lut[1] is for [32, 64] nA, lut[2] is for [64, 128] nA

Switchable output

simulated external Capacitor - should be set to buffer size of target: fast transients can’t be fully monitored by shepherd

GPIO to Target

GPIO Implementation

Pin-Name	2nd FN	Ctrl	Dir	Pru-Mon
GPIO 0		dir1-pin	Rx-Tx	yes
GPIO 2		dir1-pin	Rx-Tx	yes
GPIO 3		dir1-pin	Rx-Tx	yes
GPIO 1		dir1-pin	Rx-Tx	yes
GPIO 4			always RX	yes
GPIO 5			always RX	yes
GPIO 6			always RX	yes
GPIO 7	uart rx		always RX	yes
GPIO 8	uart tx	dir2-pin	Rx-Tx	yes
BAT OK			always TX	(yes)
SWD1 CLK	jtag TCK		always TX
SWD1 IO	jtag TDI	pDir1-pin	Rx-Tx
SWD2 CLK	jtag TDO		always TX
SWD2 IO	jtag TMS	pDir2-pin	Rx-Tx

Sampling frequency of gpio-monitor

legacy -> 160 kHz … up to 1.5 … 2.9 MHz
intermediate -> relatively stable 4.03 MHz, min: 602 kHz, max: 4.55 MHz
current code … tbd

Electrical side

translator: 74LVC2T45GS
470 R line resistor and 100k PU on both sides
analog switch: PI5A4158, ~ 34 pF line-capacitance, [< 20 nA leakage]
previous switch: > 300 pF, < 1 nA leakage

Performance-data

not available atm
previous switch limited to ~ 200 kHz
capacitance on line is ~ 1/10, resistance ~ 1/2 -> 2 MHz should be fine

Logging of system parameters while recording

io calls
cpu usage
nw usage
ram usage

Time-sync

shepherd V1 started with +- 2.4 us error
improvements on PRU-Level helped to push the boundaries
also important: hardware-accelerated network-switch
tbd: cisco-switch in TUD has layer 3 routing and >doubled spec -> could improve sync