Dernière modification le 03/08/2024
Difficulté
Difficile
Durée
4 heure(s)
Coût
600 EUR (€)
Description
Building a photovoltaic tracker from a plate lifter
Sommaire
Sommaire
- 1 Description
- 2 Sommaire
- 3 Étape 1 - Sizing : moment of inertia measurement
- 4 Étape 2 - Engine resistance test
- 5 Étape 3 - Sizing: measurement of angles and displacement
- 6 Étape 4 - Install the hydraulic cylinder on the horizontal axis
- 7 Étape 5 - Raise the plate lifter maximum angle
- 8 Étape 6 - Tester les commandes du verin et du moteur rotatif
- 9 Étape 7 - Hydraulic cylinder for vertical axis rotation
- 10 Étape 8 - Etalonner les moteurs
- 11 Étape 9 - Coder le tracking en "dur"
- 12 Étape 10 - Coder le tracking avec une IA
- 13 Commentaires
Matériaux
tracker:
BRD plate lifter: 150€
hydraulic cylinder "Actionneur linéaire 12V DC , 1320LBS(6000N) 20 pouces (500mm) moteur électrique" : 69€ on aliexpress, available on amazon a bit more expensive
Hbridge L298n 7a: 11€ delivered on aliexpress ("Moteur d'entrainement PWM 160W 7A 12V 24V, Module de commande L298, Signal de commande logique, optocoupleur, frein")
2 chain cogs 92T: "Pignon arrière 25H JO98/108/138 maillons 55T 65T 68T 70T 80T 92T, pour 47CC 49CC Mini Moto RL facades D343 Pit Pocket Bike" : 40€
2 cogs "Pignon pour scooter électrique 8T 9T 11T 13T 25H 410 420, pour moteur à courant continu 25H JOMotor MY1020 BM1109 MY1016Z MY1018"
10€
2 cogs "Pignon de moteur électrique pour Pit-Bike, pignon de moteur à courant continu, pièces RL, D343, 9T, 11T, 13T, 25H, JOMotor 25H"
4€
Engine "Moteur à engrenages CC à vis sans fin autobloquant, couple de bain, moteur de boîte de vitesses turbo en métal, inversé, basse vitesse, DC 12V, 24V, 200kg.cm" 62€
Module
Photovoltaic module Voltech 2mx1m 375W: 200€
Control:
raspberry: environ 100€
Outils
facom measuring tape: 20€
sovietic mass kit: 60€
spirit level: 5€
bracket: 6€
grinder: 50€
drill: 50€
welding machine: 100€
Étape 1 - Sizing : moment of inertia measurement
We set two axis for our module:
one axis Oz perpendicular to the module plane (vertical when the module is flat) and one axis Ox paralelle to the module plane (horizontal when the module is flat)
Caractéristics of the module:
mass: m=21,2 kg
Length: L=1,8m
Width: 1m
Wikipedia theory tells us intertia moment along Oz axis is :
Jdelta=1/12*m*L
Jdelta=1/12*1,8*21,2=3,18Nm
Nb: we suppose the module is a bar because here the rotation axis is the axis paralell to the plane of the rectangle formed by the module and not the perpendicular axis
We now experimentally verify:
Center the module on the plate lifter and put it horizontally
verify there is no wind
check the level on which the mast of the plate lifter is put
fix a horizontal lanmark at the bottom of the module
set a 500g mass at the extremity of the module
measure the distance between the position at equilibrium and the position with the mass
We have:
Jdelta=d*F
with d distance in meter to the axis of rotation
F force applied to the solid (here mg with m solid mass in kg and g gravitational constant)
We then have
Jdelta=0,9*0,5*9,8=4,41Nm
We measure the rotation: in our case, the distance of rotation at the extremity of the module varies between 3 and 10cm. The important variations
are due to frictions on the axis that can force or slide more freely
The order of magnitude of the inertia momen is verified
For the moment of inertia along axis Ox, it is more difficult to verify experimentally, because the axise of the plate lifter doesnt allow
to have a position at equilibrium with the module mass (which will finish on the stop under its own weight)
Therefore, we will accept the theory:
The theory says
Jdelta=1/12*m(b²+c²) with m mass of the module, b lenght of the short side of the module and c length of the long side of the module
Jdelta=1/12*21,2*(1,8²+1²)=7,49NM
This theoretical result is strongly lower than the weigth of the module, so we will size based on the necessary force to lift the weigth of the module (the axis being submitted to the weight):
F=mg=21,2*9,8
In order of magnitude 200Nm (1m of lever arm in order of magnitude)
We have now the caracteristics to size the engines of our module along two rotation axis
But the trackers have a consequent wind exposure
If we want to take into account the resistance to the wind, we must measure the force applied to the module according to wind speed:
Fp=1/2*ρ*v²*S*Cp
With ρ air density equal to 1,2 kg/m3 in order of magnitude for standard temperature and pressure conditions
v wind speed in m/s
S object surface in m²
Cp pressure coefficient without dimension égal to 2 for a rectangular metal plate
We have:
Fp=1/2*1,2*1,8*2*v²=2,16*v²
Chatgpt can give us the abacus of wind speeds in km/h and their conversions in m/2 and the generic name in meteorology
Calm: Less than 1 km/h (Less than 0.3 m/s)
Very light breeze: 1-5 km/h (0.3-1.5 m/s)
Light breeze: 6-11 km/h (1.6-3.0 m/s)
Gentle breeze: 12-19 km/h (3.4-5.4 m/s)
Moderate breeze: 20-28 km/h (5.5-7.9 m/s)
Fresh breeze: 29-38 km/h (8.0-10.7 m/s)
Strong breeze: 39-49 km/h (10.8-13.8 m/s)
Moderate wind: 50-61 km/h (13.9-16.9 m/s)
Near gale: 62-74 km/h (17.2-20.6 m/s)
Gale: 75-88 km/h (20.8-24.4 m/s)
Strong gale: 89-102 km/h (24.7-28.3 m/s)
Storm: 103-117 km/h (28.6-32.5 m/s)
Hurricane: At least 118 km/h (At least 32.8 m/s)
So we have a Force Fp that can vary from
An order of magnitude of 20N for a ligth breeze
to
An order of magnitude of 2000N for a storm
(NB: the gravity force of 1kg is approximately 10N so the force of 2000N corresponds in order of magnitude to the gravity force of 200kg).
The inertia moments on the axis are of the same order of magnitude (20Nm and 2000Nm) because the dimensions of the module are an order of magnitude of 1 m
If you want to build a tracker that can resist to storm conditions, it is advised to size the tracker consequently
On the one hand with sufficient ties in the ground, and on the other hand with an adapted frame -see caravan tips at the end of this stage-, and deactivate
when there are stronger winds.)
Wind resistance sizing is one of the reasons trackers are expensive et less generalized than standard photovoltaic installations
The step motors and servomotors that have an adequate torke to resist to important winds are expensive, et that can be understandable for motors designed for step precision
(For example here: https://www.distrelec.fr/fr/automatisation/moteurs-et-entrainements/moteurs-pas-pas-et-servocommandes/c/cat-L3D_525513 )
For hydraulic cylinders, the driving force is generally in orders of magnitude fitting to resist to storms.
For our low-tech tutorial, we know that core drilling machines have torque in an order of magnitude fitting to resist to storms (1W corresponds to 1 N multiplying 1 meter per second, so a core drilling machine of 2000W should have an adequate torque of more or less 2000Nm).
After verification on the technical caracteristics of drills, impact screwdriver, and after a manual verification of resistances (brake,clutch) when the engine is not powered, this type of engine doesnt fit.
To size a resistance to winds between the storm and hurricane we will proceed differently:
We find on aliexpress a affordable prices (70€) step motors or endless screw motors with torque in an order of magnitude of 20Nm(200kgcm). We will use thi engine with two reductors 1:10 to get a resisting torque of 2000Nm.
Recall: the torque expressed in Nm is a rotation force produced by the motor that can be computed with the same principle of the inertia moment: the force in newton x the distance to the rotation axis
Belt or chain transmission permit to reduce this torque, and that can be calculated very easily
C1=(R1/R2)*C2
With C1 torque on pulley 1
R1 pulley 1 radius
C2 torque on pulley 2
R2 pulley 2 radius
The radius is directly proportional to the number of teeth of a gear wheel (a bike cog or a motorcycle cog for example), we can easily calculate the transmission reduction by dividing the number of teeth of the big cog by the number of teeth of the little cog (verify they have the same teeth standards)
Therefore a transmission 80 teeth/10 teeth (80T/10T) will produce a reduction of 1:8, like here on amazon for an electric bike ("Keenso Kit Chaîne et Pignon 80T 25H 34mm 3 Trous Pignon 10T H Trou Chaîne Pignon 146 Maillons Chaîne") for 30€.
We will notice that a standard bike transmission have a maximum transmission reduction of 50 teeth/11 teeth so a reduction of 1:5, which is not enough for efficient reduction witouth multiplying pulleys
It is difficult to find transmissions with reduction of 1:10, but we find some on aliexpress and we will easily adapt a bicycle drive on which we will weld cogs, which will allow us a reduction of 1:100 with only one axis added in addition to the principal axis and the engine axis
So we will use an engine step motor or an endless screw engine (resistance unpowered test at reception) commanded by a raspberry pi (but the plate lifter has wheels, so we will take the precaution to put away the tracker when there's a storm ;))
Carvan tips: to size the width of the metal frame, modern norms dont seem to be really made on scientific basis but on commercial basis. We will therefore measure the width of old material (made in the 70s). For example, the frame of this old caravan accredited on its registration document for a 750kg weigth has a metal frame of 5mm width. On can then make a rule of 3 to verify the width of our frame fits.
This is rustic style (strenght of material calculs is rigorous and complex), but it allows to have a firest approximation "a visto de nas" as we say in gascon
Étape 2 - Engine resistance test
See videos
The rotative engine ("Moteur à engrenages CC à vis sans fin autobloquant, couple de bain, moteur de boîte de vitesses turbo en métal, inversé, basse vitesse, DC 12V, 24V, 200kg.cm" on aliexpress) takes well 20Nm being powered and not being powered.
Recall: torque = moment of intertia generated by the motor
=force*distance to motor axis
here 4kg at 0,5m=9,8*4*0,5=20Nm (in order of magnitude)
Étape 3 - Sizing: measurement of angles and displacement
We will first look at the solar trajectory
We measure typically the sun position along two coordinate systems:
the equatorial system with coordinates expressed in:
right ascension equivalent to terrestrial longitude measured in hours minutes seconds
declination equivalent to terrestrial latitude measured in degrees minutes seconds
The horizontal system with coordinates expressed in:
azimut degree
altitude degree or height
The solar trajectory abacus (for example available here https://www.astrolabe-science.fr/diagramme-solaire-azimut-hauteur ) give us the sun trajectories in a day (generally several typical days of several seasons) expressed in horizontal degrees.
To read this type of grap
If we follow the graph inserted in this tutorial and from the link above, when we follow for example the red curb for paris, we can read :
The 21st of december, when we look at the south, the sun follows a trajectory that starts at -50° azimut (along the East on the horizontal axis) at dawn, then when the sun goes west during the day (we follow the red curb), it takes height until it reaches 17° height (on the vertical axis) at noon (position 0° azimut on the horizontal axis) and then goes down again to 0° height when it sets (along the West on the horizontal axis)
For Paris, we have:
One degree azimut varying from -130° to +130° according to the hour and the season
One degree altitude or height varying from 0° to 64° according to the hour and the season
For our tracker,
To calculate our horizontal displacement (along the Oz vertical axis if the module is put down on the ground), on don't really have constraints on the plate lifter used since the axis turns 360° without problems.So we can follow the sun from -130° azimut to +130° azimut without problem.
To calculate our vertical displacement (along the Ox horizontal axis if the module is put on the ground), we have a constraint on the maximal angle.
I dont have a decimeter at hand, so I will use pythagore (see photo):
64cm*150cm*134cm
socatoa:
sinus phi=opposé/hypothenus
sinus phi=134/150
sinus phi=134/150
phi=1,1046 rad
phi=1,1046*180/pi=63°
We have a constrainte for our plate lifter accepting angles along the Ox axis from 0° to 63°
When the tracker is at its maximum angle (63°), we are perpendicular to the sun when the sun is at an angle phi of phi=180-63-90=27°
When the sun has an angle lower than 27°, the tracker can not follow and keep being perpendicular to the sun
We see however that the stop is guaranteed by the spring (on the photo we sse the mark on the paint). We can win a bit of amplitude on the stop by drilling and making a notch in the jib).
The manual measurement of the displacement between the tube axis on which is fixed the crank handle and the back of the module when the plate lifter leans at its maximum angle gives us 42cm. (see photo)
Étape 4 - Install the hydraulic cylinder on the horizontal axis
We will fix a jib on the fix part of the plate lifter that turns with the vertical axis, so we can fix a rod on which we will fix the hydraulic cylinder which will be able to turn with the vertical axis so it can adjust the angle on the horizontal axis.
We begin by fixing the metal jib welding to the fix part regarding the vertical axis. It requires to sand the paint before the welding is done (see photo). We do arc welding here.
Then we drill a metal rod we will screw to the jib (see photo).
We drill et we fix a metal rod we screw to the mobile part that allows to adjust the angle on the horizontal axis, ie the arm that permits to carry the plate or the photovoltaic module (see photo).
We then fix the hydraulic cylinder to the two metal rods. The hydraulic cylinder is equiped with fixing that adjust to the angle taken by the rods on which it is fixed (see photo)
Notice we have a metal rod with a right angle that avoid this fix is totally free. It will stop on a part of the rod, which permits then to calibrate more precisely the hydraulic cylinder amplitude).
We then test the course of the hydraulic cylinder. We will pay attention, because the stop of the plate lifter corresponds to a 40cm course for the hyrdaulic cylinder and it can extend up to 50cm. (see photos)
Étape 5 - Raise the plate lifter maximum angle
After observing the critical elements, we will modify the jib to avoid a stop when the sun has a altitude degree lower than 27°.
To do so, we will:
-let the spring go by scooping the jib (to avoid the spring be a stop)
-raise the angle by lowering the fix of the spring drilling the jib
-raise the angle cutting the edges of the jib
See photos:
1/2: observation top/below
3/4: jib disassembly
4/5 : observation dessous, angle max à l'equerre
On arrive ainsi à un angle maximum de quasi 90° et on peut donc suivre le soleil sous tous les angles!
Étape 6 - Tester les commandes du verin et du moteur rotatif
Pour controller le verin, on va utiliser un raspberry pi, l'ordinateur monocarte le plus répandu. Il est doté de d'une série 40 pins, qu'on peut connecter à divers appareils, appelés "controlleur GPIO".
A first reading of a few tutorials and available libraries to use it a some time taken to test wrong hypothesis to install correctly using the good versions leads me to talk a few possible options:
-operating system:
- dietpi images (debian) available here: https://dietpi.com/#download (see my other tutorial for how to activate wifi : https://wiki.lowtechlab.org/wiki/Serveur_orangepi-raspberry_nextcloud_en_photovolta%C3%AFque_autonome)
- rapsberry pi os (default) images available here: https://www.raspberrypi.com/software/operating-systems/
Archives and operating systems versions usefull for retrocompatibility (every reader caring for low tech computer is encouraged to keep local copies of these archives and to share them in torrent!):
*dietpi:
https://dietpi.com/downloads/images/
*raspberry pi os:
https://downloads.raspberrypi.org/raspbian/images
si vous utilisez wheezy,
echo "deb http://legacy.raspbian.org/raspbian/ wheezy main contrib non-free rpi" >> /etc/apt/sources.list
NB: ChatGPT gives us the following rapsberry release dates:
- Raspberry Pi Model B : 29 february 2012
- Raspberry Pi Model A : 4 february 2013
- Raspberry Pi Model B+ : 14 july 2014
- Raspberry Pi Model A+ : 10 november 2014
- Raspberry Pi 2 Model B : 2 february 2015
- Raspberry Pi Zero : 30 november 2015
- Raspberry Pi 3 Model B : 29 february 2016
- Raspberry Pi Zero W : 28 february 2017
- Raspberry Pi 3 Model B+ : 14 march 2018
- Raspberry Pi 3 Model A+ : 15 november 2018
- Raspberry Pi 4 Model B : 24 june 2019
- Raspberry Pi 400 : 2 november 2020
- Raspberry Pi Pico : 21 january 2021
- Raspberry Pi Zero 2 W : 28 october 2021
We hope this is true (it comes from chatgpt), but you can verify on what's written on the pcb of your card.
To verify the version of your raspberry under raspberry pi os if you dont know what version it is (see photo):
sudo usermod -a G gpio pi pinout
Adjust with the operating system that you think is relevant based on your relevancy filters. You have a list of old os versions of dietpi and raspberry pi os in case new versions would now ork anymore (and i repeat: every reader caring of low tech computer is encouraged to download and keep local copies of these archives and share them in torrent!)
To install, as usually, download balenaetcher, flash a usb key with the donwloaded image, boot. The default login/password on dietpi are root/dietpi and for raspberry pi os pi/raspberry (take care to the default qwerty on raspberry pi os at boot). To configure keyboard, locales, timezone and wifi, do:
sudo dietpi-configunder dietpi and
sudo raspi-configunder raspberry pi.
-driver utilisé pour controler le gpio
- RPi.GPIO(dev independant) ou RPi.GPIO2(dev redhat, repo recent)
Essais infructueux avec pip et le depot pypi (erreur de compilations etc.). Installer en passant en root avec la commande
sudo -s
le plus simple est alors d'installer une version déjà compilée avec apt:
sudo apt install python3-rpi.gpio
faire un test pour voir si ca fonctionne bien (toujours en etant utilisateur root):
python3 import RPi.GPIO
See the sourceforge documentation, more explicit than the pypi one: http://sourceforge.net/p/raspberry-gpio-python/wiki/Home/
- gpiozero
under dietpi do
sudo apt install python3 python3-venv python3-pip python3 -m venv venv source venv/bin/activate pip install lgpio gpiozero
The first point to understand is the numbering of the pins: The pins have all a number that goes from 1 to 40 following an order from bottom to top and from left to right, and each pin has also a gpiio number which is different from the pin number.
To do so, we can find information on internet: https://wiki.lowtechlab.org/wiki/Serveur_orangepi-raspberry_nextcloud_en_photovolta%C3%AFque_autonome or type the command under raspberry pi os (see image)
sudo usermod -aG gpio votre_user pinout
In this tutorial, wether for gpiozero or RPi.GPIO, we will use the GPIO numbering and not the pin numbering
We plug the GPIO 2 and 4 (+5V) to the +5V pins of the HBridge We plug the GPIO 6 and 7 (Ground) to the GND pin of the HBridge We plug the GPIO 23 and 16 to the switch 3 of the HBridge To test the "enable" of the HBridge, we plug the GPIO 20 and 21 to EAN of HBridge
We then use the following scripts:
import time
import RPi.GPIO as GPIO
import gpiozero
def forwardzero(wait):
led16=gpiozero.LED(16) #motor1
led23=gpiozero.LED(23) #motor1
led20=gpiozero.LED(20) #enable
led21=gpiozero.LED(21) #enable
led1=gpiozero.LED(1)
led7=gpiozero.LED(7)
try:
print("forwardzero")
#case where Hbridge needs a enable signal
led21.off()
led20.on()
#zero signal off on switch 2
led1.off()
led7.off()
#positive signal off on switch 1
led23.on()
led16.off()
#let signal active during wait time
for k in range(wait):
print(k)
time.sleep(1)
#put signal off on the switch
led23.off()
led20.off()
except KeyboardnInterrupt:
print("keyboard interrupt")
except Exception as err:
print(err)
finally:
print("zero")
#zero signal off on switch 2
led1.off()
led7.off()
#signal off on switch 1
led16.off()
led23.off()
#signal off on enable pin
led20.off()
led21.off()
def backwardzero(wait):
led1=gpiozero.LED(1)
led7=gpiozero.LED(7)
led16=gpiozero.LED(16) #motor1
led23=gpiozero.LED(23) #motor1
led20=gpiozero.LED(20) #enable
led21=gpiozero.LED(21) #enable
try:
print("backwardzero")
#signal off on switch 1
led16.on()
led23.off()
#case where Hbridge needs a enable signal
led20.off()
led21.on()
#signal on switch 2
led1.off()
led7.off()
#wait
for k in range(wait):
print(k)
time.sleep(1)
#signal off on switches
led16.off()
led21.off()
except KeyboardInterrupt:
print("keyboardinterrupt")
except Exception as err:
print(err)
finally:
print("zero")
#signal off on switch 1
led16.off()
led23.off()
#signal off on switch 2
led1.off()
led7.off()
#signal off on enable pin
led20.off()
led21.off()
#pin16 gpio23
#pin36 gpio25
def forward(wait):
GPIO.setmode(GPIO.BCM)
GPIO.setup(16,GPIO.OUT) #motor1
GPIO.setup(23,GPIO.OUT) #motor1
GPIO.setup(20,GPIO.OUT) #enable
GPIO.setup(21,GPIO.OUT) #enable
GPIO.setup(1,GPIO.OUT) #motor2
GPIO.setup(7,GPIO.OUT) #motor2
try:
print("forward")
GPIO.output(1,GPIO.HIGH)
GPIO.output(7,GPIO.LOW)
GPIO.output(21,GPIO.HIGH)
GPIO.output(20,GPIO.LOW)
GPIO.output(23,GPIO.LOW)
GPIO.output(16,GPIO.LOW)
for k in range(wait):
print(k)
time.sleep(1)
GPIO.output(1,GPIO.LOW)
GPIO.output(21,GPIO.LOW)
except KeyboardInterrupt:
print("keyboard interrupt")
except Exception as err:
print(err)
finally:
print("zero")
GPIO.output(1,GPIO.LOW)
GPIO.output(7,GPIO.LOW)
GPIO.output(20,GPIO.LOW)
GPIO.output(21,GPIO.LOW)
GPIO.output(23,GPIO.LOW)
GPIO.output(16,GPIO.LOW)
#GPIO.cleanup() on my sbc it doesnt remove the +3V
def backward(wait):
GPIO.setmode(GPIO.BCM)
GPIO.setup(20,GPIO.OUT) #enable
GPIO.setup(21,GPIO.OUT) #enable
GPIO.setup(1,GPIO.OUT) #motor2
GPIO.setup(7,GPIO.OUT) #motor2
GPIO.setup(16,GPIO.OUT) #motor1
GPIO.setup(23,GPIO.OUT) #motor1
try:
print("backward")
GPIO.output(21,GPIO.HIGH)
GPIO.output(20,GPIO.LOW)
GPIO.output(16,GPIO.LOW)
GPIO.output(23,GPIO.LOW)
GPIO.output(7,GPIO.HIGH)
GPIO.output(1,GPIO.LOW)
for k in range(wait):
print(k)
time.sleep(1)
GPIO.output(7,GPIO.LOW)
GPIO.output(21,GPIO.HIGH)
except KeyboardInterrupt:
print("keyboardinterrupt")
except Exception as err:
print(err)
finally:
print("zero")
GPIO.output(20,GPIO.LOW)
GPIO.output(21,GPIO.LOW)
GPIO.output(16,GPIO.LOW)
GPIO.output(23,GPIO.LOW)
GPIO.output(1,GPIO.LOW)
GPIO.output(7,GPIO.LOW)
#GPIO.cleanup() on my sbc it doesnt remove the +3V
forward(2) #rotation horary motor2
backward(2) #rotation antihorary motor2
forwardzero(10) #hydraulic cylinder extension motor1 111 max
backwardzero(10) #hydraulic cylinder retractation motor1 107 max
GPIO.BCM allows to use GPIO numbering GPIO.HIGH sends a +3V signal in the concerned pin GPIO.LOW sets the signal to 0V (GND) gpiozero does the same with .on() and .off() methods
Update 31.05.24: you were waiting for it, here's today update with the HBridge made in Europe. It works, see video! Update of calibration quickly at recepetion of something to measure angles a bit more handy.
Notice the pins 20 and 21 are not plugged to the HBridge. The "PWM" (pulse width modulation) is not used to "activate" (enable in the code) the engine because jumbers on ENA pins of the Hbridge are enough (here we dont need to adjust motor speed)
Étape 7 - Hydraulic cylinder for vertical axis rotation
Fix pulleys and 1:100 transmission for rotation motor
We begin by recycling two bycicle drives from second hand bikes we cut with a grinder taking care to keep the metal rod of the frame.
we will weld 92T cogs on it.
We then weld the 8T cogs on it.
We weld a 8T cog on a locked cog adapted to the motor axis.
We cut a metal rod and we weld a flat metal plate along the axis of the plate lifter on which we will screw the bike metal rods attached to the bike drives and alos a metal rod on which we will fix the motor.
We assemble and we screw.
Update when receiving the 3rd chain and waiting how to fix the chain tensions.
Update of 11.6.24: the constraints of the axis of the plate lifter make it difficult to fix a big cog to get a good reduction on the last transmission chain. Even we have several reductions with other pulleys, the plate lifter is not rotating (see video).
update of 16.6.24: additional hydraulic cylinder ordered. Estimated delivery 28 juin-2juillet
On va donc entrainer la rotation avec deux verins. Update à réception des verins fin juin (en stage et non dispo pour update ces prochaines semaines)
La tension de chaine etant mauvaise, on remplace les transmissions par un verrin hydraulique permettant de faire la rotation mais sur un angle réduit (environ 70° faute de verrin telescopique à plusieurs brins)
Étape 8 - Etalonner les moteurs
Le code mis à jour est le suivant: on définit des dictionnaire qui associe chaque angle recherché à un temps d'activation du moteur (à tester à la main et à mesurer)
#Etalonnage
#dict_angle_rotation= sun_degre_azimut:motoractivationtime
dict_angle_verin={1:0,
2:0,
3:0,
4:0,
5:0,
6:0,
7:0,
8:0,
9:0,
10:0,
11:1,
12:1,
13:2,
14:3,
15:4,
16:5,
17:6,
18:7,
19:8,
20:9,
21:10,
22:11,
23:12,
24:13,
25:14,
26:15,
27:16,
28:16,
29:17,
30:18,
31:19,
32:20,
33:21,
34:22,
35:23,
36:24,
37:25,
38:27,
39:28,
40:29,
41:30,
42:32,
43:33,
44:34,
45:35,
46:37,
47:38,
48:39,
49:41,
50:42,
51:44,
52:45,
53:47,
54:48,
55:50,
56:51,
57:52,
58:54,
59:56,
60:58,
61:59,
62:59,
63:59,
64:60,
65:63,
66:64,
67:65,
68:66,
69:68,
70:68,
71:71,
72:73,
73:75,
74:77,
75:79,
76:81,
77:82,
78:83,
79:85,
80:86,
81:87,
82:89,
83:91,
84:94,
85:96,
86:98,
87:100}
dict_angle_rotation={
0:41,
1:42,
2:43,
3:44,
4:45,
5:46,
6:47,
7:48,
8:49,
9:50,
10:52,
11:53,
12:54,
13:55,
14:56,
15:57,
16:58,
17:59,
18:60,
19:61,
20:63,
21:64,
22:65,
23:66,
24:67,
25:68,
26:69,
27:70,
28:71,
29:72,
30:73,
31:75,
32:76,
33:77,
34:78,
35:79,
36:82,
37:82,
38:82,
39:82,
40:82,
41:82,
42:82,
43:82,
44:82,
45:82,
46:82,
47:82,
-1:39,
-2:37,
-3:35,
-4:33,
-5:32,
-6:30,
-7:29,
-8:27,
-9:26,
-10:25,
-11:23,
-12:22,
-13:19,
-14:18,
-15:17,
-16:16,
-17:15,
-18:14,
-19:6,
-20:2,
-21:1,
-22:0,
}
Étape 9 - Coder le tracking en "dur"
Le code mis à jour est le suivant
import time
import RPi.GPIO as GPIO
import gpiozero
import ephem
import datetime
def forwardzero(wait):
led16=gpiozero.LED(16) #motor1
led23=gpiozero.LED(23) #motor1
led20=gpiozero.LED(20) #enable
led21=gpiozero.LED(21) #enable
led1=gpiozero.LED(1)
led7=gpiozero.LED(7)
try:
print("forwardzero")
#cas où le hbridge necessite un signal enable
led21.on()
led20.off()
#signal à zero sur interupteur 2
led1.off()
led7.off()
#mettre un signal sur l'interrupteur 1
led23.on()
led16.off()
#laisser le signal actif le temps de wait
for k in range(wait):
print(k)
time.sleep(1)
#eteindre le signal sur l'interrupteur
led23.off()
led21.off()
except KeyboardnInterrupt:
print("keyboard interrupt")
except Exception as err:
print(err)
finally:
print("zero")
#signal à zero sur interupteur 2
led1.off()
led7.off()
#signal à zero sur l'interrupteur 1
led16.off()
led23.off()
#signal à zero sur la fiche enable
led20.off()
led21.off()
def backwardzero(wait):
led1=gpiozero.LED(1)
led7=gpiozero.LED(7)
led16=gpiozero.LED(16) #motor1
led23=gpiozero.LED(23) #motor1
led20=gpiozero.LED(20) #enable
led21=gpiozero.LED(21) #enable
try:
print("backwardzero")
#signal zero sur interupteur 1
led16.on()
led23.off()
#cas où le hbrige necessite un signal sur enable
led20.off()
led21.on()
#signal sur interrupteur 2
led1.off()
led7.off()
#wait
for k in range(wait):
print(k)
time.sleep(1)
#signal off sur interrupteur
led16.off()
led21.off()
except KeyboardInterrupt:
print("keyboardinterrupt")
except Exception as err:
print(err)
finally:
print("zero")
#signal à zero sur interupteur 1
led16.off()
led23.off()
#signal à zero sur l'interupteur 2
led1.off()
led7.off()
#signal à zero sur la fiche enable
led20.off()
led21.off()
#pin16 gpio23
#pin36 gpio25
def forward(wait):
GPIO.setmode(GPIO.BCM)
GPIO.setup(16,GPIO.OUT) #motor1
GPIO.setup(23,GPIO.OUT) #motor1
GPIO.setup(20,GPIO.OUT) #enable
GPIO.setup(21,GPIO.OUT) #enable
GPIO.setup(1,GPIO.OUT) #motor2
GPIO.setup(7,GPIO.OUT) #motor2
try:
print("forward")
GPIO.output(1,GPIO.HIGH)
GPIO.output(7,GPIO.LOW)
GPIO.output(21,GPIO.HIGH)
GPIO.output(20,GPIO.LOW)
GPIO.output(23,GPIO.LOW)
GPIO.output(16,GPIO.LOW)
for k in range(wait):
print(k)
time.sleep(1)
GPIO.output(1,GPIO.LOW)
GPIO.output(21,GPIO.LOW)
except KeyboardInterrupt:
print("keyboard interrupt")
except Exception as err:
print(err)
finally:
print("zero")
GPIO.output(1,GPIO.LOW)
GPIO.output(7,GPIO.LOW)
GPIO.output(20,GPIO.LOW)
GPIO.output(21,GPIO.LOW)
GPIO.output(23,GPIO.LOW)
GPIO.output(16,GPIO.LOW)
#GPIO.cleanup() chez moi, ca n'enleve pas les +3V
def backward(wait):
GPIO.setmode(GPIO.BCM)
GPIO.setup(20,GPIO.OUT) #enable
GPIO.setup(21,GPIO.OUT) #enable
GPIO.setup(1,GPIO.OUT) #motor2
GPIO.setup(7,GPIO.OUT) #motor2
GPIO.setup(16,GPIO.OUT) #motor1
GPIO.setup(23,GPIO.OUT) #motr1
try:
print("backward")
GPIO.output(21,GPIO.HIGH)
GPIO.output(20,GPIO.LOW)
GPIO.output(16,GPIO.LOW)
GPIO.output(23,GPIO.LOW)
GPIO.output(7,GPIO.HIGH)
GPIO.output(1,GPIO.LOW)
for k in range(wait):
print(k)
time.sleep(1)
GPIO.output(7,GPIO.LOW)
GPIO.output(21,GPIO.HIGH)
except KeyboardInterrupt:
print("keyboardinterrupt")
except Exception as err:
print(err)
finally:
print("zero")
GPIO.output(20,GPIO.LOW)
GPIO.output(21,GPIO.LOW)
GPIO.output(16,GPIO.LOW)
GPIO.output(23,GPIO.LOW)
GPIO.output(1,GPIO.LOW)
GPIO.output(7,GPIO.LOW)
#GPIO.cleanup() chez moi ca n'enleve pas les +3V
#forward(41) #rotation horaire motor2
#backward(90) #rotation antihoraire motor2
#forwardzero(10) #verin extension motor1 111 max
#backwardzero(2) #verin retractation motor1 107 max
#forwardzero(90)
#Etalonnage
#dict_angle_verin= sun_degre_horizontal:motoractivationtime
#dict_angle_rotation= sun_degre_azimut:motoractivationtime
dict_angle_verin={1:0,
2:0,
3:0,
4:0,
5:0,
6:0,
7:0,
8:0,
9:0,
10:0,
11:1,
12:1,
13:2,
14:3,
15:4,
16:5,
17:6,
18:7,
19:8,
20:9,
21:10,
22:11,
23:12,
24:13,
25:14,
26:15,
27:16,
28:16,
29:17,
30:18,
31:19,
32:20,
33:21,
34:22,
35:23,
36:24,
37:25,
38:27,
39:28,
40:29,
41:30,
42:32,
43:33,
44:34,
45:35,
46:37,
47:38,
48:39,
49:41,
50:42,
51:44,
52:45,
53:47,
54:48,
55:50,
56:51,
57:52,
58:54,
59:56,
60:58,
61:59,
62:59,
63:59,
64:60,
65:63,
66:64,
67:65,
68:66,
69:68,
70:68,
71:71,
72:73,
73:75,
74:77,
75:79,
76:81,
77:82,
78:83,
79:85,
80:86,
81:87,
82:89,
83:91,
84:94,
85:96,
86:98,
87:100}
dict_angle_rotation={
0:41,
1:42,
2:43,
3:44,
4:45,
5:46,
6:47,
7:48,
8:49,
9:50,
10:52,
11:53,
12:54,
13:55,
14:56,
15:57,
16:58,
17:59,
18:60,
19:61,
20:63,
21:64,
22:65,
23:66,
24:67,
25:68,
26:69,
27:70,
28:71,
29:72,
30:73,
31:75,
32:76,
33:77,
34:78,
35:79,
36:82,
37:82,
38:82,
39:82,
40:82,
41:82,
42:82,
43:82,
44:82,
45:82,
46:82,
47:82,
-1:39,
-2:37,
-3:35,
-4:33,
-5:32,
-6:30,
-7:29,
-8:27,
-9:26,
-10:25,
-11:23,
-12:22,
-13:19,
-14:18,
-15:17,
-16:16,
-17:15,
-18:14,
-19:6,
-20:2,
-21:1,
-22:0,
}
def sun_position(time_now,lat,lon):
now_here = ephem.Observer()
now_here.lat = lat
now_here.lon = lon
#PyEphem only processes and returns dates that are in Universal Time (UT), which is simliar to Standard Time in Greenwich, England, on the Earth's Prime Meridian
# Europe/Paris is GMT+2
#tester angles pyephem sur mesures réelles
utc_now=datetime.datetime.utcnow()
#is_dst=datetime.datetime(year=utc_now.year,month=utc_now.month,day=utc_now.day).dst()
#time_diff=datetime.timedelta(hours=(1 if not is_dst else 2))
now_here.date = time_now #+datetime.timedelta(hours=time_diff) #'2007/10/02 00:50:22'
sun.compute(now_here)
sun_degre_azimut=int(sun.az*180/3.141592653589793)
sun_degre_horizontal=int(sun.alt*180/3.141592653589793)
return(sun_degre_horizontal,sun_degre_azimut)
tracker_degré_horizontal=0
tracker_degré_azimut=0
def init():
global tracker_degré_horizontal
forwardzero(111)
tracker_degré_horizontal=0
def track(time_now,lat,lon):
global tracker_degré_horizontal
global tracker_degré_azimut
init()
(sun_degre_horizontal,sun_degre_azimut)=sun_position(time_now,lat,lon)
sun_degre_azimut=min(sun_degre_azimut,87)
sun_degre_horizontal=max(sun_degre_horizontal,-22)
sun_degre_horizontal=min(36,sun_degre_horizontal)
backwardzero(dict_angle_verin[sun_degre_horizontal])
tracker_degré_horizontal=sun_degre_horizontal
backward(82)
forward(dict_angle_rotation[sun_degre_azimut-tracker_degré_azimut])
tracker_degré_azimut=sun_degre_azimut
#test Agen
lat=44.2
lon=0.6
backward(4)
forward(4) #placer le tracker direction sud angle horizontal
backwardzero(4)
forwardzero(41)
#time.sleep(100)
while True:
track(datetime.datetime.now(),lat,lon)
time.sleep(20*60) #activer le tracking toutes les 20 minutes
Étape 10 - Coder le tracking avec une IA
On va maintenant coder une "IA lowtech" pour le côté pédagogique (du ML pour machine learning, comme utilisé massivement depuis une quinzaine d'année dans de nombreux secteurs d'activité, cad pas l'ia au sens chatgptesque du terme en 2024). On pourrait dire que l'ia lowtech est l'ia dont les résultats ne relèvent pas de la pensée magique, dont les data et le code sont open source, non volés, dont les data sont à nous, par nous et pour nous et sur laquelle on a la main (ce dernier point est essentiel mais l'expérience de sortir du rang me pousse au pessimisme à ce sujet car la vérification s'il y a interférence ou pas dans les résultats du machine learning est assez difficile à détecter)
On branche une webcam, on enregistre les images comme données d'entrée, on traite l'image pour en faire un tableaux de chiffres correspondant à des variables avec lesquelles on va chercher à corréler avec un signal positif ou négatif (tourner le moter dans un sens ou moteur à l'arret).
Ici, les données d'entrées sont uniquement les 640*480*3=921600 variables des pixels des images de la vidéo (921600 colonnes/variables par lignes, à partir desquelles on cherche une corrélation avec le signal positif 1 ou 0 de la dernière colonne).
Pour faire fonctionner le tracker, ca ne fonctionnera pas bien, il faudrait faire du "feature engineering" (nom compliqué pour dire rajouter des colones de variables plus proabablement corrélées au signal positif) en rajoutant la luminosité et/ou la date et l'heure, sur des échantillons de vidéos couvrant toutes les saisons sur plusieurs années.
Si vous voulez apprendre les bases sur lesquelles reposent ce code, je recommande le cours "Applied Data Science with Python" de l'université du michigan dans lequelvous apprendrez des bases de python,pandas, et machine learning "no bullshit".
import cv2
import time
import numpy as np
import os
import argparse
import pandas as pd
import matplotlib.pyplot as plt
import numpy as np
import seaborn as sns
import sklearn.model_selection
import sklearn.metrics
import sklearn.decomposition
from sklearn.neighbors import KNeighborsClassifier
from sklearn.naive_bayes import GaussianNB
from sklearn.ensemble import RandomForestClassifier
from sklearn.ensemble import GradientBoostingClassifier
from sklearn.neural_network import MLPClassifier
def enregistrer_video(out_file,fps,capture_duration):
# Open the webcam
cap = cv2.VideoCapture(0) # Use 0 for default webcam
fourcc = cv2.VideoWriter_fourcc(*'XVID') # Codec (e.g., XVID)
frame_width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH)) # Get webcam frame width
frame_height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT)) # Get webcam frame height
# Check if webcam opened successfully
if not cap.isOpened():
print("Error opening webcam")
exit()
# Create the VideoWriter object
out = cv2.VideoWriter(out_file, fourcc, fps, (frame_width, frame_height))
# Start time for tracking duration
start_time = time.time()
while time.time() - start_time < capture_duration:
# Capture frame-by-frame
ret, frame = cap.read()
# Check if frame captured successfully
if not ret:
print("Error capturing frame")
break
# Write the frame to the video file
out.write(frame)
# Display the captured frame (optional)
cv2.imshow('Webcam Video', frame)
# Check if the user wants to quit (press 'q')
if cv2.waitKey(1) & 0xFF == ord('q'):
break
# Close resources
cap.release()
out.release()
cv2.destroyAllWindows()
print(f"1 minute video saved successfully as {out_file}!")
def charger_images_video(video_filename):
"""
Charge les images vidéo d'un fichier.
Args:
video_filename: Le chemin d'accès au fichier vidéo.
Returns:
Un tableau NumPy contenant les images vidéo (3D array: frames, rows, cols).
"""
# Ouvrir la vidéo avec OpenCV
cap = cv2.VideoCapture(video_filename)
# Vérifier l'ouverture réussie
if not cap.isOpened():
print("Erreur d'ouverture du fichier vidéo:", video_filename)
return None
# Liste vide pour stocker les images vidéo
images_list = []
# Lire les images vidéo image par image
while True:
ret, frame = cap.read()
# Vérifier la lecture de l'image
if not ret:
break
# Convertir l'image en nuance de gris (optionnel pour la normalisation)
# frame = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY) # Décommenter si nécessaire
# Ajouter l'image à la liste
images_list.append(frame)
# Fermer la capture vidéo
cap.release()
return images_list
def flatten_images_video(images_array):
# Initier une liste de resultat
result=[]
# Iterer sur le numpy array en input
for k in images_array:
result.append(k.flatten())
# convertir la liste np.array
result=np.asarray(result)
return result
# Exemple d'utilisation
# Define video parameters
out_file = "webcam_video_1min.mp4" # Output video filename
fps = 20.0 # Frames per second
capture_duration = 60 # Seconds
#enregistrer_video(out_file,fps,capture_duration)
images_video = charger_images_video("webcam_video_1min.mp4") # Fonction pour charger les images vidéo
print(images_video)
print(len(images_video))
# Flattening and normalizing
images_video = flatten_images_video(images_video)
print(images_video[0])
print(len(images_video[0]))
# compression/normalisation
scaling_factor = 1.0 / 255.0 # Divide by 255 to normalize between 0 and 1
images_video = images_video * scaling_factor
print(images_video[0])
print(len(images_video[0]))
# checking data shape
#np.set_printoptions(threshold=np.inf) # Set threshold to infinity
#print(np.array2string(images_video[0], suppress_small=True))
#print(len(images_video[0]))
# Get the array shape
image_shape = images_video.shape
# Print the dimensions
print("Image shape", image_shape)
print("Number of dimensions:", len(image_shape))
print("Image height:", image_shape[0])
print("Image width:", image_shape[1])
print("Number of color channels:", image_shape[2])
# Total number of elements (height x width x color channels)
total_elements = image_shape[0] * image_shape[1] * image_shape[2]
print("Total elements:", total_elements)E
# création du dataset pour dire "oui" pour tourner à gauche (par exemple)
# ND: c'est à cette étape que les géants de la tech emploient des kenyans sous payés
# dans une forme d'esclavage moderne
# il s'agit de définir, pour chaque image, si on doit activer le moteur vers
# la gauche (cad définir un signal positif pour que la machine fasse des corrélations
# positives avec cette image)
positives=np.zeros_like(images_video)
#Si les 14 premieres images définissent un signal positif, on fera:
#en réalité il faudra traiter des segments de vidéos positivement en définissant
#chaque image auxquelles on va associer le signal positif
positives[0:14] = 1
class ML():
@staticmethod
def ml(X,y,classifier):
"process machine learning on X data set, y yes/no data with classifier"
# dataset
#X = images_video
#y = positives
# classifier model training
clf = ML.dico_classifier[classifier]()
X_train, X_test, y_train, y_test = sklearn.model_selection.train_test_split(
X, y, random_state=0)
clf.fit(X_train, y_train)
# predictions
_predicted = clf.predict(X_test)
# scores
_accuracy, _precision, _recall = ML.compute_scores(y_test, _predicted,
classifier)
# confusion matrix
ML.compute_confusion_matrix(y_test, _predicted, _accuracy, classifier)
# courbes precision-recall
ML.plot_precision_recall(clf, X_test, y_test, classifier, _predicted)
# roc
roc_auc_clf = ML.plot_roc(clf, X_test, y_test, classifier)[0]
return pd.DataFrame(data=(_accuracy, _precision, _recall, roc_auc_clf),
index=['accuracy', 'precision', 'recall', 'AUC'],
columns=[classifier])
@staticmethod
def compute_scores(y_test, _predicted, classifier):
"compute machine learning scores"
_accuracy = sklearn.metrics.accuracy_score(y_test, _predicted)
_precision = sklearn.metrics.precision_score(y_test, _predicted)
_recall = sklearn.metrics.recall_score(y_test, _predicted)
print(str(classifier) + ' Accuracy: {:.2f}'.format(_accuracy))
print(str(classifier) + ' Precision: {:.2f}'.format(_precision))
print(str(classifier) + ' Recall: {:.2f}'.format(_recall))
return (_accuracy, _precision, _recall)
@staticmethod
def compute_confusion_matrix(y_test, _predicted, _accuracy, classifier):
"compute confusion matrix"
confusion_clf = sklearn.metrics.confusion_matrix(y_test, _predicted)
df_clf = pd.DataFrame(confusion_clf,
index=list(range(0, 2)),
columns=list(range(0, 2)))
plt.figure(figsize=(5.5, 4))
ax_heatmap=sns.heatmap(df_clf, annot=True, vmin=0, vmax=11, cmap="Blues")
plt.title(str(classifier) + ' \nAccuracy:{0:.3f}'.format(_accuracy))
plt.ylabel('True label')
plt.xlabel('Predicted label')
return df_clf,ax_heatmap
@staticmethod
def plot_precision_recall(clf, X_test, y_test, classifier, _predicted):
"plot precision recall curve"
_precision = sklearn.metrics.precision_score(y_test, _predicted)
_recall = sklearn.metrics.recall_score(y_test, _predicted)
y_score_clf = clf.predict_proba(X_test)
y_score_df = pd.DataFrame(data=y_score_clf)
precision, recall, thresholds = sklearn.metrics.precision_recall_curve(
y_test, y_score_df[1])
closest_zero = np.argmin(np.abs(thresholds))
closest_zero_p = precision[closest_zero]
closest_zero_r = recall[closest_zero]
plt.figure()
plt.xlim([0.0, 1.01])
plt.ylim([0.0, 1.01])
result,=plt.plot(precision, recall)
plt.title(
str(classifier) +
' Precision-Recall Curve \nprecision :{:0.2f}'.format(_precision) +
' recall: {:0.2f}'.format(_recall))
plt.plot(closest_zero_p,
closest_zero_r,
'o',
markersize=12,
fillstyle='none',
c='r',
mew=3)
plt.xlabel('Precision', fontsize=16)
plt.ylabel('Recall', fontsize=16)
plt.show()
return result
@staticmethod
def plot_roc(clf, X_test, y_test, classifier):
"plot roc curve"
y_score_clf = clf.predict_proba(X_test)
y_score_df = pd.DataFrame(data=y_score_clf)
fpr_clf, tpr_clf, _ = sklearn.metrics.roc_curve(y_test, y_score_df[1])
roc_auc_clf = sklearn.metrics.auc(fpr_clf, tpr_clf)
plt.figure()
plt.xlim([-0.01, 1.00])
plt.ylim([-0.01, 1.01])
result,=plt.plot(fpr_clf,
tpr_clf,
lw=3,
label=str(classifier) +
' ROC curve (area = {:0.2f})'.format(roc_auc_clf))
plt.xlabel('False Positive Rate', fontsize=16)
plt.ylabel('True Positive Rate', fontsize=16)
plt.title('ROC curve ' + str(classifier) +
' \nAUC:{0:.3f}'.format(roc_auc_clf),
fontsize=16)
plt.legend(loc='lower right', fontsize=13)
plt.plot([0, 1], [0, 1], color='navy', lw=3, linestyle='--')
plt.show()
return roc_auc_clf,result
dico_classifier = { 'knn': KNeighborsClassifier,
'naiveb': GaussianNB,
'randomforest': RandomForestClassifier,
'gtree': GradientBoostingClassifier,
'neural': MLPClassifier}
@staticmethod
def plot_heatmap(dataframe):
"plot heatmap of accuracy, precision, recall, AUC"
plt.figure()
sns.heatmap(dataframe, annot=True, vmin=0, vmax=1, cmap="Blues")
plt.title('scores des classifiers ')
plt.ylabel('scores')
plt.xlabel('modeles')
plt.show()
#process machine learning for all classifiers in dico_classifier
#and plot a heatmap of their accuracy, precision, recall, AUC
df_result = pd.DataFrame(data=(0, 0, 0, 0),
columns=['init'],
index=['accuracy', 'precision', 'recall', 'AUC'])
for clf in ML.dico_classifier:
print(clf)
result_ml = ML.ml(images_video, positives, clf)
df_result = pd.merge(df_result,
result_ml,
right_index=True,
left_index=True)
df_result.drop('init', axis=1, inplace=True)
ML.plot_heatmap(df_result)
return df_result
Conclusion: Voilà, maintenant que vous savez coder une IA, vous pouvez la critiquer d'autant mieux,et promouvoir les lowtech en connaissance de cause.
Vous noterez que les algorithmes d'ia sont open sources et assez faciles à utiliser en tant que développeur "simple utilisateur".
Et aussi que sans data, l'ia ne sert absolument à rien.
C'est pour cette raison que les géants de la tech veulent toujours plus de données et emploient des gens dans des conditions proches de l'esclavage dans de nombreux pays pour les traiter avant d'entrainer leurs modèles.
Tout comme pour les "données personnelles", la question clé des ia repose sur les données.
Voir l'excellente conf de benjamin bayart " Géopolitique de la data (Benjamin BAYART) " sur youtube ou en vidéo ici.
Vous pouvez aussi faire un tracker lowtech, mais aussi adapter le code pour créer un véhicule autonome lowtech avec 4 datasets/signaux positifs distincts pour entrainer l'activation de "tourner à gauche","accélérer", "tourner à droite", "freiner". C'est ce qu'a fait George Hotz en proclamant qu'il suffisait d'une trentaine d'heures de vidéos de conduite en enregistrant avec des capteurs pour avoir les signaux positifs correspondant aux images enregistrées pour que le machine learning fonctionne.
Evidemment, on espere qu'il n'y aura pas de hack ou que le système n'a pas de controle commande à distance sur ce type d'algorithme.
Published
Français
English
Deutsch
Español
Italiano
Português