**1. Introduction**

In this demo, I will make a simple Machine Learning - Artificial neural network system using Arduino ESP32. Certainly, I will not use ESP32 for training process; instead, i will use Python + numpy for training process. After training, I will use the result weights will be used by ESP32 for output calculation based on input.

Artificial neural network (ANN) systems are inspired by the biological neural networks. These systems can learn to do tasks by considering examples, without task-specific programming.

**One of most famous methods for ANN training is Backpropagation. You can refer here.**

In this demo: ESP32 with 2 buttons (B1 and B2) with possible states: pressed (0) or released (1) and there is 1 LED output with possible state state on (1) or off (0). The LED output can be set based on the state of input.

B1 | B2 | LED |
---|---|---|

pressed | pressed | off |

pressed | released | on |

released | pressed | on |

released | released | off |

**Figure: state of LED based on state of buttons**

**Figure: ANN structure of demo**

**2. Hardware**

Here we set input pins as INPUT_PULLUP so the schematic is simple and we can re-use the schematic in Demo 21.

**Figure: hardware connections**

[ESP32 GIO12 - BUTTON 1 - GND]

[ESP32 GIO13 - BUTTON 2 - GND]

[ESP32 GIO14 - LED - GND]

**3. Software**

**3.1 ANN is implemented by Python + numpy**

I used the code here.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 | import numpy as np epochs = 10000 # Number of iterations inputLayerSize, hiddenLayerSize, outputLayerSize = 2, 3, 1 L = 0.1 # learning rate X = np.array([[0,0], [0,1], [1,0], [1,1]]) # Buttons states array Y = np.array([ [0], [1], [1], [0]]) # LED states array def sigmoid (x): return 1/(1 + np.exp(-x)) # activation function # weights on layer inputs Wh = np.random.uniform(size=(inputLayerSize, hiddenLayerSize)) Wz = np.random.uniform(size=(hiddenLayerSize,outputLayerSize)) for i in range(epochs): H = sigmoid(np.dot(X, Wh)) # calculate forward part Z = np.dot(H,Wz) # E = Y - Z # calculate error dZ = E * L # delta Z Wz += H.T.dot(dZ) # calculate backpropagation part dH = dZ.dot(Wz.T) * sigmoid_deriv(H) # Wh += X.T.dot(dH) # update hidden layer weights print("**************** error ****************") print(E) print("***************** output **************") print(Z) print("*************** weights ***************") print("input to hidden layer weights: ") print(Wh) print("hidden to output layer weights: ") print(Wz) |

**Figure: training process and weights output**

**3.2 Arduino ESP32**

The Arduino code just do the forward part to calculate the output when getting the inputs.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 | #include <math.h> #define B1 12 #define B2 13 #define LED 14 int X[1][2] = {{1,0}}; /*these matrices was calculated by python */ float W1[2][3] = {{0.74000854, 4.47769531, -0.98692059}, {0.83034991, 4.48772758, -0.55733578}}; float W2[3][1] = {{-6.17234487}, {4.8835918}, {1.28875386}}; float Wo1[1][3]; float sum = 0; float Y = 0; /*sigmoid function*/ float sigmoid (float x) { return 1/(1 + exp(-x)); } void setup() { Serial.begin(115200); pinMode(B1, INPUT_PULLUP); pinMode(B2, INPUT_PULLUP); pinMode(LED, OUTPUT); digitalWrite(LED, LOW); } void loop() { X[0][0] = digitalRead(B1); X[0][1] = digitalRead(B2); printf("B1 = %d, B2 = %d\n", X[0][0], X[0][1]); /* calculate forward part based on weights */ //hidden layer for(int i=0; i<1; i++) { for(int j=0;j <3; j++) { for(int k=0; k<2; k++) { sum += X[i][k]*W1[k][j]; } Wo1[i][j] = sigmoid(sum); sum = 0; } } //output layer for(int i=0; i<1; i++) { for(int j=0;j <1; j++) { for(int k=0; k<3; k++) { Y += Wo1[i][k]*W2[k][j]; } } } printf("Y = %f\n", (Y)); Y = round(Y); digitalWrite(LED, int(Y)); Y = 0; delay(1000); } |

**4. Result**

## 5 comments:

sigmoid derivative function is missing

def sigmoid_deriv(x):

f = sigmoid(x)

return f * (1 - f)

source : https://stackoverflow.com/questions/10626134/derivative-of-sigmoid

Great post.

Very helpfull to start experiments.

Lot of your other posts are also well done.

Thx.

Thank friend,

this is feed forward part so you will not see derivative.

I am wrtting new blog about machine learning.

It is http://www.fossreview.com/

Hope you like it :)

Regards

I want this type of one.beacuse in recent days i searched this type of blog finally i got.thanks for this blog.

ccna Training in Chennai

ccna Training institute in Chennai

Python Training in Chennai

Python Classes in Chennai

Angularjs Training in Chennai

ccna Training in Velachery

ccna Training in Tambaram

You provide the concept is very different and useful. Thank you for sharing with us and continuing the great blogging.

Unix Training in Chennai

Unix shell scripting Training in Chennai

Tableau Training in Chennai

Excel Training in Chennai

Corporate Training in Chennai

Oracle Training in Chennai

Power BI Training in Chennai

Unix Training in Chennai

Unix shell scripting Training in Chennai

This is significant for clients who need to break down information from alternate points of view and the investigation of one viewpoint prompts business addresses that should be analyzed from different viewpoints. data science course in pune

Post a Comment