Diseño y construcción de un clasificador 4.0 basado en Machine Learning para el control de calidad de productos del sector agrícola.

Abstract

The objective of this degree work was the implementation of an electronic prototype with the capacity to supervise, classify and control the quality of products in the agricultural sector, through the application of soft computing techniques and machine learning algorithms that contributed to the identification of images in real time. The following methods were used to implement this system: deductive, heuristic, and experimental. The development of the degree work consisted of the following phases: First, the computer-aided design (CAD) of the structural delineation of the conveyor belt and the actuators was carried out. Second, the web interface was planned and implemented using the P5.js JavaScript library that allowed for real-time image recording via a webcam. Then, an automatic learning management model was created using the JavaScript library ml5.js for artificial intelligence that facilitated the implementation of the "k nearest neighbors" algorithm to create a model through a training stage. For the establishment of communication and sending of processed control data, from the neuron to the ESP8266 microcontroller, a central server was configured for the transport of MQTT messages (broker). Finally, 10 tests were performed for each of the variables (Good, Regular, and Discarded Product). It was concluded that the average efficiency of the system is 90% with a database of 45 to 50 images for each variable. It is recommended to make a good training process considering that the success of the operation is directly proportional to the number of images captured for better learning.