Dispositivo de ultra-bajo costo para la clasificación de licores usando tecnología TinyML

Objective: To design and implement a liquor identification system using TinyML, employing an MQ-135 sensor and an ESP32 microcontroller from Espressif Systems. The objective is to optimize local processing for high precision and low latency, thereby validating its viability as an accessible solution in resource-constrained environments.Methodology: The methodology employed an Artificial Neural Network (ANN) to classify liquors based on volatile compounds detected by the sensor. A dataset comprising 6,000 measurements was collected, and the ANN model was trained in MATLAB® R2018a. For performance evaluation, the dataset was split into 70% for training, 15% for validation, and 15% for testing, with the Mean Squared Error (MSE) used as the primary metric.Results:The deployed ANN model converged after 375 epochs, achieving a minimal Mean Squared Error (MSE) of 1.05×10⁻⁹ and a correlation coefficient (R) of 1.0. Synthesized onto the ESP32 microcontroller, the model utilized only 24.8% of Flash memory and 6.3% of RAM. The inference time recorded for execution was 2.43 ms, yielding a classification accuracy of 84.4%. This cost-efficient solution, priced at 15 USD, outperforms existing commercial options.Conclusions: It achieves an exceptionally fast inference time of 2.43 ms with a minimal operational energy consumption ranging from 5 mA to 10 mA. This performance, coupled with the low fabrication cost (15 USD), contrasts sharply with commercial solutions often exceeding hundreds of dollars. Consequently, the system is highly viable for implementation in embedded quality control and food industry applications.