Modelo de aprendizado de máquina para detecção e diagnóstico de falhas operacionais em sistemas de refrigeração

Abstract

Since the invention of the first vapor compression refrigeration machines up to modern systems, refrigeration has become essential for food preservation. Given this importance, ensuring the operational continuity of refrigeration systems through planned maintenance allows businesses and industries to avoid economic losses and sanitary risks. In this context, traditional approaches present important limitations, such as reliance on manual inspections, point based measurements, subjectivity in analysis, and a predominantly reactive nature. The use of generalized mathematical models tends to show low performance in scenarios with unbalanced data and difficulties in formulation due to the constructive variability of refrigeration systems. However, a transition to data driven predictive maintenance based on machine learning can overcome these limitations, as such methods can anticipate failures, increase reliability, and enable more effective strategies. This work aims to build a machine learning model capable of automatically detecting and diagnosing faults in a small cold storage chamber. The proposed system uses pressure and temperature sensors to detect operational deviations (faults), classifying the state as normal operation when no anomalies are present or into three possible fault states when behavior exceeds the operational limits established for the equipment. Data was collected in real time with internal temperature ranging from -10 ºC to -20 ºC and external temperature between 18 ºC and 35 ºC, representing different operating conditions of the mini cold chamber. In the implementation of the solution, two different classifiers were used one parametric and the other nonparametric, with the intent of identifying the best learning model. Among the algorithms evaluated, the Support Vector Machines classifier achieved the best performance on the experimental dataset, reaching an F1-score of 99.4%, outperforming the Multi-Layer Perceptron. Consequently, the results demonstrate that a machine learning tool can detect and diagnose anomalies in a real environment by monitoring pressure and temperature variables. [resumo fornecido pelo autor]