Estimativa de torque de carga por meio de imagens geradas a partir do fluxo magnético no entreferro de um motor de indução trifásico

• Ano de defesa: 2024
• Autor(a) principal: Fontenele, Jhone
• Orientador(a): Dias, Cleber Gustavo
• Banca de defesa: Dias, Cleber Gustavo , Flauzino, Rogério Andrade , Pereira, Fabio Henrique , Alves, Wonder Alexandre Luz
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Nove de Julho
• Programa de Pós-Graduação: Programa de Pós-Graduação em Informática e Gestão do Conhecimento
• Departamento: Informática
• País: Brasil
• Palavras-chave em Português: estimativa de torque; preprocessamento de imagem; processamento de sinal; redes neurais convolucionais; motores de indução trifásicos; máquinas rotativas; inception V3
• Palavras-chave em Inglês: torque estimate; image preprocessing; signal processing; convolutional neural networks; three-phase induction motors; rotational machines
• Área do conhecimento CNPq: CIENCIA DA COMPUTACAO::SISTEMAS DE COMPUTACAO
• Link de acesso: http://bibliotecatede.uninove.br/handle/tede/3513

The induction motor is an important asset in industrial processes, being part of the technological transformation in society since its inception. It has been the subject of study and improvement for several decades. Among the variables of interest aimed at monitoring its operational conditions, the torque applied to the motor shaft stands out. Such information plays a significant role in various applications, especially for assessing its performance and operational failures. Direct measurement of load torque usually requires the mechanical coupling of a torque meter between the machine’s shaft and the load, which makes installation difficult due to limited equipment access, particularly for large motors. For instance, this poses a challenge. To estimate the load torque of a three-phase induction motor using signals collected from a Hall effect probe installed in the motor air gap, this study employed a signal preprocessing technique. It discretized 100,000 samples obtained over a 10-second interval for each experiment, with a sampling frequency of 10kHz. For estimating torque in time intervals shorter than 10s, the signal was fragmented into smaller windows of 100ms, 200ms, and 400ms. This process generated images in three different sizes based on the orientation of quantized sample values in 255 shades of gray. These images served as inputs for an inception-type convolutional neural network. The study evaluated different conditions and hyperparameter values, resulting in 54 combinations. Model validation was conducted by analyzing the estimated load torque against measured load torque using a load cell after training the convolutional neural network model. The proposed approach successfully estimated load torque using three optimizers across almost the entire motor load operational range, spanning from 1.5% to 93.9% of the rated load. Four model configurations achieved a mean absolute percentage error MAPE of less than 3.5%. Specifically, two models for a 40x50 pixel image achieved MAPEs of 3.7% and 3%, one model for a 40x25 pixel image achieved a MAPE of 3.5%, and one model for a 50x80 pixel image achieved a MAPE of 3.3%. This approach was experimentally validated on a 7.5kW three-phase induction motor.