Redes neurais convolucionais na análise de vibrações para identificação de falhas de ignição em motores veiculares a combustão
Carregando...
Data
2022-09-01
Autores
Salmazo, Vinícius de Araújo
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
A manutenção é uma atividade que sempre existiu e que com o passar do tempo adquiriu novas dimensões, estando em constante processo de desenvolvimento. No atual cenário, as tecnologias têm evoluído rapidamente, viabilizando estudos e aplicações de novas abordagens para identificação e predição de falhas em sistemas mecânicos. Dentre as diversas invenções tecnológicas desenvolvidas pelo homem, o motor de combustão interna (MCI) se destaca pelos avanços que trouxe para a sociedade moderna. Uma de suas possíveis falhas é a falha de ignição, a qual possui diversas causas e pode ser crítica em relação ao aumento nas emissões de poluentes e perda de desempenho do motor. Trabalhos recentes têm abordado diferentes métodos para identificação da ocorrência desse tipo de falha, alguns empregando técnicas de Inteligência Artificial e Aprendizado de Máquina. Este trabalho tem como objetivo avaliar a aplicação de Redes Neurais Convolucionais (RNC) na identificação de falhas de ignição em MCIs. Para isso, um conjunto de aquisição de baixo custo foi confeccionado empregando um acelerômetro MEMS capacitivo. Sinais de vibração foram coletados a partir de dois MCIs similares de quatro cursos e ignição por centelha operando em quatro condições distintas. Os dados foram preparados no domínio da frequência em duas formas de entrada: utilizando gráficos e em formato de matrizes compactas das amplitudes. Hiperparâmetros (HPs) foram definidos para a RNC utilizando gráficos como entrada enquanto que para a RNC com dados em forma de matrizes alguns HPs foram selecionados por gridsearch. Treinando as RNCs até a convergência com dados do primeiro motor, acurácias superiores a 99% foram alcançadas para as duas redes classificando dados do banco de teste do mesmo motor. Ao classificar dados do segundo motor, o desempenho geral ficou entre 68% e 72%. Em relação ao tempo de classificação, a rede utilizando dados em formato compacto apresentou desempenho sete vezes superior, rotulando dados individuas em um tempo médio inferior a 418µs. Os resultados demonstram grande potencial de aplicação das RNCs em conjunto com o sistema desenvolvido para a tarefa proposta.
Maintenance is an activity that has always existed and that over the time it has acquired new dimensions, being in a constant process of development. Currently, technologies have been evolving quickly, enabling studies and applications of new approaches to identify and predict failures in mechanical systems. Among various technological inventions developed by man, the internal combustion engine (ICE) stands out due to the advances that it has brought to modern society. One of its possible failures is the misfire, which has several causes and can be critical in relation to the increase in pollutant emissions and loss of engine performance. Recent works have addressed different methods to identify the occurrence of this type of failure, some of them applying Artificial Intelligence and Machine Learning techniques. This work aims to evaluate the application of a Convolutional Neural Network (CNN) to identify the occurrence of misfires in ICEs. For this, a low cost acquisition set was developed using a capacitive MEMS accelerometer. Vibration signals were collected from two similar four-stroke spark-ignition ICEs operating under four different conditions. The data were prepared in the frequency domain for two types of input: using graphs and in a compact matrix of amplitudes format. The hyperparameters (HPs) were manually defined for the CNN using graphs as input while for the CNN with matrix data some HPs were selected by gridsearch. By training the CNNs till the convergence with data from the first engine, accuracies greater than 99% were achieved for both networks classifying data from the test dataset of the same engine. When sorting data from the second engine, the overall performance was between 68% and 72%. Regarding the classification time, the network using data in the compact format presented a seven times better performance, labeling individual data in an average time of less than 418µs. The results demonstrate great potential for the application of CNNs with the developed system for the proposed task.
Maintenance is an activity that has always existed and that over the time it has acquired new dimensions, being in a constant process of development. Currently, technologies have been evolving quickly, enabling studies and applications of new approaches to identify and predict failures in mechanical systems. Among various technological inventions developed by man, the internal combustion engine (ICE) stands out due to the advances that it has brought to modern society. One of its possible failures is the misfire, which has several causes and can be critical in relation to the increase in pollutant emissions and loss of engine performance. Recent works have addressed different methods to identify the occurrence of this type of failure, some of them applying Artificial Intelligence and Machine Learning techniques. This work aims to evaluate the application of a Convolutional Neural Network (CNN) to identify the occurrence of misfires in ICEs. For this, a low cost acquisition set was developed using a capacitive MEMS accelerometer. Vibration signals were collected from two similar four-stroke spark-ignition ICEs operating under four different conditions. The data were prepared in the frequency domain for two types of input: using graphs and in a compact matrix of amplitudes format. The hyperparameters (HPs) were manually defined for the CNN using graphs as input while for the CNN with matrix data some HPs were selected by gridsearch. By training the CNNs till the convergence with data from the first engine, accuracies greater than 99% were achieved for both networks classifying data from the test dataset of the same engine. When sorting data from the second engine, the overall performance was between 68% and 72%. Regarding the classification time, the network using data in the compact format presented a seven times better performance, labeling individual data in an average time of less than 418µs. The results demonstrate great potential for the application of CNNs with the developed system for the proposed task.
Descrição
Palavras-chave
Inteligência Artificial, Redes Neurais Convolucionais, Motor de Combustão Interna, Falha de Ignição, Artificial Intelligence, Convolutional Neural Networks, Internal Combustion Engine, Misfire