Avaliação de cenários para detecção de falhas utilizando redes neurais artificiais
Arquivos
Data
2021-08-02
Autores
Bianchin, Luiz Daniel 
Orientadores
Martins, Tiago Dias
Tipo
Trabalho de conclusão de curso
Título da Revista
ISSN da Revista
Título de Volume
Resumo
As indústrias estão se tornando cada dia mais complexas conforme as necessidades sociais, políticas e econômicas se desenvolvem, logo, é necessário que haja um desenvolvimento concomitante de técnicas e estratégias que consigam manter os processos industriais com alta rentabilidade, responsabilidade ambiental e excelentes parâmetros de qualidade e segurança. O controle de processos em âmbito industrial passou por várias etapas no passado, passando de controle humano para o controle com a confiabilidade computacional. As redes neurais artificiais (RNA) surgem como uma boa alternativa de modelo caixa preta para análise de falhas dentro de processos industriais. Para treinar as RNAs, os dados provenientes de simulação do processo Tennessee Eastman foram processados através de Análise de Componentes Principais (ACP), onde somente 23 variáveis mostraram-se suficientes para explicar 85% da variância do problema. Foram treinadas diversas estruturas de RNAs considerando diferentes métodos de otimização, funções de ativação, número de camadas e neurônios. O melhor resultado foi obtido com a RNA cuja estrutura possui 23 neurônios na camada de entrada, 10 neurônios na primeira camada intermediária, 20 neurônios na segunda camada intermediária e 1 neurônio na camada de saída (23-10-20-1), treinada com o algoritmo de Levenberg-Marquardt, contendo a função de ativação tangente hiperbólica em ambas as camadas intermediárias e a função linear na camada de saída. Essa RNA foi selecionada buscando alta acurácia e baixo nível de falsos negativos. Os resultados da RNA classificatória foram analisados através da Receiver Operating Characteristic Curve (Curva ROC) e obtiveram uma acurácia de 80% em média para as 8 falhas simuladas no sistema. Por fim, este trabalho mostrou que a combinação entre ACP e RNAs pode ser uma alternativa promissora para a obtenção de modelos para detecção de falhas.
Industries are becoming increasingly complex as social, political and economic needs develop, so there needs to be a concomitant development of techniques and strategies that are able to maintain industrial processes with high profitability, environmental responsibility and excellent quality parameters and security. Industrial process control has gone through several steps in the past, moving from human control to control with computational reliability. Artificial neural networks appear as a good black box alternative model for failure analysis within industrial processes. To train the ANNs, the data from the Tennessee Eastman process simulation were processed through Principal Component Analysis (PCA), where only 23 variables were sufficient to explain 85% of the problem's variance. Several ANN structures were trained considering different optimization methods, activation functions, number of layers and neurons. The best result was obtained with the ANN whose structure has 23 neurons in the input layer, 10 neurons in the first intermediate layer, 20 neurons in the second intermediate layer and 1 neuron in the output layer (23-10-20-1), trained with the Levenberg-Marquardt algorithm, containing the hyperbolic tangent activation function in both intermediate layers and the linear function in the output layer. This ANN was selected looking for high accuracy and low level of false negatives. The classification ANN results were analyzed using the Receiver Operating Characteristic Curve (ROC curve) and obtained an average accuracy of 80% for the 8 simulated failures in the system. Finally, this work showed that the combination of PCA and ANNs can be a promising alternative for obtaining models for fault detection.
Industries are becoming increasingly complex as social, political and economic needs develop, so there needs to be a concomitant development of techniques and strategies that are able to maintain industrial processes with high profitability, environmental responsibility and excellent quality parameters and security. Industrial process control has gone through several steps in the past, moving from human control to control with computational reliability. Artificial neural networks appear as a good black box alternative model for failure analysis within industrial processes. To train the ANNs, the data from the Tennessee Eastman process simulation were processed through Principal Component Analysis (PCA), where only 23 variables were sufficient to explain 85% of the problem's variance. Several ANN structures were trained considering different optimization methods, activation functions, number of layers and neurons. The best result was obtained with the ANN whose structure has 23 neurons in the input layer, 10 neurons in the first intermediate layer, 20 neurons in the second intermediate layer and 1 neuron in the output layer (23-10-20-1), trained with the Levenberg-Marquardt algorithm, containing the hyperbolic tangent activation function in both intermediate layers and the linear function in the output layer. This ANN was selected looking for high accuracy and low level of false negatives. The classification ANN results were analyzed using the Receiver Operating Characteristic Curve (ROC curve) and obtained an average accuracy of 80% for the 8 simulated failures in the system. Finally, this work showed that the combination of PCA and ANNs can be a promising alternative for obtaining models for fault detection.