Loading...
Publication
Restricted boltzmann machine based autoencoders for the classification of faults in rotational mechanical systems
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 9.5 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Contexto O objetivo deste trabalho foi investigar a possibilidade de usar Redes Neuronais Profundas (DNN) para classificar falhas (na forma de sinais anómalos representando vibrações) em rolamentos de esferas, e comparar os resultados com os de algoritmos tradicionais. Métodos O conjunto de dados usado neste trabalho foi recolhido de sensores colocados nos rolamentos de esferas de um eixo conectado a um motor, sendo insuficiente para treinar eficientemente uma DNN. Para superar esse problema, o conjunto de dados foi aumentado, primeiro separandoo em conjuntos disjuntos para minimizar a contaminação dos conjuntos de treino e teste, e em seguida um método de janela deslizante foi aplicado a esses conjuntos para gerar novos conjuntos de treino e teste. Dois modelos de DNN baseados em Máquinas de Boltzmann Restrita (RBM) foram usados neste trabalho, com duas variantes para cada um desses modelos, usando números nítidos ou fuzzy para os parâmetros. Para quantificar os resultados, foram utilizadas as métricas Exatidão e área sob a curva (AUC). Os resultados foram analisados em conjunto com algoritmos tradicionais comuns, usados para estabelecer uma base de comparação. Resultados Os resultados obtidos com assim DNN foram promissores, com o melhor modelo atingindo uma exatidão média de 99.678%. Ainda assim, os resultados obtidos com os algoritmos tradicionais foram melhores, com o melhor modelo alcançando uma exatidão média de 99.98%. Para a outra métrica utilizada para analisar os resultados (AUC), o cenário é semelhante, com o melhor modelo de DNN a alcançar uma AUC média de 0.99995, e o melhor modelo tradicional uma pontuação perfeita de 1. Conclusão Os resultados obtidos revelam que, embora as DNN possam ser eficazes em problemas de classificação, elas não são a melhor escolha para problemas como este (conjuntos de dados constituídos por dados numéricos tabulares). Isso deve-se ao desenvolvimento, afinação e volume de dados de treino necessários em comparação com os algoritmos tradicionais, que além de obter resultados ligeiramente melhores, o fizeram com tempos de treino significativamente menores e quase sem afinação. Para outras aplicações, se houver dados de treino suficientes, as DNN frequentemente demonstram melhor desempenho, uma vez que quando os algoritmos tradicionais atingem um limite no desempenho, as DNN podem continuar a melhorar se tiverem dados de treino adicionais.
Background The aim of this work is to investigate the possibility of using Deep Neural Networks (DNN) to classify faults (in the form of anomalous signals representing vibrations) in ball bearings, and to compare the results with those of traditional algorithms. Methods The dataset used in this work was collected from sensors placed on the ball bearings of a shaft connected to a motor and was too small to effectively train a DNN. To overcome this problem, the dataset was augmented by first separating it into disjoint sets to minimize train-test contamination, then a sliding window method was applied to these sets to generate new training and test sets. Two Restricted Boltzmann Machine (RBM) based DNN models were used in this work, and for each of these models two variants, using either crisp or fuzzy numbers for the parameters. To quantify the results, the metrics Accuracy and area under the curve (AUC) were used. The results were analyzed together with those of common traditional algorithms, used to establish a base for comparison. Results The results obtained with the DNN were promising, the best model achieving a mean accuracy of 99.678%. Yet, the results obtained with the traditional algorithms were even better, with the best model achieving a mean accuracy of 99.98%. For the other metric used to analyze the results (AUC), the scenario is similar, with the best DNN model achieving a mean AUC of 0.99995, and the best traditional model a perfect score of 1. Conclusion The results obtained reveal that while DNN can be effective at classification problems, they are not the best choice for problems such as this (datasets consisting of tabular numeric data). This is due to the development, tuning, and volume of training data required compared to the traditional algorithms, which not only obtained slightly better results, but did so with significantly lower training times, and almost no tuning. For other applications, if there is enough training data, DNN routinely show better performance, since when traditional algorithms hit a plateau in performance, DNN can continue to improve if they are provided with additional training data.
Background The aim of this work is to investigate the possibility of using Deep Neural Networks (DNN) to classify faults (in the form of anomalous signals representing vibrations) in ball bearings, and to compare the results with those of traditional algorithms. Methods The dataset used in this work was collected from sensors placed on the ball bearings of a shaft connected to a motor and was too small to effectively train a DNN. To overcome this problem, the dataset was augmented by first separating it into disjoint sets to minimize train-test contamination, then a sliding window method was applied to these sets to generate new training and test sets. Two Restricted Boltzmann Machine (RBM) based DNN models were used in this work, and for each of these models two variants, using either crisp or fuzzy numbers for the parameters. To quantify the results, the metrics Accuracy and area under the curve (AUC) were used. The results were analyzed together with those of common traditional algorithms, used to establish a base for comparison. Results The results obtained with the DNN were promising, the best model achieving a mean accuracy of 99.678%. Yet, the results obtained with the traditional algorithms were even better, with the best model achieving a mean accuracy of 99.98%. For the other metric used to analyze the results (AUC), the scenario is similar, with the best DNN model achieving a mean AUC of 0.99995, and the best traditional model a perfect score of 1. Conclusion The results obtained reveal that while DNN can be effective at classification problems, they are not the best choice for problems such as this (datasets consisting of tabular numeric data). This is due to the development, tuning, and volume of training data required compared to the traditional algorithms, which not only obtained slightly better results, but did so with significantly lower training times, and almost no tuning. For other applications, if there is enough training data, DNN routinely show better performance, since when traditional algorithms hit a plateau in performance, DNN can continue to improve if they are provided with additional training data.
Description
Keywords
Deep neural networks (dnn) Restricted boltzmann machines rbm) Autoencoder (ae) Classifier