This paper demonstrates the differences between popular transformation-based input representations for vibration-based machine fault diagnosis. This paper highlights the dependency of different input representations on hyperparameter selection with the results of training different configurations of classical convolutional neural networks (CNNs) with three common benchmarking datasets. Raw temporal measurement, Fourier spectrum, envelope spectrum, and spectrogram input types are individually used to train CNNs. Many configurations of CNNs are trained, with variable input sizes, convolutional kernel sizes and stride. The results show that each input type favors different combinations of hyperparameters, and that each of the datasets studied yield different performance characteristics. The input sizes are found to be the most significant determiner of whether overfitting will occur. It is demonstrated that CNNs trained with spectrograms are less dependent on hyperparameter optimization over all three datasets. This paper demonstrates the wide range of performance achieved by CNNs when preprocessing method and hyperparameters are varied as well as their complex interaction, providing researchers with useful background information and a starting place for further optimization.
This paper demonstrates various data augmentation techniques that can be used when working with limited run-to-failure data to estimate health indicators related to the remaining useful life of roller bearings. The PRONOSTIA bearing prognosis dataset is used for benchmarking data augmentation techniques. The input to the networks are multi-dimensional frequency representations obtained by combining the spectra taken from two accelerometers. Data augmentation techniques are adapted from other machine learning fields and include adding Gaussian noise, region masking, masking noise, and pitch shifting. Augmented datasets are used in training a conventional CNN architecture comprising two convolutional and pooling layer sequences with batch normalization. Results from individually separating each bearing’s data for the purpose of validation shows that all methods, except pitch shifting, give improved validation accuracy on average. Masking noise and region masking both show the added benefit of dataset regularization by giving results that are more consistent after repeatedly training each configuration with new randomly generated augmented datasets. It is shown that gradually deteriorating bearings and bearings with abrupt failure are not treated significantly differently by the augmentation techniques.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.