Bayesian Hyper-parameter Optimization in One-dimensional Convolutional Autoencoder for Monitoring Bearing Health Status

Abstract

This study introduces a framework that integrates hyperparameter optimization techniques for feature representation and abnormality detection in machine health diagnostics. The framework leverages one-dimensional convolutional autoencoders (1D-CAEs) and multilayer perceptron (MLP) as the key components. We propose a systematic approach that begins with the development of the 1DCAE model to encode characteristics into a latent vector. Subsequently, we employ an MLP to effectively detect and identify abnormalities within the encoded data. Our study presents a novel contribution by integrating an adjusted hybrid stopping criterion into the Bayesian optimization framework to improve the efficiency and robustness of anomaly detection. The accuracy and training time of these approaches were examined and evaluated using the bearing degradation data under real-life scenarios. The findings of our study provide strong evidence that the utilization of the Bayesian optimization method, along with the suggested stopping criterion, surpasses conventional tuning techniques in both training time and accuracy. This achievement sets a novel standard for optimizing hyperparameters in the prognostics and health management field. The superiority observed in our study extends beyond the fidelity of the derived models, encompassing the accelerated identification of ideal hyperparameters and holding practical value for real-world applications for early detection of abnormality.