Accelerometer-Based Bearing Condition Indicator Estimation Using Semi-Supervised Adaptive DSVDD

Abstract

To prevent unexpected and costly asset downtime, an accurate estimate of the assets condition is needed. In such prognostics setting typically a Condition Indicator (CI) score is calculated based on measurement data. When a future CI surpasses some predefined threshold an alarm can be automatically triggered. Recently, in the literature (data-driven) Deep Learning (DL) methods were proposed to estimate the CI. However, to the best of our knowledge such DL algorithms were never validated in terms of the quality of the CI they produce. In this work the CI estimation performance of Deep Support Vector Data Descriptor (DSVDD) was validated and compared to autoencoder and convolutional neural network alternatives using data from a rolling element bearing setup that is measured by an acceleration sensor. Furthermore, since there can be a distributional shift between data from the training and test runs, which might lead to a degradation of the model performance, both the effect and computational complexity of various SVDD model adaptation strategies on the model performance are studied. Only adapting the final layer of the model gives a performance comparable to that when the full model is adapted to the target domain while requiring less calculations. The paper also proposes a simple and easy to calculate center adaptation strategy. This procedure had a slightly reduced CI estimation performance compared to the previous two adaptation alternatives but does not require any DL training.