Investigation on Preprocessing and Transformation of Vibration Signals for Deep Learning Based Diagnosis of Rotating Machinery

Abstract

Large-scale rotating machinery requires a reliable diagnosis method that accurately predicts health state, since these systems are frequently operated in safety-related and mission-critical systems (e.g., turbines in power plants). Among various methods for rotor-system diagnosis, the data-driven approach has received considerable interest from industry and academia. Specifically, the number of research papers on deep learning based rotor system diagnosis has risen steeply in the past few years. Interest is driven, in part, by the fact that deep learning algorithms are applicable to complex systems without the need for a complete comprehension of the physics of the system. However, powerful performance of these diagnosis methods can only be achieved with the use of optimal preprocessing techniques for each target system. Thus, this dissertation focuses on developing preprocessing and transformation steps for a deep learning based diagnosis system for rotating machinery. This work specifically focuses on fluid-film bearing rotor systems.

The dissertation investigates three thrusts of preprocessing and transformation of vibration signals: 1) study of the optimal vibration image size, considering filter size, 2) research into a label-based, mini-batch gradient descent method with filter sensitivity analysis, and 3) investigation of a retraining scheme for minor classes in imbalanced data problems.

The first research thrust investigates the size of input images for convolutional neural network (CNN) based diagnosis. As a fluid-film bearing rotor system presents directional dependent health states, vibration images that consider both the temporal and the spatial correlations of omnidirectional regeneration (ODR) signals are suggested. Using the generated images, the results show that the ratio of image size to filter size affects the overall performance. Thus, the optimal range of size ratio for the vibration image is derived in this work by analyzing the performance of various ratios.

The second research thrust suggests a label-based, mini-batch gradient descent method. As the conventional random mini-batch method generates biased mini-batches in several cases, which leads to decreased overall performance, the proposed method can reduce the bias between mini-batches. In addition, various label-based, mini-batch combinations were studied in this work and their performance deviation was analyzed by filter sensitivity analysis. The result shows that the quantity of properly sensitive filters clearly improves the overall performance of the network.

Finally, the last research thrust proposes a retraining scheme for minority class data in imbalanced data set problems. The proposed two-phase approach uses equally labeled mini-batches, proposed in the second thrust, with oversampling of the minor class samples. Furthermore, in the second phase of training, filters with high sensitivity are frozen and filters with low sensitivity are retrained to represent the minor class samples. The resulting method shows increased performance by improving the recognition of the minority class samples in several imbalanced data set problems.