Classification of rotary machine fault considering signal differences

Abstract

Machine learning for the diagnosis of rotary machines takes priority in generating a training data set through the machine's past data. The training data set uses features that have physical and statistical meaning of vibration signals. A training data is formed on the assumption that the normal condition of the facility is almost similar over time. However, many industrial power plants perform regular O/H (overhaul), and the vibration level of the machine's normal condition is likely to change depending on the O/H results. The vibration level is one of the important factors representing the condition change of rotating machines and is difficult to ignore easily. This paper is a study on a method that can be used for feature-based machine learning with training data formed from past data whose vibration level of the rotating machine has changed due to the influence of maintenance. Data acquisition was made through lab-scale defect simulation test devices, and experimental equipment was simulated before and after O/H with several faults that could occur in rotating machines. The signal named "delta signal" refers to a signal that sets each normal data as a reference signal, matches the fault signal through phase synchronization and resampling, and subtracts it to leave only a difference. The algorithms used in machine learning used genetic algorithm (GA) based feature selection and support vector machines (SVM) for learning and classification. According to the experiment, it was confirmed that in raw signal learning, the similarity by the learned condition (label) decreased due to the influence of maintenance, but the method using delta signal decreased the effect by maintenance, increasing the similarity within the same learned condition.