Orbit Based Deep Learning Methods For Fault Monitoring And Diagnosis Of Rotating Machinery

With the quick development of advanced technologies,rotating machinery such as centrifugal compressors and steam turbines widely used,in various industrial fields like petrochemicals and oil & gas,has been designed to be more complicated,larger and more intelligent.As the key component of a rotating machine,a rotor is prone to fail under extremely harsh environments such as superhigh speed and corrosion,resulting in unplanned breakdown of the system.Therefore,the fault monitoring and diagnosis of rotating components is of paramount importance to ensure the safety and reliability of rotating machines.This thesis intends to exploit the orbit-based deep learning methods to monitor and diagnose the faults of a rotating machine.Axis trajectories or orbits generated by shaft vibration signals contain a great amount of important information related to machine failure modes.Each failure mode is generally represented by the trajectory shape corresponding to individual failure mechanisms.As such,the healthy status of the machine can be monitored with faults identified through tracking the shape change of the orbit images.Traditional data-driven methods usually require complicated feature engineering and largely rely on expert inputs for feature selection,which unavoidably brings subjective uncertainty to impact the model accuracy.To address this issue,this study develops orbit mechanism-driven deep learning methods to automatic feature extraction for accurate fault monitoring and diagnosis of rotatory machines,mainly including the following three aspects:(1)This study systematically investigates the mechanism of four commonly observed failure modes in rotatory machines,namely rotor rubbing,oil whirling,rotor unbalance,and misalignment,as well as the characteristics of their mapped orbit images.The related technologies including signal denoising,fault detection,image preprocessing,and deep learning are described to provide theoretical fundamental for the development of orbit mechanism-driven new models.(2)In order to enhance model generalization,monitoring robustness,and ability of handling multi-dimensional heterogeneous data,this study presents a novel orbit mechanismdriven convolutional long short-term memory(Conv LSTM)model for early fault warming of rotatory machines.The new model is able to effectively capture both temporal and spatial features from an orbit image series.A generalized flowchart is developed to automate the implementation of the proposed methodology for fault monitoring.Two cases with simulated and real-world monitoring data are employed to demonstrate the feasibility and availability of the proposed model.Numerical results have indicated that the proposed methodology provides accurate fault identifying for rotating machinery based on the orbit image series.(3)In order to address the issue of insufficient samples on the image-based fault diagnosis,this study presented a deep convolutional generative adversarial network(DCGAN)based on orbit mechanism,so-called Orbit GAN,to enhance the images respresentativity for accurate fault diagnosis of rotatory machines.A deep convolutional recognition network,Orbit CNN,is constructed to diagnose various fault modes of general rotating machines based on the enhanced orbit images.Techniques including standardization,regularization and cross validation are utilized to avoid the potential over fitting in the model training.A generalized procedure is proposed to implement the proposed methodology for automatic fault identification in various rotating machines.A comparison study with real-life monitoring data and fault events of three various rotatory machines is conducted to illustrate the effectiveness,generalization and robustness of the proposed methodology and procedure.