| Cutting-edge intelligent manufacturing is the heart of competition in the future man-ufacturing industry.It serves as the core competitiveness among countries.Researchers,scholars,professors,and experts in the world have been devoting research on intelli-gent operation and maintenance of advanced,sophisticated,and cutting- edge industrial assets.For large- scale and complicated electromechanical systems,the transmission sys-tem(gearboxes,bearings)is often exposed to extreme environmental variables,such as fluctuating speed,heavy load,high temperature,and frosty weather states.Such states make the critical components of the transmission system prone to frequent damages and various sorts of failures.As a result,unexpected malfunctions and even catastrophic acci-dents may be unavoidably led to extra maintenance costs.To prevent accidents and reduce Operation and maintenance(O&M)costs,effective fault detection and diagnosis(FDD)and health condition monitoring are needed to provide early alerts as well as enable better maintenance plans.The critical components of mechanical systems often work in complicated operating conditions.The time- varying vibration signals of the critical components are generated by local defects,signal transmission path effects,fluctuating operating conditions,etc.To accommodate the time-varying and non-stationary operating conditions,especially ex-tracting the symptoms among the time-varying and non-stationary signal,signal process-ing technology of non-linear and non-stationary appealed the researchers’ and scholars’ attention in the fields of fault diagnostics and health condition monitoring.In recent years,non-linear and non-stationary signal processing methods have been developed and applied to various types of industrial equipment.However,there are still challenges for effectively extracting the fault symptoms and adaptive health monitoring for the core components in a mechanical system.Specifically,the current time-varying signal representation ap-proaches cannot handle the demodulation analysis of fast,strong frequency modulated,overlapped,and non-stationary signals caused by complicated operating conditions.To this end,this dissertation attempts to investigate both theory and applications of advanced methods including time- frequency energy distribution representation,separation methods of multi-component and overlapped signals,threshold self- setting for health condition monitoring,and transfer learning among different domains under varying operating con-ditions.The principal research work and contributions are summarized as follows,(1)Current time- frequency analysis techniques for analyzing strong frequency mod-ulated non- stationary signals caused by complicated operating conditions have limita-tions such as low energy concentrations of time- frequency representation and poor per-formance of time- frequency readability and reconstruction ability of specific components.A new time- frequency analysis method called Second- order synchro-extracting transform(SSET)based on a short time Fourier architecture is proposed.The proposed method of-fers further estimation of the local instantaneous frequency to accommodate the fast and strong frequency modulation signals.Thus,the issue of insufficient estimation accuracy of the local instantaneous frequency within traditional synchroextracting transform can be properly addressed.(2)To separate the multi- component signal,including fast,strong frequency modu-lation and overlapped components,a time-frequency method called Recursive mapping demodulation high- order synchro-extracting transform(RMDHSET)is proposed.The proposed method principally deals with the issues of separating of overlapped compo-nents and insufficient estimation accuracy of the local instantaneous frequency.The pro-posed method is utilized and validated by the fault detection and diagnostic for planetary gearbox.(3)Most current data- driven algorithms for prognostics and health management im-plementations require a substantial amount of historical health operating data,as well as failure data.Nevertheless,for the rotating machinery like wind turbines as an expen-sive critical asset,failure data are often exceptionally sparse due to the long life cycles of these assets.As well,the issue of threshold setting for health management is restricted by the current unsupervised deep learning algorithms.Consequently,a threshold self-setting adaptation health condition monitoring method is proposed to overcome the above constraints.Furthermore,to quantitatively evaluate and track the fault severity of a wind turbine with different measurements,a sample discrepancy method is developed for eval-uating the health conditions.Eventually,by tracking two real wind turbine datasets from real wind farms,the effectiveness of the proposed method is validated.(4)Traditional signal processing methods for fault detection and diagnosis of the critical components of mechanical transmission systems are heavily dependent on a clear understanding of the behavior characteristics of gearboxes or bearings and human experi-ence from real practice.To accommodate the complicated operating conditions and mas-sive amounts of data,an operating condition adaptation neural network model is proposed for establishing an end to end learning manner with a diagnostic algorithm for planetary gearboxes.In the end,the proposed methods are verified by an experimental planetary gearbox platform under time- varying operating conditions.(5)Most current transfer learning models pay attention to the discrepancy between source and target domains,instead of the operating condition data.To implement FDD under different time- varying operating conditions,a novel knowledge transfer network with fluctuating operational condition adaptation for bearing fault detection and diagnosis is proposed.The proposed model enables the capacity of FDD under fluctuating opera-tional conditions for rolling bearing fault detection.Lastly,experiment results in bearings are used to demonstrate the strength of the proposed method. |
PDF Full Text Request