Research on Fault Diagnosis and Prognosis for Key Components of Rotating Machinery


Student thesis: Doctoral Thesis

View graph of relations


  • Changqing SHEN


Awarding Institution
Award date5 Jun 2014


Over the past decades, rotating machinery has been widely used in a range of mechanical transmission systems, including automobile transmission systems, high-speed trains and slurry pumps. Unexpected failures in such machinery can cause breakdowns, leading to significant economic losses and, even wor se, human casualties. Components like bearings, gearboxes and impellers, are commonly used in rotating machinery to support their rotating shafts, transmit torque and deliver liquid respectively. Statistical surveys suggest that these components are frequently failed and the major causes of machine breakdown. Hence, to avoid catastrophe and minimise machinery downtime, the machine fault diagnosis and its prognosis are of great industrial significance. Accordingly, the development of machine fault diagnosis and prognosis methods have attracted considerable research attention in recent decades. This thesis aims to report the development of a variety of effective methods to identify vital fault indicators, perform proper fault diagnosis and predict the remaining useful life of a number of key components in rotating machinery. The followings are the descriptions of each developed effective method and its tested results.

The first method presented is the method called morphology. The mathematical morphology is a kind of nonlinear analysis method that has been developed and successfully applied to various applications, such as image processing and signal analysis. It has drawn the attention of researchers in bearing fault diagnosis recently. In this method, the determination of the length of the structure element (SE) is crucial and may significantly affect the results. To ensure a success application, some of the proposed algorithms need some prior knowledge of the raw signal, such as the periodic interval of generated impulses or a fixed SE length in single-scale morphology analysis. Besides, some of the existing techniques need to repeat the morphological transform for many times with different scales. Based on these considerations, an adaptive varying-scale morphology analysis method is introduced here. Different from the previous morphology analyses, the SEs are determined by the time interval between two adjacent local impulsive peaks. Compared to other currently using morphology analyses, the improved method needs little prior knowledge and does not require intensive calculations as that needed by conventional multi-scale morphology analysis.

The second developed method is the new Doppler transient model for the locomotive bearing fault diagnosis. A well-known scenario is that Doppler effect may significantly distort the collected acoustic signals during high movement speeds, consequently increase the difficulty in monitoring locomotive bearings online. In this thesis, a new Doppler transient model based on the acoustic theory and the Laplace wavelet are presented for the identification of fault-related impact intervals embedded in the collected acoustic signals. An envelope spectrum correlation assessment is conducted between the transient model and the real fault signal in the frequency domain to optimize the model parameters. This new method has been proven that it can reveal the parameters used for simulated transients (the time intervals of the simulated transients) embedded in the acoustic signals. Therefore, the time intervals of localized impacts generated by a defective bearing become detectable once the parameters of transients have been successfully identified.

Third, a new intelligent fault diagnosis method based on the support vector regression model has been developed. For complex machinery, it is difficult to extract the fault-induced signatures via signal processing methods as the fault-related features could be greatly overwhelmed by useless components, like noise and irrelevant vibrations generated by non-related components. Accordingly, the artificial intelligence-based fault diagnosis method is introduced here as it has the potential to tackle these kinds of problem. The intelligent fault diagnosis method is consisted of the extraction of statistical parameters from the paving of a wavelet packet transform, a distance evaluation technique and a support vector regression based generic multi-class solver. The collected signals are first pre-processed by the wavelet packet transform at different decomposition depths. Statistical parameters are then extracted from the signals via the wavelet packet transform at different decomposition depths. A distance evaluation technique is then employed to reduce the dimensionality of the feature space. Finally, a support vector regressive based generic multi-class solver is used to identify each existing fault pattern. The intelligent fault diagnosis method demonstrates that it can provide higher accuracy and robustness in machine fault diagnosis as compared to that provided by other existing methods.

The deterioration of a normal bearing is a gradual process in functional failure. Bearing failure usually occurs when an early bearing fault grows to a severe extent. Maintenance actions taken too early or late are not suitable, as they may lead to too much or too little maintenance. Hence, the fourth development is a method that can trace the size of a bearing’s defect is meaningful for timely maintenance actions. A comprehensive bearing fault diagnostic and prognostic method should not only recognize each fault pattern, but also determine the defect size of each pattern so that the deteriorating rate of a particular kind of defective bearing can be evaluated. In order to establish a systematic scheme for diagnosing early bearing fault patterns and sizes, a new two-layer structure consists of support vector regression machines (SVRMs) is proposed to accurately recognize both bearing fault patterns and defect sizes. Selected statistical features are first extracted from the raw vibration signals collected from a machine under different bearing health conditions. The two-layer SVRMs are then trained with the feature vectors by defining decision-making functions based on the regression functions which have continuous outputs for the two layers. Finally, the bearing fault patterns and sizes can be intelligently recognized by the nonlinear models created in the two layers of SVRM.

The fifth developed method is aimed to predict the remaining useful life of a particular component of a slurry pump normally used in the oil production or mining fields. The impeller is an important and frequently failed component in the slurry pump. The mixtures transported by the pumps may include liquid and different sizes of solid, such as sand and rocks. The continuous bombardment to the impeller surface by these sand and rocks is definitely making the pump works in a very abrasive and erosive environment. Long duration working in such environment definitely causes the impeller suffering from continuous wear. To monitor the wear rate on impeller, an effective data driven technique for estimating the practical remaining useful life of impeller has been developed based on the estimation of the failure probability density function that can be obtained by using support vector machine (SVM) again. Selected spectral energy features are extracted from the raw signals under different wear conditions that have known values of remaining useful life. Then an intelligent impeller health status probability estimation model has been constructed by using the one-against-one SVM. Finally, by identifying the relationship of each health status probability to the historical remaining useful life values, the SVM is able to predict the remaining useful life of a wearing impeller.

A total of five contributions have been briefly introduced here. With the help of these five methods, different degrees in the abilities of mechanical component fault diagnosis and prognosis have been developed for rotating machinery. These methods are all adaptive to the monitored machine components which have different mechanical characteristics and operating in different working conditions. The results generated from simulations and industrial case studies, they proves that the aforementioned five methods yield satisfactory performance in fault diagnosis and prognosis and are superior to other conventional methods.

    Research areas

  • Signal Processing, Vital Feature Extraction, Transient Model, Parameter Optimization, Rotating Machinery, Fault Diagnosis, Remaining Useful Life Prediction