Effective Vibration-Based Signal Processing Methods for Condition Monitoring and Fault Diagnosis of Rolling Bearings


Student thesis: Doctoral Thesis

View graph of relations



Awarding Institution
Award date21 Aug 2018


Rotating machinery is widely used in industrial areas, and the rolling bearings are key components commonly used in the rotating machinery. As the failure of a bearing could cause cascading breakdowns of the mechanical system and then leading to economic loss or even human catastrophes, thus bearing’s health conditions have significant impacts on the reliability, availability, maintainability, and safety of the whole machines. Identification of different bearing fault types is another important task because it can help the operators replace the faulty components instead of replacing the whole rotating part, thus improving the maintenance efficiency. One of the primary methods for these purposes is based on the analysis of the measured vibration signals because such data are information-rich.

However, there still exists several problems. In real applications, it would be better if one technique can not only monitor the working conditions of the bearing but also can identify the fault type immediately after receiving a false alarm. In practice, the measured vibration signal from a faulty bearing is buried with strong background noises especially when the bearing is in the early-fault status. Moreover, the defect on the surface will cause shocks when the rolling elements pass over the defect, and it will also increase the probability of fault occurrence on other surfaces, leading to multiple faults commonly happened in bearings. To address these problems, four different methods are proposed in this research. The first two methods are proposed for condition monitoring and fault diagnosis for bearing multiple faults by using statistical techniques and signal processing methods; and the last two methods are proposed for hidden period extraction under strong background noises, which can be used for fault identification on the early state. These methods are itemized as follows:

(1) A TQWT-based integrated multiscale statistical process control method is proposed for bearing condition monitoring and fault diagnosis. This method utilizes the property of tunable Q-factor wavelet transform (TQWT) in representing the oscillatory behaviors in waveforms, and two novel health indicators are constructed based on the dominant wavelet coefficients and the remaining coefficients. Shewhart control charts on multiscale wavelet coefficients are constructed to find the out-of-control scales which contain fault-related information. With the assistance of Hilbert envelope analysis method, the fault type can be identified from the envelope spectrum of the reconstructed signal which is generated by those out-of-control scales.

(2) A wavelet-based statistical approach for bearing condition monitoring and diagnosis of multiple faults is proposed in this research. To ensure the independence across the decomposed scales, an orthonormal discrete wavelet transform (ODWT) is applied to obtain its energy dispersions at multiple levels. A health indicator based on the energy dispersions is thus proposed, the statistical properties of the proposed indicator under both the normal and faulty conditions are investigated, based on which a generalized likelihood ratio test (GLRT) is developed. To improve the detection sensitivity, an exponentially weighted moving average (EWMA) control chart is then constructed to detect faults at an early stage. The main scope of this method is to solve condition monitoring, fault diagnosis, early fault detection, and multiple fault diagnosis simultaneously.

(3) A noise resistant correlation method for period detection of noisy signals is proposed in this research. Two novel correlation functions are proposed based on a newly constructed periodic signal and the contaminated signal to effectively detect the target hidden period. Theoretical analysis of two proposed correlation functions is given in this research. This method is mainly proposed for hidden period detection under strong background noises, thus can be used for bearing fault diagnosis at an early-fault state.

(4) A bearing fault detection via B-spline constructed sparse method is proposed in this research. The proposed method utilizes the sparsity of faulty signals to construct a separated sparse representation (SSR) model with a tunable separation time parameter. Instead of using a predefined dictionary, a B-spline dictionary is adopted to represent the transients due to its inherent ability to model sparsity and its high flexibility. The proposed SSR model is solved by split augmented Lagrangian shrinkage algorithm (SALSA), and a power function is proposed as a criterion to detect the fault period.

The proposed methods have been validated using both simulation and experimental data. The test results have shown that the first two methods can be used to monitor the working condition of bearings and can effectively identify the multiple fault types. The last two have shown to be effective in detecting hidden period thus identifying the fault types, and both of them don’t need any prior knowledge about the analyzed data thus can also be used in other applications.