Research On Roller Bearing Fault Diagnosis Methods Based On Chemical Reaction Optimization Algorithms And Support Vector Machine

Roller bearings are key parts in rotating machines, which are widely used in industrial applications. Roller bearing fault diagnosis is therefore meaningful. Fault diagnosis of rolling bearings is actually a pattern recognition process, the key is to feature extraction and state identification.Vibration signal processing is commonly used feature extraction methods, which include time-frequency analysis methods. Time-frequency analysis methods are divided into two types of linear and nonlinear analysis method. Linear time-frequency analysis methods comprise short-time Fourier transform (STFT), Gabor transform and wavelet transform. Non-linear time-frequency analysis methods include discrete Fourier transform (DFT), Wigner-Ville distribution, Choi-Wiliam distribution, short-time Fourier transform (STFT). STFT shows the relationship between time and frequency but they are dominated by the Heisenberg uncertainty principle for high frequency components and low frequency signals. Wigner-Ville distribution are designed for linear but non-stationary data. The Fourier transform breaks the signal into a series of sine waves of different frequencies while the wavelet transform breaks the signal into its wavelets. Disadvantages of above methods are they can not meet the requirements of precise analysis of non-stationary and nonlinear signal. Wavelet transform is feasible for linear signal while roller bearing signal is nonlinear and nonstationary signal. Furthermore, it requires an adjustable window Fourier transformation in which the obtained result is fixed frequency band signal that only the frequency range and the ling frequency of signal is relevant. In fact, the signal itself has nothing to change. Hilbert-Huang transform (HHT) includes two parts are Empirical mode decomposition (EMD) and Hilbert spectral analysis (HAS). The EMD method analyzes the signal by decomposing the signal into its monocomponent, called as Intrinsic Mode Functions (IMF). This method has potentially viable for nonlinear and nonstationary data analysis, especially for time-frequency-energy representations, the end effect, mode mixing, overshoot and undershoot, negative frequencies-instantaneous frequency, and a lack of a theoretical foundation are all current drawbacks of EMD.Recently, a new adaptive time-frequency analysis method namely Local mean decomposition (LMD) is proposed by Smith. LMD principle is based on moving average in which a signal can be decomposed into a number of product functions (PFs) and a monotonic function. The different lengths of moving average (MA) which used in LMD have an important effect on decomposition results. Furthermore, the criterion for pure frequency modulation (FM) signal also has an effect on performance of this method. Comparing with EMD method, LMD has more advantages such as less iterative times, unobvious end effect and less phoniness components of the instantaneous frequency. The former of local mean which derived by LMD cannot be affected by overshoot and undershoot, which are drawbacks in EMD method. In addition, LMD local magnitude has more oscillatory characteristics than EMD upper envelope. Therefore, both the local mean and local magnitude of LMD based on the local characteristic time-scale of the signal.In this thesis, another method namely local characteristic-scale decomposition (LCD) method is researched. By using LCD method, a signal can be decomposed into several intrinsic scale components (ISCs). Every ISC involves the local characteristic of signal so characteristic information of the original signal can be extracted more accuracy and effectively. The LCD method is superior to the EMD method in reducing the end effect and the iteration time and in the accuracy of the instantaneous characteristic. LCD remedies the problem of marginal effect when signal are decomposed by EMD. With above advantages, LMD and LCD method are proposed as the pre-processing methods for roller bearing fault diagnosis in this thesis.Bearing fault diagnosis is actually a pattern recognition process with two commonly used methods which include artificial neural network (ANN) and support vector machine (SVM). The ANN has some drawbacks is that it is very difficult to determine why a particular conclusion was reached. Furthermore, it requires a large number of samples that is difficult in practice. In addition, slow convergence speed has also an effect on the computational time. SVM is a powerful machine learning method, which based on statistical learning theory and the structural risk minimization principle. SVM not only can solve the problems of overfitting, local optimal solutions, and slow convergence rates that exist in ANN, but it also have an excellent generalization capability in situations where there are a small number of samples. Furthermore, SVM can solve nonlinear, high-dimensional pattern recognition problems with a limited number of samples and represent nonlinear relationships between the input and the output. Therefore, SVM is widely used in pattern recognition as well as other areas. However, it is difficult to select proper parameters for SVM. The common method for selection SVM parameters is to use heuristic algorithms. Genetic algorithm (GA) and Particles swarm optimization (PSO) algorithm are commonly used to optimize SVM parameters. GA algorithm has a slow convergence problem is easy to lose a local optimal solution. Also, GA does not solve the optimization problem and specific variants of the problem. PSO algorithm to solve the problem when there are easy to describe, easy to implement, fast convergence and other advantages, but there are not effectively prevent premature convergence defects.Inspired by occurring of the chemical reactions, chemical reaction optimization algorithms (CROAs) were developed and applied to solve optimization, classification problems. Therefore, CROAs are used to optimize SVM parameters. The results show that CROAs combined with SVM (CROAs-SVM) are superior to the Genetic algorithm combined with SVM (GA-SVM) and the Particle swarm optimization combined with SVM (PSO-SVM) in solving classification problem. Furthermore, this dissertation will introduce using CROAs-SVM combined with Local Mean Decomposition (LMD) and Local Characteristic-scale Decomposition (LCD) to diagnose roller bearing fault.The main work is as follows:1. Time-frequency methods are presented. The theory of new time frequency analysis methods in signal processing is briefly discussed. Two new methods of time-frequency analysis, LMD and LCD, are introduced. A comparison of LMD and EMD or LCD and EMD are also investigated. The analysis results from simulation signals as well as roller bearing signal show that LMD and LCD are superior to that of EMD.2. CROAs are depicted. The limitations of heuristic algorithms, which are GA and PSO, are targeted. CROAs are proposed to remedy disadvantages of these algorithms. Principles, chemical reactions, and parameters of CROAs are presented. This thesis offers a comparison between CROAs and these methods. The advantages and disadvantages of CROAs are also discussed.3. SVM parameter optimization on CROAs is put forward. In SVM, the generalization performance and the trade-off between minimizing the training error and minimizing model complexity are decided by kernel parameters and regularization constant C. The parameters of the kernel function implicitly define the non-linear mapping from input space to high-dimensional feature space. The performance of SVM will be weakened if these parameters are not properly chosen. The traditional optimization methods to SVM parameter optimization are investigated, such as GA and PSO methods. The limitations of traditional optimization methods in SVM are solved by ACROA and CRO methods. The results show that CROAs have better performance than GA and PSO both in training speed and classification rate. The parameters of CROAs are investigated to give the best performance.4. The application of CROAs-SVM combined with LMD, LCD methods is proposed to diagnose roller bearing fault and included three applications as follows: (1) The first application describes a new method for roller bearing fault diagnosis based on LMD and ACROA-SVM. Firstly, the original modulation roller bearing vibration signals are decomposed into product functions (PFs) by using the local mean decomposition (LMD) method. Secondly, the ratios of amplitudes at the different fault characteristic frequencies in the envelope spectra of some PFs that include dominant fault information are defined as the characteristic amplitude ratios. Finally, the characteristic amplitude ratios are used as input to the ACROA-SVM classifiers, and the fault patterns of the roller bearing are identified. The result shows that the combination of this ACROA-SVM classifiers and LMD method can effectively improve the accurate rate of fault diagnosis and reduce cost time.(2) The next application investigates a novel method for roller bearing fault diagnosis based on LCD energy entropy, together with a SVM designed using an ACROA, referred to as an ACROA-SVM. Firstly, the original acceleration vibration signals are decomposed into intrinsic scale components (ISCs). Secondly, the concept of LCD energy entropy is introduced. Thirdly, the energy features extracted from a number of ISCs that contain the most dominant fault information serve as input vectors for the support vector machine classifier. Finally, the ACROA-SVM classifier is proposed to recognize the faulty roller bearing pattern. The analysis of roller bearing signals with inner-race and outer-race faults shows that the diagnostic approach based on the ACROA-SVM and using LCD to extract the energy levels of the various frequency bands as features can identify roller bearing fault patterns accurately and effectively. The proposed method is superior to approaches based on Empirical Mode Decomposition method and requires less time.(3) Finally, an application presents a novel method for roller bearing fault diagnosis based on LCD and CRO-SVM, referred to as an LCD-CRO-SVM. Firstly, the original roller bearing vibration signals are decomposed into ISCs by using the LCD method. Secondly, the ratios of amplitudes at the different fault characteristic frequencies in the envelope spectra of some ISCs are calculated. Then, these ratios are put into the CRO-SVM classifier. The experimental result shows that the combination of this CRO-SVM classifiers and LCD method obtains the higher classification accuracy and the lower cost time compared to the other methods.