Research On Adaptive Characteristic-Scale Decomposition And Its Application To Rotating Machinery Fault Diagnosis

Researches on the fault diagnosis for rotating machinery have great significance, whose key step is the extraction of fault information. However, vibration signals of the rotating machinery are mostly non-stationary. Consequently, extracting state characteristics from complicated vibration signals by an appropriate signal analysis method is always a key and hot research. As the recognized analysis processing method for non-stationary signals, time-frequency analysis method has been applied in many fields.In recent years, as complicated vibration signals can be properly processed by the Wavelet Analysis(WA), Empirical Mode Decomposition(EMD), Local Mean Decomposition(LMD), etc, these methods have been used for fault diagnosis of rotating machinery and all achieve great results. However, these methods have limitations. Recently, an adaptive time-frequency analysis method, Local Characteristic-scale Decomposition(LCD), is proposed. Based on defining a mono-component signal with physically meaningful instantaneous frequencies, Intrinsic Scale Component(ISC), and extracting the characteristic scale parameters of the data, this method can decompose signal into a sum of the several ISC components. Relative to the EMD and LMD, LCD takes more advantages in decomposition speed, the restraint of end effects, mode mixing and so on. However, LCD has shortcomings in defining the mean curve as well as mode mixing, whose theory needs to be developed and improved further. Funded by the Natural Science Foundation of China(No. 51075131), this paper further studies and improves LCD. And a new Adaptive Characteristic-scale Decomposition(ACD) is presented and applied to fault diagnosis ofrotating machinery. The main researches and innovations of ths paper are shown below.(1) Improvement on the mean curve of LCD.(1) The interpolation of the mean curve of LCD is calculated by the ligature of the two adjacent extreme points. The attribute of the interpolation mainly depends on the two adjacent extreme points, which may decrease the decomposition accuracy. Concerning this issue, the Lagrange Interpolation based Local Characteristic-scale Decomposition(LILCD) is introduced.(2) For the sifting based adaptive signal decomposition method, its key point is to define an appropriate mean curve. The appropriateness of the mean curve directly affects the accuracy and efficiency. Consequently, Generalized Local Characteristic-scale Decomposition(GLCD) is put forward.(2) Study on the methods of restraining mode mixing.(1) Differential Operator based Local Characteristic-scale Decomposition(DOLCD) is proposed. By changing the amplitude ratio of each components in the original signal, this method promotes the ability of restraining the mode mixing problem of LCD.(2) Compact Local Characteristic-scale Decomposition(CLCD) is presented. A least extreme scale is defined to measure other signal scales for restraining the mode mixing problem of LCD, and pseudo-extremes are added to homogenize the signal scales.(3) Based on the research of LCD, this paper combines the advantages of GLCD and CLCD, borrows ideas from the sifting based signal decomposition methods such as EMD and LCD, and then defines a new mono-component signal with physically meaningful instantaneous frequencies, i.e., Intrinsic Compact-scale Component(ICC). Finally, a new signal decomposition method, Adaptive Characteristic-scale Decomposition, is proposed. Meanwhile, the evaluation criterion of ICC is also given. In the sifting procedure for separating certain order component, a set of ICCs is obtained generated by using different mean curves and compact coefficients The optimal ICC for this order sifting will be selected from the candidate ICCs using the evaluation criterion of ICC, which guarantees ACD outperforms LCD.(4) Because the moving average adopted by General Local Frequency Demodulation(GLFD) cannot reflect the volatility and change rule of signals. Especially, when the values of General Local Frequency(GLF) and General Local Amplitude(GLA) are abrupt, the moving average can result in an obvious deviation. To solve this problem, Improved General Local Frequency Demodulation(IGLFD) is proposed. This method adopts Robust Locally Weighted Regression(RLWR) to smooth GLF and GLA curves, and inserts Pseudo-endpoints(PEs) into GLF and GLA to optimize the choice of the smooth span value.(5) The methods mentioned above are validated by simulated signals and experimental signals. The results show their validity.