Fault Feature Extraction And Diagnosis Of Rolling Element Bearing Under Non-steady-state Conditions

Rolling bearing is one of the most important components of rotating machinery. Its main function is to provide a low friction operating condition for the rotating shaft, and the damage of these components may cause the equipment serious damage. Therefore, it is particularly important to study the feature extraction and the on-line fault diagnosis of rolling bearing. In the process of starting, stopping and running of rotating machinery, there will be varying degrees of speed fluctuation and load fluctuation, which makes the mechanical equipment operating in non-steady-state conditions. Traditional time-domain feature extraction method and the frequency domain feature extraction method can't be applied to the non-steady-state conditions. So, we need to find ways to solve the non-steady-state conditions of rolling bearing fault feature extraction and diagnosis.In this thesis, according to the fault diagnosis of rolling bearing under the non-steady-state conditions, three methods are proposed.1. This thesis proposed a dynamic FDA based fault feature extraction method. This method by means of introducing delay window way to capture the rolling bearing vibration signal cyclostationarity, and a rolling bearing fault feature discriminant space are established by the FDA, and then the k nearest neighbor(KNN), SVM and decision tree method are used to classify the fault. The CWRU and QPZZ-II fault experimental platform are used to simulate the variable load and variable speed conditions. The experimental results show that the method can effectively realize the online fault diagnosis of bearings for each fault condition when the delay window is choose properly.2. The dynamic FDA method is effective but need to select a large delay to complete fault diagnosis, a dynamic FDA feature extraction method based on EMD is proposed to reduce the amount of information. One-dimensional vibration signal was decomposed into multilayer IMF component by the method of EMD. Select the time delay, and a multidimensional dynamic matrix is constructed by the IMF component, then a feature discriminant space is constructed by the FDA analysis on the multidimensional dynamic matrix. The experimental results show that the method can use historical fault to complete the feature extraction and fault diagnosis of variable load and variable speed data, diagnosis with high precision.3. Due to the IMF components in the same kind of fault have multi cluster problem in the second method, a local dynamic FDA feature extraction algorithm is proposed based on LPP. The local computation between class and within class scatter by using this algorithm to complete FDA analysis, the feature discriminant space is constructed by using multidimensional dynamic matrix which is composed by IMF component. This method is applied in variable load and variable speed experiment, and can complete fault classification by k nearest neighbor, SVM and decision tree classification methods. Compared with the previous two methods, this method has stronger robustness, stability and higher diagnostic accuracy.