Rotor Fault Feature Based On Manifold Learning Research For Dimension Reduction Method

Fault diagnosis technology can avoid rotating machinery major or catastrophic accidents.However, influenced by complex working environment factors, through the right way to establish fault mechanism model to reveal the causal relationship between the state of fault characteristics and categories, there are still a considerable distance. Therefore, based on process historical data driven fault identification research get the attention. Feature extraction from the vibration signal in time domain, frequency domain and time-frequency domain, thus to construct a describe fault state of the original high-dimensional feature space. But data mining research shows that, high dimension data set is not conducive to the realization of intelligent recognition. Therefore, it was caused out of the question about dimension reduction of the high dimensional failure data set.Because of the vibration signals in rotating machinery fault has very strong nonstationarity, the relationship between the fault characteristics are nonlinear. So using traditional linear dimension reduction, such as principal component analysis and Fisher discriminant analysis to solve the problem cannot get very good result on dimension reduction.At present, the dimension reduction method based on manifold learning is widely studied,manifold learning is a method can reveal complex data structure in data driven approach. This article is based on manifold learning dimension reduction of rotor fault feature set work mainly include the following contents:1) Aiming at the questions high dimension and low precision of the recognition for rotating machinery fault diagnosis, a intelligent fault diagnosis methods based on kernel supervised locality preserving projection and K nearest neighbor weighted by feature selection Relief F algorithm(RWKNN) was proposed. KSLPP can effectively extract nonlinear information in original feature data set, at the same time make full use of class information in dimension reduction projection, make the sample minimize the dispersion within class,maximum the separation between classes. Then, the sensitive low dimension feature data set fed into K nearest neighbor weighted by feature selection Relief F algorithm to recognize the fault type. RWKNN can highlight the contribution rate of different features for classification,strengthen the sensitive characteristics, weaken the irrelevant features, improve the classification accuracy and robustness. At last, the validity of the proposed method was verified by the typical fault vibration signal of rotor.2) Aiming at the dimension reduction of the small sample size fault data set, a new method in dimension reduction was proposed based on the combination of principalcomponent analysis and kernel local Fisher discriminant analysis. This method first used PCA to extract key information and dimension reduction of the data set, then the gaussian kernel was used mapped the feature subset to a high-dimensional liner space, in this space local Fisher discriminant analysis was applied to a train most discrimination classification feature set. Finally, a small sample size rotor fault data feature set were employed to verify this method. According to the result of dimension reduction, clear space between various faults categories, small distance in the similar class. This method provide an effective way to solve the problem of small sample size rotor fault data set classification.3)In order to precise and efficient identify different types with different degree of fault rolling bearing. Put forward a intelligent fault diagnosis methods based on localized fisher discriminant score. At first, the feature extractions of time domain, frequency domain and time–frequency from vibration signal, then use localized fisher discriminant score select the most sensitive feature subset from the original feature set, finally, the sensitive low dimension feature data set fed into least squares support vector machine algorithm to recognize the fault type. The proposed method was verified by the typical fault vibration signal of rolling bearing.according to the example result, the feature subset select by localized fisher discriminant score can reveal the differences among the different types with different degree.4)Based on the virtual instrument technology, the C# and.Net platform were used to exploit a rotor test rig test system. The system can automatic control the motor speed, acquire vibration signal, vibration domain signal transform to frequency domain singal, real-time display of the time domain signal and frequency domain, and the third chapter of the rotating machinery fault diagnosis model was embedded into this system, make the system realize intelligent fault diagnosis function.