The Study On Feature Extraction And Feature Selection In Rotor Fault Intelligent Diagnosis

The significance of aeroengine fault diagnosis is: first, it can confirm the fault part and the degree of fault rapidly and accurately, and it is good for insuring the safety of aviation, reducing the manpower and material resources in servicing, shortening the time of aircraft stop, and improving the usage rate of aircraft; second, it is the necessary means and precondition for realizeing on-condition maintenance. In this paper, based on the technique of morden equipment fault diagnosis, we extract the features of 3 classes rotor faults (imbalance, rub-impact and oil whirling) in aeroengine, and select these features by Genetic Algorithm, and use structure self-adaptive ANN to diagnose the 4 classes rotor faults intelligently.First, in this paper, we study the technique of extracting fault features. For rotating machinery's vibration signal, we extract the traditional spectrum features, and obtain the feature vector. Second, we use wavelet analysis technique to study the rotor fault's feature extraction method based on multi-resolution analysis and continuous wavelet analysis theory, and advance an auto-extracting method for rotor fault signal energy features in wavelet multi-scale space, and extract the texture features of the scalogram image based on continous wavelet transform. Then, we use examples to prove that these methods are correct and valid.Second, we study the feature selection technique based on Genetic Algorithm for these extracted features. Because the effect of fitness function on genetic algorithm os very great, in this paper we advance 4 fitness funtion, which is based on advanced distance criterion, mean- variance, Fisher-criterion and most near neighbour class method. Finally, we validate these methods by an example, and prove that the feature selection technique based on genetic algorithm is feasible.Third, we use ZT-3 multi-functions rotor simulation experimental setup to obtain 95 samples including 26 imbalance samples, 29 rub-impact samples, and 40 oil whipping samples. Firstly, we use the traditional spectrum feature extraction method, the auto-extraction method for wavelet energy features and the extracting method for continous wavelet scalogram image texture feature, to extract 17 spectrum features, 10 wavelet energy features and 20 continous wavelet scalogram image texture features. Then, we use the feature selection method based on GA to select these features. At last, we construct the structure self-adaptive compositive NN, and use it to diagnose the selected features intelligently. The results adequately prove that the methods of feature extraction and feature selection advanced in this paper are rational and effective.