Study On Neural Networks Rule Extraction Method And Its Application In Rotor Fault Diagnosis

Neural networks have been widely applied in fault diagnosis of Aero-engine's performance, wear and vibration, because it has strong nonlinear mapping and error-tolerance ability, the distributed knowledge expression, and implied parallel computation. However, neural networks have an inherently defect: knowledge is implied in lots of connect wights, so that it is difficult to be understood and explained. This has seriously limited the development of neural networks in intelligent diagnosis. Therefore, it is needed to extract rules from neural networks. This is also the instant need to the development of neural networks intelligent diagnosis and neural network expert systems. On the basis of the existing method of neural networks rule extraction, in this paper a new approach based on the functional opinion.is given to extract the rules from neural networks, and it is applied to carry out the intelligent diagnosis of rotor faults.(1) In this paper, an overview of the present studies on neural networks rule extraction methods is summarized, some methods of neural networks rule extraction at home and abroad which are based on the functional opinion or the structural opinion are intrduced, and the meanings and defects of the existing methods are pointed out aiming at the intelligent diagnosis of Rotor faults(2) In this paper, heprinciple and the diagnosis process of neural networks expert system are discussed, neural networks expert system and rule expert system are compared with each other, their advantages and defects are pointed out respectively. It also points out that extracting rules from neural networks is the objective need of the development of neural networks expert system.(3) According to the function opinion, in this paper, a new method of extracting rules from neural networks is put forward. In feature selection, the entropy method is introduced which has been used extensively in data mining; In discreting continuous attributes, a classic method which is put forward by S. H. Nguyen and Skowron is introduced, this method can deal with discreted attributes as continuous attributes, it can reduce the complexity; In structural design of neural networks, genetic algorithm is introduced, and the automatic optimization of neural networks'structure is realized in order to ensure the trained neural networks have the best generalization ability; In rule extracting, a new rule extraction method is advanced for ensuring the integrity and the priorities of etracted rules. The correctness and validity of the new method is verified by Iris Data and Crowd Classification Data.(4) Fault samples are acquired from ZL-3 rotor experimental test-bed and aero-engine rotor experimental rig. These faults include imbalance, rubbing and film whirling. The new method is used to extract fault diagnosis knowledge from lots of fault samples. The result indicates that the extracted rules are correct and easily to be understood.