A Research On A Fault Diagnosis Method With The Combination Of The Rough Sets Theory And ANN

Nowadays, fault diagnosis has developed to the intelligent phase on which the emphasis of research has turned to the rising Computational Intelligence from the traditional Artificial Intelligence (AI). Some theories of Computational Intelligence, such as Artificial Neural Network (ANN) and Rough Set (RS), have been widely applied in fault diagnosis. ANN is a self-adapt non-linear dynamics system. It has abilities of learning and parallel computing, and it can be used to classify, self-organize and association memory. RS theory is a mathematical tool of depictingincomplete and uncertain knowledge. It can analysis and process efficiently all kinds of imprecise, inconsistent and incomplete information, discover connotative knowledge, and reveal potential rules. RS theory can make up the limitations of ANN, which is the poor ability of determining the redundancy or usefulness of the knowledge and the long training time. So the combination of ANN and RS theory is more significant.This thesis mainly studies the method of discretization of continuous attributes and reduction of attributes based on the analysis of existing fault diagnosis method with the combination of ANN and RS theory. The main achievements are as follows:1. We put forward a method that combines Competitive Neural Network with support degree of condition attributes to discretize the continuous attributes. The correct result of reduction shows that this method can objectively describe the distribution of data.2. When there are many condition attributes in the fault diagnosis system, a strategy of establishing the partitioning decision tables then reduction is adopted. In this thesis, the feasibility andthe time complexity are analyzed. The result shows that the method is feasible and it can make reduction more simplify and shorten the working time.3. Combining an example of axletree fault diagnosis, we give the processing course of the fault data with the above method. Then, a BP Neural Network is designed with the GUI of MATLAB. It is continually trained with the learning sample sets. When the training is finished, we simulate the testing sample sets and get the correct result. The example shows that the fault diagnosis method we put forward is feasible.