Research On Pattern Recogntion And Fault Diagnosis With Novel Rough Set Neural Network

Pattern recognition and fault diagnosis based on the rough sets theory and neural networks is studied in this dissertation. Rough set theory in the noise environment and in the real region is generalized, and as the sametime, the methods of combine rough set theory with neural networks are proposed. The main contents of the dissertation are organized as follow:At first, a relation of nearness instead of indiscernibility is proposed for increasing the robustness of decision system which consists of noise pollution data. A real nearness class is presented by a given threshold that can decrease the number of the general nearness class. Some definitions and properties of this tolerance rough set are given and a new definition of degree of nearness is formulated. Attributesreduction under the nearness relation is effective and the decision under the noising data is tolerant.Then nearness rough membership function based on nearness relationship rough sets are defined and its properties are investigated. As a rough factor, nearness rough membership function is integrated into a kind of neural networks. And the new neural network can decrease the influence of noise and make it convergent rapidly. Subsequently, the questions on the neural network for fault pattern recognition are studied.In order to overcome the disadvantage of the neural network that it has lower learning precision for small train set, a rough neural network based on support vector machine is presented. In this kind of networks, rough neurons locate in the hidden-layer and they consist of three parts which generated by two hyperplanes partition universe. The hyperplanes are obtained by support vector machines. This neural network is characteristic of fixed configuration, good understandability, simple computation and exact accuracy.An ellipsoidal basic function neural network based on rough K-means is also proposed in this dissertation. This network's structure is similar to the RBF neural network rather than the conventional ellipsoidal unit neural network. In the new network, the ellipsoidal unit functions are used in the hidden layer, and the weights between the hidden nods and the output nods are all connected. A method of rough K-means is used for obtaining the centers of ellipsoidal basic functions and the way of deciding the threshold. The new neural network can not only partition the input space locally, but also make it limitary and bounded. So the network has the better capability of function approximation and pattern recognition. A real rough set space and the concepts of real lower and upper approximation corresponding to real-valued attributes is studied. A rhombus neighborhood for SOM is proposed, and the combination of SOM and rough sets theory is explored in the dissertation. According to the distance between the weight of winner node and the input vector in the real rough sets space, some new weights learning rules are defined. The modified method makes the output classification clearer and the intervals between different classes larger.Furthermore, in order to handle the real attribute values conveniently, a real rough set theory based on the general neighborhood relation is proposed. The theory...