An Application Of Immune Feedforward Neural Network In Fault Diagnosis

With the rapid development of modern industry and technology, the security, reliability and validity become more and more important. While traditional fault diagnosis can not meet the requirement, a new method is found in the field of fault diagnosis when Artificial Neural Network (ANN) is prompt progressing. But it is very complex to design neural network for the specifical problem. The network structure, activation function and training method are usually acquired according to the designer's experience and many tries, it makes that the solution is not optimum and reduces the performance.Immune Algorithm (IA) is a kind of optimization algorithms which comes from immune system's characteristic of biologic system. It makes antibody optimize continually by using cross and mutation, adjusting based on antibody concentration, and then gets the optimum antibody. So in this paper, the ANN, which is optimized by IA, is used in the fault diagnosis of water-quality monitor to get better for practical application and raise the level of fault diagnosis.The paper studies the measuring principle of water-quality parameters in the monitor system, and establishes the dynamic model of fault diagnosis about water-quality. Back Propagation (BP) network is used in the fault diagnosis of water-quality monitor. A method is presented to optimize BP network by IA, in which the network structure, activation function and training method are encoded as an individual, in the purpose of optimum solution. The problem has been availably solved that there isn't a guided rules to specify the network structure, activation function and training method. Hardware in the project is based on computer and PCI 9112 DAQ card, and software is based on virtual instrument Labwindows/CVI. The software design is finished for optimal BP via IA in Labwindows/ CVI by calling ANN function of Matlab. The interface of water-quality monitoring system is accomplished with the function of ANN training off line and fault diagnosis on line. Finally, the performance is compared and analyzed between optimized network by IA and experiential network. The results show that the convergence rate and the mean squared-error are both better than the experiential network.