| Computational intelligence simulates human's intelligence to solve problems. It is based on computational or mathematical model, and characterized by distributed parallel computation. The application of theories and methods in computational intelligence such as neural networks, fuzzy logic, genetic algorithms, and rough sets theory to fault diagnosis of machinery was studied. According to the requirements in engineering, an Intranet-based on-line condition monitoring and fault diagnosis system for large-scale machinery was developed.The prediction of machinery condition based on neural networks was studied.As far as the nonstationarity during the long period operation of machinery was concerned, the application of Adaptive Linear Element (Adaline) neural network to prediction of nonstationary time series was studied. The relationship between Adaline and Auto Regressive (AR) model was analyzed, and the method to determine the number of input neurons in Adaline prediction model according to BIC criteria was presented. The effect of the adaptive learning rate on prediction was also analyzed. From the simulation, it can be concluded that Adaline is qualified for the on-line prediction of nonstationary time series.The General Regression Neural Network (GRNN) was utilized to predict the vibration of rotating machine. The basics about GRNN were introduced. The BIC method generalized from AR model was adopted to determine the number of input neurons in GRNN prediction model. The GRNN was applied to single-step and multi-step ahead prediction of the vibration time series of a rotating machine, and its performance was compared with that of 3-layers perceptrons network with error back propagation training algorithm (BPNN). It is indicated that the GRNN is more appropriate for prediction of time series than the BPNN, and the performance of GRNN is qualified even with sparse sample data.The reasoning in diagnosis based on computational intelligence is the main topic. Considering the ability of rough sets theory to analysis of incompleteness and uncertainty in data, and that of (fuzzy) neural networks to pattern recognition, a solution of integration was presented for intelligent diagnosis. Diagnosis decision system was reducted based on rough sets theory to find the key conditions for diagnosis, so that the cost can be reduced, and the efficiency can be raised. In order to avoid matching the fault symptoms with the identification conditions artificially, (fuzzy) neural network was designed for diagnosis according to the optimal decision system.For the continuous quantitative diagnosis data such as the measurement, and the result of signal processing, a new hybrid system of Self-Organizing Map (SOM)/Fuzzy c-Means (FCM), rough sets theory, and Adaptive Neuro-Fuzzy Inference System (ANFIS) was presented. Firstly, the continuous attributes in diagnosis decision system were discretized with SOM or FCM. Then, reducts were found based on rough sets theory, and the key conditions for diagnosis were determined. Lastly, according to the chosen reduct, the ANFIS was designed for fault diagnosis. The diagnosis of a 4135 diesel engine verified the feasibility of the presented solution in engineering application. With enough samples, it can be applied to other machinery.For the discrete qualitative diagnosis data such as the diagnosis cases, and the rules summarized by experts, because the obtained decision system was incomplete usually, the integration of generalized rough sets theory and neural network was presented. With fault diagnosis of rotating machinery as example, the similarity-based generalized rough sets theory was applied to reduction of incomplete decision system to find necessary conditions for diagnosis. Based on the optimal decision system obtained from the reducts, BP neural network was designed for fault identification. The application of the reducted decision system to the neural fault classifier indicated that rough sets based reduction reduces the dimension ofinput to neural network, and raise...
|
PDF Full Text Request