Industrial Process Data Mining Based On Fuzzy Inference System

The rapid progress in digital data collection and storage technology of industrial process has led to fast growing tremendous and amount of data stored in databases, data warehouses, or other kinds of data repositories. The useful information and knowledge, extracted from large amounts of data of industrial process, can provide powerful decision support for online monitoring, fault diagnosis, modeling, control strategy designing, and predication and so on.The first important task of industrial process data mining is to select and build effective and suitable data mining methods to process industrial data. The fuzzy inference system based data mining method can use the same structure (fuzzy IF-THEN rules) for both description and predication tasks. The models extracted are easy to understand for operators and make decision for managers. For a complex nonlinear industrial process, this method can be used in a natural way to rank the importance of the input variables such that the most relevant variables are selected to describe the dynamic behavior of system. Furthermore, the flexibility in defining the membership functions can help to develop process operating models in different granularity space and mine the relationships and rules among process variables to solve the actual problems of industrial process, effectively. The main contributions are described as follows:(1) In order to resolve the conflicts of convergence speed and oscillation existed in the traditional gradient-based fuzzy inference system learning method, and the selecting issue of coefficients in the momentum learning method, an improved gradient-based real time learning algorithm (G-RTL) for parameter optimization of fuzzy inference system is developed. Comparing with these two kinds of methods mentioned above, the proposed method has higher accuracy and convergence speed in the same condition of learning rate coefficients, by introducing dynamic error transfer factor associated with the mean squared error. Moreover, the learning process is steady. Therefore, it is more suitable for online learning. The method is applied to the classical truck backer-upper control problem and compared with approximation informance of BP network. The simulation results show the effectiveness of the proposed method.(2) Aiming at the features of high dimensions and nonlinear of industrial process data, an adaptive method for mining fuzzy rules is developed based upon the normalized variance information (NV-AMFR). Starting from a very simple initial structure, the parameters defining fuzzy inference system are adaptively updated by means of the G-RTL learning algorithm, based upon data mining and Mamdani model. Then the information of normalized variance and confidence measurement obtained from this phase can be utilized to determine in which region of the input space the density of fuzzy rules should be enhanced and for which variable the number of fuzzy sets used to partition the corresponding domain should be increased. Hence, this procedure allows mining the relationships and rules among process variables in different granularity space. Furthermore, the method can also rank the importance of the input variables in a natural way such that the most relevant variables are selected to describe the dynamic behavior of system. Consequently, this produces a new and more appropriate model structure. Some simulations demonstrate the efficiency of the proposed method. The method is applied to the problem of nonlinear function approximation.(3) According to inconsistent, incomplete, and historical data properties existed in the industrial process databases, a most neighboring diffusion method for extrapolating the missing fuzzy rules is developed (ND-EMR). Based upon fuzzy inference system and priori knowledge of sampling data distribution, the centers of optimal output fuzzy sets and the confidence measurements of fuzzy rules are determined by means of the G-RTL learning algorithm, simultaneously. Then the information obtained can be utilized to extrapolate these missing fuzzy rules over the regions not covered by sampling data, and build a complete fuzzy rule base. Therefore, non-predictable problem of system could be resolved in these regions effectively. The proposed method is applied to the predictive problem of chaotic timeeries and compared with WM method. Simulation results show that the method is effective and suitable for the non-predictive case.(4) A novel data mining based method for complex industrial process intelligent control is developed, based upon fuzzy T-S prediction model and several data mining techniques. Fuzzy T-S prediction model can be identified quickly and accurately by means of the G-RTL algorithm. For constrained nonlinear optimal control of batch process, the control problem of a complex nonlinear system can be transformed into that of several local linear sub-systems based on the identified model, by using the parallel distributed compensation algorithm (PDC) and the Pontryagin's minimum principle (PMP). Thus, a simple and effective optimal fuzzy control strategy is offered. The proposed method has been illustrated on the modeling and optimal control of a fed-batch reactor. Simulation results show the method not only has higher modeling accuracy but also can obtain more desired product quantity.