A Study On Neural Networks Based Process Identification And Control

Neural networks technique is an important part of intelligent control theory. But, there are still many problems in the application of neural networks used for process identification and control. Based on the analysis of the approximating capability and the generalization capability of neural networks, many improved strategies to Back-Propagation learning algorithm is studied. Many subject matters, such as the nonlinear system identification used neural networks, neural network PID control strategy, neural network internal mold control and neural network generalized predictive control, are studied in detail. Some new improved methods and control strategies are brought forward separately. At last the neural network is applied in engineering practice successfully.The approximating capability and the generalization capability of the neural network is studied. On the one hand it points out that the function approximation theory is not sufficiently to form the theoretical basis of the mathematical approximating of the feedforward networks, on the other hand it points out that the generalization problem originated in its open system and creating function. The problems are impossibly to avoid absolutely. And, differences between neural networks and artificial intelligence are compared.Based on the analysis of the questions of Back–Propagation learning algorithm, a comparative study on some typical improved learning methods based on numerical optimization and those based on grads direction is presented. It is indicated by analysis and application that the RPROP method has more quick convergence rate and preferable adaptability. And the influence of the choice of neuron activation function and nonlinear cost function to learning process is also analyzed.The feature and applicability of a few kinds of neural networks model is analyzed. A new type of dynamic recursive neural networks used in process identification is presented and it is obvious improved on the problem of poor generalization capability and slow study speed of Elman network. On the side of the combination with traditional control method, a new simple neural network PID controller is presented, in which the differentiation that the system's output against the manipulated variable is approximated by the difference of it to correct the weight value. And simultaneously a method used in single neuron self-adapting control is put forward.Based on the discussion of contradict model's exist condition and learning method, a Internal Model Control strategy that may used in on-line self-adapting adjust and self-collected is presented. The strategy settles the problem in identification and control of the no self-balance object by using feedback compensatory sangfroid implement. And it also settled the unable on-line problem due to the slow training speed through separating control process and on-line study process. Stability of strategy is analyzed.Some improved measures and new control strategies are presented aiming at the algorithm structure that neural network used for generalization prediction control. Firstly, a simplification algorithm with quicker computational speed is presented. Secondly, a neural network prediction control method using non-derivative optimization is submitted, which meet both the constraint control problem and simplify of the computation. Thirdly, a compound control strategy is also presented in which neural network prediction control combined to traditional PID control. Lastly, based on the research of multivariate algorithm of neural network generalization control, the strategy used in power plant is demonstrated by simulation. The application about neural networks used in process control is studied in light of auctorial engineering practice. Neural networks for process identification and control has been used in CAE2000 (a computer aided engineering software project) and the LN2000 (a Distributed Control System) which are developed by ourselves.