Research For Expert Knowledge Applied In Neural Network And Its Application For Soft Sensor Modeling

As an important soft sensor modeling method, Neural network modeling have a wide range of application which doesn’t need precise mathematical formula and has better nonlinear mapping capability. However, the neural network technology still has great promotion space in training algorithm and model training. The traditional BP algorithm used in neural network structure training traps into local optima easily and has a slow convergence speed. A new G-AEA Algorithm based on improved AEA (Alopex-based Evolutionary Algorithm) is presented to replace the traditional BP algorithm in this paper. A certain percentage of good individuals in AEA population are selected as memory population in G-AEA Algorithm. Deep mining on memory population individual information is carried out according to genetic algorithm evolution mechanism. Memory population is used to update the AEA original population, and in turn enhances the global search ability of the algorithm. At the same time, the diversity of evolution population is strengthened through genetic algorithm operation such as crossover and mutation. G-AEA algorithm retains the inspired mode and annealing mechanism of the AEA algorithm and also enhances mining effective information of population. The performance of the proposed algorithm is tested by benchmark functions and compared with two well-known algorithms at the same conditions. The results show that the G-AEA algorithm is outstanding in convergence precision and stability.The performance of the model is largely dependent on complete enough data as the traditional neural network is the data-based modeling. However, it is difficult to get sufficient data in the actual process industry and the data obtained usually contains noise influence. The model cannot get enough information through training when the data is missing. The model lack of physical basis and the extrapolation effect is not ideal. In view of this situation, this paper presents a method combining mechanism or expert experience with the neural network modeling to improve the generalization ability of neural network. In the training process of the model, sensitivity analysis on key variables in every iteration is conducted to verify whether the corresponding output trends are consistent with and the mechanism analysis. Adaptive punishment will be put on the objective function of the model according to the violation degree. The improved method is verified in the mathematics model and industrial field problems. Simulation results of the crystallization kinetics model and other industrial field model show this method can effectively improve the generalization ability of neural network, especially when the number of training samples is small.