改进的协同遗传BP算法在动态流量软测量技术中的研究与应用

Research and Application in Soft Dynamic Flow Measurement Technology through the Improved Co-evolutionary Genetic and BP Algorithm

传统的流量测量是采用物理流量计进行测量,物理流量计具有价格昂贵,不易维修的缺点。采用神经网络技术对流量进行动态测量不仅可以降低价格,更有易于维修的优点,在液压技术中有十分重要的意义。在动态流量软测量中算法效率低及容易陷入局部极小点是目前动态流量软测量面临的主要难题,针对这个问题提出了协同遗传BP算法。通过遗传算法可以克服容易陷入局部极小点的问题,而BP算法又有在局部迅速查询到极值点的优点,因此采用遗传算法和BP算法混合的协同遗传BP算法。算法对种群进行了合理的结构设置,采用实数编码,以网络训练误差的倒数作为适应度函数,通过采用协同思想,对种群进行选择、变异、交叉、有代沟的替代等操作,加强了两个种群之间的竞争,对于更优种群的产生起到了促进作用。同时,在不影响训练精度的情况下,科学地减少了训练样本数量,以此来提高训练速度减少训练时间。通过理论论证说明了算法的可行性,通过试验验证了算法的性能,试验结果表明该算法较好地克服了易陷入局部极值点的问题,并且比传统的神经网络方法节约时间,效率提高9.03%,能更好地适应动态流量软测量的需要。

Traditionally, the physical flowmeter is used to measure the flow. However, the physical flowmeter is not only expensive but also is difficult to repair. Using neural network technology for dynamic flow measurement has the advantages of lower prices and easy maintenance, thereby having important significance in hydraulic technology. Low efficiency of algorithm and easily failing into local minimum point is the mainly problem in present dynamic soft flow measurement. In view of this problem, an improved co-evolutionary genetic and BP algorithm is proposed. Genetic algorithms can overcome the problem of easily falling into local minimum point, and BP algorithm has the advantage of finding out the local minimum point quickly, so the genetic algorithm mixed with BP algorithm is adopted and named co-evolutionary genetic algorithm （CGA）-BP. The algorithm makes a rational structure setting of the stocks, real number coding is adopted, and the reciprocal of network training error is taken as the fitness function. Through adopting the collaborative thinking, making choice, variations, crossing and substitution with generation gap among stocks strengthens the competition between two stocks, thereby promoting the production of more superior stocks. Without affecting the precision of training, it reduces the number of training samples scientifically, thus improving the training speed and reducing training time. Theoretical demonstration shows the feasibility of the algorithm. Its performance is verified through test. Test results show that the algorithm can overcome the problem of easily falling into local minimum point and save time by 9.03% compared to the traditional genetic algorithm. It can better cater for the need of dynamic soft flow measurement.