动态流量软测量中BP算法平坦区问题

Flat Area Problem of BP Algorithm on Soft Measurement System of Dynamic Flow

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。目前使用比较广泛的神经网络BP算法存在易陷入局部极小点、收敛速度慢的难题,而收敛速度慢有一部分原因是由于存在平坦区而产生的。针对BP算法收敛速度慢以及陷入平坦区难以逃离的难题,提出了附加奇数的动量BP算法。算法首先用公式求出权值修正的初始值,而后使用增加幂指数的梯度下降函数将权值修正函数值根据误差反馈值作动态改正,使之在遇到平坦区时增大梯度下降步伐,及时跳出平坦区。同时,通过降低计算量来提高计算速度、节省计算时间,加快收敛速度。通过理论证明了算法的正确性。试验结果表明,附加奇数的动量BP算法在保证网络较高的收敛率、达到训练精度的情况下,比传统的动量BP算法节省了9.65%的时间,训练步数也减少了31.13%,更加适合动态流量软测量中网络训练的实时性要求。

Traditionally, the physical flowmeter is used to measure the flow. However, the physical flowmeter is expensive and difficult to repair. Using neural network technology for dynamic flow measurement has the advantage of lower price and easy maintenance, which has important significance in hydraulic technology. The widely used BP algorithm has the problems of easy falling into local minimum point and low convergence speed. In view of the problems, an improved momentum back propagation （BP） algorithm with an odd parameter is proposed. First, a starting weight value is given. Then, the excitation function is improved to change the weight value in order to increase the convergence speed and gradient. At the same time, the number of useless test samples is decreased. The correctness of the proposed algorithm is proved by theory. The test result indicates that the new algorithm can save training time about 9.65% and the training step is also decreased about 31.13% than traditional BP algorithm while satisfying the requirement of convergence and training precision. The proposed algorithm is suitable, for real-time dynamic soft flow measurement.