基于改进CS-BPNN的松茸发酵过程软测量建模

Soft sensor modeling of matsutake fermentation based on improved CS-BPNN

摘要:

针对松茸发酵过程中关键参量难以实时在线检测的难题，提出了一种基于改进布谷鸟算法（CS）与改进BP神经网络（BPNN）相结合的松茸菌丝生物量软测量建模方法。首先采用两阶段动态发现概率法对传统CS进行改进，平衡CS的全局搜索与局部搜索能力；然后引入附加动量和动态调整学习率对BPNN进行改进，提高BPNN参量的修正精度；最后，通过CS算法获取BPNN的初始权值和阈值，并由权值修正公式（附加动量与动态学习率相结合）对权值进行动态修正。仿真结果表明，改进的CS-BPNN软测量模型在预测精度提高了6%以上，能够实现松茸发酵过程实时在线测量的需求。

Abstract:

In view of the fact that the key parameter in the fermentation process of matsutake is difficult to be detected online，a new method of soft sensor modeling of mycelium biomass based on improved cuckoo algorithm (CS) and improved BP neural network (BPNN) is proposed. Firstly, the traditional two-phase dynamic discovery probability method is used to improve the global CS search and local search capabilities; Then the BPNN is improved by introducing additional momentum and dynamic adjustment learning rate to improve the correction accuracy of BPNN parameters; Finally, the initial weights and thresholds of BPNN are obtained by the CS algorithm, and the weights are dynamically modified by the weight correction formula (a combination of additional momentum and dynamic learning rate) to overcome the traditional BPNN soft-sensing model easy to fall into the local minimum, slow convergence and other issues. The simulation results show that the improved CS-BPNN soft-sensing model can improve the prediction accuracy of the model by more than 6%, which can meet the demand of real-time on-line measurement during the fermentation of matsutake.