改进的模糊Modular神经网络在既有建筑可靠性鉴定中的应用

Application of Improved Fuzzy Modular Neural Network for Reliability Appraisal of Existing Buildings

在Takagi-Sugeno模糊逻辑系统的基础上,提出了改进的模糊Modular神经网络模型(IF-MNN),并将该模型应用于既有建筑的可靠性鉴定。改进的模型是将传统的模糊Modular神经网络模型中的单输出改进为多输出。这种改进的多输入多输出的模糊Modular神经网络模型具有预测性能好、训练学习速度快的优点,它的系统门网络采用模糊C均值聚类算法代替K-means算法,专家网络的训练中引进了先进的Levenberg-Marquardt算法。在应用改进的模糊Modular神经网络模型对既有建筑进行可靠性鉴定的过程中,综合考虑了各主要因素对既有建筑可靠性鉴定等级的影响,并将经量化处理的影响因素作为网络的外部输入,将网络计算得到的4个输出值分别作为样本对应于不同可靠性等级的隶属度,建筑可靠性鉴定的最终评判等级为最大隶属度所对应的等级。训练和预测样本的计算结果证明了改进的模糊Modular神经网络模型在既有建筑可靠性鉴定中应用的可行性和有效性。

Based on the Takagi-Sugeno fuzzy logic system, the theory of improved fuzzy modular neural network （IFMNN） and its application for reliability appraisal of existing buildings were proposed in this paper. The singular output of the conventional modular neural network was changed into multiple outputs in IFMNN. The multiple-input-multiple-output IFMNN has good predicting capability and quick training speed. Fuzzy cmeans algorithm was used to train the system gating network of IFMNN instead of K-means algorithm. Also, Levenberg-Marquardt algorithm was introduced in the training of expert network of IFMNN. Varied factors having the influence on the grade of reliability appraiser were taken into considerations in the process of reliability appraisal of the existing buildings using fuzzy modular neural network, in which the quantified factors were reckoned as outside input. In addition, the four outputs computed by the network were equivalent to the membership degree of reliability grades, and the final grade of the reliability appraiser was correspondent to the grade of the biggest membership degree. The computing results from training and predicting samples indicate that applying the improved modular neural network into the reliability appraisal of existing buildings is practical and effective.