鞠萍华　柯磊　冉琰　朱晓　李松涛

摘   要：為了提高对机械零件失效概率的预测精度，提出一种基于GRA和AHP的广义回归神经网络零件失效概率预测方法.在分析机械零件失效概率影响因素的基础上，首先利用灰色关联分析法（Grey Relational Analysis，GRA）分析影响机械零件失效概率的主要因素，通过层次分析法（Analytic Hierarchy Process，AHP）构建机械零件失效概率的评价指标层次体系，评估各个指标对于零件失效概率的权重;结合各个指标权重与初始值，以获取各指标的加权评价值;最后通过广义回归神经网络（Generalized Regression Neural Network，GRNN）建立以各指标加权评价值来预测机械零件失效概率的预测模型.利用本文方法所建立的预测模型对某企业数控转台的上齿盘失效概率进行预测，并与传统的GRNN神经网络预测模型、BP神经网络预测模型和回归预测模型进行对比，结果显示本文所建立的模型预测误差小于0.8%、残差在-0.2%～0.2%范围内，均优于对比模型的预测结果，表明所建立的预测模型具有更高的精度和更强的稳健性，适合于零件失效概率的预测.

关键词：广义回归神经网络;灰色关联分析;层次分析法;加权评价值;预测

中图分类号：TH165+.4                                 文献标志码：A

Failure Probability Prediction Method on Parts of Generalized Regression

Neural Network Based on GRA and AHP

JU Pinghua，KE Lei，RAN Yan，ZHU Xiao，LI Songtao

（State Key Laboratory of Mechanical Transmissions，Chongqing University，Chongqing 400044，China）

Abstract：To improve the prediction precision of failure probability of machine parts，failure probability prediction method of generalized regression neural network based on GRA and AHP was proposed. The main influence factors on failure probability of mechanical parts were analyzed by grey relational analysis method based on the analysis of influence factors on failure probability of mechanical parts. The hierarchy model of evaluation index for failure probability of each mechanical part was constructed and the weight of each index was evaluated by analytic hierarchy process. Then，the weight and initial value of each index were combined to obtain the weighted evaluation value of each index. Finally，the generalized regression neural network was used to establish a predictive model by using weighted evaluation value of each index to predict the failure probability of mechanical parts. This optimization method was applied to predict the failure probability of upper gear disk in numerical control rotary table. The prediction results of traditional generalized regression neural network ，BP neural network and regression analysis method were compared. The result shows that the prediction error of the proposed model is less than 0.8%，and the residual error is in the range of -0.2% and 0.2%，which is better than the comparison models. Meanwhile，the model established by using the proposed method in this paper has higher accuracy and stronger stability，which is suitable for the prediction of failure probability of parts.