中文摘要：

工业生产中，多零件装配产品通常具有多系统特性——多几何要素影响的特点。建立能够准确表征装配产品特性与零件几何要素间关系的数学模型，实现装配产品特性的预测，将对工业生产有着重要意义。对于存在多几何要素影响且要素影响程度不同的装配产品系统，直接建立系统的BP神经网络预测模型，由于输入神经元和隐含层神经元较多，将导致神经网络结构十分复杂、学习与训练时间增加、收敛速度慢、预测精度不高等问题。提出利用灰色系统理论的灰关联分析方法，对影响装配产品系统特性的多几何要素进行灰色关联分析，得到多几何要素影响下的系统几何要素与装配产品特性之间的关联程度，确定影响装配产品特性的主要几何要素。在此基础上，利用影响装配产品特性的主要几何要素建立系统的BP神经网络，简化BP神经网络模型，同时保证模型能较真实的反映系统的特性，实现对多几何要素影响下的装配产品特性的准确预测。以典型的多几何要素影响下的装配产品——液压偶件系统为例，通过对液压偶件系统多几何要素关联程度的研究，以影响系统的主要几何要素为输入量，构造简化的BP神经网络的预测模型，试验结果表明，利用主要几何要素建立的预测模型结构相对简单、收敛快、预测精度高。

英文摘要：

In the industrial production, assembled product with many parts usually has the feature of multiple system characteristics and multiple geometric elements effect. To get the exactly math model of the system so as to predict the system characteristics are essentially important for the manufacture process. Because of complexity of the assembled product system with multiple geometric elements and different effect degree, building the BP neural network forecasting model of the system directly along with increment of input neurons and hidden layer neurons leads to very complicated structure of neutral network, increase of study and training time, slow convergence rate, and low precision forecasting. A new method is proposed to build the forecasting model. After analyzing the multiple geometric elements of the system, the grey correlation model is used to obtain the main geometric elements. Then the main geometric elements are used to built the BP neural network and simplify the BP neural network model. The model can truly reflect the feature of the system and can achieve high-precision forecasting for the assembled product system characteristics with multiple geometric elements. In this way, the characteristics predicting of hydraulic valve system is achieved. The hydraulic valve, an assembled product with multiple geometric elements, is taken as example. Through the study on the correlation degree of multiple geometric elements of the hydraulic valve system, the main geometric elements that influence the system are used as input to build a simplified forecasting model of BP neural network. Experimental results indicate that the forecasting model features simple structure, quick convergence and high-precision forecasting.