基于主成分分析和BP神经网络的综合安全评价模型

A renovated safety evaluation model based on the principal component analysis and BP neural network

提出了基于主成分分析法和BP神经网络的综合安全评价模型。利用主成分分析法的数据降维功能对评价指标进行特征提取,选用合适的主成分作为神经网络的输入,构建网络拓扑结构。在建立多层次道路危险货物运输企业安全评价指标体系的基础上,采用MAT-LAB对企业安全现状进行实证仿真分析。结果表明,基于主成分分析法和BP神经网络的综合安全评价模型的评价结果优于其他评价模型。研究表明,本文得到的综合模型能够更加客观、准确地反映评价对象的实际情况。

The present paper intends to present its authors＇ newly renovated safety evaluation model based on the analysis of the principal components and BP neural network. As is seen in traditional evaluation practice, it is often the case that quite a few problems are to be dealt with, such as large quantity of nodes in the input level, the information overlap among the nodes as well as the lengthy computational routine of low precision. To solve all the above mentioned problems, we have brought about a renovated safety evaluation model of comprehensive nature, which intends to integrate the Principal Component Analysis （PCA） and BP neural network into a unified one. In the proposed model, it becomes a new evaluation routine to apply all the functions of feature extraction and dimension reduction of data to pretreating the evaluation indices of input data. And, then, in accordance with the assessing criteria of the principal component, some other chief second-level influential components can be extracted as new indices of input data for network topology building. And, finally, based on the establishment of the safety evaluation index penalty, we have tried to apply such influential factors to the safety evaluation practice of dangerous goods transportation. As we know, such a system usually consists Of 5 levels and 27 evaluation indices with calculation devices including SPSS 13.0 and MATLAB. Besides, the whole model has to use simulation analysis of the current conditions of dangerous goods transportation. Moreover, it is also necessary to collect and evaluate some 40 sampling data for PCA from the actual evaluation values on such dangerous goods transportation. However, with the newly proposed evaluation model, we have only to use 7 principal components. And, consequently, such 7 principal components and 20 standardized sample data can also be put into BP network for training and learning. Thus, repeated experimental tests show that this model is expected to achieve better evaluation results than the other conventional ones with its error limit just from 0.5 to 1.2 percent. Therefore, it can be seen that the proposed model can produce more objective and accurate evaluation results by overcoming shortcomings of the traditional models, which is worthy to be put forward as a reference model for safety evaluation.