基于BP神经网络和证据理论的超声检测缺陷识别  

Flaw Identification in Ultrasonic Testing Based on BP Neural Network and Evidence Theory

作　　者：王力 周志杰 赵福均 

机构地区：[1]火箭军工程大学,西安710025

出　　处：《电光与控制》2018年第1期65-69,共5页Electronics Optics & Control

基　　金：国家自然科学基金(60736026)

摘　　要：针对超声检测缺陷性质识别中准确率较低的问题,研究了一种基于BP神经网络和证据理论进行超声检测缺陷识别的方法。首先,提出了一种基于BP神经网络和证据理论的融合模型,利用BP神经网络进行特征层融合,将其输出作为证据源的概率分布函数。其次,在决策层融合中针对传统D-S证据理论易出现证据冲突的情况,考虑到不同传感器获取数据的可靠性差异,给出获取证据源可靠性因子的方法。通过引入可靠性因子λ衡量不同证据源的可靠性,使得所有证据源经过可靠性评估后再进行数据融合。最后,通过超声检测手段获取某航空材料的缺陷数据,并对提出的方法进行了验证。研究结果显示,该方法能够更加准确地进行缺陷识别,与传统D-S证据理论相比提高了缺陷识别的准确性。In order to solve the problem of low accuracy in ultrasonic characterization of defect detection, a new method based on BP neural network and D-S evidence theory is proposed to identify defects in ultrasonic inspection. Firstly, a fusion model based on BP neural network and evidence theory is proposed.The BP neural network is used to fuse the feature layer, and its output is taken as the basic probability distribution function of the evidence source. Secondly, in the fusion of decision-making level, the validity of the evidence source is evaluated by introducing an effective factor A to evaluate the reliability of the evidence source. Therefore, all the evidence sources are fused after the reliability evaluation. Finally, the defect data of a certain type of aeronautical material are obtained by means of ultrasonic testing, and the method proposed in this paper is verified. The results show that：compared with the traditional D-S evidence theory, this method can identify the flaws more accurately and improve the accuracy of flaw recognition.

关 键 词：数据融合 证据理论 神经网络 超声检测 证据可靠性 

分 类 号：O213.2[理学—概率论与数理统计]