摘要
大型回转体超声探伤中由于需要实时处理大量数据,且全面的缺陷特征信息难以获得,导致缺陷类别在线识别困难.对多个超声波探头获取的同一缺陷的互补特征信息,利用BPNN的并行计算能力分别进行缺陷类别的局部决策,再采用D-S理论实现缺陷类型的融合识别.为使神经网络更适合于在线数据处理和缺陷识别,对标准BP算法进行了改进,在不增加计算量和存储量的前提下,避免了网络陷入局部最小,提高了网络的收敛速度.将改进后的BP网络的非线性建模能力与D-S证据理论的不确定性推理能力进行有机结合,使论据理论的支持度的分配避免了主观性,从而提高了决策的确定度和识别的可靠性.采用已知缺陷的回转体工件进行在线融合识别机制进行测试,其结果说明了该方法的可行性及有效性.
During ultrasonic testing on large rotors, it is difficulty to classify the defects online due to the requirement of instant processing capability on huge amount of data and the incomplete information of the defects. Multi-ultrasonic sensors are used for acquisition of the complementary defect information on rotors in this paper. For the parallel computing performance of BPNN （Backward Propagation Neural Network）, it is adopted to realize local classification of the defects with the information form the sensors. The results of local decisions are combined with D-S theory （ Dempster-Shafter Evidence Theory） to classify the defects. To processing data and classify defects online, a BP algorithm is modified to avoid it from trapping into local minimum and to accelerate the convergence rate, on the premise of without increasing computational complexity and memory quantity. Because the nonlinearity modeling ability of BPNN is combined with the indeterminacy reasoning capability of the D-S theory, the distribution of support degree is abstained from subjectivity, thus the validity of decision-making progress and the reliability of defect classification are increased. The online fusion and identification mechanism is tested with known defects, and the results have proved the feasibility and the validity of the proposed method.
出处
《应用基础与工程科学学报》
EI
CSCD
2008年第2期247-254,共8页
Journal of Basic Science and Engineering
基金
四川省国际科技合作与交流研究计划项目资助(2007H12-017)
关键词
缺陷类别
在线
特征提取
改进算法
数据融合
defect classifications
online
feature extraction
modified algorithm
data fusion
