摘要
为实现磁记忆检测中对缺陷承受的拉伸载荷、宽度及深度进行识别,提取出缺陷处磁记忆信号切向分量的最大差值X1、峰值X2、曲线下围面积X3、梯度的峰谷值X4;法向分量的峰谷值X5、曲线下围面积X6、缺陷梯度X7以及梯度的峰峰宽X8等8个特征参数,采用主成分分析法（PCA）与遗传算法优化神经网络（GA-BP）法有机结合对数据处理,该算法可以有效地逼近力磁耦合作用下的复杂非线性关系,实现了对磁记忆信号的定量识别。该算法不仅降低了数据冗余度、提高了计算效率,还使得神经元的权值和阈值更加稳定,避免陷入局部最优解,解决了输出结果不稳定的缺点。其中法向分量所提取指标（X5,X6,X7）和切向分量提取的指标（X1,X4）对第1主成分影响较大,而指标（X2,X3,X8）对第2主成分影响较大,这些指标都可以反映缺陷的信息。检测实例表明,该算法对宽度的预测精度最低,而对深度的预测精度最高。
In order to identify the tensile load, width and depth of the defect in the magnetic memory test, 8 characteristic parameters of magnetic memory signal at defect were extracted. It contains the maximum difference of the tangential component X1, the peak X2, the area under the curve X3, the peak value of the gradient X4.The peak-to-valley value X5 of the normal component, the area under the curve X6, the defect gradient X7, and the peak-to-peak width X8. The use of principal component analysis（PCA） and genetic algorithm-optimized neural network（GA-BP） method combined with the data processing, the algorithm can effectively approximate the complex nonlinear relationship under the coupling of magnetic and magnetic forces, and realize the quantitative identification of magnetic memory signals. The algorithm not only reduces the data redundancy, improves the computational efficiency, but also makes the weight and threshold of the neuron more stable, avoids falling into the local optimal solution, and solves the shortcomings of unstable output. The index extracted by the normal component（X5, X6, X7） and the tangential component（X1, X4） has a greater influence on the first principal component, while the index（X2, X3, X8） influences the second principal component. These indicators can reflect the defect information. The detection example shows that the algorithm has the lowest prediction accuracy for width and the highest accuracy for depth prediction.
出处
《电子测量与仪器学报》
CSCD
北大核心
2018年第10期190-196,共7页
Journal of Electronic Measurement and Instrumentation
基金
国家自然科学基金(51675155,51635010,51722502)资助项目
关键词
磁记忆
特征参数
主成分分析
GA-BP神经网络
定量识别
magnetic memory
characteristic parameters
principal component analysis
GA-BP neural network
quantitative identification
