BP神经网络在红外热波无损检测定量识别中的应用

Application of BP neural network to quantitative identification in thermal wave NDT

为实现红外热波检测对缺陷的定量识别,应用BP神经网络,拟合函数关系来实现定量识别。在脉冲热像中,以最佳检测时间和表面最大温差为输入量,以缺陷的深度和直径为输出量,利用BP神经网络拟合输入量与输出量之间的关系。借助数值计算的方法,提供样本训练神经网络,并进行了30次随机计算。通过结果分析,发现使用BP神经网络进行计算具备以下特点:网络收敛速度并不决定计算的精度;网络训练过程中,是否达到计算目标误差不会对计算精度带来较大影响;该方法具有较好的计算稳定性。针对计算结果分布特点,设计计算方法,对数据中的较大误差点进行剔除,最后使用取均值的方法减小获得较大误差的风险,提高计算精度。计算结果表明,在4个参数的计算中,最大误差为3.32%,最小误差为0.1%,证明BP神经网络方法可以用于实现缺陷的定量识别计算。

In order to solve the problem of quantitative identification,BP neural network was implied to fit function to achieve quantitative identification.In pulsed thermography,the highest temperature difference and the best testing time was taken as input,and defect depth and diameter was taken as output,and then BP neural networks was used to fit the function relationship of them.Training samples of the neural network was provided by numerical method,and 30 times randomized computations were carried out.As the result analysis showed,the method had following characteristics： the precision of the calculation did not depend on network convergence speed;in network training process,the calculation precision error was not relative with whether calculation target was achieved;the calculation method had good stability.And then,according to distribution characteristics of calculation results,calculation method was designed to eliminate data with big calculating error.Finally,method of taking the average was taken to reduce the risk of a greater error,and to improve accuracy of calculation.As results show,in four parameters calculation,the biggest error is 3.32%,and the minimum error is 0.1%.And this proves that the method could be used to achieve defect quantitative identification calculation.