基于深度学习的天然气钢制管道缺陷检测方法研究

在天然气钢质管道腐蚀缺陷超声检测中,常规模式识别采用人工提取回波信号的方法,存在主观性强、普适性低的问 题。 基于此,本文提出用一维卷积神经网络(one-dimensional convolutional neural network,1D-CNN)提取回波信号的特征结合随 机森林(random forest RF) 分类的方法。 首先根据实验获取的回波信号的噪声情况,用小波包变换(wavelet packet transform WPT)对信号进行去噪;并用变分模态分解( variational model decomposition VMD)对去噪后的信号进行分解和重构以获得平滑 的信号;最后将处理好的回波信号进行 1D-CNN 网络特征提取和随机森林分类。 实验结果表明,基于 VMD-1D-CNN-RF 的天然 气钢质管道缺陷检测方法针对人造缺陷的识别准确率为 85. 71%,针对天然气站场的管道缺陷识别准确率为 71. 05%,表明无需 专家识别也可初步判别管道状况。

Research on defect detection method of natural gas steel pipeline based on deep learning

In the ultrasonic detection of corrosion defects of natural gas steel pipelines, the conventional pattern recognition method adopts the method of manually extracting echo signals, which has the problems of strong subjectivity and low universality. Based on this, this paper proposes a method to extract the features of echo signals by using one-dimensional convolutional neural network and classify the features by combining with random forest. Firstly, according to the noise of the echo signal, the wavelet packet transform is used to denoise the signal. The denoised signal is decomposed and reconstructed by variational modal decomposition to obtain a smooth signal. Finally, the processed echo signals are extracted by 1D-CNN network features and classified by random forest. The experimental results show that the identification accuracy of the method based on VMD-1D-CNN-RF is 85. 71% for artificial defects and 71. 05% for pipeline defects in natural gas stations, indicating that the pipeline condition can be preliminarily identified without expert identification.