Page 31 - 无损检测2025年第一期
P. 31

试验研究
              试验研究

              DOI:10.11973/wsjc240215


                   基于贝叶斯优化的 BiLSTM 小管径弯管漏磁


                                                     缺陷识别




                                                齐卜弘,张 颖,赵鹏程,王加旭
                                         (常州大学 安全科学与工程学院,常州 213164)

                       摘  要:为提升对小管径弯管的缺陷识别效果,提出了一种基于贝叶斯优化的BiLSTM缺陷识
                   别方法。创建了φ114 mm×8 mm(壁厚)弯管三维漏磁仿真有限元模型,获取了不同类型缺陷的数
                   据信号,提取了信号波形的多模态特征作为特征样本库,利用贝叶斯优化方法调整了BiLSTM关键
                   超参数,并通过试验验证了该方法的识别效果。试验结果表明,贝叶斯优化的BiLSTM可以较为准
                   确地识别小管径弯管的缺陷种类,改进后的模型准确率达到了96. 07%。该方法对小管径弯管的漏
                   磁缺陷识别具有较高可行性。
                       关键词:小管径弯管;漏磁检测;贝叶斯优化;BiLSTM;缺陷识别
                       中图分类号:TG115.28      文献标志码:A    文章编号:1000-6656(2025)01-0001-08

                       Magnetic leakage defect identification of small-diameter pipe elbow based on

                                                Bayesian optimized BiLSTM


                                       QI Buhong, ZHANG Ying, ZHAO Pengcheng, WANG Jiaxu
                          (School of Safety Science and Engineering, Changzhou University, Changzhou 213164, China)
                      Abstract: In order to improve the defect recognition effect of small-diameter pipe elbows, a Bayesian optimized
                   BiLSTM defect recognition method was proposed. A three-dimensional magnetic leakage simulation finite element model of
                   φ114×8 mm (wall thickness) pipe elbow was established, data signals of different types of defects were obtained, multimodal
                   features of signal waveforms were extracted as a library of feature samples, and the key hyperparameters of BiLSTM were
                   adjusted by using Bayesian optimization, and the recognition effect of the method was verified by experiments. The results
                   showed that the Bayesian optimized BiLSTM can identify the defect types of small-diameter pipe elbows more accurately,
                   and the accuracy of the improved model reached 96.07%. The method was highly feasible for the identification of magnetic
                   leakage defects in small-diameter pipe elbows.
                      Key words: small-diameter pipe elbow; magnetic flux leakage; Bayesian optimization; BiLSTM; defect identification

                  目前,小管径管道应用广泛,管道长时间服役后                         用漏磁检测技术,但小管径弯管处的磁导率分布不
              会出现腐蚀、老化等问题,进而引发管道运输事故,                           均匀,磁场强度相对于直管而言更难达到饱和,其产
                                      [1]
              带来经济损失和环境污染 。小管径管道检测常采                            生的漏磁场分布复杂且易发生信号畸变。为了提高
                                                                数据分析的准确率,减少不必要的工作量,有必要对
                                                                                 [2]
                 收稿日期:2024-05-15                                数据进行优化处理 。
                 基金项目:中国石油天然气股份有限公司—常州大学创新联合体                        长短期记忆网络(Long short-term memory,
              科技合作项目(KC20210301)
                                                                LSTM)能弥补一般循环神经网络(Recurrent neural
                 作者简介:齐卜弘(2000—),女,硕士研究生,研究方向为特种设备
                                                                network,RNN)模型中的爆炸及消失等不足 。杜
                                                                                                         [3]
              健康监测及智能诊断
                                                                        [4]
                 通信作者:张 颖(1972—),教授,博士,研究方向为特种设备                小磊等 将深层小波卷积自编码分析与LSTM
                                                                                            [5]
              健康监测及智能诊断,aezy163@163.com                         相结合进行运用;赵志宏等 提出了一种基于
                                                                                                           1
                                                                                         2025 年 第 47 卷 第 1 期
                                                                                                  无损检测
   26   27   28   29   30   31   32   33   34   35   36