【作者】 陈天璐；

【导师】 阙沛文；

【作者基本信息】 上海交通大学 ， 精密仪器及机械， 2007， 博士

【摘要】 海底管道检测技术及离线信号分析是目前国际无损检测领域的研究热点。开展管道检测相关技术、方法和手段的研究具有重大的理论和现实意义。论文依托国家“十五”863项目“海底管道内爬行器及其检测技术”,一方面以爬行器中超声传感器阵列获取的损伤信号为研究对象,提取了回波到达时刻特征,并通过特征级和决策级的融合,获取了损伤的尺寸、类型和等级等信息;另一方面,尝试将脉冲涡流这种新的检测技术应用于管道无损检测领域,提出了用于管道检测的脉冲涡流探头单元和阵列设计思路,在时域和频域提取了若干脉冲涡流信号的特征,并将其用于管道损伤的分类中,为研制具有我国自主知识产权的管道智能检测器的离线数据分析评估系统提供了关键技术。首先,根据项目需求和海底恶劣复杂的工况,为了最大可能的消除可以消除的不确定性,得出有效合理的分析结果,构建了离线数据处理和多级融合的策略。该策略首先采用不同的方法,对多种传感器阵列的检测信号分别进行独立的分析,提取出信号和(或)损伤的特征,然后进行特征级和决策级的融合,以消除不同类型传感器和不同技术之间的干扰,降低对归一化和配准等融合前提的要求,从不同角度更全面地掌握损伤状况。在基于超声信号的损伤信息提取部分,应用经验模态分解和Hilbert变换等时频分析方法,对非线性非稳态的超声射频信号进行分解、模态选择、重构和变换,可从含强噪声且混叠的信号中精确的提取出各次回波的到达时刻,进而获得准确的剩余壁厚。该方法不仅可明显提高管道剩余壁厚以及内部裂缝尺寸和位置的测量精度,还在很大程度上减小了超声衍射渡越时间法的盲区。在对脉冲涡流技术的研究中,首先,提出了用于管道检测的脉冲涡流探头单元和阵列布置的设计思路。其次,在时域提取出脉冲涡流差分信号的下降点,进而将信号分段,并分别提取出三对形状特征量;在频域提取出两个频谱特征量。此后,对特征量的鲁棒性和普适性以及相互关系进行了分析、比较和总结,得到了一些有价值的结论,丰富了信号解释和损伤分析的可用资源,为快速高效的检测和控制提供了理论依据。在实验中,还发现了信号的双峰现象。此现象未见论文报道。最后,鉴于特征量在数量上和种类上的优势,通过合理集成和组合,实现了损伤的二维和三维快速精确分类。推荐了最佳的二维和三维分类组合,并通过实验,验证了最佳组合的稳定性和普适性。除了基于脉冲涡流信号特征的损伤分类,在对管道损伤的原因和类别进行分析的基础上,基于独立分量分析得到的信号分量,构建了互信息矩阵,提取出了超声射频信号的互信息特征。基于这些特征,利用神经网络对损伤进行了分类,得到了满意的分类结果。在损伤尺寸融合部分,采用神经网络和基于层次分析思想的加权平均两种方法,实现了同类和异类传感器信息的特征级融合,并对两种融合方法的精度、速度和计算负担进行了比较分析。其中,基于层次分析的加权平均法从实际检测信息出发,不需要任何先验信息,计算负担轻,精度较神经网络方法稍差。建议根据实际需要选择合适的方法或将两种方法结合起来以提高融合性能。在损伤等级评估部分,研究了基于不确定性推理的决策级融合方法,分析了证据理论的失效问题,比较了现有的解决方法,从证据源本身出发,提出了基于焦元相似度和证据可信度的两种加权平均方法。通过证据的加权平均预处理,解决了失效问题。此外,基于改进的证据理论,构建了管道损伤等级评价策略,通过综合来自多个评估规则的信息,可以得到更合理可靠的融合结果。为模拟样机编制了管道损伤等级评估和三维显示软件。本文提出的超声信号特征提取和各种融合方法也可用于其它类似情况的信号处理中。更多还原

【Abstract】 Currently, offshore pipeline detection technology and offline signal analysis are research hotspots in the international non-destructive testing (NDT) field. The research on technologies, methods, and approaches for pipeline detection makes great theoretical and realistic sense. Based on“863”of the high technology research and development program“Offshore pipeline detection device and inspection technology”, this dissertation extracted the echo arrival time feature from ultrasonic signal provided by large amounts of ultrasonic sensors built in the detection device, obtained defect dimension, type, and level information by feature level and decision level fusion approaches. In addition, this dissertation attempted to apply a new technology, pulse eddy current (PEC), to the area of pipeline nondestructive testing by designing the structure of a PEC probe and probe array for pipeline detection preliminarily, by extracting a lot of new signal features, and by combining different features for defect classification. All these provide key technologies for the offline data analysis system of the program.Considering the project demand and the harsh and complicated offshore environment, in order to eliminate the uncertainty as much as possible, and to achieve reasonable and effective result, an offline data processing strategy with multiple fusion levels was built. This strategy analyzes data from different types of sensors separately by using different approaches and then combines the extracted signal and/or defect features on feature and decision levels. Thus, the negative interplay between different types of sensors and technologies can be removed, the prerequisites of fusion such as normalization and association can be reduced, and the overall status of a defect can be mastered from multiple points of view.In the section of defect information extraction based on ultrasonic signal, the nonlinear and non-stable ultrasonic signal was decomposed, selected, reconstructed, and transformed by using time-frequency analysis approaches such as Empirical Mode Decomposition (EMD) and Hilbert transform. The arrival time of each echo was extracted from noisy and overlapped ultrasonic signal for getting accurate pipeline wall thickness. This method can not only improve the measurement precision of pipeline wall thickness and the dimension and location of inner crack distinctively, but can also diminish the blind area of ultrasonic time-of flight technology.In the research on PEC technology, the idea of special PEC probe and probe array for pipeline detection was proposed preliminarily. Then, a descending point of PEC differential signal was extracted in time domain and so the signal can be divided into several segments. Three pairs of shape features from each segment respectively and two spectrum features were extracted. Thereafter, the robustness and generalization of new features were analyzed, compared, and summarized and some valuable conclusions were obtained. These conclusions enriched the useable resources for PEC signal explanation and provided theoretic basis for fast and effective detection. Furthermore, by plenty of experiments, a“dual peak”signal phenomenon was discovered which has not been reported at present. Finally, as a result of large numbers of and various types of features, defects can be classified in 2-D and 3-D spaces quickly and accurately by combining different features properly. The best 2-D and 3-D combinations were recommended and the stability and generalization of the two recommended combinations were validated.Besides PEC signal features, ultrasonic signal can be used for defect classification as well. After the introduction of the cause and category of pipeline defects, a mutual information matrix was built by using all the components decomposed from ultrasonic signal (by using the independent component analysis approach). Then, the mutual information features of ultrasonic signal can be extracted. Based on these features, a defect can be classified correctly into certain category by using neural network fusion.In the section of defect dimension fusion, two approaches were proposed for feature level fusion based on the information from same or different types of sensors. Their accuracy, speed, and computation burden were compared and analyzed. The weighted addition approach referring the hierarchy analysis idea is based on real measurement data. It does not need any prior information and so has less computation burden but less accuracy as well. It is wise to select an appropriate method or combine two methods together for the improvement of fusion efficiency.In the section of defect level assessment, some research was conducted on the fusion methods based on uncertainty reasoning. Then, the invalidation problem of evidence theory was analyzed and some typical modification methods were compared. Two weighted addition approaches based on basic probability assignment (BPA) similarity and evidence reliability respectively were proposed and the invalidation problem was solved by weighted averaging all the evidences before combining them. On the basis of revised evidence theory, a strategy for pipeline defect level assessment was built. A more reasonable and more reliable result can be deduced by integrating information from several assessment rules. A 3-D simulation display and level assessment software for pipeline defects was made.All the approaches regarding ultrasonic signal feature extraction and information fusion can be used in other similar signal processing cases.更多还原