【作者】 胡文刚；

【导师】 刚铁；

【作者基本信息】 哈尔滨工业大学 ， 材料加工工程， 2012， 博士

【摘要】 随着现代工业的不断发展，焊接构件在锅炉、压力容器、铁路、海洋造船、能源、航空航天等工业领域都得到了广泛的应用。目前对于工业领域中的在役焊接结构，常规的手动超声检测仍然被广泛采用。但是，常规超声检测方法无法实现检测数据的自动存储，检测效率低且容易发生漏检。尤其对于缺陷性质的识别一直是常规超声检测方法中的难题。常规超声检测无法实现缺陷三视投影图像的自动直观显示，对缺陷的识别诊断完全依靠检测者的经验，使得检测结果因人而异，这些都给实际的现场在役检测带来诸多不便。针对常规手动超声检测中存在的问题，本文首先开发了一套基于视频图像定位的超声检测成像系统，该系统不但成本低廉、结构简单、便于携带，而且能够实现焊缝缺陷连续扫查，检测数据的自动存储，以及缺陷的三视投影成像，为后续的缺陷智能识别提供了可靠的数据。利用开发的基于视频图像定位的超声检测成像系统对含有气孔、夹渣、裂纹、未焊透和未熔合五类实际焊接缺陷的对接焊缝试件进行了超声检测，实时存储了每个缺陷体的超声回波信号，并获得了缺陷的三视投影成像图，直观的给出了缺陷的位置、大小、分布和取向等形态分布特征，为缺陷特征值的提取提供了可靠的数据来源。采用人工智能方法进行缺陷的识别与判断，最关键的是获取能够反映不同缺陷类别的特征值。通过分析不同缺陷反射体的特性，从多值域的角度出发，分别在时域、频域、时频域、几何统计学和缺陷的位置形态上提取了用于缺陷智能识别的缺陷特征值，实现了实际焊接缺陷的多值域特征提取。缺陷特征值是缺陷智能识别的关键，正确选择有效的特征量是保证识别系统具有良好的分类性能的前提。为此，本文构建了基于欧氏距离的特征值评价准则，对提取的所有缺陷特征值进行了评价与优化，获取了缺陷的最佳特征值子集，实现了特征空间的降维，为缺陷识别提供了有效的特征向量。构建基于BP神经网络的智能化模式识别分类器，研究了基于缺陷超声信号特征与形态特征的多值域特征分类识别新方法，对五类实际焊接缺陷进行了智能识别。与常规的基于超声信号特征的缺陷识别结果相比，该方法充分利用了缺陷的位置形态信息，有效地提高了缺陷的识别率。以Dempster-Shafer(D-S)证据理论为基础，研究了超声检测中缺陷识别的双探头源数据融合新方法。该方法通过对探头信息的融合，可以更准确、更全面地获得被检缺陷的信息。进一步构建了基于BP神经网络与D-S证据理论联合的双探头源缺陷智能识别分类器，实现了五类实际焊接缺陷的智能识别。与单一探头数据下的识别结果相比，该方法融合了双探头源的互补信息，提高了缺陷信息的利用率，降低了单一探头识别系统的不确定性，从而进一步提高了缺陷识别的可靠性和准确率。更多还原

【Abstract】 With the development of modern industry, welded structures have been widelyused in boiler, pressure vessel, rail, marine, shipbuilding, energy, aerospace and otherindustrial fields. At present, conventional manual ultrasonic testing was widely usedto detect the welded structure in service. However, conventional ultrasonic testingmethod cannot acquire the continuous echoes of flaw, and the data cannot be storedautomatically. Thus this method is inefficient, and the flaw is easy to be missedduring the detection process. Furthermore, flaw recognition is still a difficult problemto be resolved in conventional ultrasonic testing method. The three-view projectionimage of flaw cannot be illustrated automatically and intuitively. So the recognitionand diagnosis of flaw completely relies on the experience of the operator, whichmakes the detection result vary from person to person. Thus, there are many issuesfor detecting welded structures in service by conventional ultrasonic testing.To solve the problems which exist in conventional ultrasonic testing, a manualultrasonic testing system based on video image positioning was developed in thispaper. This system not only has many characters of cheap cost, simple structure andconvenient to carry, but also can achieve a continuous scanning of weld flaw, theautomatic storage of data and the three-view projection imaging of weld flaw, whichprovides the reliable data for the intelligent recognition of weld flaw.There are many butt weld specimens, which contain five types of typical weldflaws of porosity, slag, crack, lack of penetration and lack of fusion. They wereinspected by this developed system, and ultrasonic echoes reflected from each defectwere stored in real time. Then the three-view projection image of weld flaw wasshowed, and the location, size and distribution of weld flaw were characterizedconveniently and intuitively, which provides a reliable data source for featureextraction of weld flaw. For intelligent recognition of defect by using artificialintelligence method, the most critical thing is to obtain the feature which can reflectthe characteristic of different defects. By analyzing the characteristics of differentflaw reflectors, flaw features were extracted in time domain, frequency domain andtime-frequency domain. Furthermore, geometric features, statistical features andmorphological features were also extracted. Thus, the flaw features were extracted inmulti-domain in this paper.Flaw feature plays an important role in intelligent recognition of weld flaw.Correct selection of efficient feature vector is necessary to ensure a goodperformance of classification for the recognition system. Thus, evaluation criterion based on Euclidean distance was constructed to evaluate and optimize the feature.Then the optimum feature subset was obtained, and the dimensionality reduction offeature space was brought out. The results provided efficient feature vectors for theflaw recognition.Intelligent recognition classifier was constructed base on back propagation (BP)neural network. Then a new method of flaw recognition based on multi-feature ofultrasonic signal feature and morphological feature was studied. And it was applied toidentify the five types of weld flaws. This method made full use of the morphologicalinformation of weld flaw. Thus, compared to the conventional recognition resultbased on ultrasonic signal feature, the recognition rate of weld flaws was improvedeffectively.Based on Dempster-Shafer (D-S) evidence theory, a new method of flawrecognition based on data fusion of dual-probe sensor was studied. More accurate andcomprehensive flaw information was obtained by the data fusion of dual-probeinformation. Then an intelligent pattern classifier based on BP neural network andD-S evidence theory was developed to carry out the flaw recognition of five types ofweld flaws. This method combined the complementary information of dual-probesensor, improved the utilization of defect information, and reduced the systematicuncertainty of single probe. Thus, compared to the recognition result based on singleprobe, the reliability and accuracy of flaw recognition was improved effectively.更多还原