張士華 黃松嶺 孫永泰 史永晉 王宏安

摘 要：漏磁檢測是在管道內(nèi)檢測中應(yīng)用最廣泛的一種無損檢測技術(shù)，檢測數(shù)據(jù)量化與分析是氣難點。在技術(shù)方面針對課題重點研究的關(guān)鍵技術(shù)開展了一系列研究，提出了油氣管道漏磁檢測數(shù)據(jù)的分類和量化方法，并基于此研發(fā)出一套漏磁檢測數(shù)據(jù)分析軟件。漏磁檢測中缺陷量化困難的原因在于缺陷的形態(tài)對漏磁場的形態(tài)有復(fù)雜的非線性的影響，繼而影響對漏磁信號的定量解釋，因此，根據(jù)缺陷的開口形狀將缺陷進行分類，對于實現(xiàn)將其準確量化是十分必要的。再者，由于實際檢測條件的限制，往往只能通過空間離散的漏磁感應(yīng)強度信號的一維分量推算缺陷的三維形態(tài)，這本身不適合使用精確的數(shù)學(xué)或者統(tǒng)計模型加以描述。使用神經(jīng)網(wǎng)絡(luò)對缺陷進行量化，是漏磁檢測缺陷量化領(lǐng)域近20年來的一個研究熱點。根據(jù)課題研究內(nèi)容以及檢測器設(shè)計指標，提出了一種基于改進徑向基函數(shù)網(wǎng)絡(luò)的量化算法，它以缺陷漏磁場信號的特征量為輸入，輸出向量為缺陷的三維外形參數(shù)。徑向基函數(shù)網(wǎng)絡(luò)是一種局部最佳逼近網(wǎng)絡(luò)，但漏磁檢測中漏磁感應(yīng)強度信號與缺陷外形之間強烈的非線性關(guān)系，往往更要求所選用的網(wǎng)絡(luò)能夠識別兩者間的內(nèi)在聯(lián)系，并使得面對新的數(shù)據(jù)時仍有合理的量化結(jié)果。為此，對徑向基函數(shù)網(wǎng)絡(luò)做出基于泛化能力優(yōu)化的改進，提出新的評價函數(shù)，并采用能夠迅速適應(yīng)新樣本的在線學(xué)習(xí)算法，實驗驗證表明，的確能大幅提高網(wǎng)絡(luò)的泛化能力。在實際工程檢測管道中，多缺陷聚集會明顯影響漏磁場的形態(tài)，軸向槽缺陷漏磁場與兩個坑狀缺陷信號波形極為相似，緩變?nèi)毕萋┐艌鲂盘栕兓厔葺^小，這對定量漏磁檢測的實用化是不容忽視的問題。討論了不同類型缺陷漏磁場形態(tài)和強度的影響，并測試了量化神經(jīng)網(wǎng)絡(luò)對缺陷間隔變化的適應(yīng)能力。研究以分類和量化算法為核心，研發(fā)一套漏磁檢測數(shù)據(jù)分析系統(tǒng)。該系統(tǒng)配合內(nèi)檢測器已項目中投入測試，對牽拉實驗數(shù)據(jù)分析的結(jié)果驗證了所提出算法的確具有優(yōu)秀的量化性能。

關(guān)鍵詞：漏磁檢測 缺陷分類 缺陷量化 多缺陷聚集 數(shù)據(jù)分析系統(tǒng)

Abstract：The magnetic flux leakage（MFL） is the most generalized method for in-pipe inspection. A method of classification and quantification of defects in MFL inspection is proposed， and a data analysis system is developed based on this method. The pattern of magnetic flux leakage has a complex non-linear relationship with the shape of defects， which makes it a difficult problem to make quantitative analysis to the magnetic flux leaked.Furthermore， in reality testing conditions， usually only the component in one direction is detected for quantification. Such problems do not adapt to accurate mathematical models. Utilizing neural network as a quantification method has become a focus in MFL inspection during the last 20 years. A method of quantification based on modified radial base function neural network （RBFNN） is proposed. RBFNN promises locally optimal approximation， but the non-linear relationship between magnetic flux pattern and the defect shape requires a strong capability to recognize their inner connection， to better deal with generalized samples.Anon-line trainingalgorithm to determine the number of nodes in hidden layer is proposed， and new merit function based on optimized generalization is employed to train the central vectors and widths. Both of them， verified by experiments， can greatly enhanced the generalized capability of RBFNN. Corrosions usually appear as multi-defect assemblies in pipelines. The relationship between magnetic flux leakage and the pattern of multi-defect assembly is discussed. And different neural network models are employed to solve the inverse problem for multi-defect assembly. Based on the methods stated above， a data analysis expert system is developed. This system works coordinating with in-line inspector and is tested in a submerged pipeline in-service testing project. Results prove that the modified methods gives accurate predicts to a wide range of defects.

Key Words：Magnetic Flux Leakage Inspection；Classification of Defects；Quantification of Defects；Multi-defect Assembly；Data Analysis System

閱讀全文鏈接（需實名注冊）：http：//www.nstrs.cn/xiangxiBG.aspx？id=65390&flag=1