王斯昊,何蘭芳,李亮
摘 要:隨著地球物理數(shù)據(jù)和地質(zhì)信息的不斷積累,神經(jīng)網(wǎng)絡(luò)作為一種數(shù)據(jù)驅(qū)動(dòng)的建模工具在地球物理中的應(yīng)用越來(lái)越廣泛,在數(shù)據(jù)預(yù)處理和圖像識(shí)別等方面表現(xiàn)出色。為了梳理神經(jīng)網(wǎng)絡(luò)在勘探地球物理的應(yīng)用現(xiàn)狀與發(fā)展趨勢(shì),本文回顧了神經(jīng)網(wǎng)絡(luò)的發(fā)展歷程、基本原理和主要技術(shù)框架,總結(jié)了神經(jīng)網(wǎng)絡(luò)在勘探地震、電磁法、重磁勘探中的應(yīng)用,同時(shí)分析了神經(jīng)網(wǎng)絡(luò)的優(yōu)勢(shì)與不足。神經(jīng)網(wǎng)絡(luò)在地球物理中的可行性已被學(xué)者們廣泛認(rèn)可,但在實(shí)際應(yīng)用中仍然存在一系列難題有待深入探索。
關(guān)鍵詞:神經(jīng)網(wǎng)絡(luò);地球物理勘探;應(yīng)用進(jìn)展; 發(fā)展趨勢(shì)
地球物理是通過(guò)觀測(cè)數(shù)據(jù)去探測(cè)不同尺度的未知地質(zhì)目標(biāo)。對(duì)于絕大多數(shù)地球物理方法而言,探測(cè)目標(biāo)的特征屬性和觀測(cè)數(shù)據(jù)之間并不存在線性對(duì)應(yīng)關(guān)系,而是一種非線性映射,地球物理反演就是用正演模型去模擬這一映射[1],但需要對(duì)觀測(cè)數(shù)據(jù)進(jìn)行甄選。隨著地球物理數(shù)據(jù)的海量增長(zhǎng),人工甄選和提取數(shù)據(jù)已變得越來(lái)越不現(xiàn)實(shí),這為深度學(xué)習(xí)在地球物理中推廣應(yīng)用帶來(lái)了全新的機(jī)遇[2-3]。如何利用神經(jīng)網(wǎng)絡(luò)提高地球物理數(shù)據(jù)處理效率和準(zhǔn)確度是地球物理普遍關(guān)心的問(wèn)題。此外,與傳統(tǒng)地球物理數(shù)據(jù)處理方法相比,神經(jīng)網(wǎng)絡(luò)的優(yōu)勢(shì)在于不需要對(duì)映射進(jìn)行先驗(yàn)限制,只要網(wǎng)絡(luò)經(jīng)過(guò)訓(xùn)練就可直接建立目標(biāo)映射[4-6]。但是,神經(jīng)網(wǎng)絡(luò)主要有“黑盒子”性質(zhì),且需大量數(shù)據(jù)驅(qū)動(dòng)。由于地球物理問(wèn)題的多解性及神經(jīng)網(wǎng)絡(luò)的不確定性,神經(jīng)網(wǎng)絡(luò)應(yīng)用需要有足夠多合適的訓(xùn)練數(shù)據(jù),同時(shí)需要據(jù)網(wǎng)絡(luò)在實(shí)際數(shù)據(jù)的表現(xiàn)進(jìn)行評(píng)估[6]。
1? 神經(jīng)網(wǎng)絡(luò)的發(fā)展歷程
20世紀(jì)40年代,伊利諾伊大學(xué)McClloch和芝加哥大學(xué)Pitts基于數(shù)學(xué)和一種稱為閾值邏輯的算法創(chuàng)造了一種神經(jīng)網(wǎng)絡(luò)的計(jì)算模型[7],并由此開創(chuàng)了人工神經(jīng)網(wǎng)絡(luò)(簡(jiǎn)稱為神經(jīng)網(wǎng)絡(luò))學(xué)科(圖1)。20世紀(jì)60年代,由于多層神經(jīng)網(wǎng)絡(luò)訓(xùn)練的可能性被否定,70年代,神經(jīng)網(wǎng)絡(luò)領(lǐng)域的研究基本處于停滯狀態(tài)。80年代,反向傳播算法的提取開啟了神經(jīng)網(wǎng)絡(luò)的第二次研究高潮[8]。隨著計(jì)算機(jī)和網(wǎng)絡(luò)技術(shù)的飛速發(fā)展,地球科學(xué)已進(jìn)入大數(shù)據(jù)時(shí)代,神經(jīng)網(wǎng)絡(luò)迎來(lái)新一輪研究和應(yīng)用高峰。卷積神經(jīng)網(wǎng)絡(luò)(CNN)是目前最廣泛應(yīng)用的深度神經(jīng)網(wǎng)絡(luò),主要特點(diǎn)為通過(guò)卷積運(yùn)算操作來(lái)有效地提取數(shù)據(jù)特征,在學(xué)習(xí)到數(shù)據(jù)特征的同時(shí)能極大地減少網(wǎng)絡(luò)參數(shù)數(shù)量。由于卷積神經(jīng)網(wǎng)絡(luò)的飛速發(fā)展,已在各個(gè)行業(yè)廣泛應(yīng)用,在地球物理中的應(yīng)用已從地震勘探拓展到地球電磁學(xué)和位場(chǎng)方法等多個(gè)領(lǐng)域,并有綜述文獻(xiàn)發(fā)表[2-3]。
2? 卷積神經(jīng)網(wǎng)絡(luò)基本原理與技術(shù)框架
神經(jīng)網(wǎng)絡(luò)的核心可以表示為:y =g(wxT +b),其中y為輸出,x為輸入,w為加權(quán)系統(tǒng),b為偏置參數(shù),g為激活函數(shù)。即神經(jīng)網(wǎng)絡(luò)包含3個(gè)重要步驟:加權(quán)、偏置和非線性激活。神經(jīng)網(wǎng)絡(luò)系統(tǒng)通常包含輸入層、輸出層和隱藏層。復(fù)雜神經(jīng)網(wǎng)絡(luò)系統(tǒng)可理解為多個(gè)單一系統(tǒng)的組合,包含輸入層、輸出層和多個(gè)隱藏層。數(shù)據(jù)流從輸入層至輸出層的過(guò)程稱為正向傳播,反過(guò)來(lái)稱為反向傳播。
為克服傳統(tǒng)神經(jīng)網(wǎng)絡(luò)運(yùn)算量大、過(guò)擬合、參數(shù)過(guò)多等問(wèn)題,學(xué)者們?cè)趥鹘y(tǒng)神經(jīng)網(wǎng)絡(luò)的基礎(chǔ)上發(fā)展了卷積神經(jīng)網(wǎng)絡(luò)。CNN的基本原理是讓神經(jīng)網(wǎng)絡(luò)模仿人類在識(shí)別圖像時(shí)的局部感受機(jī)制,神經(jīng)元(卷積核)檢測(cè)圖像是否滿足卷積核的特征并輸出相應(yīng)的特征圖像,最后通過(guò)全連接層對(duì)該圖像進(jìn)行分類(圖2)。通過(guò)反向傳播不斷地調(diào)整卷積核的權(quán)值,得到所需要的圖像識(shí)別器。CNN通過(guò)局部感知、共享權(quán)值、空間或時(shí)間上的池采樣來(lái)充分利用數(shù)據(jù)本身包含的局部特性,以優(yōu)化網(wǎng)絡(luò)結(jié)構(gòu),保證一定程度上的位移和變形的不變性[9]。
3? 卷積神經(jīng)網(wǎng)絡(luò)在地震勘探中的應(yīng)用
20世紀(jì)40年代以來(lái),地震勘探一直是勘探地球物理的主戰(zhàn)場(chǎng),也是現(xiàn)代科技試驗(yàn)田和前沿陣地。在1991年SEG年會(huì)上,神經(jīng)網(wǎng)絡(luò)被列為“90年代技術(shù)突破”四類新興技術(shù)之一。經(jīng)過(guò)30年的發(fā)展,以卷積神經(jīng)網(wǎng)絡(luò)為代表的神經(jīng)網(wǎng)絡(luò)已在地震勘探廣泛應(yīng)用,在初至拾取與結(jié)構(gòu)識(shí)別2個(gè)方面發(fā)展相對(duì)完善。
3.1? 初至拾取
初至拾取是估計(jì)地震事件位置及層析成像或矩張量反演等其他工作流程所必需的先期工作,本質(zhì)上是一個(gè)模式識(shí)別過(guò)程,人工識(shí)別初至需耗費(fèi)大量人力和時(shí)間,隨著地震數(shù)據(jù)的大量生成和積累,人工拾取已無(wú)法完成,因此許多學(xué)者不斷地探索自動(dòng)化拾取初至的方法。Murat等 首次將人工神經(jīng)網(wǎng)絡(luò)成功地應(yīng)用于噪聲背景下的初至波提取,此后人工神經(jīng)網(wǎng)絡(luò)甚至卷積網(wǎng)絡(luò)在初至拾取的有效性被多個(gè)團(tuán)隊(duì)所驗(yàn)證[11-19]。神經(jīng)網(wǎng)絡(luò)的優(yōu)勢(shì)在于一旦訓(xùn)練完成,處理效率相較于人工處理有極大提高。McCormack等發(fā)現(xiàn)訓(xùn)練好的網(wǎng)絡(luò)三維數(shù)據(jù)集拾取效率相較人工拾取提高了8倍。在這兩位學(xué)者之后,相關(guān)研究主要基于2個(gè)方面:提高神經(jīng)網(wǎng)絡(luò)的訓(xùn)練和處理速度與拾取準(zhǔn)確度[11,14,15,19]。訓(xùn)練速度方面,利用邏輯模糊網(wǎng)絡(luò)能保證不損失準(zhǔn)確度的同時(shí)提升訓(xùn)練速度[11]。Kahrizi等發(fā)現(xiàn)在準(zhǔn)確度上MLP(多層感知器)表現(xiàn)出色,但在訓(xùn)練速度上RBF(徑向基)神經(jīng)網(wǎng)絡(luò)更有優(yōu)勢(shì)[15]。準(zhǔn)確度方面,Yuan等與Duan等指出前人對(duì)于地震相拾取的研究都是基于單道的,利用卷積神經(jīng)網(wǎng)絡(luò)我們可進(jìn)行多道數(shù)據(jù)分析;Zhe等利用級(jí)聯(lián)算法拾取地震初至,級(jí)聯(lián)算法的優(yōu)勢(shì)在于其收斂速度快且不需要調(diào)整網(wǎng)絡(luò)結(jié)構(gòu)[16]。Maity等人設(shè)計(jì)了一種新的混合自動(dòng)拾取網(wǎng)絡(luò)結(jié)構(gòu),發(fā)現(xiàn)神經(jīng)網(wǎng)絡(luò)在低信噪比情況下,表現(xiàn)比其他方法更穩(wěn)定[14]。Yuan等指出人工神經(jīng)網(wǎng)絡(luò)很少使用波形的空間相干特征,為此提出利用CNN進(jìn)行初至拾取,但并未對(duì)比前人人工神經(jīng)網(wǎng)絡(luò)的結(jié)果[17]。提取哪些地震屬性對(duì)初至拾取更有幫助也是未來(lái)值得研究的問(wèn)題之一。
3.2? 基于圖像處理的地下結(jié)構(gòu)識(shí)別
從地震圖像中圈定斷層是地震構(gòu)造解釋、儲(chǔ)層描述和布井的關(guān)鍵步驟。地球科學(xué)家通常用一個(gè)或多個(gè)地震屬性來(lái)判斷三維地震圖像中的斷層,若所用屬性不能很好地將斷層與周圍非斷層區(qū)域區(qū)分開來(lái),有可能導(dǎo)致錯(cuò)誤解釋結(jié)果。卷積神經(jīng)網(wǎng)絡(luò)的優(yōu)勢(shì)在于無(wú)需選擇地震屬性,直接建立原始地震信號(hào)與斷層分布之間的映射,并且隨著數(shù)據(jù)增多,CNN對(duì)噪聲和不同模式的斷層識(shí)別能力會(huì)越來(lái)越強(qiáng)。Wu等提出一種基于卷積神經(jīng)網(wǎng)絡(luò)二維地震圖像的斷層自動(dòng)解釋方法。對(duì)于同一地震圖像相比于常規(guī)方法的結(jié)果,CNN能夠給出更清晰、更連續(xù)斷層的特征,甚至在更復(fù)雜的情況下,CNN仍然能夠很好地預(yù)測(cè)斷層分布[20]。Di等利用CNN原始疊后地震振幅建立地震信號(hào)與斷層分布之間的映射關(guān)系,并將CNN與其他方法進(jìn)行對(duì)比。結(jié)果表明,CNN的分類結(jié)果最接近手動(dòng)解釋結(jié)果[21]。Guitton等設(shè)計(jì)了一個(gè)3D卷積神經(jīng)網(wǎng)絡(luò)(3DCNN),它在自動(dòng)估計(jì)地震體特征方面非常有效,可幫助完成分類任務(wù)[22]。前人已證明CNN在斷層特征識(shí)別的有效性。為提升網(wǎng)絡(luò)性能,Shi等在每次卷積后都對(duì)結(jié)果進(jìn)行正則化處理[23];Wu等采用U型網(wǎng)絡(luò)實(shí)現(xiàn)三維地震圖像的斷層識(shí)別(圖3),并使用了二元交叉熵?fù)p失函數(shù)平衡實(shí)際地震數(shù)據(jù)中0(非斷層)和1(斷層)之間的高度不平衡,這樣能使網(wǎng)絡(luò)對(duì)有斷層的地震數(shù)據(jù)更為敏感[24]。隨著地震數(shù)據(jù)的增長(zhǎng),需要新的平臺(tái)對(duì)數(shù)據(jù)進(jìn)行管理。Huang等建立了一個(gè)基于云的地震數(shù)據(jù)分析平臺(tái),該平臺(tái)可管理地震數(shù)據(jù)、計(jì)算地震屬性、進(jìn)行特征提取和選擇,并應(yīng)用深度學(xué)習(xí)軟件包(TensorFlow和Caffe)來(lái)推斷地下結(jié)構(gòu)特征,為進(jìn)一步工作打下堅(jiān)實(shí)基礎(chǔ)[25]。CNN能夠有效地從原始地震數(shù)據(jù)中識(shí)別斷層,但離實(shí)際應(yīng)用還有一定距離。今后需要評(píng)估已訓(xùn)練模型的有效性。Shi等指出,今后可利用貝葉斯分割網(wǎng)絡(luò)(Bayesian Segnet)對(duì)模型的不確定性進(jìn)行評(píng)估以提高推斷地下結(jié)構(gòu)的可信度[23]。
4? 神經(jīng)網(wǎng)絡(luò)在重磁勘探中的應(yīng)用
相對(duì)于地震勘探而言,重力、磁法和電法面對(duì)的數(shù)學(xué)物理問(wèn)題與地質(zhì)模型更為復(fù)雜,雖然神經(jīng)網(wǎng)絡(luò)在非地震勘探中發(fā)展歷程和地震相似,但進(jìn)展更為緩慢,仍有很大的拓展空間。對(duì)于重磁位場(chǎng)數(shù)據(jù),主要是利用位場(chǎng)數(shù)據(jù)反演地下重磁異常結(jié)構(gòu)。
4.1? 神經(jīng)網(wǎng)絡(luò)在重磁反演的應(yīng)用
利用位場(chǎng)數(shù)據(jù)反演地下重磁異常結(jié)構(gòu)也是學(xué)者們感興趣的問(wèn)題。Poulton等通過(guò)電磁橢圓率數(shù)據(jù)反演地下導(dǎo)體的位置、深度和面積,并對(duì)比了不同結(jié)構(gòu)神經(jīng)網(wǎng)絡(luò)的計(jì)算速度與分類準(zhǔn)確率??傮w看來(lái),對(duì)于維度超過(guò)100的輸入向量,自組織映射到反向傳播的混合網(wǎng)絡(luò)雖然速度較慢,但測(cè)試效果最好[26]。Elawadi 等利用反向傳播神經(jīng)網(wǎng)絡(luò)估計(jì)微重力數(shù)據(jù),探測(cè)美國(guó)Medford洞穴在地下的深度,估算參數(shù)與鉆探資料吻合良好。Elawadi 等指出神經(jīng)網(wǎng)絡(luò)優(yōu)勢(shì)在于反演效率更高,對(duì)大型復(fù)雜數(shù)據(jù)集提供靈活解決方案,同時(shí)對(duì)噪聲容忍性更高[27]。Bescoby等與Al-Garni等利用人工神經(jīng)網(wǎng)絡(luò)(ANN)對(duì)地下磁性體參數(shù)進(jìn)行反演,結(jié)果表明ANN對(duì)于實(shí)測(cè)數(shù)據(jù)的表現(xiàn)令人滿意[28-29]。Al-Garni利用模塊化神經(jīng)網(wǎng)絡(luò)反演傾斜堤壩模型參數(shù),對(duì)于印度安得拉邦卡里姆納加爾地區(qū)露頭石英脈狀巖體磁異常和秘魯馬可納地區(qū)的馬可納磁異常。結(jié)果表明,與大多數(shù)常規(guī)方法相比,該技術(shù)與實(shí)測(cè)資料吻合較好[30]。由于實(shí)際地質(zhì)模型的復(fù)雜性,常需建立不同網(wǎng)絡(luò)來(lái)反演模型,這是神經(jīng)網(wǎng)絡(luò)在重磁勘探中的發(fā)展較為緩慢的主要原因。
4.2? 神經(jīng)網(wǎng)絡(luò)在地電學(xué)中的應(yīng)用
神經(jīng)網(wǎng)絡(luò)在地電學(xué)的應(yīng)用主要分為3方面:大地電磁測(cè)深時(shí)間序列處理、直流電測(cè)深模型反演和大地電磁反演。神經(jīng)網(wǎng)絡(luò)在一維直流電測(cè)深模型的反演已相當(dāng)成熟[31-32]。El-Qady等指出,對(duì)于一維反演問(wèn)題神經(jīng)網(wǎng)絡(luò)的結(jié)果已足夠讓人滿意。Srinivas等在前人的基礎(chǔ)上選擇比較了不同網(wǎng)絡(luò)在一維直流電測(cè)深視電阻率反演的表現(xiàn),如徑向基網(wǎng)絡(luò)、廣義回歸網(wǎng)絡(luò)和前饋反向傳播網(wǎng)絡(luò)。選取10組合成數(shù)據(jù)對(duì)網(wǎng)絡(luò)進(jìn)行測(cè)試,發(fā)現(xiàn)前饋反向傳播網(wǎng)絡(luò)表現(xiàn)最佳[38]。二維模型的反演問(wèn)題仍有待研究[14,28,36]。Neyamadpour等對(duì)反向傳播、彈性傳播等不同算法進(jìn)行比較,發(fā)現(xiàn)在訓(xùn)練過(guò)程中,彈性傳播算法收斂更快、誤差更小。對(duì)于二維直流電測(cè)深模型反演,彈性反向傳播算法可能是一個(gè)較有效的選擇。
神經(jīng)網(wǎng)絡(luò)在大地電磁反演也有應(yīng)用。普遍認(rèn)為一維反演需要不同神經(jīng)網(wǎng)絡(luò)學(xué)習(xí)不同的一維模型[4,36-37]。實(shí)現(xiàn)自動(dòng)化反演需要多個(gè)網(wǎng)絡(luò)組成的模塊化神經(jīng)網(wǎng)絡(luò),部分網(wǎng)絡(luò)負(fù)責(zé)對(duì)曲線分類,部分負(fù)責(zé)對(duì)曲線回歸[37-38]。對(duì)于二維問(wèn)題,有學(xué)者傾向于先對(duì)電磁場(chǎng)數(shù)據(jù)分類再回歸,但二維模型如何分類仍有待研究[39-41]??偟膩?lái)說(shuō),一維大地電磁反演較為成熟,二維反演的模型分類及如何設(shè)計(jì)網(wǎng)絡(luò)仍有待研究,三維問(wèn)題的反演結(jié)果不盡人意(圖4)[40]。
神經(jīng)網(wǎng)絡(luò)還能用于大地電磁時(shí)間序列噪聲識(shí)別與去噪分析。電導(dǎo)率結(jié)果的可靠性主要取決于電磁場(chǎng)數(shù)據(jù)的信噪比,采用人工處理的方法需要觀察時(shí)間序列后手動(dòng)刪除異常數(shù)據(jù),耗時(shí)費(fèi)力。學(xué)者們?yōu)榻鉀Q該問(wèn)題提出許多人為消除電磁干擾的方法,相較于神經(jīng)網(wǎng)絡(luò),其他方法無(wú)法同時(shí)在時(shí)域和頻域壓制噪聲[42]。神經(jīng)網(wǎng)絡(luò)去噪的可行性已得到認(rèn)可[43,44],它的優(yōu)勢(shì)在于能夠快速、穩(wěn)健地剔除時(shí)間序列中的噪聲,在低信噪比情況下,神經(jīng)網(wǎng)絡(luò)的表現(xiàn)與人為剔除結(jié)果具有一致性[44]。未來(lái)應(yīng)結(jié)合無(wú)監(jiān)督式分類器與監(jiān)督式循環(huán)網(wǎng)絡(luò)進(jìn)行去噪,無(wú)監(jiān)督式學(xué)習(xí)可幫助數(shù)據(jù)分類并節(jié)省時(shí)間。
5? 主要認(rèn)識(shí)
神經(jīng)網(wǎng)絡(luò)在地球物理圖像識(shí)別和數(shù)據(jù)處理中已取得不俗成果,但在反演中的應(yīng)用還有待探索,而神經(jīng)網(wǎng)絡(luò)在分類中表現(xiàn)更為出色。神經(jīng)網(wǎng)絡(luò)本身更適合分類問(wèn)題,回歸問(wèn)題會(huì)弱化神經(jīng)網(wǎng)絡(luò)非線性模擬的優(yōu)勢(shì),且對(duì)標(biāo)簽數(shù)據(jù)逆歸一化,也需假定預(yù)測(cè)數(shù)據(jù)和訓(xùn)練數(shù)據(jù)為同一分布,這對(duì)于地球物理的實(shí)測(cè)數(shù)據(jù)要求過(guò)于苛刻。神經(jīng)網(wǎng)絡(luò)的優(yōu)勢(shì)主要有2點(diǎn):①多個(gè)線性分類器組合給予其強(qiáng)大的非線性模擬能力,面對(duì)大量數(shù)據(jù)神經(jīng)網(wǎng)絡(luò)的表現(xiàn)顯著優(yōu)于傳統(tǒng)算法;②神經(jīng)網(wǎng)絡(luò)的訓(xùn)練和預(yù)測(cè)是分開的,使得神經(jīng)網(wǎng)絡(luò)在處理大量數(shù)據(jù)時(shí)速度和準(zhǔn)確度具有壓倒性優(yōu)勢(shì)。神經(jīng)網(wǎng)絡(luò)的缺陷也很明顯,當(dāng)出現(xiàn)不同分布的數(shù)據(jù),網(wǎng)絡(luò)需要不斷訓(xùn)練來(lái)建立新的映射,對(duì)神經(jīng)網(wǎng)絡(luò)的應(yīng)用造成一定阻礙。面對(duì)多種網(wǎng)絡(luò)結(jié)構(gòu)和算法,必須理解其原理,針對(duì)問(wèn)題滿足不同需求。誤差分析對(duì)于神經(jīng)網(wǎng)絡(luò)的實(shí)際應(yīng)用尤為重要,可幫助人們理解在哪些數(shù)據(jù)上“犯錯(cuò)”,以便按照生產(chǎn)實(shí)踐要求進(jìn)行優(yōu)化??傊?,神經(jīng)網(wǎng)絡(luò)在地球物理中的可行性已被學(xué)者們廣泛認(rèn)可,但在很多實(shí)際應(yīng)用領(lǐng)域有待探索。
參考文獻(xiàn)
[1]? ? Zhdanov M S. Geophysical inverse theory and regularization problems[M].Annals of Glaciology,2002,32(1):321-327.
[2]? ? 王昊,嚴(yán)加永,付光明,等.深度學(xué)習(xí)在地球物理中的應(yīng)用現(xiàn)狀與前景[J].地球物理學(xué)進(jìn)展,2020,35(2):0642-0655
[3]? ? 王家映.人工神經(jīng)網(wǎng)絡(luò)反演法[J].工程地球物理學(xué)報(bào),2008,5(3):? 256-265
[4]? ? Poulton M M. Neural networks as an intelligence amplification tool: A review of applications[J].Geophysics,2002, 67(3):979-993.
[5]? ? Sandham? W A. Leggett, M. Geophysical applications of artificial? ?neural networks and fuzzy logic[M]. Kybernetes,2003,21
[6]? ? Baan M ,Jutten C. Neural networks in geophysical applications[J].? Geophysics,2000, 65(4):1032-1047.
[7]? ? McCulloch W S, Pitts W. A logical calculus of the ideas immanent in nervous activity[J]. The bulletin of mathematical biophysics,1943, 5(4):115-133.
[8]? ? Hinton G E .Osindero, S. A fast learning algorithm for deep belief nets[J]. Neural Computation,2006, 18(7):1527-1554.
[9]? ? Lecun Y, Bottou L, Bengio Y,et al. Gradient-based learning applied to document recognition[J]. Intelligent Signal Processing,2001,86(11):306-351.
[10]? 付文博,孫 濤,梁 藉,等.深度學(xué)習(xí)原理及應(yīng)用綜述[J].計(jì)算機(jī)科學(xué),2018,45(6):11-15.
[11]? Chu P, Mendel J M. First break refraction event picking using fuzzy logic systems[J].IEEE International Conference on Fuzzy Systems,1993, 2(4):255-266
[12]? Murat M E, Rudman A J. AUTOMATED FIRST ARRIVAL PICKING: A NEURAL NETWORK APPROACH1[J]. Geophysical Prospecting,1992, 40? ? ? (6): 587-604.
[13]? McCormack, M D. FIRST-BREAK REFRACTION EVENT PICKING AND SEISMIC DATA TRACE EDITING USING NEURAL NETWORKS[J].Geo? ? physics,1993, 58(1):67-78.
[14]? Maity D, Aminzadeh, F, Karrenbach M. Novel hybrid artificial neural network based autopicking workflow for passive seismic data[J]. Geophysical Prospecting,2014, 62(4):834-847.
[15]? Kahrizi A, Hashemi H. Neuron curve as a tool for performance evaluation of MLP and RBF architecture in first break picking of seismic data[J]. Journal of Applied Geophysics,2014, 108:159-166.
[16]? ?Zhe L, Rong S, Weixiu X, et al. The Improvement of Neural Network Cascade-Correlation Algorithm and Its Application in Picking Seismic First Break[J]. Advances in Petroleum Exploration and Development,2013, 5(2):41-51.
[17]? ?Yuan S, Liu J, Wang S, et al. Seismic Waveform Classification and First-Break Picking Using Convolution Neural Networks[J]. IEEE Geoscience and Remote Sensing Letters,2018, 15(2):272-276.
[18] Zhu W, Beroza G C. PhaseNet: a deep-neural-network-based seismic arrival-time picking method[J]. Geophysical Journal International,2018, 216(1):261-273.
[19] Duan X, Zhang J. Multitrace first-break picking using an integrated seismic and machine learning methodPicking based on machine learning[J]. Geophysics,2020, 85(4):269-277.
[20] Wu X, Shi, Y. Fomel, S.,Liang, L. Convolutional neural networks for fault interpretation in seismic images[J]. In 2018 SEG International Exposition and Annual Meeting,2018:1946-1950
[21] Di H, Wang Z, AlRegib G. Seismic Fault Detection from Post-Stack Amplitude by Convolutional Neural Networks[J]. In 80th EAGE Conference and Exhibition 2018,2018:1-5.
[22] A Guitton, 3D Convolutional Neural Networks for Fault Interpretation[J].In 80th EAGE Conference & Exhibition 2018,2018:1-5
[23] Shi Y, Wu X, Fomel S. Automatic salt-body classification using a deep convolutional neural network[J]. In 2018 SEG International Exposition and Annual Meeting,2018:1971-1975
[24] Wu X, Liang L, Shi Y, Fomel, S. (2019). FaultSeg3D: Using synthetic data sets to train an end-to-end convolutional neural network for 3D seismic fault segmentation CNN for 3D fault segmentation[J]. Geophysics, 84(3):35-45.
[25] Huang L, Dong X, Clee T E. A scalable deep learning platform for identifying geologic features from seismic attributes[J]. Geophysics,2017,36(3):249-256.
[26] Poulton M M.“Location of Subsurface Targets in Geophysical? ? ? ? Data Using Neural Networks.” Geophysics,57(12):1534-1544.
[27] Elawadi E, Salem A, Ushijima K. Detection of cavities and tun? ? ?nels from gravity data using a neural network[J]. Exploration? ? ? ? ? ?Geophysics, 2001,(1):1-4.
[28] Bescoby D J, Cawley G C, Chroston P N. Enhanced interpretation of magnetic survey data from archaeological sites using arti? ? ?ficial neural networks[J]. Geophysics,2006, 71(5):45.
[29] Al-Garni, Mansour A. “Interpretation of Magnetic Anomalies Due to Dipping Dikes Using Neural Network Inversion[J].” Arabian Journal of Geosciences,2014,8(10):8721-8729.
[30] Al-Garni, Mansour A. “Interpretation of Some Magnetic Bodies Using Neural Networks Inversion[J].” Arabian Journal of Geosciences,2009,2(2):175-184.
[31] Calderón-Macías C, Sen M K, Stoffa, P L. Artificial neural networks for parameter estimation in geophysics[J]. Geophysical Prospecting,2000, 48(1):21-47.
[32] El-Qady G, Ushijima K. Inversion of DC resistivity data using neural networks[J]. Geophysical Prospecting,2001,49(4):417-430.
[33] Srinivas Y, Raj A S, Oliver D H. Muthuraj, D., et al.A robust behavior of Feed Forward Back propagation algorithm of Artificial Neural Networks in the application of vertical electrical sounding data inversion[J]. Geoscience Frontiers,2012,3(5):729-736.
[34] Singh U K, Tiwari. Inversion of 2-D DC resistivity data using rapid optimization and minimal complexity neural network[J]. Nonlinear Processes in Geophysics,2010, 17(1): 65-76.
[35] Neyamadpour A, Taib S, Abdullah. Using artificial neural networks to invert 2D DC resistivity imaging data for high resistivity contrast regions: A MATLAB application[J]. Computers & Geosciences,2009, 35(11):2268-2274.
[36] Hidalgo H,Gomez-Trevino E,Swiniarski R.Neural network approximation of an inverse functional[J].In Proceedings of 1994? ? ? ?IEEE International Conference on Neural Networks,1994,5:3387 -3392.
[37] Zhang L, Poulton. Borehole electrical resistivity modeling using neural networks[J]. Geophysics,2002, 67(6):1790-1797.
[38] Poulton M M, Birken R A. Estimating one-dimensional models from frequency-domain electromagnetic data using modular neural networks. IEEE Transactions on Geoscience and Remote Sensing,1998,36(2):547-555.
[39] Zhang Y, Paulson K V, Magnetotelluric inversion using regularized Hopfield neural networks[J]. Geophysical Prospecting,1997,45(5):725-743.
[40] Shimelevich M I, Obornev E A. Formalized Inversion of Geophysical Data Using Neural Network Technologies with Application to the Tasks of Geoelectrics and Gravimetry[J]. In Engineering and Mining Geophysics 2019 15th Conference and Exhibition,2019,2019:1-5.
[41] Shimelevich M I, Obornev M A. Rapid neuronet inversion of 2D magnetotelluric data for monitoring of geoelectrical section parameters[J]. Annals of Geophysics,2007, 50(1).
[42] Popova I V, Ogawa Y. Application of a Modified Hopfield Neural Network to Noisy Magnetotelluric Data[J]. Izvestiya-Physics of The Solid Earth,2007, 43(3):217-224.
[43] Carbonari R, Maio R D, Piegari, E. Filtering of noisy magnetotelluric signals by SOM neural networks[J]. Physics of the Earth and Planetary Interiors,2018,285:12-22.
[44] Manoj C, Nagarajan N. The application of artificial neural networks to magnetotelluric time-series analysis[J]. Geophysical Journal International,2003,153(2):409-423.
Application of Neural Network in Exploration Geophysics
Wang Sihao,He Lanfang,Li Liang
(Institute of Geology and Geophysics, Chinese Academy of Sciences, Beijing,100029,China)
Abstract:With the development of the geophysical and geological computing, neural network, a data-driven modeling tool, has been widely used in geophysics. Great progress has been made in geophysical data processing and image recognition based on neural network. We briefly reviews the state of the art and the way ahead of the neural network and its apllication in geophysics in this paper. The development history, basic principles and main technical framework of neural network, summarizes the application of neural network in exploration seismic, electromagnetic, gravity and magnetic exploration are presented. The advantages and disadvantages of neural network are discussed. The feasibility of neural network in geophysics has been widely recognized by scholars, but there are still a series of problems to be further explored in practical application.
Key words: Nerual network;Geophysical exploration; Application progress; Development trend