張, 鄭曉東, 李勁松, 路交通, 曹成寅, 隋京坤
1 中國石油勘探開發(fā)研究院, 北京 100083 2 中石化石油工程地球物理有限公司, 北京 100029
?
基于SOM和PSO的非監(jiān)督地震相分析技術
1 中國石油勘探開發(fā)研究院, 北京 100083 2 中石化石油工程地球物理有限公司, 北京 100029
地震相分析技術是儲層預測的一種重要方法,可以用來描述有利沉積相帶的分布規(guī)律.傳統(tǒng)的地震相聚類分析方法對大數(shù)據(jù)的處理運算速度較慢,且容易陷入局部極小值,造成聚類分析的結(jié)構不準確.本文提出基于自組織神經(jīng)網(wǎng)絡(SOM)和粒子群優(yōu)化方法(PSO)相結(jié)合的地震相分析技術,利用自組織神經(jīng)網(wǎng)絡能夠保持原始地震數(shù)據(jù)的拓撲結(jié)構特性的特點,將大量冗余樣本壓縮為小樣本數(shù)據(jù),再通過粒子群的全局尋優(yōu)能力改善K均值聚類的效果.理論模型和實際應用表明該方法能既有效實現(xiàn)數(shù)據(jù)壓縮,又能提供較為準確的全局解,在地震相預測中兼顧計算效率和計算精度.
自組織神經(jīng)網(wǎng)絡;粒子群算法;非監(jiān)督地震相分析;聚類
在地震儲層特征描述和檢測的技術中,地震相分析(Nivlet, 2007)是一種不可或缺的方法.人工地震相分析(Saggaf et al., 2003)需要投入大量的時間,更要求解釋人員具有足夠的經(jīng)驗,并結(jié)合一定的分析方法才能完成.隨著勘探精度的提高,高密度、寬方位采集技術的應用,地震數(shù)據(jù)進入大數(shù)據(jù)時代,如何從海量的地震數(shù)據(jù)中提取地質(zhì)特征信息成為研究的熱點,近年來非監(jiān)督地震相分析方法(de Matos et al., 2007)逐漸受到重視,其主要借鑒模式識別的原理,通過由地震數(shù)據(jù)得到的地震屬性以及其他一些輔助信息來刻畫地質(zhì)體.非監(jiān)督地震相分析方法(例如,Marroquín et al., 2008; Roy et al., 2010; 李芳等, 2014)完全基于數(shù)據(jù)驅(qū)動,極大地降低了人為因素的干擾,即便對于不熟悉當?shù)氐刭|(zhì)情況的人員,也能得到一個較為客觀準確的結(jié)果.
從模式識別的角度來講,地震數(shù)據(jù)具有連續(xù)性、冗余性和一致性的特點.地震相分析本質(zhì)上就是對地震數(shù)據(jù)進行分類,既可以是有監(jiān)督的,也可以是無監(jiān)督的.無監(jiān)督的分類也叫聚類,常見的有K均值聚類方法,其主要是以k為參數(shù),將n個對象分為k簇,使簇內(nèi)具有較高的相似度.而相似度的計算主要是用每個對象到簇中心的歐幾里德距離來度量.Coléou(2003)利用K均值對地震相進行聚類分析,但是該方法(韓家煒和坎伯,2001;張學工,2010)需要預先定義生成簇的數(shù)目,易受“噪聲”和孤立點的影響,且不能保持數(shù)據(jù)的拓撲結(jié)構.對于連續(xù)、低維、高噪聲的地震數(shù)據(jù)來說,K均值聚類不能取得較好的效果.自組織映射是由Kohonen(1982,1990)提出的一種非監(jiān)督模式識別方法,其主要的思想是將數(shù)據(jù)投影到一個低維空間,以獲得更為直觀的理解,但它只能根據(jù)經(jīng)驗來判別類別數(shù),選擇最優(yōu)地震屬性來刻畫地震數(shù)據(jù)中的地質(zhì)特征.劉力輝等(1996)利用自組織神經(jīng)網(wǎng)絡進行地震微相的劃分.陸文凱和牟永光(1998)利用自組織神經(jīng)網(wǎng)絡來追蹤地震同相軸.Steeghs和Drijkoningen(2001)利用聯(lián)合時頻分析的方法來描述由地下反射信息中的微小變化會引起頻率成分的波動.穆星(2005)利用自組織神經(jīng)網(wǎng)絡優(yōu)選幾何屬性對地震相進行自動識別.Marcílio等(2007)引入小波變換的方法來識別瞬時地震道的每個地質(zhì)信息段中的奇異值,該方法更易于實現(xiàn)SOM聚類.粒子群算法是由Kennedy和Eberhart(1997)、Eberhart和Shi(2001)對鳥類的覓食行為的研究時提出的一種基于群智能的優(yōu)化方法計算. 岳碧波等(2009)通過三點濾波的方法改進粒子群的更新速度,從而使粒子更快速收斂.朱童(2011)通過前后粒子的相互作用改進了粒子群方法,從而提高了收斂速度.Liu等(2011)提出基于粒子群的多屬性動態(tài)聚類方法,該方法主要利用群體智能優(yōu)化方法來消除K均值聚類中奇異值對中心點選擇的影響.
本文提出基于自組織神經(jīng)網(wǎng)絡和粒子群優(yōu)化的K均值聚類地震相分析方法,利用SOM網(wǎng)絡將高維的地震數(shù)據(jù)投影到低維空間,有效地保持樣本空間的結(jié)構,在尋找聚類中心點時借鑒群體智能全局尋優(yōu)的思想,利用粒子群方法優(yōu)化自組織神經(jīng)網(wǎng)絡中神經(jīng)元的聚類,最后將聚類的結(jié)果返回到原始空間中,完成地震相的分析.理論模型和實際應用均表明:利用自組織神經(jīng)網(wǎng)絡能夠有效地實現(xiàn)數(shù)據(jù)樣本空間的縮減,極大地減小數(shù)據(jù)的存儲空間,縮短運算時間,再利用粒子群算法能夠提供一個較為準確的全局解,兩者的結(jié)合可以兼顧計算效率和精度,能夠較快較好完成地震相分析.
芬蘭學者Kohonen在研究聯(lián)想記憶和自適應學習機時提出自組織映射,這種網(wǎng)絡的學習機制與人類的大腦皮層上的分區(qū)自組織現(xiàn)象具有很多的相似性,是一種無監(jiān)督的競爭性學習方法.如圖1a SOM網(wǎng)絡由輸入層和輸出層構成,層內(nèi)無連接,層間全連接.輸出層的神經(jīng)元呈矩形或六邊形規(guī)則地排列在同一層上.輸入層中的學習樣本按順序地輸入到網(wǎng)絡空間中進行訓練,輸入樣本的每個維度通過權值與網(wǎng)絡上的神經(jīng)元相連,而神經(jīng)元之間沒有相互連接,通過歐式距離來度量神經(jīng)元對于輸入樣本的敏感度,最為敏感的神經(jīng)元成為最佳匹配單元.假設訓練樣本x(t)的維度為N,計算i位置上的神經(jīng)元向量的連接權值mij(t)與訓練樣本xi(t)之間歐幾里德距離,即
(1)
選擇最近的神經(jīng)元作為最佳匹配單元,學習速率為α,鄰域函數(shù)為h,更新鄰近神經(jīng)元,更新準則為
mij=mij+αh(mwinner,mij)(xi(t)-mij).
(2)
圖1b中的紅色神經(jīng)元是最佳匹配單元,以它為中心,其一定范圍內(nèi)的臨近的神經(jīng)元都會受到影響,而區(qū)域外的神經(jīng)元不會受到影響,這個區(qū)域就構成了相互作用的鄰域.該鄰域函數(shù)可定義為
或
(3)
其中,(i,j)和(m,n)分別表示最佳匹配單元和鄰近神經(jīng)元的位置,其形狀見圖1b.
在訓練過程中,半徑r的選擇由大變小,以確保收斂,訓練結(jié)束后的網(wǎng)絡具有很好的一致性.這種算法對于鄰域函數(shù)的改變表現(xiàn)出很好的一致性.同時,對于半徑和學習速率隨著訓練次數(shù)的增加而減小.如圖2所示,Radius軸上半部分的神經(jīng)元處于興奮狀態(tài),即神經(jīng)元的連接的權重會受到一定程度的改變;Radius軸下半部分的神經(jīng)元處于抑制狀態(tài),神經(jīng)元的連接權重沒有變化.
自組織神經(jīng)網(wǎng)絡聚類的原理是經(jīng)過足夠多的學習后,每個訓練樣本與SOM網(wǎng)絡上神經(jīng)元形成了一個確定的映射關系,即每一個樣本對應網(wǎng)絡中的一個神經(jīng)元,這個神經(jīng)元稱為像,對應的樣本,叫做原像.每一原像只有一個像,而每個像對應著樣本空間中的多個原像.由于神經(jīng)元的個數(shù)有限,當樣本足夠多時,訓練好的神經(jīng)元網(wǎng)絡通過權值矩陣保留原始樣本的數(shù)據(jù)結(jié)構,從而有效地實現(xiàn)樣本數(shù)壓縮,這就為大數(shù)據(jù)情況下地震相的聚類分析提供了一個基礎.
圖1 自組織神經(jīng)網(wǎng)絡結(jié)構圖(a)及神經(jīng)元權值調(diào)整示意圖(b)Fig.1 The structure of Self-Organizing Map and the sketch map of adjusting weight
圖2 鄰域函數(shù)圖(橫軸代表鄰域半徑,縱坐標代表神經(jīng)元的響應函數(shù);左圖和右圖是兩種不同的神經(jīng)元響應函數(shù),R代表鄰域半徑)Fig.2 Neighborhood function
利用自組織神經(jīng)網(wǎng)絡實現(xiàn)地震數(shù)據(jù)的樣本壓縮后,將采用聚類方法對地震相進行劃分.盡管自組織神經(jīng)網(wǎng)絡具有一定的聚類能力,但由于相鄰神經(jīng)元具有較高的相似性,利用該網(wǎng)絡進行聚類分析,往往達不到預設的類別數(shù).本文利用一種粒子群優(yōu)化聚類方法對經(jīng)過自組織神經(jīng)網(wǎng)絡處理的樣本進行聚類劃分.
Kennedy和Eberhart對鳥類的覓食行為研究時發(fā)現(xiàn)鳥群在飛行過程中經(jīng)常會突然改變方向,散開或聚集.盡管行為不可預測,但整體上總保持著很好的一致性,同時,個體與個體之間也保持著最適宜的距離.假設鳥群在覓食的過程中,不知道食物的位置,但已知距離食物的范圍.搜索的目標就是尋找離食物最近的鳥,以找到其周圍的食物.由于粒子群算法采用了不同于遺傳算法的隨機搜索策略,因此在解決全局尋優(yōu)問題時表現(xiàn)出極優(yōu)的搜索效能.
在基本粒子群算法中,粒子群由n個粒子組成,每個粒子的位置代表優(yōu)化問題在D維空間中搜索潛在的解.每個粒子根據(jù)它的位置通過優(yōu)化函數(shù)計算出一個適應值,通過速度來決定其飛行方向和距離,粒子根據(jù)如下三條原則來更新自身狀態(tài):(1)保持自身的慣性;(2)按自身的最優(yōu)位置來改變狀態(tài);(3)按群體的最優(yōu)位置來改變狀態(tài).
設粒子的群體規(guī)模為M,則第i(i=1,2,…,M)個粒子位置可表示為Xi,它所經(jīng)歷的最好位置記為pbest[i],它的速度用Vi表示,群體中最好粒子的位置的索引號用g表示,其位置表示為gbest[i].粒子更新自己的速度和位置的方式如下
(4)
Xi=Xi+Vi,
(5)
粒子群算法最初是用來處理優(yōu)化問題,其搜索過程是從一組解迭代到另一組解,采用同時處理群體中多個個體的方法,具有本質(zhì)的并行性;采用實數(shù)進行編碼,直接在問題域上進行處理,無需轉(zhuǎn)換;與遺傳算法類似,粒子群算法也是多點搜索,其解的質(zhì)量不依賴于初始點的選??;各粒子的移動具有隨機性,可搜索不確定的復雜區(qū),具有更有效的全局搜索能力.
K均值聚類作為一種經(jīng)典算法,通過給定樣本類別數(shù)和中心數(shù)實現(xiàn)數(shù)據(jù)的聚類分析.然而傳統(tǒng)的K均值聚類算法具有兩個固有的缺點:(1)初始值的隨機化會導致不同的聚類結(jié)果,有時甚至存在無解的情況;(2)該算法基于梯度下降法,因此不可避免地陷入局部極小值(劉靖明等,2005).粒子群優(yōu)化方法具有較強的全局尋優(yōu)能力,將其與K均值方法結(jié)合,將K均值方法中類內(nèi)中心點的求取替換為粒子群算法中粒子的搜尋,由此,可以有效地避免傳統(tǒng)K均值方法的缺點,且能夠提高精度與速度.
粒子群方法作為一種全局優(yōu)化方法,結(jié)合K均值聚類,把多個神經(jīng)元聚合的中心作為粒子,有速度、位置以及適應度函數(shù)構成.這里選擇類內(nèi)距離作為適應度函數(shù)
)1/2,
mwj為第wj類的樣本中心,mi為類內(nèi)樣本.當類內(nèi)神經(jīng)元距離達到最小時,適應度函數(shù)最小.粒子群優(yōu)化方法提高了尋優(yōu)能力,有效避免陷入局部極小值.這種方法不但能夠克服基于SOM聚類得到樣本數(shù)要少于實際類別的問題,又能夠大大地縮短運行時間.
① 優(yōu)選能夠反映地質(zhì)目標的敏感屬性,對地震屬性進行預處理,初始化SOM網(wǎng)絡;
② 將地震屬性逐個代入網(wǎng)絡中進行訓練,根據(jù)式(1)計算神經(jīng)元與樣本的距離,確定最佳匹配單元,根據(jù)式(2)更新權值;
③ 如達到一定的迭代次數(shù)或是權值穩(wěn)定不再改變,則訓練完成,否則t=t+1,回到②;
圖3 算法流程圖Fig.3 Flow diagram of improved method
⑤ 按照式(4)和式(5)更新所有粒子的速度和位置,再次進行聚類,并記錄屬性樣本點與神經(jīng)元的對應關系,如果滿足迭代條件則算法結(jié)束,如果不滿足,帶入步驟④重新計算;
⑥ 將最終得到的神經(jīng)網(wǎng)絡聚類結(jié)果反映射到原來的樣本空間,得到已分好類的屬性樣本,此時地震相聚類完成.
圖4a為一個四層介質(zhì)模型,第一、三和四層的速度分別為4000m·s-1,4000m·s-1和5000m·s-1,第二層的速度在橫向上發(fā)生變化,分別為3000m·s-1,3300m·s-1和3700m·s-1. 為了證明該方法對于地層橫向介質(zhì)巖性發(fā)生變化的有效分析,本文采用一個主頻為30Hz,4ms采樣的雷克子波,正演得到如圖4b所示合成地震記錄剖面.用本文中介紹的方法將每一道數(shù)據(jù)代入到圖3的算法中進行聚類分析(該模型將時間樣點默認為屬性),圖5a為利用本文方法得到的結(jié)果,橫向上的三種變化清晰地通過三種類別展現(xiàn)出來,其中,不同的顏色代表不同的類別.如圖5b所示,該模型在迭代初期,目標函數(shù)就近似于零,但隨著迭代次數(shù)增加,目標函數(shù)值始終為零,說明該方法的穩(wěn)定性好且收斂.總之,本文方法能夠識別這種由于儲層橫向上的變化導致的地震相的不同,最終將其區(qū)分開來,即使在邊界處也被很好地區(qū)分開來,同時該方法具有較好收斂特性及很好的穩(wěn)定性.
圖4 地質(zhì)模型(a)及其合成地震記錄(b)Fig.4 Geological model (a) and its synthetic seismogram (b)
圖5 聚類結(jié)果(a)和收斂函數(shù)(b)Fig.5 Cluster results (a) and convergence function (b)
圖6 基于SOM和粒子群屬性動態(tài)聚類結(jié)果(a) SNR=25 dB; (b) SNR=10 dB; (c) SNR=2 dB; (d) SNR<1 dBFig.6 Cluster results of seismic attribute using PSO-SOM
圖7 某商業(yè)軟件聚類結(jié)果(SNR<1)Fig.7 Cluster results with some commercial software(SNR<1)
為了驗證本文算法的實用性,選取實際工區(qū)數(shù)據(jù)加以處理,并通過工區(qū)內(nèi)井資料加以驗證.研究區(qū)塊位于塔里木盆地中部,良里塔格族的礁灘相碳酸鹽巖儲層為主要目的層段,埋深約為5000m,主要巖性為泥質(zhì)灰?guī)r,地震資料的主頻約為20Hz.圖8為該研究工區(qū)的一條地震剖面,圖中圓圈所圈位置是塔中的I號坡折帶,形成于奧陶世中晚期.該地層發(fā)育了一套縱向上多旋回疊置、橫向上多期次加積的較大規(guī)模的礁灘相沉積體系,主要為裂縫-孔洞性儲層類型,是一套主力產(chǎn)層,分布于臺緣坡折帶上(趙文智等,2013).
圖9為研究區(qū)域的地層殘余厚度圖,由于礁灘孔隙型儲層主要受沉積相帶和成巖作用控制,本文主要研究方法是在沉積層序解釋的基礎上,利用地層殘余厚度、古地貌、巖相古地理和地震屬性來分析儲層和沉積相的宏觀展布,利用古地形厚度圖能夠判別礁灘儲集體空間展布特征.根據(jù)鉆探井資料顯示,工區(qū)內(nèi)有6口鉆井(見圖9中黑色圓點所示),其中,W2井、W21井、W22井和W23為工業(yè)油氣井,W25井、W28井為干井或低效井.利用本文提出的方法對有利的儲集相帶進行預測,具體步驟如下.
(1)對目的層段提取了十多種地震屬性,并結(jié)合區(qū)域的地質(zhì)情況和測井資料進行分析.由于縫洞性碳酸鹽巖受斷裂展布的控制,最終優(yōu)選具有明確的地質(zhì)意義的瞬時振幅、瞬時頻率、瞬時相位、曲率4種屬性作為樣本輸入.圖10a為瞬時振幅屬性,紅色代表振幅異常區(qū)域,表征該區(qū)域可能存在巖性或是流體變化;圖10b為瞬時頻率,對于裂縫存在的吸收衰減具有一定的檢測作用;圖10c為瞬時相位,對地震相的邊緣具有很好的描述特性;圖10d為相干曲率,對于斷裂具有很好的預測作用.
(2)對地震數(shù)據(jù)中的空值與野值進行剔除,對優(yōu)選的地震屬性樣本進行預處理,由于不同屬性的數(shù)量級相差較大,為了便于后面的處理,將數(shù)據(jù)進行歸一化處理.
圖8 地震剖面Fig.8 Seismic section
圖9 時間厚度圖Fig.9 The figure of time thickness
(3)初始化SOM網(wǎng)絡中的神經(jīng)元,輸入屬性樣本進行學習,通過網(wǎng)絡權值有效地保留樣本的拓撲結(jié)構,在此基礎上再利用粒子群優(yōu)化方法進行聚類分析,極大地減少了計算時間并將所得結(jié)果與單純利用SOM或粒子群多屬性動態(tài)方法進行比較.
圖11左下角為地震屬性在SOM圖上的聚類結(jié)果.從圖11左下角圖中可知,網(wǎng)絡權值被劃分為紅色、綠色和藍色三類,將神經(jīng)元與樣本聚類結(jié)果的對應關系返回到原空間,就得到右側(cè)的屬性樣本的結(jié)果.由于SOM上的紅色區(qū)域?qū)臉颖緮?shù)據(jù)較少,在聚類結(jié)果上并沒有得到很好的顯示.然而,實際上一個地區(qū)地震相數(shù)目要在5~12之間,這說明SOM網(wǎng)絡相鄰聚類結(jié)果往往會低于真實數(shù)據(jù).造成這種現(xiàn)象的主要原因是單純利用SOM網(wǎng)絡聚類,網(wǎng)絡上臨近的神經(jīng)元權重具有很大的相似性,容易造成類別劃分不明顯,使得聚類結(jié)果顯示出某種不確定性,造成劃分的類別不能達到當初的期望.
圖12、13和14所示分別為利用商業(yè)軟件、粒子群、和本文方法所得到的結(jié)果.圖中深藍色代表臺內(nèi)洼地,土黃色表示礁灘相地層發(fā)育,深綠色代表丘灘相地層,紅色代表有利的油氣聚集帶,深褐色代表斷層,黃色代表斜坡.圖12、13和14中礁灘相(土黃色)在空間上都顯示有一定的展布特征,表征這三種方式均能在一定程度上刻畫碳酸鹽巖礁灘相的展布,藍色的洼地分布也展現(xiàn)在地震相圖中.但圖12的礁灘相的邊界范圍與洼陷模糊不清,所以這種刻畫并不準確.圖13和14對礁灘相的邊界和斷裂帶刻畫都比較清晰.
圖10 地震屬性圖(a) 瞬時振幅; (b) 瞬時頻率; (c) 瞬時相位; (d) 相干曲率.Fig.10 Seismic attribute map(a) Instantaneous amplitude; (b) Instantaneous frequency; (c) Instantaneous phase; (d) Coherent curvature.
圖11 自組織神經(jīng)網(wǎng)絡分類結(jié)果和地震屬性聚類圖Fig.11 Cluster results in SOM and cluster results of seismic attribute using SOM
圖12 某商業(yè)軟件結(jié)果Fig.12 Cluster results with some commercial software
圖13 粒子群聚類結(jié)果Fig.13 Cluster results of seismic attribute using PSO
圖14 基于SOM和粒子群屬性動態(tài)聚類結(jié)果Fig.14 Cluster results of seismic attribute using PSO-SOM
圖15 計算時間對比Fig.15 Computing time
再來對比油氣的預測結(jié)果,從鉆錄井資料獲悉W2井、W21井、W22井及W23均為產(chǎn)油井,而在圖12某商業(yè)軟件處理結(jié)果中只有W21、W23落在油氣聚集區(qū)(紅色區(qū)域),W2及W22落在非油氣聚集區(qū)(藍色區(qū)域),與鉆錄井結(jié)果有偏差;在圖13粒子群聚類結(jié)果中只有W2井、W22井、W23井落在油氣聚集區(qū)(紅色區(qū)域),而W21井落在非油氣聚集區(qū)(黃色區(qū)域),這與實際情況不符;在圖14本文算法處理結(jié)果中,W2井、W21井、W22井和W23均落在油氣顯示相帶(紅色區(qū)域),W25井和W28井落在含油氣性不好區(qū)域(褐色相帶),都能夠與井資料較好吻合.實際資料處理加以井資料驗證,說明本文算法不僅對丘灘相、斜坡以及油氣分布有利相帶的刻畫更加清晰,而且油氣預測與錄井資料也比較吻合,表明本文方法有效性及實用性.
同時,為了對比本文算法在運算效率方面的優(yōu)勢,將商業(yè)軟件處理時間與本文算法處理時間進行了比較,如圖15所示.本文通過利用自組織神經(jīng)網(wǎng)絡保持原始數(shù)據(jù)拓撲結(jié)構特性的特點,將大量冗余樣本壓縮為小樣本數(shù)據(jù),運算時間縮短為商業(yè)軟件處理的1/4,運算效率得到極大的提高.由此表明該方法具有一定的工業(yè)應用價值,尤其是將其應用到大數(shù)據(jù)處理上面.
常規(guī)地震相分析受到海量地震數(shù)據(jù)的限制,導致計算效率較慢.本文首次提出了基于自組織神經(jīng)網(wǎng)絡和粒子群多屬性動態(tài)聚類結(jié)合方法的地震相分析技術,既減小了計算工作量,又具有全局尋優(yōu)能力,有效避免了陷入局部極小值,能取得較為準確的聚類.在模型數(shù)據(jù)和實際資料的應用中,以及在信噪比較差情況下,都取得了較好的聚類效果,較好揭示地震相特征,證明了本文方法的可行性及有效性.
Coléou T, Poupon M, Azbel K. 2003. Unsupervised seismic facies classification: A review and comparison of techniques and implementation.TheLeadingEdge, 22(10): 942-953, doi: 10.1190/1.1623635.
de Matos M C, Osorio P L M, Johann P R S. 2007. Unsupervised seismic facies analysis using wavelet transform and self-organizing maps.Geophysics, 72(1): P9-P21, doi: 10.1190/1.2392789.
Eberhart R C, Shi Y H. 2001. Particle swarm optimization: developments, applications and resources. ∥Proceedings of the IEEE Congress on Evolutionary Computation. Seoul: IEEE, 81-86. Han J W, Kamber M. 2001. Data Mining: Concepts and Techniques (in Chinese). Beijing: China Machine Press.Kennedy J, Eberhart R C. 1997. A discrete binary version of the particle swarm algorithm. ∥ Proceedings of 1997 IEEE International Conference on Systems, Man, and Cybernetics. Orlando, FL: IEEE, 4104-4109.
Kohonen T. 1982. Self-organized formation of topologically correct feature maps.BiologicalCybernetics, 43(1): 59-69, doi: 10.1007/BF00337288.
Kohonen T. 1990. The self-organizing map.ProceedingsoftheIEEE, 78(9): 1464-80.
Li F, Wang S D, Chen X H, et al. 2014. Oil-gas prediction with multi-attributes fusion based on fuzzy logic.OilGeophysicalProspecting(in Chinese), 49(1): 197-204.
Liu J M, H L CH, Hou L W. 2005. Cluster analysis based on particle optimization algorithm.SystemsEngineering-Theory&Practice(in Chinese), (6): 54-58.
Liu L Y, Yang M Y, Xiong J L, et al. 1996. Seismic analysis with self-organizing neutral network.OilGeophysicalProspecting(in Chinese), (S2): 90-94. Liu X F, Zheng X D, Xu G C, et al. 2011. Locally linear embedding-based seismic attribute extraction and applications.AppliedGeophysics, 7(4): 365-375, doi: 10.1007/s11770-010-0260-2.
Lu W K, Mou Y G. 1998. Seismic event automatic tracking using self-organizing neural network.GeophysicalProspectingforPetroleum(in Chinese), 37(1): 77-83.
Marroquín I D, Braulr J J, Hart B S. 2009. A visual data-mining methodology for seismic facies analysis: Part 1: Testing and comparison with other unsupervised clustering methods.Geophysics, 74(1): P1-P11, doi: 10.1190/1.3046455.
Mu X. 2005. Method of seismic facies auto-classification by the seismic geometrical attribute and the self-organizing neutral network.GeologicalScienceandTechnologyInformation(in Chinese), 24(3): 109-112.
Nivlet P. 2007. Uncertainties in seismic facies analysis for reservoir characterisation or monitoring: Causes and consequences.Oil&GasScienceandTechnology-Revuedel′IFP, 62(2): 225-235, doi: 10.2516/ogst:2007019.
PSO-based multi-attribute dynamic clustering technology and its application.SEGTechnicalProgramExpandedAbstracts, 1913-1917, doi: 10.1190/1.3627581.
Roy A, Matos M, Marfurt K J. 2010. Automatic seismic facies classification with kohonen self organizing maps-a tutorial.GeohorizonsJournalofSocietyofPetroleumGeophysicists, 6-14. Saggaf M M, Toks?z M N, Marhoon M I. 2003. Seismic facies classification and identification by competitive neural networks.Geophysics, 68(6): 1984-1999, doi: 10.1190/1.1635052.
Steeghs P, Drijkoningen G. 2001. Seismic sequence analysis and attribute extraction using quadratic time-frequency representations.Geophysics, 66(6): 1947-1959, doi: 10.1190/1.1487136.
Yue B B, Peng Z M, Hong Y G, et al. 2009. Wavelet inversion of pre-stack seismic angle-gather based on particle swarm optimization algorithm.ChineseJ.Geophys. (in Chinese), 52(12): 3116-3123, doi: 10.3969/j.issn.0001-5733.2009.12.021. Zhang X G. 2010. Pattern Recognition (Third Edition) (in Chinese). Beijing: Tsinghua University Press.Zhao W Z, Shen A J, Pan W Q, et al. 2013. A research on carbonate karst reservoirs classification and its implication on hydrocarbon exploration: Cases studies from Tarim Basin.ActaPetrologicaSinica(in Chinese), 29(9): 3213-3222.
Zhu T, Li X F, Li Y Q, et al. 2011. Seismic scalar wave equation inversion based on an improved particle swarm optimization algorithm.ChineseJ.Geophys. (in Chinese), 54(11): 2951-2959, doi: 10.3969/jissn.0001-5733.2011.11.025.
附中文參考文獻韓家煒, 坎伯. 2001. 數(shù)據(jù)挖掘: 概念與技術. 北京: 機械工業(yè)出版社. 李芳, 王守東, 陳小宏等. 2014. 基于模糊邏輯的多屬性融合油氣預測方法. 石油地球物理勘探, 49(1): 197-204.
劉靖明, 韓麗川, 侯立文. 2005. 基于粒子群的K均值聚類算法. 系統(tǒng)工程理論與實踐, (6): 54-58.
劉力輝, 楊夢巖, 熊金良等. 1996. 應用自組織神經(jīng)網(wǎng)絡劃分地震微相. 石油地球物理勘探, (S2): 90-94.
陸文凱, 牟永光. 1998. 用自組織神經(jīng)網(wǎng)絡實現(xiàn)地震同相軸自動追蹤. 石油物探, 37(1): 77-83.
穆星. 2005. 利用地震幾何屬性和自組織神經(jīng)網(wǎng)絡進行地震相的自動識別. 地質(zhì)科技情報, 24(3): 109-112.
岳碧波, 彭真明, 洪余剛等. 2009. 基于粒子群優(yōu)化算法的疊前角道集子波反演. 地球物理學報, 52(12): 3116-3123, doi: 10.3969/j.issn.0001-5733.2009.12.021.
張學工. 2010. 模式識別(第三版). 北京: 清華大學出版社.
趙文智, 沈安江, 潘文慶等. 2013. 碳酸鹽巖巖溶儲層類型研究及對勘探的指導意義——以塔里木盆地巖溶儲層為例. 巖石學報, 29(9): 3213-3222.
朱童, 李小凡, 李一瓊等. 2011. 基于改進粒子群算法的地震標量波方程反演. 地球物理學報, 54(11): 2951-2959, doi: 10.3969/j.issn.0001-5733.2011.11.025.
(本文編輯 胡素芳)
Unsupervised seismic facies analysis technology based on SOM and PSO
ZHANG Yan1, ZHENG Xiao-Dong1, LI Jin-Song1, LU Jiao-Tong2, CAO Cheng-Yin1, SUI Jing-Kun1
1ResearchInstituteofPetroleumExploration&Development,Beijing100083,China2SinopecGeophysicalCorporation,Beijing100029,China
Seismic facies, as the mappable 3D seismic units composed of groups of reflections whose parameters differ from those of adjacent facies units, represent seismic reflections to macro characteristics of sedimentary facies. Seismic facies analysis technique is to describe and interpret the seismic reflection parameters, such as configuration, continuity, amplitude, and frequency, within the stratigraphic framework of a depositional sequence. As a key step in the seismic interpretation workflow, seismic facies analysis determines so much information on depositional process, environment and ultimately can predict potential reservoir only from seismic data in the absence of well data. When the geological information is incomplete or nonexistent, seismic facies analysis is called non-supervised and is performed through unsupervised learning or clustering algorithms. Although unsupervised seismic facies analysis is an effective technique for reservoir prediction, the big seismic data are processed slowly with the traditional methods.In order to overcome the defects of traditional ways which easily fall into the minimum value and lead to the inaccuracy of the cluster of seismic data, this paper proposes a new method to analyse seismic facies combining the Self-Organizing Map (SOM) and the Particle Swarm Optimization (PSO). In this paper, we firstly select the sensitive attribute which can reflect the geological target and normalize the seismic attribute and initialize the SOM network. The reason why we choose SOM is that it can compress a large number of redundant seismic data into a smaller number. As one of the most promising mathematical techniques applied to non-supervised pattern classification, SOM has the characteristics of keeping the topology structure of the original samples. Secondly we will train the seismic attribute one by one in the network, compute the distance between neuron and sample according to Euclidean distance, confirm the optimum matching unit, and update the weight according to renewing criterion. If it reaches to a certain iteration or the weight trends to stabilization, the training is finished, otherwise, return to last step. After the previous data compression, we will improve the K-means cluster using the global optimization of the PSO, which is initialized with a group of random particles (solutions) and then search for optima by updating generations. In every iteration, each particle is updated by following two “best” values. The first one is the best solution (fitness) it has achieved so far, which is called pbest displaying the best location. Another “best” value that is tracked by the particle swarm optimizer is the best value, obtained so far by any particle in the population, which is a global best and called gbest indicating the best swarm. Based on the well trained SOM network, we can find out a proper clustering divide using PSO optimizing method directly, which minimize the fitting degree from which we can get the minimum Euclidean distance then we record down the pbest and gbest. On the basis of results in the last step, we can update all the particles′ velocities and locations and cluster them again and keep a record of the corresponding relationship between attribute samples and neurons. If it reaches to the iterative condition, then quits the algorithm, otherwise, returns to last step to recompute them. Finally, we reverse mapping the clustering results into the original samples space to acquire the well classified attribute samples.In the theoretical model, we design a four layers medium model with a horizontal velocity change in the second layer which means the formation lithology variation in the lateral. With the Ricker wavelet forward modeling, we get the synthetic seismogram and then use the algorithm in the last paragraph mentioned to cluster them, which defaults the time samples as the attribute. Based on our method, three kinds of variations in the lateral can be clearly displayed, from which different colors represent the different classifications. Meanwhile, this method has a very good stability and convergence when the iteration times increase the objective function value still is near zero. For the purpose of testing the robustness to noise, we add noise of different Signal to Noise Ratio (SNR) to model, including SNR=25 dB, SNR=10 dB, SNR=2 dB and SNR<1 dB. From the results, we can find that when SNR>1 the clustering performance is very good and the horizontal variation is discriminated very well by the distinct boundaries. Even if SNR<1 we still can detect the changes basically, and the results can be referenced for our research although there are some clustering errors. Especially we select the SNR<1 models to be processed by certain commercial software from which the clustering is completely disordered and disappointing. It indicates that we can get stable results using our method when the seismic data quality is bad. According to the application to real data from Tarim Basin, the seismic facies map based on our method and SOM are better than the commercial software, the border and fault zone of reef facies are depicted more clearly. From comparing the seismic faices to the wells located in the area, we can find out that oil wells W2, W21, W22 and W23 are distributed in the red color area which implies the potential oil reservoir and dry wells W25 and W28 are located in the brown belt which doesn′t have oil production processed by our method. At the same time, our improved algorithm can greatly shorten the calculation time from comparison of consuming time between our algorithm and commercial software.The traditional seismic facies analysis methods are usually restricted by the massive seismic data because of very low computational efficiency. In our paper, we try to solve the problem and propose a new multi-attribute clustering method combining the SOM and PSO. We make full use of the SOM advantage of compressing redundant seismic data into a smaller number and keeping the original topology structure, and then improve the K-mean clustering by the PSO global optimizing characteristic. The theoretical model and real data show that our algorithm can realize the compression of the seismic data effectively, and provide a more accurate global solution. For seismic facies prediction, it does well in both the calculation efficiency and the accuracy.
Self-organizing feature map (SOM); Particle Swarm Optimization (PSO); Unsupervised seismic facies analysis; Clustering
10.6038/cjg20150933.
Zhang Y, Zheng X D, Li J S, et al. 2015. Unsupervised seismic facies analysis technology based on SOM and PSO.ChineseJ.Geophys. (in Chinese),58(9):3412-3423,doi:10.6038/cjg20150933.
10.6038/cjg20150933
P631
2014-06-11,2015-07-07收修定稿
國家重大專項(2011ZX05004-003)和國家自然科學基金(40504110)聯(lián)合資助.
張,女,1990年生,中國石油勘探開發(fā)研究院在讀博士研究生,主要從事地震儲層預測方面的研究工作.E-mail:xiaoyan_zy@163.com