Pet.Sci.(2013)10:81-90 DOI /s x 81 A new diagnostic method for identifying working conditions of submersible reciprocating pumpin

Size: px

Start display at page:

Download "Pet.Sci.(2013)10:81-90 DOI /s x 81 A new diagnostic method for identifying working conditions of submersible reciprocating pumpin"

审为房
5 years ago
Views:

1 DOI 0.007/s x 8 A new dagnostc method for dentfyng workng condtons of submersble recprocatng pumpng systems Yu Delang, Zhang Yongmng, Ban Hongme 2, Wang Xnmn 3 and Q Wegu 2 3 Abstract: structure, dfferent characterstcs n loadng and unloadng processes of the submersble lnear motor were model proposed had an excellent performance n falure dagnoss of the submersble pumpng system. Key words: Introducton The submersble recprocatng pumpng system s a new type of pumpng unt used n ol felds. The submersble recprocatng pump s drven by a lnear motor to lft whch s powered by a submersble cable, to control the Dfferent from conventonal pumpng systems, n ths pumpng system a submersble plunger pump s drectly connected to a lnear permanent magnet synchronous motor at the well bottom to lft crude ol (Yu et al, 20a; 20b). Snce the output power of the lnear motor acts drectly on the pump for lftng crude ol, the submersble pump has hgh energy converson effcency (Rossn et al, ntroduced nto the submersble pumpng system compared wth conventonal pumpng unts, thus completely avodng the lnear submersble motor and the plunger pump are exposed to severe well bottom condtons, falures such as lnear motor damage may occur, whch serously shortens the Therefore, t s mportant to fnd an effectve way to detect of the submersble pumpng system. Wth more and more unts have been wdely used for ths type of pumpng system. Ths has major sgnfcance for energy savng and effcent producton (Yu et al, 20a; 20b). such as paraffn deposts n ol tubng, sand producton, ts n-servce lfe, as well as the contnuous runnng tme, dynamcally process the abnormal condtons of the pumpng system and to perform onlne dagnoss, thus provdng early

2 82 falure warnng. For conventonal beam pumpng unts, an ntellgent patterns (Gbbs and Neely, 966; Foley and Svnos, 989; recprocatng pump s a new type of rodless pumpng unt, (or dsplacement) to obtan a surface dynamometer card. few papers on the analyss of ts typcal falure nformaton. The pumpng system would not be runnng under abnormal condtons for a long tme n practce, so t s dffcult to nformaton and runnng state have an observably nonlnear relaton (Zhou and Bennett, 998; Polycarpou and Trunov, 2000; Polycarpou, 200; Zhang et al, 2005), especally n the transtory stage of startng or stoppng. For these reasons, 999). In order to solve the above problems, we propose a submersble pumpng systems based on a support vector of the loadng/unloadng process of a submersble lnear are extracted. Then the sample data of the characterstc pumpng systems s determned by observng and dentfyng compared. 2 Falures and ther characterstcs of the submersble pumpng system of several submersble pumpng systems n several ol shortage n the pump and gas obstructon may occur after probably cause damage to submersble lnear motors and pumps, affectng the ol producton. The submersble recprocatng pumpng system s dfferent s under the pump body, whle ts standng valve s above the pump body. Furthermore, the submersble pumpng system s a type of rodless pumpng unt and so ts dynamometer card cannot be measured. In ths paper, the load changes under pumpng system. relatvely hgh, some gas may be separated from the fluds gas may be compressed and the exhaust valve cannot open soluton gas and compressed gas are left n the clearance valves between the beam pumpng unt and the submersble pumpng unt, the gas obstructon for the motor load s the same. The more the gas there s, the longer the loadng and T UU, loadng T DU, unloadng tme durng the T UD T DD, P UE, average power P DE ) are extracted from loadng tme, unloadng tme and power of the lnear motor, whch serve as Influence of lqud shortage n the pump If the s started, the sucton valve cannot mmedately open, so I UE, average motor current durng I DE, varance of the motor current durng I UV, varance of the motor current durng the I DV T UD, and T DD ) are extracted Influence of leakage the pressure between the sucton and exhaust valves gradually goes up, and the pressure underneath the sucton valve valve, supportng the plunger of the recprocatng pump and the exhaust valve and the upward movement of the plunger, the pressure between the sucton and exhaust valves wll velocty, t wll reproduce the support effect and lead to and exhaust valves where the exhaust valve s not tght durng

3 83 of the sucton valve. Therefore, the unload rate becomes slow and the movement of the plunger decelerates durng the latter pump leads to premature loadng. 3) When the travellng and the supermposed effect of these cases n solaton. 4) When and unloadng processes are not hysterestc, the maxmum load cannot reach the theoretcal maxmum. Therefore, the T UU, T DU, T UD, T DD, I UE, I DE, P UE, P DE ) are extracted from loadng tme, unloadng tme and Influence of sand producton When the submersble pump s nstalled at a locaton of the tubng strng whch s surrounded wth unconsoldated formaton or the producton pressure s too hgh, sand wll move nto the wellbore from the reservor. If some fne sand gets nto the pump, t wll generate addtonal resstance to the plunger n a certan nduce a drastc change n the pump load n a short tme. I UV, I DV ) are extracted Influence of paraffn deposts Paraffn deposted n the ol tubng creates an addtonal resstance to the plunger durng operaton. Ths wll ncrease the load durng the located at the lower part of the pump, the load change of the (I UE, I DE, P UE, P DE ) are extracted from load fluctuaton n a deposts. The analyses of the pump structure and the operaton the submersble recprocatng pumpng system. 3 Extracton of characterstc quanttes The submersble recprocatng pump s operated n an ntermttent mode. Each runnng cycle ncludes four nterval. Changes of the motor voltage and current n a cycle research focus. are observed n the motor current curve, whch are the startng current of the lnear motor. The startng current may when extractng characterstc parameters. Therefore, t s before processng data. The method s shown n Fg. 2, whch uses t 2 t ) before t and t 2 Voltage U, V Current, A Current A Tme t, s Tme t, s Fg. Voltage and current of the lnear motor 0 t t 2 Peak value of the startng current Tme t, s Frst mnmum value Fg. 2 Startng current of the lnear motor pumpng system and the load of the submersble motor, and the data such as operatng voltage, current and power may mrror the change of the load of the lnear motor. Therefore, t s necessary to extract the characterstc parameters of the condtons. In an operatonal cycle of the lnear motor, the motor 4 4

4 84 I t t U UE U () t () U t where U t U s the I UE s the average I k I (0 k ) (2) UU UU UE UU I k I (0 k ) UD UD UE UD where k UU and k UD I UU and I UD are the threshold values of the loadng current and the T UU mn S s.t. I I S{, 2,, t } S UU U T (6) UD max Q (7) where T UU, T UD (6) are the loadng tme and td IDE D () t (8) td t where D t D s the I DE s the average I DU, I and k DD DU, k DD (3) (0< k DU <, 0< k DD <). k DU and k DD are the loadng and unloadng coeffcents of the motor current durng the I and I DU DD are the threshold of loadng current In a smlar way, T and T DU DD are the loadng tme and the P UE t U PU () t tu t tu I ( ( t) I ) UV U UE tu t td I ( ( t) I ) DV D DE td t 2 2 (3) (4) (5) (9) (0) () (2) where P UE and P DE are the average motor power values durng I UV and I DV are 4 Dagnoss of workng condtons based on SVM 4. Bayes decson rule based on SVM In actual ol producton, downhole condtons of producton wells are dfferent, each submersble recprocatng condtons, so t s necessary to propose a method for a statstcal learnng machne whch maps low-dmensonal, lnearly non-separable data to a hgh dmensonal space and constructs an optmal hyperplane to classfy the data (Chen For a set of non-lnearly separable samples, {( x, )} N y, y or y, N samples n all, f there s a nonlnear mappng. 0 0 ( ), R n m x R, ( m0 n0) (3) hyperplane can be expressed as follows: f ( x) ( x) b (4) where and (x) are m 0 dmensonal vector; b s offset. On the optmal hyperplane, 2/ should be the largest. Lagrange method. N T mn J C (5), b, 2 s.t. y ( x b) (6) where ; C s the penalty coeffcent. When C ncreases, the degree obtaned: Lagrange s multpler, { } N, meets the constrants: N s.t. 0 C, y 0 (, 2,, N) (7) (8)

5 85 T kx ( xj) ( x) ( xj) can be obtaned from the orgnal nput space: N f ( x) yk( x, x) b (9) valued classfcaton usually can be obtaned by usng a dagnostc method s to construct a super-sphere n hgh samples, and refuses other samples. For an h postve class, that s y S, and the correspondng category s, then f( x) k ( x, x), xs,2,, h (20) Then we obtan hf (x sample category (b s neglected for t s a constant and does as follows: kxx (, ) f( x) f ( x) kxx (, )d x f( x)dx (2) Because f ( x) 0 and f ( x)dx, f ( x) can be sed as a densty functon of some probablty dstrbuton to construct the Bayes classfer. Condtonal probablty s estmated as follows: f ( x) Px ( )= (22) f ( x)dx Pror probablty estmaton can be expressed as: P( )= f ( x)dx 2 f ( x)d x f ( x)d x f ( x)dx h (23) Px ( )= Px ( ) P( ) Px ( 2) P( 2) Px ( h) P( h) (24) estmaton can be expressed as: Px ( ) P( ) P( x) Px ( ) f ( x) f ( x)d x f ( x)d x f ( x)dx Px ( ) 2 h (25) Because the total probablty densty does not affect P(x as follows: f ( x) P( x) f( x)d x f2( x)d x fh ( x)dx (26) (,2,, h) sample categores can been judged. If f( x) max { f j( x)}, then x (27) j, 2,, h h sngle-valued x, b ) 2, b 2 ) c, b c ) f (x) f 2 (x) f c (x) Fg. 3 max 4.2 Constructng an nput space vector In a cycle of operaton of the pumpng unt, several are extracted. Because the orders of magntude of these x xmn xˆ 2 (28) xmax xmn where ( xˆ [, ]), and x s the x max and x mn are the maxmum and mnmum values of the characterstc G [ Tˆ, Tˆ, Tˆ, Tˆ, Iˆ, Iˆ, Iˆ, Iˆ, Pˆ, Pˆ ] UU UD DU DD UV DV UE DE UE DE T (29) 4.3 Dagnoss of workng condtons of the submersble pumpng system One of the mportant problems n the dagnoss of the expressed as follows:

86 2 2 kx (, x) exp( x x / ) k l k l (30) where. I UE, I DE, P and P UE DE are nherent n the runnng of the submersble pumpng system, and they only vary wth the sample data.

6 kx (, x) exp( x x / ) k l k l (30) where. I UE, I DE, P and P UE DE are nherent n the runnng of the submersble pumpng system, and they only vary wth the sample data. But T UD, T DU, T DD, I UV and I DV are not only related wth sample data but also related wth k UU, k UD, k and k DU DD, so the selecton of the loadng/unloadng threshold wll affect the accuracy of the producton, paraffn deposts and normal runnng, orgnal C and of machne for multple condtons of the submersble pumpng system, as shown n Fg. 3. system. 5 Smulaton experments 5. Smulaton of workng condton dagnoss based on SVM flud crculaton n an ol well. The platform conssted of a smulated crculaton system, a flow control devce, a submersble recprocatng pumpng system and ts the expermental platform are shown n Fgs. 4 and 5, respectvely. Electrcal parameters of the cylnder-shaped lnear motor used were as follows: outer dameter 0 mm, length 6 m, rated voltage 660 V, the maxmum nput current N, maxmum thrust 35,000 N, maxmum travel of the slder 23 cm. Fg. 6 shows the cylnder-shaped lnear motor and the recprocatng pump, the red one s the lnear motor, and the blue one s the recprocatng pump. Fg. 7 shows the control cabnet of the submersble recprocatng pumpng system. In experments, of the lnear motor and ts current and power and other vector, the frst 230 nput vectors were used to construct tranng samples for the classfer, and the remanng 70 Control cabnet Manual valve Pressure meter 2 Flow meter 2 Pressure meter Electrc valve Flow meter Manual valve 2 for the submersble pumpng system was traned by the 230 samples and tested by the 70 samples. k UU, k UD, k DU and k DD were manually-set parameters, f they changed, the tranng samples would change, and C and would also change. The msjudgment rate of each falure wth dfferent parameters s shown n Table, F : gas obstructon, F 2 n the pump, F 3 4 : standng valve : sand producton, F 8 : paraffn deposts, F 9 : normal runnng (when the output of F to F 8 value, the runnng state may be determned as F 9 ). The values of relevant parameters shown n Table are only part of the expermental data. It was clear that when k UU =0.85, k UD =0.80, k DU =0.80, k DD =0.75, the msjudgment rate of each falure could be controlled n an expected M Crculatng pump Lqud storage tank Submersble recprocatng pump Fg. 4 Schematc of the smulaton platform Fg. 5 Photo of the smulaton platform

So the msjudgment rate of F 9 was hgher than those of others. system s not only related to the model parameters, but also related to the number of tranng samples.

7 87 Fg. 6 Lnear motor and recprocatng pump range. It should be noted that the normal runnng state F 9 was not an ndvdual classfer, ts output came from other the correspondng classfers were traned and tested. So the msjudgment rate of F 9 was hgher than those of others. system s not only related to the model parameters, but also related to the number of tranng samples. Ths paper on the msjudgment rate at the same model parameters. an ncrement of 0, were used to estmate the dependence of Fg. 7 Control cabnet of the submersble recprocatng pumpng system condton dagnoss decreases wth an ncrease n the sand producton, the msjudgment rate decreases at an almost Table k UU, k UD, k DU, k DD (0.70, 0.80, 0.75, 0.70) (0.75, 0.80, 0.75, 0.75) (0.80, 0.85, 0.75, 0.80) (0.85, 0.80, 0.80, 0.75) (0.90, 0.85, 0.90, 0.80) F F F F F F F F F

8 Msjudgment rate, % F F2 F7 F8 Msjudgment rate, % F3 F4 F5 F Number of tranng samples (a) Number of tranng samples (b) Fg. 8 the number of tranng samples reaches 200 or so, and then the msjudgment rate changes slghtly. Under the condton when the number of tranng samples s less than 50. The valve and the both valves, the change of the msjudgment rate are 250 tranng samples or so, and then fluctuate a lttle. rate decreases slowly when the number of tranng samples Therefore, the msjudgment rate can reach a desred value 5.2 Comparatve analyss of SVM- and LVQ-based methods Usng the same expermental data, the msjudgment rates of the dagnostc method based on the learnng vector for classfer ncluded the frst 230 nput vectors, and the structure of LVQ was : the number of the neurons n the nput layer was the vector dmenson of characterstc neurons n the compettve layer was 50 (the desgn and s not dscussed here). The classfer based on LVQ for system was traned and tested, respectvely. The msjudgment rate (lsted n Table 2) was also related to the tranng sample Fault Table F F F F F F F F condtons are not dscussed here), the effect of the sample when the model parameters are the same. The tranng sample a constant rate when the sample number s below 70, and then decreases slowly wth an ncrease n the tranng sample

9 89 Msjudgment rate, % SVM LVQ Number of tranng samples Fg. 9 decreases at a faster rate and fnally reaches a mnmum condtons of the submersble recprocatng pumpng system. 6 Conclusons of the submersble recprocaton pumpng. Based on the structural characterstcs of the pumpng unt, a new way recprocatng pumpng unt was proposed and then tested by the data obtaned from the smulaton platform. ) Based on the structural and runnng characterstcs of the submersble recprocatng pumpng unt, the characterstcs of the loadng/unloadng of the submersble n ths paper could be accurately descrbed by the loadng/ unloadng characterstcs. 2) The loadng/unloadng characterstcs of the k UU, k UD, k DU, k DD and other characterstc T UU, T UD, T DU, T DD, I UV, I DV, I UE, I DE, P UE, P DE were obtaned, whch were used submersble recprocatng pumpng unt. recprocatng pumpng unt on the smulaton platform. These recprocatng pumpng unt. The test results were compared References 2th Internatonal Conference on Pattern Recognton and Neural wth event-classes. IEEE Transactons on Relablty (): 3-4 Nonlnear Phenomena (3): vector machne. Transactons of Chna Electrotechncal Socety (8): (n Chnese) 3-26 Journal of Chna Coal Socety (): (n Chnese) Fol ey W L and Svnos J G. Expert advsor program for rod pumpng. Journal of Petroleum Technology (4): (n Chnese) 8(): 9-98 Kre bel U. Parwse classfcaton and support vector machnes. In: based on boostng Bayesan classfer. Transactons of Chna Electrotechncal Socety (5): 79-8 (n Chnese) Transactons of Chna Electrotechncal Socety (6): (n Chnese) 2(2): (6): 9-4 (paper SPE 25420) 384 Control (4): nonlnear dynamcal systems usng a learnng approach. IEEE

10 90 44(4): trees for computer-based systems. IEEE Transactons on Relablty (4): on ndcator dagram of twn screw refrgeraton compressor. Internatonal Journal of Refrgeraton (4): Wu L, Zhu Y and Yuan L. Novel method for ntegrated dagnoss of of Chna Electrotechncal Socety (4): 45-5 (n Chnese) (): Proceedngs of the CSEE. 20a. 3(27): (n Chnese) Yu D, Q W, Deng S, et al. Submergence forecastng of a submersble Snca. 20b. 32(3): (n Chnese) Control (3): Chnese) Internatonal Journal of System Scence (): (Edted by Sun Yanhua)

SVM [6] PCA+SVM 79.75% 9 FERE FERE. PCA LDA Adaboost SVM 5 1 SVM Moghaddam [6] M (x,y ) x R N y x y {0,1} M f ( x) = y α k( x, x ) + b x k f(x) = 1 x

SVM [6] PCA+SVM 79.75% 9 FERE FERE. PCA LDA Adaboost SVM 5 1 SVM Moghaddam [6] M (x,y ) x R N y x y {0,1} M f ( x) = y α k( x, x ) + b x k f(x) = 1 x 100084 PCA Fsher FLD Adaboost SVM 9 FERE Adaboost 1. Gollomb [1] SEXNE 30 30 Cottrell [] BP Edelman [3] Alce [4] PCA PCA [5] Moghaddam [6] (SVM) RBF Fsher FLD FERE 3.4% Shakhnarovch [7] Adaboost 78% ±15