CN104675380A - Online oil-drilling drill string monitoring system and fault diagnosis method - Google Patents

Abstract

The invention discloses an online oil-drilling drill string monitoring system and a fault diagnosis method and belongs to the technique of oil drilling monitoring. The online oil-drilling drill string monitoring system comprises a PCB (printed circuit board), an upper computer, a torque sensor and an acceleration sensor, the torque sensor and the acceleration sensor are respectively connected with a drill string and the PCB and transmit collected torque signals and vibration signals to the PCB, and the PCB communicates with the upper computer with data subjected to pre-processing through a Bluetooth module. By the arrangement, transplantation is good, drilling state of the drill string can be monitored in real time, and abnormal conditions can be subjected to fault diagnosis accurately and reliably, so that convenience is brought to drilling personnel to take remedial measures timely, and occurrence of dangerous accidents is effectively prevented.

Description

Oil drilling drill string on-line monitoring system and method for diagnosing faults

Technical field

The invention belongs to oil drilling monitoring technology, particularly drilling tool monitoring system and method for diagnosing faults thereof in Process of Oil Well Drilling.

Technical background

At present, domestic forecast analysis is carried out to engineering accident contingent in drilling tool drilling process, mainly come by the various ANOMALOUS VARIATIONS of creeping into engineering parameter of measuring and analysis, its test macro accuracy is not high and major part can only diagnose out contingent fault by comparing before and after the data of measurement according to artificial experience, this method not only efficiency is low, and the degree of accuracy is not high, easily cause erroneous judgement.Therefore propose a kind of energy and monitor the system that drilling tool creeps into state in time, accurately, reliably, the abnormal conditions that can occur when the very first time creeps into drilling tool make warning and the fault diagnosing out drilling tool to occur exactly is very necessary, are conducive to the generation avoiding causing security incident because drilling tool is abnormal.

Summary of the invention

The object of this invention is to provide a kind of oil drilling drill string on-line monitoring system and method for diagnosing faults, fault diagnosis can be carried out for the abnormal conditions of drilling tool in drilling process and alarm, improve the benefit solving engineering abnormal problem, effectively avoid the generation of peril.

Technical scheme of the present invention is: oil drilling drill string on-line monitoring system, it is characterized in that, comprise pcb board, host computer and the torque sensor be connected with drill string respectively and acceleration transducer, torque sensor is connected with pcb board again respectively with acceleration transducer, and the torque signal collected and vibration signal are flowed to pcb board, pretreated data are carried out communication by bluetooth module and host computer by pcb board.

Described torque sensor is the strain-type torque sensor be made by the two-way foil gauge of two panels, and foil gauge is attached on rectangle metal foil film, and the length of side of metal foil film is equal with drill string girth.

The torque sensor of described torque sensor exports between ground wire and torque sensor signal output line and forms torque output signal, and is connected with pcb board respectively.

The surface-mounted integrated circuit that described pcb board is is core with C8051F005 single-chip microcomputer, comprises that 7.2V turns 5V potential circuit, 5V voltage turns 12V potential circuit, 5V voltage turns ± 5V potential circuit, torque signal amplifying circuit, vibration signal bias voltage circuit, Keysheet module circuit, C8051F005 single-chip microcomputer and peripheral circuit, alarm module circuitry.

The Z axis signal output lead of described acceleration transducer, Y-axis signal output lead, X-axis signal output lead gather the radial direction of drill string, the vibration signal of vertical and horizontal respectively, and three signal output leads are connected with pcb board respectively.

The input of described bluetooth module is provided with bluetooth module power line, bluetooth module ground wire, bluetooth module data downloading wire and bluetooth module DOL Data Output Line, and is connected with pcb board respectively; Wherein bluetooth module power line and bluetooth module ground wire are the power line of bluetooth module, power to bluetooth module, and the cooperating of bluetooth module data downloading wire and bluetooth module DOL Data Output Line makes bluetooth module transfer data to host computer.

The state that described host computer creeps into through the man-machine interface on-line monitoring drill string of labview Software for Design.

Oil drilling drill string on-line fault diagnosis method, it is characterized in that the method for diagnosing faults optimizing RBF neural based on self adaptation Dynamic Clonal Selection Algorithm, concrete method for diagnosing faults is as follows:

One, the central point in self adaptation Dynamic Clonal Selection Algorithm optimization RBF neural and width parameter, the connection weights between least square method determination hidden layer and output layer, concrete steps are as follows:

1: data normalization process

RBF neural requires that the data of input are positioned at interval [0,1], otherwise transfer function cannot process, and therefore the data of input is normalized, if the maximum value in node data is x _imax, minimum value is x _imin, i ∈ (1,2, m), normalization formula is as follows:

${\hat{x}}_i=\frac{x_i-x_{i\min }}{x_{i\max }-x_{i\min }};$

2: floating-point encoding, initialization of population and decomposition

1) floating-point encoding

G will be established for antibody population Ab (g)={ ab ₁, ab ₂, ab _nbe the N tuple of antibody, antibody ab _j=(z ₁ ^j, z ₂ ^j, z _l ^j), i-th gene z of a jth antibody _i ^jfor the floating number in [0,1] interval, z _i ^jwith x _i ^jmapping relations as follows:

x _i ^j＝a _i+z _i ^j*(b _i-a _i)

Claim z _i ^jx _i ^jcoding, x _i ^jz _i ^jdecoding;

2) initialization of population and decomposition

The maximum number D of setting hidden layer node, the size L of population size, evolutionary generation G, e-learning speed ξ (ξ ∈ [0,1]); Hidden center c is initially the floating number in input amendment data area, the floating number on center width is initialized as [Isosorbide-5-Nitrae] interval;

Population decomposition is Ab (g) by antibody and the descending sort of antigen affinity size, and the Ab (g) after sequence is resolved into three populations according to the ratio of 1:6:3, is Ab respectively _m(g), Ab _n(g), Ab _r(g); Ab _mg () is the highest population of affinity, Ab _ng () is the population of medium affinity, Ab _{r (g)}for population (Ab (g) the ∈ S that affinity is minimum ^n*L, Ab (g)=Ab _m(g) ∪ Ab _n(g) ∪ Ab _r(g), m+n+r=N);

3: the foundation of object function

Decoded data message is built RBF neural, and calculate the overall error of whole network, overall error is the smaller the better, and the quadratic sum therefore choosing desired output and actual both output difference is minimum as object function; When object function is minimum, RBF neural training terminates, and its expression formula is:

$E=\frac{1}{2}\underset{j=1}{\overset{M}{\mathrm{\Σ}}}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{({\hat{y}}_{\mathrm{ji}}-y_{\mathrm{ij}})}^2$

Wherein, M is number of training, and N is output layer nodes, be M the N number of neuronic desired output of sample, y _jibe M the N number of neuronic real output value of sample;

4: self adaptation dynamic clonal selection operates

1) clone

Antagonist ab ∈ Ab _m(g) ∪ Ab _n(g), clone by following formula:

C＝∪ _iC _i＝I×ab,i＝1,2,···,n _c

Wherein, I is n _cdimension unit row vector, n _cfor the clone sizes of antibody ab, C _i=ab is a clone of ab;

2) evolutionary operator

If the sequence of antibody ab ∈ Ab (g) is rank (ab), population scale is Q, by affinity standardization, as shown in the formula:

$F\left(\mathrm{ab}\right)=\frac{Q-\mathrm{rank}\left(\mathrm{ab}\right)}{Q}$

By following formula, change of scale is carried out to F (ab)

T _sc(ab)＝η·exp(-ρ·F ²(ab))

Wherein, η and ρ is constant, general 0 < η < 0.5,2≤ρ≤10; Can find out, F (ab) larger T _sc(ab) less;

3) elite's clonal vaviation T _em()

High affinity antibody ab ∈ Ab _mg () is positioned at majorized function f (x _i) near peak value, therefore at its locally optimal solution of less space search; To ab ∈ Ab _mthe clone C of (g) _i∈ C, by Probability p _m∈ (0,1) carries out following mutation operation:

C _i'＝C _i+T _sc(ab)·(2δ-1)

Wherein, δ is the random number series vector in l dimension [0,1] interval; p _mget the small value, general p _m=1/l;

4) clone's intersection T _hm()

If ab ∈ is Ab _nthe clone C of (g) _i∈ C, random selecting antibody ab' ∈ Ab _m(g), carry out interlace operation by following formula:

C _i'＝C _i×(1-T _sc(ab)·δ)+ab'×(T _sc(ab)·δ)

Wherein, δ is the random number series vector in l dimension [0,1] interval;

5) Immune Clone Selection

To ab antibody through T _em() or T _hmclone group C'=∪ C after () operation _i' carry out affinity and assess { F (C') }=∪ _if (C _i'), if there is F (C _i')=maxF (C') >F (ab), then ab:=C _i', f (ab) :=f (C _i'), thus upgrade antibody population, realize information exchange;

6) dynamic self-adapting algorithm

T _sc(ab) except relevant with antibody affinity, the same impact by η and ρ; For elite's antibody population Ab _mg () is mainly used in searching for locally optimal solution, therefore η gets the small value, and makes it be η ₁; Antibody population Ab _ng () is mainly used in exploitation global optimum solution space, therefore η takes large values, and makes it be η ₂; If optimal solution is constant within some generations, then η ₁to be multiplied by randomly or divided by 1.3, otherwise η ₁get back to its initial set value; In like manner, Ab is worked as _m(g) ∪ Ab _nwhen the average criterion functional value of () is substantially constant within some generations g, then η ₂to be multiplied by randomly or divided by 1.3, otherwise η ₂get back to its initial set value; Wherein, 1.3 is a good auto-adaptive parameter;

The hunting zone impact of ρ value on medium affinity antibody is comparatively large, on-the-fly modifies in the following way to ρ value:

Increase ρ value every some generations, return initial setting value after increasing to certain value, circulation is carried out successively, is shown below;

$\mathrm{\&rho;}(g+1)=\left\{\begin{array}{c}\mathrm{\&gamma;}\&CenterDot;\mathrm{\&rho;}\left(g\right),\mathrm{if\; rem}(g,g_D)=0\\ {\mathrm{\&rho;}}_0,\mathrm{if\; rem}(g,g_{D\max })=0\end{array}\right.$

Wherein, ρ ₀for initial value, γ >1 is Dynamic gene, the remainder that rem (x, y) is x/y, g _dfor interval algebraically, g _dmaxfor largest interval algebraically;

5: weight computing

After central point and width are determined, the RBF in RBF neural structure and hidden layer output valve are determined, what differ between hidden layer and output layer is the matrix product factor, utilize least square method to obtain the weights W of each element in matrix;

1) input all training datas, obtain the quadratic sum of square error:

$E\left(n\right)=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left(y_i\left(n\right)\widehat{-}{y}_i\left(n\right)\right)}^2$

Wherein, n represents iterations

2) following adjustment is made to connection weights:

$w_i(n+1)=w_i\left(n\right)+\mathrm{\&xi;}\&CenterDot;\frac{\&PartialD;E\left(n\right)}{\&PartialD;w_i}$

Wherein, ξ is the learning rate set in neutral net, ξ ∈ [0,1];

3) when E is less than specified threshold ε, calculate and stop, in this patent, get ε=0.0001;

6: repeat Step3 ~ Step5, neural network training, makes initial weight constantly revise evolution, obtain globally optimal solution;

Two, fault diagnosis

By the neutral net that the input of the data of Real-time Collection trains, draw oil drilling fault diagnosis result about drilling tool in drilling process; Wherein, neutral net input number of nodes is 2, is input TORQ and drill string acceleration calculation value respectively; Output node number is 3, and represent bit freezing, 010 with 001 respectively and represent drilling string broken off, 100 and represent drill stem washout, zero defects represents with 000.

Through state that the man-machine interface on-line monitoring drill string of labview Software for Design creeps in host computer, the method simultaneously applied based on self adaptation Dynamic Clonal Selection Algorithm optimization RBF neural carries out fault diagnosis to the sample data of Real-time Collection, if there is unusual condition to occur when drill string creeps into, system sends reports to the police and show fault diagnosis result in labview front panel.

Patent transplantability of the present invention is good, and the state that energy Real-Time Monitoring drill string creeps into also can carry out fault diagnosis to abnormal conditions accurately and reliably, is convenient to drilling well staff and adopts remedial measures in time, effectively prevent the generation of peril.

Usefulness of the present invention:

1) the present invention is by torque sensor, acceleration transducer and be that the pcb board of core is connected to form data acquisition unit with single-chip microcomputer, and transplantability is good, real-time;

2) the present invention is by Bluetooth wireless transmission data to host computer, realizes remote monitoring;

3) the present invention utilizes self adaptation Dynamic Clonal Selection Algorithm optimization neural network, makes neutral net not easily be absorbed in local optimum, and makes neural network parameter optimum, thus improves the precision of fault diagnosis.

Accompanying drawing explanation

Fig. 1 is Fundamentals of Supervisory Systems figure in the present invention;

Fig. 2 is the Troubleshooting Flowchart of self adaptation Dynamic Clonal Selection Algorithm optimization neural network in the present invention;

In figure: 1 torque sensor, 2 torque sensors export ground wire, 3 acceleration transducers, 4Z axis signal output lead, 5Y axis signal output lead, 6X axis signal output lead, 7PCB plate, 8 bluetooth module power lines, 9 bluetooth module ground wires, 10 bluetooth module data downloading wires, 11 bluetooth module DOL Data Output Lines, 12 bluetooth modules, 13 torque sensor signal output lines, 14 host computers.

Detailed description of the invention

In order to can statement this patent clearly, 1,2 pair of this patent be further described by reference to the accompanying drawings.

Oil drilling drill string provided by the present invention creeps into on-line monitoring system, comprises torque sensor 1, torque sensor exports ground wire 2, acceleration transducer 3, Z axis signal output lead 4, Y-axis signal output lead 5, X-axis signal output lead 6, pcb board 7, bluetooth module power line 8, bluetooth module ground wire 9, bluetooth module data downloading wire 10, bluetooth module DOL Data Output Line 11, bluetooth module 12, torque sensor signal output line 13, host computer 14.

Torque sensor 1 was both connected with drill string respectively with acceleration transducer 3, was connected respectively again with pcb board 7.Torque sensor exports between ground wire 2 and torque sensor signal output line 13 and forms torque output signal, and is connected with pcb board 7.

Three output lines of acceleration transducer 3 and Z axis signal output lead 4, Y-axis signal output lead 5 are connected with pcb board 7 with X-axis signal output lead 6, and it represents the radial direction of drill string, the vibration signal of vertical and horizontal respectively.

The torque signal of the drill string of collection and vibration signal are flowed to pcb board 7 by torque sensor 1, acceleration transducer 3, by single-chip microcomputer pretreatment pcb board 7, pretreated data are carried out wireless telecommunications by bluetooth module 12 and host computer 14.

The input of bluetooth module has four lines and bluetooth module power line 8, bluetooth module ground wire 9, bluetooth module data downloading wire 10 and bluetooth module DOL Data Output Line 11 to be connected with pcb board respectively.Wherein, bluetooth module power line 8 and bluetooth module ground wire 9 are the power line of bluetooth module, power to bluetooth module.The cooperating of bluetooth module data downloading wire 10 and bluetooth module DOL Data Output Line 11 makes bluetooth module can transfer data to host computer 14.

Torque sensor is the strain-type torque sensor be made by the two-way foil gauge of two panels, and foil gauge is attached on rectangle metal foil film, and the length of side of metal foil film is equal with drill string girth.The surface-mounted integrated circuit being core with C8051F005 single-chip microcomputer is made up of supply voltage circuit, torque signal amplifying circuit, vibration signal bias voltage circuit, Keysheet module circuit, C8051F005 single-chip microcomputer and peripheral circuit, alarm module circuitry etc., integrated circuit is not only nursed one's health signal, and sound and light alarm is carried out to abnormal signal, remind staff to note the generation of abnormal conditions.

The state that host computer 14 creeps into through the man-machine interface on-line monitoring drill string of labview Software for Design, the method simultaneously applied based on self adaptation Dynamic Clonal Selection Algorithm optimization RBF neural carries out fault diagnosis to the data sample of Real-time Collection, if there is unusual condition to occur when drill string creeps into, system sends reports to the police and show fault diagnosis result in labview front panel.

Oil drilling drill string creeps into on-line fault diagnosis method, is the method for diagnosing faults optimizing RBF neural based on self adaptation Dynamic Clonal Selection Algorithm.Concrete method for diagnosing faults is as follows:

One, the central point in self adaptation Dynamic Clonal Selection Algorithm optimization RBF neural and width parameter, the connection weights between least square method determination hidden layer and output layer, concrete steps are as follows:

Step1: data normalization process

RBF neural requires that the data of input are positioned at interval [0,1], otherwise transfer function cannot process, and therefore the data of input is normalized, if the maximum value in node data is x _imax, minimum value is x _imin, i ∈ (1,2, m), normalization formula is as follows:

Step2: floating-point encoding, initialization of population and decomposition

1) floating-point encoding

G will be established for antibody population Ab (g)={ ab ₁, ab ₂, ab _nbe the N tuple of antibody, antibody ab _j=(z ₁ ^j, z2 ^j, z _l ^j), i-th gene z of a jth antibody _i ^jfor the floating number in [0,1] interval, z _i ^jwith x _i ^jmapping relations as follows:

x _i ^j＝a _i+z _i ^j*(b _i-a _i)

Claim z _i ^jx _i ^jcoding, x _i ^jz _i ^jdecoding.

2) initialization of population

The maximum number D of setting hidden layer node, the size L of population size, evolutionary generation G, e-learning speed ξ (ξ ∈ [0,1]).Hidden center c is initially the floating number in input amendment data area, the floating number on center width is initialized as [Isosorbide-5-Nitrae] interval.

3) population decomposition

Population decomposition is Ab (g) by antibody and the descending sort of antigen affinity size, and the Ab (g) after sequence is resolved into three populations according to the ratio of 1:6:3, is Ab respectively _m(g), Ab _n(g), Ab _{r (g)}.Ab _mg () is the highest population of affinity, Ab _ng () is the population of medium affinity, Ab _{r (g)}for population (Ab (g) the ∈ S that affinity is minimum ^n*L, Ab (g)=Ab _m(g) ∪ Ab _n(g) ∪ Ab _r(g), m+n+r=N).

Step3: the foundation of object function

Decoded data message is built RBF neural, and calculate the overall error of whole network, overall error is the smaller the better, and the quadratic sum therefore choosing desired output and actual both output difference is minimum as object function.When object function is minimum, RBF neural training terminates, and its expression formula is:

$E=\frac{1}{2}\underset{j=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{({\hat{y}}_{\mathrm{ji}}-y_{\mathrm{ij}})}^2$

Wherein, M is number of training, and N is output layer nodes, be M the N number of neuronic desired output of sample, y _jibe M the N number of neuronic real output value of sample.

Step4: self adaptation dynamic clonal selection operates

1) clone

Antagonist ab ∈ Ab _m(g) ∪ Ab _n(g), clone by following formula:

C＝∪ _iC _i＝I×ab,i＝1,2,···,n _c

Wherein, I is n _cdimension unit row vector, n _cfor the clone sizes of antibody ab, C _i=ab is a clone of ab.

2) evolutionary operator

If the sequence of antibody ab ∈ Ab (g) is rank (ab), population scale is Q, by affinity standardization, as shown in the formula:

$F\left(\mathrm{ab}\right)=\frac{Q-\mathrm{rank}\left(\mathrm{ab}\right)}{Q}$

By following formula, change of scale is carried out to F (ab)

T _sc(ab)＝η·exp(-ρ·F ²(ab))

Wherein, η and ρ is constant, general 0 < η < 0.5,2≤ρ≤10.Can find out, F (ab) larger T _sc(ab) less.

3) elite's clonal vaviation T _em()

High affinity antibody ab ∈ Ab _mg () is positioned at majorized function f (x _i) near peak value, therefore at its locally optimal solution of less space search.To ab ∈ Ab _mthe clone C of (g) _i∈ C, by Probability p _m∈ (0,1) carries out following mutation operation:

C _i'＝C _i+T _sc(ab)·(2δ-1)

Wherein, δ is the random number series vector in l dimension [0,1] interval; p _mget the small value, general p _m=1/l.

4) clone's intersection T _hm()

If ab ∈ is Ab _nthe clone C of (g) _i∈ C, random selecting antibody ab' ∈ Ab _m(g), carry out interlace operation by following formula:

C _i'＝C _i×(1-T _sc(ab)·δ)+ab'×(T _sc(ab)·δ)

Wherein, δ is the random number series vector in l dimension [0,1] interval.

5) Immune Clone Selection

To ab antibody through T _em() or T _hmclone group C'=∪ C after () operation _i' carry out affinity and assess { F (C') }=∪ _if (C _i'), if there is F (C _i')=maxF (C') >F (ab), then ab:=C _i', f (ab) :=f (Ci _'), thus upgrade antibody population, realize information exchange.

6) dynamic self-adapting algorithm

T _sc(ab) except relevant with antibody affinity, the same impact by η and ρ.For elite's antibody population Ab _mg () is mainly used in searching for locally optimal solution, therefore η gets the small value, and makes it be η ₁; Antibody population Ab _ng () is mainly used in exploitation global optimum solution space, therefore η takes large values, and makes it be η ₂.If optimal solution is constant within some generations, then η ₁to be multiplied by randomly or divided by 1.3, otherwise η ₁get back to its initial set value; In like manner, Ab is worked as _m(g) ∪ Ab _nwhen the average criterion functional value of () is substantially constant within some generations g, then η ₂to be multiplied by randomly or divided by 1.3, otherwise η ₂get back to its initial set value.Wherein, 1.3 is a good auto-adaptive parameter.

The hunting zone impact of ρ value on medium affinity antibody is comparatively large, on-the-fly modifies in the following way to ρ value:

Increase ρ value every some generations, return initial setting value after increasing to certain value, circulation is carried out successively, is shown below.

$\mathrm{\&rho;}(g+1)=\left\{\begin{array}{c}\mathrm{\&gamma;}\&CenterDot;\mathrm{\&rho;}\left(g\right),\mathrm{if\; rem}(g,g_D)=0\\ {\mathrm{\&rho;}}_0,\mathrm{if\; rem}(g,g_{D\max })=0\end{array}\right.$

Wherein, ρ ₀for initial value, γ >1 is Dynamic gene, the remainder that rem (x, y) is x/y, g _dfor interval algebraically, g _dmaxfor largest interval algebraically.

Step5: weight computing

After central point and width are determined, the RBF in RBF neural structure and hidden layer output valve are determined, what differ between hidden layer and output layer is the matrix product factor, utilize least square method to obtain the weights W of each element in matrix.

1) input all training datas, obtain the quadratic sum of square error:

$E\left(n\right)=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left(y_i\left(n\right)\widehat{-}{y}_i\left(n\right)\right)}^2$

Wherein, n represents iterations

2) following adjustment is made to connection weights:

$w_i(n+1)=w_i\left(n\right)+\mathrm{\&xi;}\&CenterDot;\frac{\&PartialD;E\left(n\right)}{\&PartialD;w_i}$

Wherein, ξ is the learning rate set in neutral net, ξ ∈ [0,1].

3) when E is less than specified threshold ε, calculate and stop, in this patent, get ε=0.0001.

Step6: repeat Step3 ~ Step5, neural network training, makes initial weight constantly revise evolution, obtain globally optimal solution.

Two, fault diagnosis

By the neutral net that the input of the data of Real-time Collection trains, draw oil drilling fault diagnosis result about drilling tool in drilling process.Wherein, neutral net input number of nodes is 2, is input TORQ and drill string acceleration calculation value respectively; Output node number is 3, and represent bit freezing, 010 with 001 respectively and represent drilling string broken off, 100 and represent drill stem washout, zero defects represents with 000.

The present invention be exactly in order to dangerous situation when preventing drill string to creep into generation and correctly cannot judge the problems such as the fault that drill string occurs, oil drilling drill string provided by the invention creeps into on-line monitoring system and method for diagnosing faults, its transplantability is good, and the state that energy Real-Time Monitoring drill string creeps into also can carry out fault diagnosis to abnormal conditions accurately and reliably, be convenient to drilling well staff adopt remedial measures in time, effectively prevent the generation of peril.

Claims (8)

1. an oil drilling drill string on-line monitoring system, it is characterized in that, comprise pcb board (7), host computer (14) and the torque sensor (1) be connected with drill string respectively and acceleration transducer (3), torque sensor (1) is connected with pcb board (7) again respectively with acceleration transducer (3), and the torque signal collected and vibration signal are flowed to pcb board (7), pretreated data are carried out communication by bluetooth module (12) and host computer (14) by pcb board (7).

2. oil drilling drill string on-line monitoring system according to claim 1, it is characterized in that, described torque sensor (1) is the strain-type torque sensor be made by the two-way foil gauge of two panels, foil gauge is attached on rectangle metal foil film, and the length of side of metal foil film is equal with drill string girth.

3. oil drilling drill string on-line monitoring system according to claim 1 and 2, it is characterized in that, the torque sensor of described torque sensor exports between ground wire (2) and torque sensor signal output line (13) and forms torque output signal, and is connected with pcb board respectively.

4. oil drilling drill string on-line monitoring system according to claim 1, it is characterized in that, the surface-mounted integrated circuit that described pcb board (7) is is core with C8051F005 single-chip microcomputer, comprises that 7.2V turns 5V potential circuit, 5V voltage turns 12V potential circuit, 5V voltage turns ± 5V potential circuit, torque signal amplifying circuit, vibration signal bias voltage circuit, Keysheet module circuit, C8051F005 single-chip microcomputer and peripheral circuit, alarm module circuitry.

5. oil drilling drill string on-line monitoring system according to claim 1, it is characterized in that, the Z axis signal output lead (4) of described acceleration transducer (3), Y-axis signal output lead (5), X-axis signal output lead (6) gather the radial direction of drill string, the vibration signal of vertical and horizontal respectively, and three signal output leads are connected with pcb board (7) respectively.

6. oil drilling drill string on-line monitoring system according to claim 1, it is characterized in that, the input of described bluetooth module (12) is provided with bluetooth module power line (8), bluetooth module ground wire (9), bluetooth module data downloading wire (10) and bluetooth module DOL Data Output Line (11), and is connected with pcb board respectively; Wherein bluetooth module power line (8) and bluetooth module ground wire (9) are the power line of bluetooth module, power to bluetooth module, the cooperating of bluetooth module data downloading wire (10) and bluetooth module DOL Data Output Line (11) makes bluetooth module can transfer data to host computer (14).

7. oil drilling drill string on-line monitoring system according to claim 1, is characterized in that, the state that described host computer (14) creeps into through the man-machine interface on-line monitoring drill string of labview Software for Design.

8., according to the oil drilling drill string on-line fault diagnosis method in claim 1-7 described in any one, it is characterized in that the method for diagnosing faults optimizing RBF neural based on self adaptation Dynamic Clonal Selection Algorithm, concrete method for diagnosing faults is as follows:

One, the central point in self adaptation Dynamic Clonal Selection Algorithm optimization RBF neural and width parameter, the connection weights between least square method determination hidden layer and output layer, concrete steps are as follows:

1: data normalization process

RBF neural requires that the data of input are positioned at interval [0,1], otherwise transfer function cannot process, and therefore the data of input is normalized, if the maximum value in node data is x _imax, minimum value is x _imin, i ∈ (1,2 ... m), normalization formula is as follows:

${\hat{x}}_i=\frac{x_i-x_{i\min }}{x_{i\max }-x_{i\min }};$

2: floating-point encoding, initialization of population and decomposition

1) floating-point encoding

G will be established for antibody population Ab (g)={ ab ₁, ab ₂... ab _nbe the N tuple of antibody, antibody ab _j=(z ₁ ^j, z ₂ ^j..., z _l ^j), i-th gene z of a jth antibody _i ^jfor the floating number in [0,1] interval, z _i ^jwith x _i ^jmapping relations as follows:

x _i ^j＝a _i+z _i ^j*(b _i-a _i)

Claim z _i ^jx _i ^jcoding, x _i ^jz _i ^jdecoding;

2) initialization of population and decomposition

The maximum number D of setting hidden layer node, the size L of population size, evolutionary generation G, e-learning speed ξ (ξ ∈ [0,1]); Hidden center c is initially the floating number in input amendment data area, the floating number on center width is initialized as [Isosorbide-5-Nitrae] interval;

Population decomposition is Ab (g) by antibody and the descending sort of antigen affinity size, and the Ab (g) after sequence is resolved into three populations according to the ratio of 1:6:3, is Ab respectively _m(g), Ab _n(g), Ab _r(g); Ab _mg () is the highest population of affinity, Ab _ng () is the population of medium affinity, Ab _{r (g)}for population (Ab (g) the ∈ S that affinity is minimum ^n*L, Ab (g)=Ab _m(g) ∪ Ab _n(g) ∪ Ab _r(g), m+n+r=N);

3: the foundation of object function

Decoded data message is built RBF neural, and calculate the overall error of whole network, overall error is the smaller the better, and the quadratic sum therefore choosing desired output and actual both output difference is minimum as object function; When object function is minimum, RBF neural training terminates, and its expression formula is:

$E=\frac{1}{2}\underset{j=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{({\hat{y}}_{\mathrm{ji}}-y_{\mathrm{ji}})}^2$

Wherein, M is number of training, and N is output layer nodes, be M the N number of neuronic desired output of sample, y _jibe M the N number of neuronic real output value of sample;

4: self adaptation dynamic clonal selection operates

1) clone

Antagonist ab ∈ Ab _m(g) ∪ Ab _n(g), clone by following formula:

C＝∪ _iC _i＝I×ab,i＝1,2,…,n _c

Wherein, I is n _cdimension unit row vector, n _cfor the clone sizes of antibody ab, C _i=ab is a clone of ab;

2) evolutionary operator

If the sequence of antibody ab ∈ Ab (g) is rank (ab), population scale is Q, by affinity standardization, as shown in the formula:

$F\left(\mathrm{ab}\right)=\frac{Q-\mathrm{rank}\left(\mathrm{ab}\right)}{Q}$

By following formula, change of scale is carried out to F (ab)

T _sc(ab)＝η·exp(-ρ·F ²(ab))

Wherein, η and ρ is constant, general 0 < η < 0.5,2≤ρ≤10; Can find out, F (ab) larger T _sc(ab) less;

3) elite's clonal vaviation T _em()

High affinity antibody ab ∈ Ab _mg () is positioned at majorized function f (x _i) near peak value, therefore at its locally optimal solution of less space search; To ab ∈ Ab _mthe clone C of (g) _i∈ C, by Probability p _m∈ (0,1) carries out following mutation operation:

C _i′＝C _i+T _sc(ab)·(2δ-1)

Wherein, δ is the random number series vector in l dimension [0,1] interval; p _mget the small value, general p _m=1/l;

4) clone's intersection T _hm()

If ab ∈ is Ab _nthe clone C of (g) _i∈ C, random selecting antibody ab ' ∈ Ab _m(g), carry out interlace operation by following formula:

C _i′＝C _i×(1-T _sc(ab)·δ)+ab′×(T _sc(ab)·δ)

Wherein, δ is the random number series vector in l dimension [0,1] interval;

5) Immune Clone Selection

To ab antibody through T _em() or T _hmclone group C '=∪ C after () operation _i' carry out affinity to assess { F (C ') }=∪ _if (C _i'), if there is F (C _i')=maxF (C ') >F (ab), then ab:=C _i', f (ab) :=f (C _{i '}), thus upgrade antibody population, realize information exchange;

6) dynamic self-adapting algorithm

T _sc(ab) except relevant with antibody affinity, the same impact by η and ρ; For elite's antibody population Ab _mg () is mainly used in searching for locally optimal solution, therefore η gets the small value, and makes it be η ₁; Antibody population Ab _ng () is mainly used in exploitation global optimum solution space, therefore η takes large values, and makes it be η ₂; If optimal solution is constant within some generations, then η ₁to be multiplied by randomly or divided by 1.3, otherwise η ₁get back to its initial set value; In like manner, Ab is worked as _m(g) ∪ Ab _nwhen the average criterion functional value of () is substantially constant within some generations g, then η ₂to be multiplied by randomly or divided by 1.3, otherwise η ₂get back to its initial set value; Wherein, 1.3 is a good auto-adaptive parameter;

The hunting zone impact of ρ value on medium affinity antibody is comparatively large, on-the-fly modifies in the following way to ρ value:

Increase ρ value every some generations, return initial setting value after increasing to certain value, circulation is carried out successively, is shown below;

$\mathrm{\&rho;}(g+1)=\left\{\begin{array}{c}\mathrm{\&gamma;}\&CenterDot;\mathrm{\&rho;}\left(g\right),\mathrm{if\; rem}(g,g_D)=0\\ {\mathrm{\&rho;}}_0,\mathrm{if\; rem}(g,g_{D\max })=0\end{array}\right.$

Wherein, ρ ₀for initial value, γ >1 is Dynamic gene, the remainder that rem (x, y) is x/y, g _dfor interval algebraically, g _dmaxfor largest interval algebraically;

5: weight computing

After central point and width are determined, the RBF in RBF neural structure and hidden layer output valve are determined, what differ between hidden layer and output layer is the matrix product factor, utilize least square method to obtain the weights W of each element in matrix;

1) input all training datas, obtain the quadratic sum of square error:

$E\left(n\right)=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\left(y_i\left(n\right)\widehat{-}{y}_i\left(n\right)\right)}^2$

Wherein, n represents iterations

2) following adjustment is made to connection weights:

$w_i(n+1)=w_i\left(n\right)+\mathrm{\&xi;}\&CenterDot;\frac{\&PartialD;E\left(n\right)}{{\&PartialD;w}_i}$

Wherein, ξ is the learning rate set in neutral net, ξ ∈ [0,1];

3) when E is less than specified threshold ε, calculate and stop, in this patent, get ε=0.0001;

6: repeat Step3 ~ Step5, neural network training, makes initial weight constantly revise evolution, obtain globally optimal solution;

Two, fault diagnosis

By the neutral net that the input of the data of Real-time Collection trains, draw oil drilling fault diagnosis result about drilling tool in drilling process; Wherein, neutral net input number of nodes is 2, is input TORQ and drill string acceleration calculation value respectively; Output node number is 3, and represent bit freezing, 010 with 001 respectively and represent drilling string broken off, 100 and represent drill stem washout, zero defects represents with 000.