CN102364501A - Method for reproducing two-dimensional defect of petroleum pipeline PSO-BP (Particle Swarm Optimization-Back-Propagation) neural network - Google Patents

Abstract

The invention aims at providing a method for reproducing a two-dimensional defect of a petroleum pipeline PSO-BP (Particle Swarm Optimization-Back-Propagation) neural network. Actually measured pipeline magnetic flux leakage data and pipeline defect data are used as experimental data of defect reconfiguration. The method comprises the steps of: with a magnetic flux leakage signal as input and a defect outline as outlet, setting a particle initial parameter, randomly initializing an initial position and an initial speed of each particle, calculating a particle fitness function numerical value , determining a past best value pbest of each particle and a global best value gbest of the whole particle swarm, updating the position and the speed of each particle, judging whether reaching the maximum iteration time or preset precision, if meeting the weight and the threshold of outputting a neutral network; and otherwise, re-comparing. The neutral network after the weight and the threshold are optimized by using a particle swarm algorithm is used for reproducing the two-dimensional defect of the pipeline and also reproducing a defect outline of the pipeline. According to the invention, the defect that the BP algorithm is easy to fall into a local minimum value can be effectively solved, and the convergence precision is improved, thus the defect of the pipeline is accurately reproduced.

Description

A kind of petroleum pipe line PSO-BP neural network two-dimensional defect replay method

Technical field

What the present invention relates to is the method that a kind of defect of pipeline detects.

Background technology

The fast development of As China's oil and gas industry, pipeline transportation have become the main mode of the land oil gas transportation of China.But along with reasons such as pipe growth in age, constructional deficiency, artificial destruction and corrosion, pipeline accident takes place again and again, not only causes great economic loss, the life security of going back serious environment pollution even jeopardizing producers.The Magnetic Flux Leakage Inspecting technology is most popular a kind of method during defect of pipeline detects, and it utilizes flaw detection principles such as ultrasound wave, leakage field, ray, under the situation that does not influence ordinary production; Through the walking of intelligent detector in pipeline, to the tube wall of oil and gas pipes or the defective of coating:, carry out online detection and analysis like distortion, damage, burn into perforation, tube wall weightlessness and variation in thickness etc.; To obtain to detect accurately and reliably data; Detect the various defectives that exist in the pipeline,, avoid blindly overhaul or keep in repair untimely for pipeline maintenance provides the science accurate data; Thereby save a large amount of maintenance costs, produce great economic benefit and social benefit.A major issue is the signal inverse problem in the leakage field Non-Destructive Testing, promptly from measuring-signal, confirms the shape of the length and width of defective, parameter such as dark or definite defective.Inverse problem is very complicated, and a method of generally using finding the solution inverse problem is to use alternative manner, but the calculated amount of this method is very big.

The BP neural network has the ability of approaching any Nonlinear Mapping through study, but the standard BP neural network exists and to be prone to be absorbed in local minimum, speed of convergence is slow and causes shortcoming such as oscillation effect.(Particle Swarm Optimization PSO), is a kind of optimizing algorithm based on iteration that is proposed first in nineteen ninety-five by Kennedy and Eberhart to particle swarm optimization algorithm.This algorithm is the simulation to the flock of birds social action; PSO algorithm and genetic algorithm are similar, are a kind of optimized Algorithm based on colony (population), and each particle is through carrying out information interaction with other particles; Adjust the evolution direction of oneself, and avoid being absorbed in local optimum.

Summary of the invention

The object of the present invention is to provide the PSO algorithm is combined with the BP neural network; Utilize particle swarm optimization algorithm to optimize the weights and the threshold value of BP neural network, solve the BP algorithm effectively and be prone to be absorbed in the shortcoming of local minimum and realize a kind of petroleum pipe line PSO-BP of quick convergent neural network two-dimensional defect replay method.

The objective of the invention is to realize like this:

A kind of petroleum pipe line PSO-BP of the present invention neural network two-dimensional defect replay method is characterized in that:

(1) will be from the measured value of actual defect of pipeline as the sample set that carries out network training, to adopt signal carry out de-noising after, as the experimental data of defective reconstruct, with the input of magnetic leakage signal as the PSO-BP neural network, defect profile is as output;

(2) initialization particle cluster algorithm parameter is provided with population scale, inertia weight w, study factor c ₁c ₂, iterations, confirm the particle dimension, the initial position of random initializtion particle and initial velocity;

(3), calculate particle fitness function value F, with the current fitness value of each particle optimal value p historical with it to each particle in the population _BestRelatively: if currency is superior to p _Best, then upgrade p _Best, otherwise keep p _BestConstant;

(4) with the p of each particle _BestGlobal optimum g with whole population _BestRelatively: if currency is superior to g _Best, then upgrade g _Best, otherwise keep g _BestConstant;

(5) upgrade particle position and speed;

(6) judge whether to reach maximum iteration time or preset precision: if reach maximum iteration time or preset the precision then weights and the threshold value of output nerve network, otherwise, forward step (3) to;

(7) neural network model that trains is used for the pipeline two-dimensional defect and reappears, magnetic leakage signal as input, is reappeared the defect of pipeline profile.

Advantage of the present invention is:

(1) because the BP neural network has the ability of approaching any Nonlinear Mapping through study, Application of Neural Network in defect of pipeline two dimension reconstruct problem, only need be utilized pipe leakage data and defective data, set up two-dimentional reconstruction model and carry out defect of pipeline reconstruct.

(2) particle swarm optimization algorithm is combined with the BP neural network; Utilize particle swarm optimization algorithm to optimize the weights and the threshold value of BP neural network; Set up PSO-BP neural network pipeline two dimension reconstruction model; Can solve the shortcoming that the BP algorithm is prone to be absorbed in local minimum effectively, improve convergence precision, thereby realized accurate reproduction defect of pipeline.

Description of drawings

Fig. 1 is a PSO-BP neural network pipeline two-dimensional defect reconstructing method process flow diagram of the present invention;

Fig. 2 is the comparison diagram of BP neural network pipeline two-dimensional defect predicted value of the present invention and actual value;

Fig. 3 is the comparison diagram of PSO-BP neural network pipeline two-dimensional defect predicted value of the present invention and actual value.

Embodiment

For example the present invention is done description in more detail below in conjunction with accompanying drawing:

In conjunction with Fig. 1～3:

1. particle swarm optimization algorithm

The proposition of particle swarm optimization algorithm receives the inspiration of flock of birds foraging behavior, and each particle position is represented a candidate solution of problem to be optimized in the algorithm.Each particle will move in solution space, and by speed its heading of decision and speed size, then through finding optimum solution by the generation search.

Suppose that in the search volume of a D dimension, each particle i has the position vector x of a D dimension _iWith velocity vector v _ix _iBe used to calculate the adaptive value of particle, the size reflection particle of adaptive value and the degree of approximation of optimum solution; And v _iThen be used for revising particle position.Particle changes the position through remembering 2 amounts, and the desired positions of the process that is particle in the process of seeking optimum solution (is designated as p _Best), another is that best particle position (is designated as g in the population _Best).

The following formula of PSO algorithm use that Kennedy and Eberhart propose upgrades particle state:

$\left\{\begin{array}{c}{v}_i(t+1)=w{v}_i+{\mathrm{<figure-callout\; id="1"\; label="c"\; filenames="DEST\_PATH\_HDA0000110963650000031.png"\; state="\{\{state\}\}">c</figure-callout>}}_1\mathrm{rand}(p_i\left(t\right)-x_i\left(t\right))+{\mathrm{<figure-callout\; id="2"\; label="c"\; filenames="DEST\_PATH\_HDA0000110963650000021.png"\; state="\{\{state\}\}">c</figure-callout>}}_2\mathrm{rand}(g\left(t\right)-x_i\left(t\right))\\ x_i(t+1)=x_i\left(t\right)+v_i(t+1)\end{array}\right.$

(1)

In the formula: what t represented is the t time iteration; p _i(t) the desired positions p of particle i after the t time iteration of expression _Best, the desired positions g of whole colony after the t time iteration of g (t) expression _BestRand is the random number between [0,1] independently; c ₁And c ₂It is the study factor.For avoiding algorithm convergence too fast, also need introduce a threshold value v _Max, be used for guaranteeing v _iBe no more than interval [v _Max, v _Max]; W is an inertia weight.

Study factor c ₁, c ₂Be used for controlling the relative influence between memory and companion's the memory of particle self, just particle flies to individual extreme value p _BestWith global extremum g _BestThe acceleration weight.Generally speaking, get c ₁=c ₂=2.

Inertia weight w is used for controlling particle, and speed was to the influence of present speed in the past, and it directly influences the overall situation and the local search ability of particle, if the inertia weight parameter is too big, population possibly missed optimum solution, causes algorithm not restrained, and perhaps can not converge to optimum solution.The value that w are found in a large amount of experiments between [0.9,0.2] on average algorithm have reasonable performance, and the linear minimizing of the value of w is than good with fixing value.Generally speaking, being provided with as follows w:

$w=w_{\max }-(w_{\max }-w_{\min })\&times;\frac{t}{t_{\max }}---\left(2\right)$

In the formula: w _MaxBe maximum inertia weight, w _MinBe minimum inertia weight, t is the current iteration number of times, t _MaxBe the algorithm maximum iteration time.

2. particle group optimizing BP neural network algorithm

The BP neural network structure is simple, and algorithm is ripe, has advantages such as accurate optimizing.But standard BP algorithm is prone to be absorbed in local minimum, speed of convergence is slow.Particle swarm optimization algorithm is a kind of optimized Algorithm based on iteration pattern, has very strong search macro ability, and simple, general-purpose, strong robustness.Therefore utilize particle cluster algorithm to optimize the weights and the threshold value of neural network, confirm neural network best weight value and threshold value, avoid the BP algorithm to be easy to be absorbed in local minimum and the slow shortcoming of speed of convergence.

Algorithm design

(1) particle dimension (Particle Dimension)

Weights or threshold value in all corresponding neural network of dimension component of each particle position vector X, promptly the dimension of particle equals the weights and the threshold value number of neural network.Therefore the search volume dimension of particle is:

D＝R×L+L×S+L+S (3)

Wherein: R is a neural network input node number, and L the number of hidden nodes, S are the output node number.

(2) fitness function

In order to make the error between neural network output and the objective function minimum, fitness function is defined as the square error index of neural network:

$F=\frac{1}{\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{s}{\mathrm{\&Sigma;}}}{(y_{i,j}-y_{i,j}^{\&prime;})}^2+e}---\left(4\right)$

In the formula: n is the training sample sum, and s is the output node number, y _{I, j}Be reality output, y ' _{I, j}Be desired output, e is zero an added decimal for avoiding denominator.Training error is more little, and the fitness function value of particle is high more, and the final result of PSO algorithm is the best weight value and the threshold value of neural network.

Embodiment:

The process of the magnetic leakage signal prediction defect geometry parameter that produces according to defective comes down to a process of setting up the mapping relations of magnetic leakage signal and defect geometry parameter.

(1) will be from the measured value of actual defect of pipeline as the sample set that carries out network training, to adopt signal carry out pre-service work such as de-noising after, as the experimental data of defective reconstruct.With the input of magnetic leakage signal as the PSO-BP neural network, defect profile (length and the degree of depth) is as output.Sample data is totally 90 groups of data, utilize preceding 80 groups as training data, the back 10 groups as test data, every group of data sampling point is 120.

(2) initialization particle cluster algorithm parameter is provided with population scale, inertia weight w, study factor c ₁c ₂, iterations, confirm the particle dimension, the initial position of random initializtion particle and initial velocity, wherein inertia weight w adopts formula (2) to calculate;

(3), calculate particle fitness function value F, the current fitness value of each particle optimal value p historical with it according to formula (4) to each particle in the population _BestRelatively, if currency is superior to p _Best, then upgrade p _Best, otherwise keep p _BestConstant;

(4) with the p of each particle _BestGlobal optimum g with whole population _BestRelatively, if currency is superior to g _Best, then upgrade g _Best, otherwise keep g _BestConstant.

(5) upgrade particle position and speed according to formula (1).

(6) judge whether to reach maximum iteration time or preset precision, if satisfy condition the then weights and the threshold value of output nerve network; Otherwise, forwarded for the 3rd step to.

(7) neural network model that trains is used for the pipeline two-dimensional defect and reappears, magnetic leakage signal as input, is reappeared the defect of pipeline profile.

Fig. 2 is the defect of pipeline predicted value that obtains after through the BP neural metwork trainings of 4 groups of samples and the comparison diagram of actual value.Dotted line is an actual value among the figure, and solid line is a predicted value.The X axle is a sampling number; The Y axle is represented depth of defect, inches (inch).Provide 4 groups of test data reconstruction result figure.Fig. 3 is the defect of pipeline predicted value that obtains after through the PSO-BP neural metwork trainings of 4 groups of samples and the comparison diagram of actual value.Dotted line is an actual value among the figure, and solid line is a predicted value.The X axle is a sampling number; The Y axle is represented depth of defect, inches (inch).Can find out that in conjunction with Fig. 2, Fig. 3 the reconstruction result of employing particle group optimizing neural net method obviously is superior to the reconstruction result of BP neural network, can effectively improve the defect of pipeline fidelity based on the two-dimentional reconstructing method of the pipeline of PSO-BP neural network.

Claims (1)

1. petroleum pipe line PSO-BP neural network two-dimensional defect replay method is characterized in that:

(1) will be from the measured value of actual defect of pipeline as the sample set that carries out network training, to adopt signal carry out de-noising after, as the experimental data of defective reconstruct, with the input of magnetic leakage signal as the PSO-BP neural network, defect profile is as output;

(2) initialization particle cluster algorithm parameter is provided with population scale, inertia weight w, study factor c ₁c ₂, iterations, confirm the particle dimension, the initial position of random initializtion particle and initial velocity;

(3), calculate particle fitness function value F, with the current fitness value of each particle optimal value p historical with it to each particle in the population _BestRelatively: if currency is superior to p _Best, then upgrade p _Best, otherwise keep p _BestConstant;

(4) with the p of each particle _BestGlobal optimum g with whole population _BestRelatively: if currency is superior to g _Best, then upgrade g _Best, otherwise keep g _BestConstant;

(5) upgrade particle position and speed;

(6) judge whether to reach maximum iteration time or preset precision: if reach maximum iteration time or preset the precision then weights and the threshold value of output nerve network, otherwise, forward step (3) to;

(7) neural network model that trains is used for the pipeline two-dimensional defect and reappears, magnetic leakage signal as input, is reappeared the defect of pipeline profile.

Images