CN115076620A - Buried oil pipeline leakage detection method based on improved UPEMD and DTWSVM - Google Patents

Abstract

The invention relates to the technical field of pipeline leakage detection and sound wave signal processing, in particular to a buried oil pipeline leakage detection method based on improved UPEMD and DTWSVM, which comprises the following steps: step 1, collecting sound wave signals of a pipeline under different working conditions by using sound pressure sensors arranged at two ends of the pipeline; step 2, decomposing the collected sound wave signals by adopting a UPEMD algorithm until the residual signals can not be decomposed; step 3, mutual information values of each mode function IMF, the leakage signal and the normal working condition signal are respectively calculated, the similarity coefficient of each IMF is calculated according to the mutual information values, and then an IMF reconstruction signal containing much leakage information is selected according to the similarity coefficient; step 4, constructing a model DTWSVM according to the deep learning DL and the twin Gemini support vector machine TWSVM; and 5, training the DTWSVM by adopting a training sample, determining the optimal value of each parameter, and obtaining a classification result by using a test sample to determine the working condition of the buried oil pipeline. The invention can effectively remove noise in signals and keep leakage information.

Description

Leak detection method of buried oil pipeline based on improved UPEMD and DTWSVM

technical field

The present invention relates to the technical field of pipeline leak detection and acoustic signal processing, in particular to a buried oil pipeline leak detection method based on improved UPEMD and DTWSVM.

Background technique

As a convenient and economical transportation method, pipelines have been widely used in various fields. Because the environment in which the buried pipeline is located is relatively complex, and the pipeline is easily corroded by rainwater, etc., the pipeline is prone to leakage. Leakage of oil pipelines not only causes waste of resources, but also has a huge impact on soil and vegetation. Therefore, leak detection of buried pipelines is of great significance.

The acoustic wave method has many advantages such as low false alarm rate, low false alarm rate, long propagation distance, good sensitivity, etc. Therefore, the use of acoustic wave method for pipeline leak detection has been widely used. Since the sound wave is easily affected by noise in the process of propagation, it is very important to de-noise the sound wave signal. UPEMD (uniform phase empirical mode decomposition, uniform phase empirical mode decomposition) is a signal processing method optimized and improved on the basis of EMD (empirical mode decomposition, empirical mode decomposition). It uses sinusoidal signals with uniform phase distribution as the Auxiliary signal, and by increasing the phase number of the sine wave to achieve better suppression of modal aliasing. However, after UPEMD decomposes the signal into several IMFs (intrinsic modal function, intrinsic modal function), there is no suitable screening method. When reconstructing the signal, if the selected reconstructed IMF contains less useful information, the error will increase, and the analysis effect of the signal will not reach the expectation.

SUMMARY OF THE INVENTION

In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a buried oil pipeline leak detection method based on improved UPEMD and DTWSVM, which solves the large error and signal analysis effect of the existing buried oil pipeline leak detection technology. Bad technical issues. The invention can improve the leak detection capability of the pipeline and accurately judge the working condition of the pipeline.

In order to achieve the above-mentioned purpose, the main technical scheme adopted in the present invention includes:

The embodiments of the present invention provide a leak detection method for buried oil pipelines based on improved UPEMD and DTWSVM. The method includes the following steps:

Step 1, using a sound pressure sensor to obtain the sound wave signal data of the oil pipeline under normal working conditions and leakage;

Step 2, using the UPEMD algorithm to decompose the acoustic wave signal data collected in step 1, until the residual cannot be decomposed, and several IMFs are obtained;

Step 3: Calculate the mutual information value of each IMF obtained in step 2 and the acoustic wave signal when the pipeline leaks and during normal operation, calculate the similarity coefficient of each IMF according to the obtained mutual information value, and select the one containing more leakage information through the similarity coefficient. IMF performs signal reconstruction;

Step 4, according to the deep learning DL and the twin support vector machine TWSVM, establish the network model DTWSVM;

Step 5: According to the reconstructed signal obtained in Step 3, the training samples and test samples of the DTWSVM model are established; the training samples are input into the DTWSVM model, and the parameters in the DTWSVM model are adjusted through supervised learning and unsupervised learning to obtain a trained DTWSVM. model; and then input the test samples into the trained DTWSVM model to judge the working conditions of the buried oil pipeline.

Further, adopt UPEMD algorithm in described step 2 to the concrete steps of acoustic wave signal decomposition as follows:

Step 2.1, calculate the cycle period T _c and the cycle number n _c of the UPEMD algorithm;

n _c = log ₂ (n) (1)

T _c =2 ^m , m = 1:n _c (2)

In the formula, n represents the total length of the signal data, and the value of the mask frequency is equal to the reciprocal of the cycle period, that is, f _c =1/T _c , and m is the reciprocal of the number of cycles.

Step 2.2. Set the number of phases n _p and the amplitude ε, where n _p ∈ N, n _p > 1, and the number of n _p must be an integer power of 2, when the value of n _p is 2, UPEMD is the mask EMD;

Step 2.3. Let the original signal x(t)=r ₀ (t), and at the same time, construct the masking signal, namely:

w(t; ε _m ; f _c ; θ _k )=ε _m ·cos(2π·f _c ·t+θ _k ) (3)

Among them, ε _m =ε×r _m-1 (t), r _m-1 (t) is the standard deviation, and the phase θ _k is evenly divided into n _p parts in the range of [0,2π], so

θ _k = 2π(k-1)/n _p , k=1:n _p (4)

Step 2.4, the expression of disturbance signal y _k (t) is:

y _k (t)=x(t)+ε _m ·cos(2π·f _c ·t+θ _k ) (5)

Step 2.5, decompose the disturbance signal y _k (t) with EMD, and define the i-th IMF component obtained by the decomposition as E _i ( ); therefore, the first component obtained by EMD decomposition is:

c _m,k (t)=E ₁ (x(t)+w(t; ε _m ; f _c ; θ _k )), k=1,2...n _p ,m=1:n _c (6 )

Step 2.6, after subtracting the masking signal w(t; ε _m ; f _c ; θ _k ) from the obtained cm, _k (t), and then performing integrated averaging to obtain IMF1, namely:

Step 2.7: Use the remainder calculated in step 2.6 as a new disturbance signal to be decomposed, and repeat steps 2.1 to 2.6 until all IMFs are decomposed, and then stop the cycle.

Further, the concrete method of described step 3 is:

Calculate the mutual information value of each IMF and the acoustic signal when the pipeline leaks and when it is working normally:

where X and Y are two random signals, p(x, y) represents the joint distribution of X and Y, and the marginal probability density functions of X and Y are p(x) and p(y) respectively;

Define the mutual information between the IMF and the leaked signal as l _i , define the mutual information between the IMF and the signal under normal working conditions as h _i , calculate the similarity coefficient β _i according to the obtained mutual information value, β _i = _{li -hi} ; i=1,2...,k;

According to the similarity coefficient, the IMF reconstruction signal with more leakage information is selected, and the IMF with the larger similarity coefficient contains more leakage information.

Further, the concrete method of described step 4 is:

On the basis of deep learning and TWSVM, a new network model DTWSVM is established by analyzing the influence of various network structures and network layers on the detection accuracy, and the optimal value of each parameter in the model is determined through model training;

The DTWSVM model is a three-layer network model including an input layer, a hidden layer and an output layer. Three TWSVMs are set in the hidden layer to extract the features of the input data, and a main TWSVM is set in the output layer to perform the pipeline conditions. to judge;

The TWSVM in the hidden layer is divided into two categories. When the input data is linear data, the TWSVM is linear, generates three pairs of hyperplanes, maps the original data to a 6-dimensional space, and records the 6-dimensional position information of the original data; input When the data is nonlinear, TWSVM is nonlinear, and the kernel function data is introduced to map the original data to the 6-dimensional space;

The original data T=[(x ₁ , y ₁ ),...,(x _m+k ,y _m+k )] is input to the input layer of the model, and the three TWSVMs in the hidden layer perform feature extraction of the data to generate a feature data set:

X=f(x)=(f(x) _1,+ ,f(x) _1,- ,f(x) _2,+ ,f(x) _2,- ,f(x) _3,+ ,f( x) _3,- ) ^T (27)

f(x) _i,j =(wi _,j ·x)+b _i,j , i=1,2,3,j=+,- (28)

In the formula, b _i,j and w _i,j are parameters in the DTWSVM model, and specific values are obtained after training;

In the case of nonlinearity, the TWSVM in the hidden layer is mapped by the kernel function:

f(x) _i,j =K(x ^T ,C ^T )z _i,j +b _i,j , i=1,2,3,j=+,- (29)

In the formula, K(x ^T , C ^T ) represents the kernel function, and zi _{, j} are the parameters in the DTWSVM model;

The parameters in the TWSVM are determined by the methods of cross-validation and grid search;

The main TWSVM of the output layer is divided into two categories, linear TWSVM and nonlinear TWSVM; the category is consistent with the TESVM in the hidden layer, and the main TWSVM is used to judge the working condition of the pipeline;

The discriminant formulas of the main TWSVM in the linear and nonlinear states are:

In the formula, K(x ^T , C ^T ) represents the kernel function.

Further, in the step 5, the training of the model adjusts various parameters in the network model through bottom-up unsupervised learning and top-down supervised learning; wherein, unsupervised learning is to adjust each parameter of the model one by one. One layer is trained, and each parameter of each layer is adjusted. During training, the output of the previous layer is used as the input of the next layer until all layers are trained; supervised learning is based on unsupervised learning. The parameters of each layer are adjusted so that the parameters of each layer reach the optimal value, thereby minimizing the error.

The beneficial effects of the present invention are as follows: in the buried oil pipeline leak detection method based on the improved UPEMD and DTWSVM of the present invention, the UPEMD algorithm based on mutual information optimization is used to denoise the collected acoustic signal, and the UPEMD decomposes the signal into several IMFs. , the similarity coefficient of each IMF is obtained according to the mutual information, and the IMF with more leakage information is selected according to the similarity coefficient to reconstruct the signal. The reconstructed signal can accurately reflect the leakage information of the pipeline and contains very little noise.

According to DL (deep learning, deep learning) and TWSVM (twin support vector machine, twin support vector machine), the present invention establishes a DTWSVM (deep twin support vector machine, deep twin support vector machine) model, the DTWSVM model combines the depth The advantages of learning and TWSVM can accurately judge the working conditions of the pipeline.

Description of drawings

Fig. 1 is the flow chart of the buried oil pipeline leak detection method based on improved UPEMD and DTWSVM of the present invention;

Fig. 2 is the flow chart that the improved UPEMD provided by the present invention denoises the signal;

Fig. 3 is the structural diagram of the DTWSVM model provided by the present invention;

Fig. 4 is the detection principle diagram provided by the present invention;

Fig. 5 is a site map provided by the present invention;

Fig. 6 is the original signal (that is, the pipeline leakage acoustic wave signal before denoising) that the sensor 1 provided by the present invention collects at the first station of the pipeline;

Fig. 7 is the partial IMF obtained by decomposition provided by the present invention and its corresponding spectrogram;

8 is a schematic diagram of similar coefficient numerical values of each IMF provided by the present invention;

9 is a graph of the leaked acoustic wave signal comparing the reconstructed signal of the present invention with the signal denoised by EMD and UPEMD; wherein, FIG. 9(a) is a graph of the leaked acoustic wave signal denoised by EMD; Fig. 9(b) is a graph of the leaked acoustic wave signal after denoising by UPEMD; Fig. 9(c) is a graph of the leaked acoustic wave signal after denoising by the method of the present invention.

Detailed ways

In order to better explain the present invention and facilitate understanding, the present invention will be described in detail below with reference to the accompanying drawings and through specific embodiments.

As shown in Figure 1, the present embodiment provides a buried oil pipeline leak detection method based on improved UPEMD and DTWSVM, comprising the following steps:

Step 1, use the sound pressure sensor to collect the sound wave signal of the oil pipeline during normal operation and leakage.

In step 2, the UPEMD algorithm is used to decompose the collected acoustic signal, and the decomposition is stopped until the residual can no longer be decomposed.

EMD is a commonly used signal processing method, but its existence of modal aliasing will cause the reconstructed signal to be mixed with more noise and other discontinuous signals, resulting in errors. UPEMD is a signal processing method optimized and improved on the basis of EMD. It uses sinusoidal signals with uniform phase distribution as auxiliary signals, and increases the phase number of sinusoidal waves to better suppress modal aliasing. The specific steps of UPEMD are as follows:

(1) First, calculate the cycle period T _c and the cycle number n _c of the UPEMD algorithm.

n _c = log ₂ (n) (1)

T _c =2 ^m , m = 1:n _c (2)

In the formula, n represents the total length of the signal data, the value of the mask frequency is equal to the reciprocal of the cycle period, that is, f _c =1/T _c , and m is the reciprocal of the number of cycles.

(2) Set the number of phases n _p and the amplitude ε. Among them, n _p ∈ N, n _p >1, and the number of n _p must be an integer power of 2. When the value of n _p is 2, UPEMD is the mask EMD.

(3) Let the original signal x(t)=r ₀ (t), and at the same time, construct the masking signal, that is

w(t; ε _m ; f _c ; θ _k )=ε _m ·cos(2π·f _c ·t+θ _k ) (3)

Among them, ε _m =ε×r _m-1 (t), r _m-1 (t) is the standard deviation, and the phase θ _k is evenly divided into n _p parts in the range of [0,2π], so

θ _k = 2π(k-1)/n _p , k=1:n _p (4)

(4) The expression of the disturbance signal y _k (t) can be obtained from the above:

y _k (t)=x(t)+ε _m ·cos(2π·f _c ·t+θ _k ) (5)

(5) The disturbance signal is decomposed by EMD, and the i-th IMF component obtained by decomposing is defined as E _i (·). Therefore, the first component obtained from the EMD decomposition is:

c _m,k (t)=E ₁ (x(t)+w(t; ε _m ; f _c ; θ _k )), k=1,2...n _p ,m=1:n _c (6 )

At this time, after subtracting w(t; ε _m ; f _c ; θ _k ) from the c _m,k (t) obtained by formula (5), and then performing an integrated average on it, IMF1 can be obtained, that is,

(6) Take the calculated remainder as a new disturbance signal to be decomposed, and repeat steps (1)-(5) until all IMFs are decomposed and stop the cycle.

In step 3, the mutual information is used to optimize the UPEMD algorithm, the similarity coefficient is calculated according to the mutual information value of each IMF, and the appropriate IMF is selected for signal reconstruction according to the similarity coefficient. The flow chart of the improved UPEMD algorithm is shown in Figure 2.

Mutual information can measure the degree of correlation between two variables or systems. The greater the mutual information value, the greater the degree of dependence between two variables or systems, and the more common information they contain. Definition X and Y are two random signals, their joint distribution is represented by p(x,y), and the marginal probability density functions of X and Y are p(x) and p(y) respectively. The mutual information can be obtained by the following formula:

When denoising the signal collected by the sensors at the first and last stations of the pipeline, the UPEMD algorithm is used to decompose the signal first, and the leakage information contained in each IMF obtained is somewhat different. The UPEMD algorithm only decomposes the signal, and does not provide a screening method for IMF. In general, all IMFs are reconstructed or the IMFs located in a specific frequency band are selected for reconstruction through frequency domain analysis. However, the signal obtained by this selection method still contains a certain amount of noise, which cannot effectively reduce the error. Therefore, when reconstructing the signal, it is necessary to determine a method that can filter out IMFs that contain a lot of leakage information, and reduce the error caused by noise. Mutual information can reflect the degree of correlation between two variables. Therefore, the present invention adopts mutual information to optimize the UPEMD algorithm, and the concrete steps are as follows:

1) Decompose the signals collected by the sensors of the first and last stations using the UPEMD algorithm to obtain k IMFs;

2) Calculate the mutual information value between each IMF and the normal operating condition signal, and define it as h _i . The more leaked information contained in the IMF, the smaller the value of _hi , and vice versa;

3) Calculate the mutual information value between each IMF and the leaked signal, and define it as l _i . When the IMF contains more _leaked information, the value of li is larger, and vice versa;

4) Define the similarity coefficient β _i =li _-hi , _i =1,2...,k. Then, the similarity coefficient corresponding to each IMF is calculated according to the mutual information values obtained in steps 2) and 3). The more _leaked information is contained in the IMF, the larger the li and the smaller the _hi , and the larger the similarity coefficient.

5) Screen the IMF according to the similarity coefficient obtained in step 4). The larger the value of the similarity coefficient β _i , the more leaked information the IMF contains. Therefore, IMFs with large similarity coefficients are selected for reconstruction to obtain signals with less noise.

Step 4, according to the deep learning DL and the twin support vector machine TWSVM, establish the DTWSVM model, and the model structure is shown in Figure 3.

The twin support vector machine TWSVM draws on the classification idea of support vector machine SVM, and classifies the samples by establishing two hyperplanes. In SVM, positive and negative samples are classified using the same hyperplane, while TWSVM establishes a hyperplane for positive and negative samples respectively for classification. A hyperplane corresponds to a quadratic programming problem, so TWSVM needs to solve two quadratic programming problems, which greatly improves its accuracy. In addition, SVM solves a quadratic programming problem at one time, while TWSVM divides a problem into two small quadratic programming problems, which improves the training speed and detection ability of the model.

TWSVM realizes the classification of data by generating two hyperplanes, and defines the data set as:

T=[(x ₁ ,y ₁ ),...,(x _m+k ,y _m+k )] (9)

In the formula, x _l ∈R ⁿ , l=1,2,…,m+k, ys=+1, _s =1,2,…,m, yq=-1, _q =m+1,m +2,...,m+k, define A=(x ₁ ,x ₂ ,...,x _m ) ^T , B=(x _m+1 ,x _m+2 ,...,x _m+k ) ^T .

It can be obtained from the above analysis that there are two quadratic programming problems that TWSVM needs to solve. A pair of hyperplanes (w ₊ ·x)+b ₊ =0 and (w ₋ ·x)+b ₋ =0 can be obtained according to equations (10)-(11).

In the formula, c ₁ and c ₂ are the penalty parameters that control the size of ξ, e ₊ and e _- are two column vectors whose constituent elements are both 1, ξ ₊ and ξ _- are slack variables that control the tolerance to noise, and e ₊ ,ξ ₊ ∈R ^m , e _- ,ξ _- ∈R ^k . From the above two equations, the constraints of one objective function are determined by the mode of the other objective function, and the two hyperplanes constrain each other. In the linear state, the Lagrangian dual problem of TWSVM is:

Wherein, H=[A e ₊ ], G=[B e ₋ ].

Based on the above, it can be obtained that under linear conditions, in order to have an optimal solution to the linear programming problem, it can be obtained from the KKT condition:

At this time, when there is a new sample input to be determined, the linear TWSVM can determine the label to which the sample belongs according to the following formula.

In the formula, |·| indicates that the absolute value operation is performed.

When the data is linearly inseparable in the low-dimensional space, SVM uses the kernel function to map it to the high-dimensional space, and then classify the data. Similar to the processing method of SVM, TWSVM also uses the kernel function to map the data, and then uses the method under linear conditions to classify the data. After introducing the kernel function, the two hyperplanes of TWSVM can be expressed as:

The two quadratic programming problems of nonlinear TWSVM are as follows:

In the nonlinear case, define:

At this time, the Langrangian dual problem corresponding to TWSVM is:

Through the above analysis, in order to have an optimal solution to the nonlinear programming problem, it can be obtained from the KKT condition

At this time, when there is a new sample input to be determined, TWSVM can determine the label to which the sample belongs according to the following formula.

where |·| is the vertical distance from the sample point to the hyperplane.

Compared with SVM (support vector machine, support vector machine), TWSVM has faster training speed, higher classification accuracy, and good generalization ability. Deep learning DL does not need to manually extract features, but automatically filters and extracts features from data. In addition, compared with traditional neural networks, deep learning has better learning ability, adaptability and portability. On the basis of deep learning and TWSVM, the present invention combines the advantages of both to form a new network model DTWSVM.

Step 5, use the training samples to train the DTWSVM, determine the optimal values of various parameters, and then use the test samples to obtain the classification results to determine the working conditions of the buried oil pipeline.

DTWSVM is a network model with input layer, hidden layer and output layer. The input layer inputs the data measured by the sensors at both ends of the pipeline. In the hidden layer of the model, three TWSVMs are set. The only difference between these three TWSVMs is the difference in parameters, which are mainly used for feature extraction. Three TWSVMs generate 6 hyperplanes, which can map the original data into a 6-dimensional space. A TWSVM is set in the output layer for classification. This model not only gives full play to the advantages of DL and TWSVM, but also overcomes the shortcomings of shallow models and has good accuracy. The model of DTWSVM is shown in Figure 3.

c _i1 , c _i2 , i=1, 2, and 3 are the parameters of the three TWSVMs in the hidden layer, respectively, and the present invention adopts the methods of cross-validation and grid search to determine their optimal values. When the pipeline leaks, the data collected by the sensor is basically nonlinear data, and the RBF kernel function (radial basis function, radial basis function) has good nonlinear mapping ability and generalization ability. The invention adopts the RBF kernel function K(x _i , x _j )=exp(-λ||x _i -x _j || ² ).

In the linear case, according to equations (10)-(11), each TWSVM in the hidden layer generates a pair of hyperplanes.

The 6 hyperplanes generated by TWSVM in the hidden layer can project the original data x collected by the sensor into the 6-dimensional space, obtain the 6-dimensional position information of the original data, and generate a new data set X.

X=f(x)=(f(x) _1,+ ,f(x) _1,- ,f(x) _2,+ ,f(x) _2,- ,f(x) _3,+ ,f( x) _3,- ) ^T (27)

f(x) _i,j =(wi _,j ·x)+b _i,j , i=1,2,3,j=+,- (28)

In the formula, b _i,j and w _i,j are parameters in the DTWSVM model, and specific values are obtained after training.

In the case of nonlinearity, each TWSVM in the hidden layer generates two surfaces by formulas (18)-(19), and projects the original data x collected by the sensor into the generated 6-dimensional space to generate a new data set X.

f(x) _i,j =K(x ^T ,C ^T )z _i,j +b _i,j , i=1,2,3,j=+,- (29)

In the formula, K(x ^T , C ^T ) represents the kernel function, and zi _{, j} are the parameters in the DTWSVM model.

According to the above analysis, the data set X is the data features extracted from the original data x by the three TWSVMs in the hidden layer, and X will be used as the input of the output layer to judge the working conditions of the pipeline.

The training of the model can be divided into two parts: bottom-up unsupervised learning and top-down supervised learning. Unsupervised learning is to train each layer of the model one by one, and adjust each parameter of each layer. During training, the output of the previous layer is used as the input of the next layer until all layers are trained. Supervised learning is to adjust the parameters of each layer of the model on the basis of unsupervised learning, so that the parameters of each layer reach the optimal value, thereby minimizing the error. The basic steps of DTWSVM are as follows:

<1> In the linear case, the parameters c _i1 , c _i2 , i=1, 2, and 3 of the three TWSVMs in the hidden layer are substituted into equations (12)-(13) for operation, and then through equation (14) )-(15) to get the parameters w _i,+ ,b _i,+ ,wi _,- ,b _i,- . In the case of nonlinearity, the parameters c _i1 , c _i2 are substituted into equations (21)-(22), and the parameters zi _,+ ,b _i,+ ,z _i,- are obtained through equations (23)-(24) ,b _i,- .

<2> In the case of linearity, according to the parameters obtained in step <1> and formula (28), convert the raw data x _l collected by the sensor into f(x _l ). In the case of non-linearity, formula (29) is used to transform the original data. The transformed data samples (f(x _l ), y _l ) are used as the input of the main TWSVM at the output layer.

<3> In the linear case, the optimal values of the parameters c ₁ , c ₂ of the main TWSVM are obtained by cross-validation and grid search methods. According to formulas (12)-(15), the values of parameters w ₊ , b ₊ , w _- , b _- can be obtained. In the case of nonlinearity, according to equations (20)-(23), the values of parameters z ₊ , b ₊ , z _- , b _- are obtained.

<4> In the case of linearity, substitute w ₊ , b ₊ , w _- , b _- obtained in step <3> into formula (16) to judge the pipeline condition. In the case of nonlinearity, z ₊ , b ₊ , z _- , b _- are substituted into equation (25) to judge the pipeline condition.

Example

The present invention adopts the aviation oil pipeline of a certain aviation oil pipeline company to carry out the leakage experiment. The aviation oil pipeline is 10km in length and 200mm in diameter. The pipeline contains No. 3 aviation kerosene, and the pressures at the beginning and end of the pipeline are 2.3kPa and 0.53kPa respectively. There is a ball valve with a diameter of 8mm and 5mm at 1km, 4km and 7km of the pipeline to simulate the generation of leakage. A PCB106B sound pressure sensor is installed at the head and end of the pipeline to collect leakage signals. The data acquisition device adopts the NI-DAQ9181 data acquisition device, which is used in conjunction with the LABVIEW software to collect data. Its schematic diagram and field diagram are shown in Figure 4 and Figure 5.

When all valves are closed, the acoustic signal in the pipeline is collected, that is, the acoustic signal under normal working conditions. Open different leakage valves at 1km, 3km and 7km of the pipeline to simulate leakage, that is, the sound wave signal when leakage occurs.

Taking the signal collected by sensor 1 as an example, the denoising experiment is carried out. Open the 8mm valve 3km away from the pipeline to simulate leakage in the pipeline. The original signal collected by sensor 1 at the first station of the pipeline is shown in Figure 6.

It can be seen from Figure 6 that the original signal contains more noise, and the signal fluctuates greatly, so it is impossible to judge whether the pipeline is leaking. Now the UPEMD algorithm based on mutual information optimization is used to denoise the original signal. First, the UPEMD algorithm is used to decompose the original signal, and the original signal is decomposed into 12 IMFs with different scales and a residual term. Figure 7 shows some of the IMFs and their spectrograms obtained by decomposition.

The similarity coefficient can reflect the amount of leaked information contained in each IMF, so the IMF with more leaked information can be selected according to the similarity coefficient. First, the mutual information value of each IMF and normal signal is calculated. Then calculate the mutual information with the leaked signal. Finally, the numerical value of the similarity coefficient is calculated. According to the definition of similarity coefficient, it can be obtained that the larger the value of similarity coefficient is, the more leakage information is contained in the IMF. After calculation, the obtained similarity coefficient values of each IMF are shown in Figure 8.

It can be seen from Figure 8 that the similarity coefficient of IMF7-IMF12 is significantly larger than that of IMF1-IMF6. Therefore, IMF7-IMF12 are selected for signal reconstruction. The reconstructed signal is compared with the signal denoised by EMD and UPEMD, and the comparison result is shown in Figure 9.

It can be clearly seen from Fig. 9(a), Fig. 9(b) and Fig. 9(c) that the curve obtained by the method of the present invention is smoother and contains less noise, and the change trend of the signal and inflection point.

8mm and 5mm valves were opened at 1km, 3km and 7km of the pipeline to simulate leakage, and 1500 sets of leakage acoustic signals were collected. In the normal working condition of the pipeline, 1500 sets of data are also collected. The UPEMD algorithm based on mutual information optimization is used to denoise the signal.

In the linear case, the parameters of the three TWSVMs in the hidden layer are: c ₁₁ =0.3, c ₁₂ =0.1, c ₂₁ =c ₂₂ =0.2, c ₃₁ =0.3, and c ₃₂ =0.25. The parameter c ₁ =c ₂ =1 of the output layer TWSVM. In the nonlinear case, c ₁₁ =0.1, c ₁₂ =0.3, c ₂₁ =c ₃₁ =0.35, and c ₂₂ =c ₃₂ =0.1. The parameters c ₁ =1, c ₂ =10 of the output layer TWSVM. The three TWSVMs in the hidden layer all use the RBF kernel function, and the values of the parameter λ are 1, 0.1, and 0.1, respectively.

The first 800 groups of denoised data are used for the training of the DTWSVM model, and each parameter in the model is adjusted through supervised learning and unsupervised learning until the error is minimized. After the training is completed, the remaining 700 groups of denoised data are input as test samples into the trained DTWSVM for pipeline condition recognition, and the recognition accuracy of DTWSVM is compared with that of TWSVM and DSVM. The comparison results are shown in Table 1. Show.

Table 1 Comparison results of recognition accuracy

According to Table 1, whether it is leaks at different locations or leaks of different sizes at the same location, the accuracy of DTWSVM for identifying pipeline leaks is significantly higher than that of TWSVM and DSVM, up to 99.6%.

The present invention optimizes the UPEMD method by using mutual information, and obtains a UPEMD filtering method based on mutual information optimization, which can effectively remove noise in the signal and greatly preserve the leakage information. In addition, the present invention combines deep learning with TWSVM to obtain a new network model DTWSVM. It has been verified by experiments that inputting the denoised signal into DTWSVM can accurately judge the pipeline working conditions, and the recognition accuracy is high, which provides convenience for the construction and management of the intelligent pipeline network.

Claims (5)

1. The buried oil pipeline leakage detection method based on the improved UPEMD and the improved DTWSVM is characterized by comprising the following steps of:

step 1, sound pressure sensors are adopted to obtain sound wave signal data of an oil pipeline under normal working conditions and during leakage;

step 2, decomposing the sound wave signal data acquired in the step 1 by adopting a UPEMD algorithm until the residue cannot be decomposed to obtain a plurality of IMFs;

step 3, calculating mutual information values of the IMFs obtained in the step 2 and sound wave signals when the pipeline leaks and normally works, calculating similarity coefficients of the IMFs according to the obtained mutual information values, and selecting IMFs containing much leakage information through the similarity coefficients to reconstruct the signals;

step 4, establishing a network model DTWSVM according to the deep learning DL and the twin Gemini support vector machine TWSVM;

step 5, establishing a training sample and a test sample of the DTWSVM model according to the reconstruction signal obtained in the step 3; inputting the training sample into a DTWSVM model, and adjusting parameters in the DTWSVM model through supervised learning and unsupervised learning to obtain a trained DTWSVM model; and then inputting the test sample into the trained DTWSVM model, and judging the working condition of the buried oil pipeline.

2. The buried oil pipeline leakage detection method based on the improved UPEMD and DTWSVM according to claim 1, characterized in that the concrete steps of decomposing the sound wave signal by using the UPEMD algorithm in the step 2 are as follows:

step 2.1, calculating the cycle period T of the UPEMD algorithm _c And the number of cycles n _c ；

n _c ＝log ₂ (n) (1)

T _c ＝2 ^m ,m＝1:n _c (2)

Where n represents the total length of the signal data and the value of the mask frequency is equal to the reciprocal of the cycle period, i.e. f _c ＝1/T _c And m is the reciprocal of the number of cycles.

Step 2.2, number n of phases _p And an amplitude epsilon is set, wherein，n _p ∈N,n _p Is > 1, and n _p The number of (2) must be an integer power of 2 when n is _p When the value of (1) is 2, the UPEMD is the mask EMD;

step 2.3, let original signal x (t) r ₀ (t) simultaneously, constructing a masking signal, namely:

w(t；ε _m ；f _c ；θ _k )＝ε _m ·cos(2π·f _c ·t+θ _k ) (3)

wherein epsilon _m ＝ε×r _m-1 (t)，r _m-1 (t) is the standard deviation, phase θ _k At [0,2 π]Is uniformly divided into n _p Portions of, therefore

θ _k ＝2π(k-1)/n _p ,k＝1:n _p (4)

Step 2.4, perturbing the signal y _k The expression of (t) is:

y _k (t)＝x(t)+ε _m ·cos(2π·f _c ·t+θ _k ) (5)

step 2.5, using EMD to scramble signal y _k (t) decomposing to obtain the ith IMF component as E _i (·); thus, the first component resulting from EMD decomposition is:

c _m,k (t)＝E ₁ (x(t)+w(t；ε _m ；f _c ；θ _k )),k＝1,2...n _p ,m＝1:n _c (6)

step 2.6 from c _m,k (t) subtracting the masking signal w (t;. epsilon.) _m ；f _c ；θ _k ) Then, the IMF1 can be obtained by performing ensemble averaging, that is:

and 2.7, taking the remainder obtained by calculation in the step 2.6 as a new disturbing signal to be decomposed, and circulating the steps 2.1-2.6 until all IMFs are decomposed, and stopping circulation.

3. The buried oil pipeline leakage detection method based on UPEMD and DTWSVM as claimed in claim 1, wherein the specific method of step 3 is:

calculating mutual information values of the IMFs and the sound wave signals when the pipeline leaks and works normally:

wherein X and Y are two random signals, p (X, Y) represents the combined distribution of X and Y, and the edge probability density functions of X and Y are p (X) and p (Y), respectively;

defining the mutual information of IMF and leakage signal as l _i Defining the mutual information of IMF and signal under normal condition as h _i Calculating a similarity coefficient beta from the obtained mutual information value _i ，β _i ＝l _i -h _i ；i＝1,2…,k；

The IMF reconstructed signal containing more leakage information is selected according to the similarity coefficient, and the IMF with larger similarity coefficient contains more leakage information.

4. A buried oil pipeline leakage detection method based on modified UPEMD and DTWSVM as claimed in claim 1, characterized in that the specific method of step 4 is:

on the basis of deep learning and TWSVM, a new network model DTWSVM is established by analyzing the influence of various network structures and network layer numbers on the detection precision, and the optimal value of each parameter in the model is determined by training the model;

the DTWSVM model is a three-layer network model comprising an input layer, a hidden layer and an output layer, wherein three TWSVM are arranged in the hidden layer to extract the characteristics of input data, and a main TWSVM is arranged in the output layer to judge the working condition of a pipeline;

the TWSVM in the hidden layer is divided into two types, when input data are linear data, the TWSVM is linear, three pairs of hyperplanes are generated, original data are mapped into a 6-dimensional space, and 6-dimensional position information of the original data is recorded; when nonlinear data are input, the TWSVM is nonlinear, and kernel function data are introduced to map original data to a 6-dimensional space;

original data T ═ x ₁ ,y ₁ ),…,(x _m+k ,y _m+k )]Inputting the data into an input layer of the model, and performing feature extraction on the data by three TWSVM in a hidden layer to generate a feature data set:

X＝f(x)＝(f(x) _1,+ ,f(x) _1,- ,f(x) _2,+ ,f(x) _2,- ,f(x) _3,+ ,f(x) _3,- ) ^T (27)

f(x) _i,j ＝(w _i,j ·x)+b _i,j ，i＝1,2,3,j＝+,- (28)

in the formula, b _i,j And w _i,j Parameters in the DTWSVM model are obtained, and specific numerical values are obtained after training;

in the nonlinear case, the TWSVM in the hidden layer is mapped by a kernel function:

f(x) _i,j ＝K(x ^T ,C ^T )z _i,j +b _i,j ，i＝1,2,3,j＝+,- (29)

in the formula, K (x) ^T ,C ^T ) Representing a kernel function, z _i,j Are parameters in the DTWSVM model;

the parameters in the TWSVM are determined to be optimal values through a cross validation and grid search method;

the main TWSVM of the output layer is divided into two categories, namely a linear TWSVM and a nonlinear TWSVM; the category is consistent with that of the TESVM in the hidden layer, and the main TWSVM is used for judging the working condition of the pipeline;

the discrimination formulas of the main TWSVM under linear and nonlinear states are respectively as follows:

in the formula, K (x) ^T ,C ^T ) Representing a kernel function.

5. A buried oil pipeline leak detection method based on modified UPEMD and DTWSVM as claimed in claim 1, characterized in that in step 5, the training of the model adjusts parameters in the network model by unsupervised learning from bottom to top and supervised learning from top to bottom; the unsupervised learning is to train each layer of the model one by one, adjust each parameter of each layer, and when training, the output of the previous layer is used as the input of the next layer until all layers are trained; the supervised learning is to adjust the parameters of each layer of the model on the basis of unsupervised learning so that the parameters of each layer reach the optimal values and further minimize errors.

Images