A fluid pipeline leakage identification method based on improved PSO-LSTSVM in compressed sensing domain

Abstract

The present invention belongs to the field of fluid pipeline leakage detection, and specifically relates to a fluid pipeline leakage identification method based on an improved PSO-LSTSVM in a compressed sensing domain. The method comprises: acquiring a fluid pipeline signal, performing compressed sensing processing on the fluid pipeline signal to obtain a compressed signal; performing feature extraction on the compressed signal to obtain a feature data set; and using an improved PSO-LSTSVM model to process feature data in the feature data set to obtain an identification result. The present invention proposes an attenuation amplitude variation strategy, through which parameters such as an inertia weight and a learning factor are dynamically varied to balance the relationship between the global search and local search capabilities of a particle swarm optimization algorithm during the search process, thereby preventing the optimization algorithm from falling into a local optimal solution and improving its convergence efficiency. Finally, the strategy is used to optimize the LSTSVM to achieve performance improvement of the recognition model.

Description

Technical Field

[0001] The present invention belongs to the field of fluid pipeline leakage detection, and specifically relates to a fluid pipeline leakage identification method based on an improved PSO-LSTSVM in a compressed sensing domain. Background Art

[0002] Pipelines are essential transportation vehicles for fluids such as natural gas, oil, and tap water, and are also crucial urban infrastructure. However, due to corrosion, cracks, dents in pipeline materials, vandalism, and environmental factors, pipeline leaks are common. These accidents not only impact people's lives but can also lead to significant economic losses and environmental disasters. Therefore, real-time monitoring of pipeline operating conditions is crucial for maintaining safe pipeline network operation, preventing resource waste, and achieving sustainable production. Currently, fluid pipeline leak detection faces two major challenges. First, as pipelines continue to expand in size, the amount of data generated by fluid pipeline detection systems is rapidly increasing, ushering in the "big data" era for fluid pipeline leak detection. However, traditional Nyquist data acquisition methods generate massive amounts of data for collecting pipeline leakage signals, resulting in data redundancy and other issues, hindering data storage, communication, and leak identification within the fluid pipeline detection system. Second, while fluid pipeline leak detection is crucial for maintaining normal pipeline operation, pipeline leak identification suffers from low accuracy, necessitating a high level of accuracy.

[0003] Currently, researchers in the field of fluid pipeline leak identification typically classify and identify fault features after leak signal processing using data-driven intelligent diagnostic methods. Machine learning approaches such as support vector machines (SVM), artificial neural networks (ANN), and k-nearest neighbors (KNN) are often employed. SVM and its improved versions are intelligent diagnostic methods based on statistical theory, with a solid theoretical foundation and excellent performance, offering significant advantages in nonlinear and high-dimensional pattern recognition. The SVM parameter combination directly affects the model's recognition performance. Many researchers have used particle swarm optimization (PSO) to optimize SVM parameter combinations. However, traditional PSO optimization algorithms often suffer from problems such as being prone to local optimal solutions and slow convergence, hindering the SVM model's ability to find the optimal parameter solution and, in turn, directly affecting the accuracy of leak identification. Furthermore, when working with large feature datasets, SVM recognition models suffer from slow recognition speed and decreased recognition accuracy, hindering the efficiency of fluid pipeline leak identification. Summary of the Invention

[0004] To solve the problems existing in the above-mentioned prior art, the present invention proposes a fluid pipeline leakage identification method based on an improved PSO-LSTSVM in a compressed sensing domain. The method comprises: acquiring a fluid pipeline signal, performing compressed sensing processing on the fluid pipeline signal to obtain a compressed signal; performing feature extraction on the compressed signal to obtain a feature data set, wherein the feature data set includes a leakage feature set and a non-leakage feature set; inputting the feature data set into an improved PSO-LSTSVM model, and training and testing the model through a 5-fold cross-validation method; acquiring a fluid pipeline signal to be detected, inputting the fluid pipeline signal to be detected into the trained improved PSO-LSTSVM model, and obtaining an identification result.

[0005] Preferably, the process of performing compressed sensing processing on the fluid pipeline signal includes: setting the compression rate and the observation matrix; obtaining the length of the processed fluid pipeline signal according to the compression rate; obtaining the shape of the observation matrix according to the original fluid pipeline signal length and the processed signal length; and using the observation matrix to compress and reduce the dimension of the fluid pipeline signal to obtain the compressed observation value of the original signal.

[0006] Preferably, extracting features from the compressed signal includes extracting the average amplitude, standard deviation, energy, kurtosis, root amplitude and effective value of the compressed signal; and combining all the extracted features to obtain a feature data set.

[0007] Preferably, the process of optimizing the parameters in the least squares twin support vector machine using the improved PSO optimization algorithm includes:

[0008] Step 1: Initialize the relevant parameters of the improved PSO optimization algorithm, including population size, particle dimension, particle position and velocity, individual optimal value of the particle swarm and global optimal value, where the particle dimension represents the number of optimization parameters;

[0009] Step 2: Calculate the fitness function value of each particle based on the relevant parameters, that is, the accuracy of the recognition model LSTSVM corresponding to each particle;

[0010] Step 3: Use the attenuation amplitude strategy to update the inertia weight, individual learning factor and global learning factor respectively;

[0011] Step 4: Update the particle's velocity and position according to the updated inertia weight, individual learning factor, and global learning factor, and increase the number of iterations by 1;

[0012] Step 5: Determine the difference between the current number of iterations and the set number of iterations. If it is less than the set number of iterations, return to step 2; otherwise, go to step 6.

[0013] Step 6: Output the global optimal particle position corresponding to the global optimal value. This particle position is the optimal parameter value of LSTSVM.

[0014] Furthermore, calculating the fitness function value of the particle includes calculating the recognition accuracy of the LSTSVM model using a 5-fold cross-validation method; the calculation process includes: dividing the feature data set into 5 subsets, 4 of which are used as training sets and 1 subset is used as a test set, alternating the training set and the test set, and obtaining the recognition accuracy of the LSTSVM model each time, and taking the average of the 5 recognition results as the accuracy of the LSTSVM model.

[0015] Preferably, the formula for updating the inertia weight, individual learning factor and global learning factor respectively using the attenuation amplitude strategy is:

[0016]

[0017]

[0018]

[0019] Among them, w u and c 1u They represent the inertia weight w, the upper limit of the amplitude of the individual learning factor c1, and c 2d represents the lower limit of the amplitude of the global learning factor c2, w d 、c 1d and c 2u is the base value of w, c1 and c2 to limit the amplitude range of the parameters, and the expressions are w d =w u -0.3, c 1d =c 1u -0.3, c 2u =c 2d +0.3, T represents the vibration period, t is the current iteration number, T max is the maximum number of iterations.

[0020] Preferably, the formula for updating the velocity and position of the particle is:

[0021] v i (t+1)=wv i (t)+c1r1(p i (t)-x i (t))+c2r2(p g (t)-x i (t))

[0022] x i (t+1)=x i (t)+v i(t+1)

[0023] Among them, v i (t+1) represents the particle velocity of the next iteration, t represents the current number of iterations, w represents the inertia weight, c1 represents the individual learning factor, c2 represents the global learning factor, r1 and r2 are both random numbers that vary within [0,1], x i (t) represents the current position of the particle, p i (t) represents the optimal position searched by the i-th particle, i.e., the individual optimal position, p g (t) represents the best position searched by all particles in the group, i.e., the global optimum, and x i (t+1) represents the particle position of the next iteration

[0024] Preferably, the improved PSO-LSTSVM model is trained to solve an optimization problem constructed according to the improved PSO-LSTSVM model, and the optimal solution of the optimization problem is obtained. The expression of the optimization problem is:

[0025]

[0026]

[0027] Among them, K represents the kernel function, A represents the positive sample set, B represents the negative sample set, and C = [A T B T ] T , w (1) 、w (2) Both represent normal vectors, b (1) 、b (2) Both represent intercepts, e represents the identity matrix, c1 and c2 represent penalty factors, and y represents the slack variable.

[0028] Beneficial effects of the present invention:

[0029] The present invention adopts compressed sensing theory to compress and collect acquired signals, reducing the amount of data required to process the signals, thereby solving problems such as data redundancy. The present invention proposes a decreased and variable amplitude (DVA) strategy, which attenuates parameters such as the inertia weight and learning factor to balance the global and local search capabilities of the particle swarm optimization algorithm during the search process, preventing the optimization algorithm from falling into a local optimal solution and improving its convergence efficiency. The present invention applies an improved PSO to the LSTSVM to optimize the model's parameter combination. The resulting LSTSVM model is used for fluid pipeline leakage identification, ultimately achieving improvements in the accuracy and speed of fluid pipeline leakage identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is an overall flow chart of an embodiment of the present invention;

[0031] Figure 2 This is a schematic diagram of the compression and dimensionality reduction process of the present invention;

[0032] Figure 3 A comparison diagram of the pipeline leakage signal of the present invention and the observed value signal after compression processing;

[0033] Figure 4 Graph showing the changes in inertia weight and learning factor of the DVAPSO of the present invention;

[0034] Figure 5 This is a performance graph of the DVAPSO algorithm of the present invention on the test function. DETAILED DESCRIPTION

[0035] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0036] A specific implementation of a fluid pipeline leakage identification method based on an improved PSO-LSTSVM in a compressed sensing domain, such as Figure 1 As shown in the figure, the method includes: first, performing compressed sensing processing on the collected pipeline leakage time domain signal, and extracting features from the compressed signal to form a feature data set; then, using the improved PSO to optimize the parameter combination of LSTSVM to further improve the accuracy of LSTSVM for pipeline leakage detection; finally, inputting the extracted feature data set into the optimized LSTSVM recognition model, and using the 5-fold cross-validation method to verify the recognition performance of the LSTSVM model.

[0037] A specific implementation of a fluid pipeline leakage identification method based on an improved PSO-LSTSVM in a compressed sensing domain includes: acquiring a fluid pipeline signal, performing compressed sensing processing on the fluid pipeline signal to obtain a compressed signal; performing feature extraction on the compressed signal to obtain a feature data set; and using an improved PSO-LSTSVM model to process feature data in the feature data set to obtain an identification result.

[0038] The process of compressed sensing processing of fluid pipeline signals includes: setting the compression rate and observation matrix, and knowing the length of the processed fluid pipeline signal according to the compression rate; obtaining the shape of the observation matrix according to the original fluid pipeline signal length and the processed signal length; and using the observation matrix to compress and reduce the dimension of the fluid pipeline signal to obtain the compressed observation value of the original signal.

[0039] Specifically, the theory of compressed sensing (CS) states that as long as the signal is sparse or the transform coefficients Θ on a certain transform domain Ψ are sparse, then a measurement matrix that is unrelated to the transform basis can be used to perform linear projection to obtain the observation vector. Its mathematical expression is:

[0040] Y=ΦX=ΦΨΘ

[0041] Among them, Y is the obtained observation vector, X is the original signal with a length of N, and Φ is the observation matrix.

[0042] The compression rate is a measure of the degree to which the original signal is compressed and collected, and its definition is:

[0043]

[0044] Where A is the sample length of the signal, and B is the signal length obtained after measurement matrix observation. Therefore, different dimensions of the measurement matrix can achieve different degrees of compressed acquisition. The larger the CR, the higher the compression ratio of the leakage signal, and the less compressed sensing domain data is measured.

[0045] The present invention compresses and reduces the dimension of the collected fluid pipeline leakage signal to obtain the observation value of the original signal. The observation value is greatly reduced in data volume, but the observation value still contains the characteristic information of the original signal. The compression and dimensionality reduction process is as shown in the attached figure. Figure 2 As shown. Taking the water supply pipeline leakage as an example, a water supply pipeline leakage signal is selected with a signal length of 20480. When the observation matrix is ​​a partial Fourier matrix and the compression rate is 50%, the original signal and the obtained observation value are shown as follows: Figure 3 As shown in the figure, the observed values ​​obtained after compression of the leakage time domain signal have a time scale of only half that of the original signal, and the final observation length is 10240. Therefore, at a 50% compression rate, after compressing and reducing the dimension of the collected fluid pipeline leakage dataset, the resulting observation dataset is only half the size of the original dataset. Because the observed values ​​still retain the characteristic information of the original signal, the feature extraction method used for the original signal can be applied to the observed value dataset, collecting a total of five sets of features to form a feature dataset for pipeline leak identification.

[0046] Feature extraction of the compressed signal includes extracting the average amplitude, standard deviation, energy, kurtosis, root amplitude and effective value of the compressed signal; all the extracted features are combined to obtain a feature data set.

[0047] The calculation expression of the average amplitude is:

[0048]

[0049] Among them, X amrepresents the average amplitude, N represents the length of sample data, x i Represents the i-th data in the sample.

[0050] The calculation expression of standard deviation is:

[0051]

[0052] The calculation expression of energy is:

[0053]

[0054] The calculation expression of kurtosis is:

[0055]

[0056] The calculation expression of the square root amplitude is:

[0057]

[0058] The calculation expression of effective value is:

[0059]

[0060] The process of optimizing the parameters of the least squares twin support vector machine using the improved PSO optimization algorithm includes:

[0061] Step 1: Initialize the relevant parameters of the improved PSO optimization algorithm, including population size, particle dimension, particle position and velocity, individual optimal value of the particle swarm and global optimal value, where the particle dimension represents the number of optimization parameters;

[0062] Step 2: Using the 5-fold cross-validation method, calculate the fitness function value for each particle, that is, the accuracy of the LSTSVM recognition model corresponding to each particle. The model accuracy is calculated by dividing the feature dataset into five subsets, with four subsets serving as training sets and one subset serving as a test set. The training and test sets are then alternated, and the recognition accuracy of the LSTSVM model is calculated for each run. The average of the five recognition results is then taken as the accuracy of the LSTSVM model.

[0063] Step 3: Use the attenuation amplitude strategy to update the inertia weight, individual learning factor and global learning factor respectively;

[0064] Step 4: Update the particle's velocity and position according to the updated inertia weight, individual learning factor, and global learning factor, and increase the number of iterations by 1;

[0065] Step 5: Determine the difference between the current number of iterations and the set number of iterations. If it is less than the set number of iterations, return to step 2; otherwise, go to step 6.

[0066] Step 6: Output the global optimal particle position corresponding to the global optimal value. This particle position is the optimal parameter value of LSTSVM.

[0067] Specifically, the PSO algorithm is an adaptive optimization algorithm based on population iterative search. The algorithm searches for the optimal solution in a complex problem space through the interaction between the whole and individuals in the population particles. The particle swarm optimization algorithm has the advantages of simple structure, few parameters, high operating efficiency, and fast convergence speed, and has therefore been widely used. The particle swarm optimization algorithm simulates the migration and foraging behavior of birds. Each individual in the group moves in a given space and continuously updates its speed and position based on its own cognitive learning information and the social information obtained from other particles. After continuous updates and iterations, the particle swarm will eventually find the global optimal position in the complex space, that is, the global optimal solution. Assume that the particles move in the D-dimensional search space, the number of particle swarms is N, and the current position and speed of the i-th particle are x, respectively. i (t) and v i (t), the optimal position searched by the i-th particle is p i (t), called individual optimality, the best position searched by all particles in the group is p g (t), called the global optimum. At the (t+1)th iteration, particle i updates its velocity v according to the following rules i (t+1) and position x i (t+1), that is

[0068] v i (t+1)=wv i (t)+c1r1(p i (t)-x i (t))+c2r2(p g (t)-x i (t))

[0069] x i (t+1)=x i (t)+v i (t+1)

[0070] Among them, v i (t+1) represents the particle velocity of the next iteration, t represents the current number of iterations, w represents the inertia weight, c1 represents the individual learning factor, c2 represents the global learning factor, r1 and r2 are both random numbers that vary within [0,1], x i (t) represents the current position of the particle, p i (t) represents the optimal position searched by the i-th particle, p g (t) represents the best position searched by all particles in the group, x i(t+1) represents the particle position at the next iteration.

[0071] In the PSO algorithm parameters, w represents the weight that affects the current velocity of the particle. The value of w determines the particle's local and global search capabilities. When the w value is large, the algorithm has a stronger global search capability, and vice versa. On the other hand, the learning factors c1 and c2 represent the weights of particles flying in the direction of a single optimal position and the global optimal position, respectively. By adjusting the size between the two, the global search capability and the late convergence capability can be balanced. However, the traditional particle swarm optimization algorithm cannot effectively balance the local and global search capabilities of the particle swarm optimization algorithm due to its fixed parameters such as inertia weight and learning factor. The algorithm always has the problem of being easily trapped in the local optimal solution.

[0072] To address the above issues, the ideal change states of the inertia weight w and learning factors c1 and c2 must meet the following two requirements:

[0073] (1) w describes the inertia of the particles. In the early stage of the iteration, w should be larger to ensure that each particle fully searches the entire space and the algorithm does not fall into the local optimal value. In the later stage of the iteration, the value of w should be appropriately small to help particles search the local area in the later stage of the search process and accelerate the convergence of the algorithm.

[0074] (2) The individual learning factor c1 and the global learning factor c2 also affect the global and local search capabilities. In the early stages of the iteration, the particle swarm hopes to have a larger c1 and a smaller c2, so that the particles can use more individual-related information to guide the particle flight trajectory in the initial stage, promoting the global search of the group. At the same time, more use of individual information is also conducive to particles jumping out of the local optimal solution and avoiding premature convergence. In the later stages of the iteration, the particle swarm hopes to have a smaller c1 and a larger c2, which is conducive to particles using more population-related information to guide flight in the later stages, promoting the group to converge quickly and find the global optimal value.

[0075] The present invention proposes a decaying amplitude strategy (DVA) for the parameters such as the inertia weight and learning factor of the PSO optimization algorithm. The specific expression is as follows:

[0076]

[0077]

[0078]

[0079] Among them, w u and c 1u They represent the inertia weight w, the upper limit of the amplitude of the individual learning factor c1, and c 2drepresents the lower limit of the amplitude of the global learning factor c2, w d 、c 1d and c 2u is the base value of w, c1 and c2 to limit the amplitude range of the parameters, and the expressions are w d =w u -0.3, c 1d =c 1u -0.3, c 2u =c 2d +0.3, T represents the vibration period, t is the current iteration number, T max is the maximum number of iterations.

[0080] Preferably, the following initial values ​​are set for the above parameters: u =0.95, c 1u =2.5, c 2d =1.5, T=2000, T max =9000, the trend of the parameter changes with the increase of iteration number is as follows Figure 4 shown.

[0081] exist Figure 4 In the process, by introducing the tangent function term, w and c1 are in a process of decreasing from high to low with amplitude attenuation, while c2 is in a process of decreasing from low to high with amplitude attenuation. This change trend satisfies the requirement that the inertia weight w and the learning factor c1 need to maintain a large value in the early stage of iteration and a small value in the later stage of iteration; it also meets the requirement that the learning factor c2 needs to be small in the early stage and large in the later stage.

[0082] The alternating phenomenon of dispersion and reorganization often occurs in social groups, such as the dispersion and aggregation behavior of birds during hunting. Inspired by this phenomenon, the present invention introduces cosine function terms to make the size of each parameter change alternately, and at the same time adds exponential function terms to make the amplitude change of the parameters continuously decrease. Figure 4 As shown by the trend of c1, in the early stages of iteration, c1 maintains a high range of attenuated amplitude fluctuations. This high range ensures the particles' initial global search capabilities. Furthermore, the vibration strategy ensures the diversity of the particles' global search, allowing them to fully utilize their cognitive learning capabilities. In the later stages of iteration, c1 is kept at a low level of fluctuation. This vibration helps particles escape local optima while ensuring that the social learning information of the particle swarm is fully volatilized. Similarly, corresponding attenuated amplitude fluctuation strategies can be set according to the changing needs of w and c2.

[0083] In order to verify the effectiveness of the improved PSO optimization algorithm of the present invention, namely the decaying variable amplitude particle swarm optimization algorithm (DVAPSO), a comparison was made with the PSO algorithm and the TVACPSO algorithm (time-varying acceleration coefficient particle swarm optimization algorithm) on four groups of test functions. Figure 5 It can be seen that DVAPSO is significantly better than PSO and TVACPSO algorithms in terms of convergence speed and optimization accuracy, which verifies the effectiveness of the DVAPSO algorithm.

[0084] There are four test functions: Sphere function F1, Ackle function F2, Rastrigin function F3, and Griewank function F4. Their expressions are:

[0085]

[0086]

[0087]

[0088]

[0089] Where D represents the dimension of the search space.

[0090] The least squares twin support vector machine (LSTSVM) is an improved version of the twin support vector machine (TSVM). LSTSVM inherits the advantage of TSVM in reducing the model's computational complexity compared to SVM. At the same time, it replaces the constraints in TSVM with equality constraints, converting the original convex quadratic programming problem into solving two linear equations, further reducing the solution complexity while ensuring model performance.

[0091] The learning process of LSTSVM can be transformed into an optimization problem. In pipeline leakage detection, the kernel function idea is usually used to improve the recognition accuracy of the model, that is, to construct a nonlinear LSTSVM. Given a positive and negative sample set Nonlinear LSTSVM requires solving the following optimization problem:

[0092]

[0093]

[0094] Where C=[A T B T ] T , K represents the kernel function; the kernel function is a Gaussian kernel function, and its expression is as follows:

[0095]

[0096] By substituting the two equality constraints into the objective function, the original problem is transformed into the following two unconstrained minimization problems:

[0097]

[0098]

[0099] Solve the two unconstrained minimization problems above, let M = [K(A,C T )e],N=[K(B,C T )e] and introduce the regularization term εI to obtain:

[0100]

[0101]

[0102] According to the solution of the above formula, the classification hyperplane can be constructed in the nonlinear space. i The classification decision function is:

[0103]

[0104] The values ​​of the penalty factors c1 and c2 and the kernel function parameter g directly affect recognition accuracy. Therefore, finding appropriate LSTSVM parameters plays a crucial role in establishing a pipeline leak detection model. This paper utilizes the decaying variable amplitude PSO algorithm (DVAPSO) to optimize the LSTSVM parameter combination, thereby obtaining the optimal LSTSVM parameter combination and achieving the optimal performance. This further improves the recognition accuracy of the LSTSVM model. Selecting accuracy based on 5-fold cross-validation in the fitness function effectively prevents overfitting and underfitting.

[0105] The present invention obtains a feature data set by extracting features in the compressed sensing domain, and its data volume is greatly reduced compared with the feature data volume of the original signal; and uses the DVAPSO algorithm to optimize the parameter combination of LSTSVM to improve the recognition accuracy of the model.

[0106] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation plans of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A fluid pipeline leakage identification method based on improved PSO-LSTSVM in the compressed sensing domain, characterized by: include: Acquire a fluid pipeline signal, perform compressed sensing processing on the fluid pipeline signal, and obtain a compressed signal; Feature extraction is performed on the compressed signal to obtain a feature dataset, which includes a leakage feature set and a non-leakage feature set. The feature dataset is input into the improved PSO-LSTSVM model, and the model is trained and tested using a 5-fold cross-validation method. The fluid pipeline signal to be detected is obtained and input into the trained improved PSO-LSTSVM model to obtain the recognition result. The improved PSO optimization algorithm is used to optimize the parameters of the least squares twin support vector machine, including: Step 1: Initialize the relevant parameters of the improved PSO optimization algorithm, including population size, particle dimension, particle position and velocity, individual optimal value of the particle swarm and global optimal value, where the particle dimension represents the number of optimization parameters; Step 2: Calculate the fitness function value of each particle based on the relevant parameters, that is, the accuracy of the recognition model LSTSVM corresponding to each particle; Step 3: Use the attenuation amplitude strategy to update the inertia weight, individual learning factor and global learning factor respectively; Step 4: Update the particle's velocity and position according to the updated inertia weight, individual learning factor, and global learning factor, and increase the number of iterations by 1; Step 5: Determine the difference between the current number of iterations and the set number of iterations. If it is less than the set number of iterations, return to step 2; otherwise, go to step 6. Step 6: Output the global optimal particle position corresponding to the global optimal value. This particle position is the optimal parameter value of LSTSVM.

2. The fluid pipeline leakage identification method based on the improved PSO-LSTSVM in the compressed sensing domain according to claim 1 is characterized in that: The process of compressed sensing processing of fluid pipeline signals includes: setting the compression rate and observation matrix; obtaining the length of the processed fluid pipeline signal according to the compression rate; obtaining the shape of the observation matrix according to the original fluid pipeline signal length and the processed signal length; and using the observation matrix to compress and reduce the dimension of the fluid pipeline signal to obtain the compressed observation value of the original signal.

3. The fluid pipeline leakage identification method based on improved PSO-LSTSVM in the compressed sensing domain according to claim 1 is characterized in that: Feature extraction of the observations includes extracting the mean amplitude, standard deviation, energy, kurtosis, square root amplitude and effective value of the observations; all the extracted features are aggregated to obtain a feature data set.

4. The fluid pipeline leakage identification method based on improved PSO-LSTSVM in the compressed sensing domain according to claim 1 is characterized in that: Calculating the fitness function value of the particle includes using the 5-fold cross-validation method to calculate the recognition accuracy of the LSTSVM model; the calculation process includes: dividing the feature data set into 5 subsets, 4 of which are used as training sets and 1 subset is used as a test set, alternating the training set and the test set, and obtaining the recognition accuracy of the LSTSVM model each time, and taking the average of the 5 recognition results as the accuracy of the LSTSVM model.

5. The fluid pipeline leakage identification method based on improved PSO-LSTSVM in the compressed sensing domain according to claim 1 is characterized in that: The formulas for updating the inertia weight w, individual learning factor c1, and global learning factor c2 using the attenuation amplitude strategy are as follows: Among them, w u and c 1u They represent the inertia weight w, the upper limit of the amplitude of the individual learning factor c1, and c 2d represents the lower limit of the amplitude of the global learning factor c2; w d 、c 1d and c 2u are the base values ​​of w, c1 and c2, respectively, to limit the amplitude range of the parameters, and the expressions are w d =w u -0.3, c 1d =c 1u -0.3, c 2u =c 2d +0.3; T represents the vibration period, t is the current iteration number, T max is the maximum number of iterations.

6. The fluid pipeline leakage identification method based on improved PSO-LSTSVM in the compressed sensing domain according to claim 1 is characterized in that: The formula for updating the particle's velocity and position is: v i (t+1)=wv i (t)+c1r1(p i (t)-x i (t))+c2r2(p g (t)-x i (t)) x i (t+1)=x i (t)+v i (t+1) Among them, v i (t+1) represents the particle velocity of the next iteration, t represents the current number of iterations, w represents the inertia weight, c1 represents the individual learning factor, c2 represents the global learning factor, r1 and r2 are both random numbers that vary within [0,1], x i (t) represents the current position of the particle, p i (t) represents the optimal position searched by the i-th particle, i.e., the individual optimal position, p g (t) represents the best position searched by all particles in the group, i.e., the global optimum, and x i (t+1) represents the particle position at the next iteration.

7. The fluid pipeline leakage identification method based on improved PSO-LSTSVM in the compressed sensing domain according to claim 1 is characterized in that: The improved PSO-LSTSVM model is trained to solve the optimization problem constructed based on the improved PSO-LSTSVM model, and the optimal solution of the optimization problem is obtained. The expression of the optimization problem is: Among them, K represents the kernel function, A represents the positive sample set, B represents the negative sample set, and C = [A T B T ] T , w (1) 、w (2) Both represent normal vectors, b (1) 、b (2) Both represent intercepts, e represents the identity matrix, c1 and c2 represent penalty factors, and y represents the slack variable.

Images