Gas-liquid two-phase flow pattern recognition method based on flow noise signal

Abstract

The flow noise signal-based gas-liquid two-phase flow pattern recognition method of the present application comprises the following steps: 1) collecting the gas-liquid two-phase flow noise signals in a pipeline system, classifying the collected flow noise signals, and constructing an original data set; 2) pre-processing the original data set to obtain the IMF components of each flow noise signal; 3) using a signal analysis method to extract three component characteristic parameters of the flow noise signal IMF; 4) processing the extracted characteristic parameters as a sample set, dividing the sample set into a training set and a test set, inputting the training set into an SVM training module, training to obtain an SVM flow pattern recognition model of the gas-liquid two-phase flow, inputting the test set as an input parameter into the SVM flow pattern recognition model to obtain a recognition result; and 5) after extracting the characteristic parameters of an unknown flow pattern noise signal, constructing a characteristic vector and inputting it into the trained SVM flow pattern recognition model to obtain a flow pattern recognition result. The method of the present application has obviously improved accuracy.

Description

Technical Field

[0001] This invention belongs to the field of flow pattern recognition technology and relates to a method for identifying the flow pattern of gas-liquid two-phase flow based on flow noise signals. Background Technology

[0002] Two-phase flow is widely used in various fields such as energy, chemical industry, metallurgy, nuclear industry, and aerospace. As one of the most common forms of two-phase flow, gas-liquid two-phase flow is widely used in the development and utilization of deep-sea oil and gas, as well as in mixed transportation processes. Flow pattern, that is, the state and structure of flow, is the most fundamental and critical parameter in gas-liquid two-phase flow. It is closely related to mass and heat transfer processes. Different properties and behaviors of two-phase flows lead to complex flow structures and flow patterns. The accuracy of flow pattern determination is of great significance for correctly predicting the pressure drop and heat transfer performance of two-phase flow, accurately measuring two-phase flow parameters, and ensuring the safe operation of equipment.

[0003] Currently, methods for identifying gas-liquid two-phase flow patterns are mainly classified into three categories: flow pattern mapping, direct observation, and indirect analysis. Direct observation methods include visual inspection and high-speed photography; indirect analysis methods include capacitance tomography, ultrasonic measurement, gamma ray analysis, optical methods, and differential pressure signal identification. However, direct observation methods are often highly subjective and require high equipment transparency, making them difficult to implement in practical engineering. Gamma rays pose a health hazard to on-site personnel due to their radioactivity. Ultrasonic measurement, capacitance tomography, and optical methods involve complex and expensive equipment, resulting in high costs and poor economic efficiency, making them unsuitable for practical engineering applications. Therefore, it is essential to develop a gas-liquid two-phase flow pattern identification method that is highly accurate, easy to operate, and economical.

[0004] In gas-liquid two-phase flow processes, the different physical properties of the two phases lead to different flow patterns and flow noise signals depending on the flow velocity. These noise signals contain rich physical and dynamic process information. Compared to flow pattern identification methods using pressure, pressure difference signals, or phase density image reconstruction, noise signals typically have sampling frequencies reaching several megahertz, resulting in shorter sampling times and better economic efficiency, making them suitable for identifying gas-liquid two-phase flow patterns in industrial processes. Therefore, developing a new method for flow pattern identification using gas-liquid two-phase flow noise signals has become a promising technological direction. Summary of the Invention

[0005] The purpose of this invention is to provide a method for identifying the flow pattern of gas-liquid two-phase flow based on flow noise signals. This method solves the problem that existing technologies are unable to accurately and quickly identify the flow pattern of gas-liquid two-phase flow in pipeline systems during deep-sea oil and gas extraction, utilization, and transportation, resulting in insufficient timeliness and accuracy in the detection of gas-liquid two-phase flow and multiphase flow parameters, and inadequate protection of system safety.

[0006] The technical solution adopted in this invention is a method for identifying the flow pattern of a gas-liquid two-phase flow based on flow noise signals, which is implemented according to the following steps:

[0007] Step 1: Collect the flow noise signal of the gas-liquid two-phase flow in the pipeline system, classify the collected flow noise signals under different flow conditions, and construct the original dataset;

[0008] Step 2: Preprocess the original dataset to obtain the three components of the IMF component of each flow noise signal;

[0009] Step 3: Using signal analysis methods, extract the characteristic parameters of the three components of the flow noise signal IMF;

[0010] Step 4: After processing the feature parameters extracted in Step 3, the sample set is divided into a training set and a test set. The training set is input into the SVM training module to train the SVM flow pattern recognition model of gas-liquid two-phase flow. The test set is used as input parameters to input into the flow pattern SVM flow pattern recognition model obtained in the test to obtain the flow pattern recognition result and verify the recognition accuracy of the trained model.

[0011] Step 5: Extract feature parameters from the unknown manifold noise signal according to the optimal feature subset determined in Step 3, construct a feature vector, and input it into the trained SVM manifold recognition model to obtain the manifold recognition result.

[0012] The beneficial effects of this invention include the following aspects:

[0013] 1) By collecting the flow noise signal of gas-liquid two-phase flow in pipeline, a method for identifying the flow pattern of gas-liquid two-phase flow in riser is proposed. The flow noise signal contains very rich dynamic process information, which is closer to the essence of flow. The signal acquisition is fast, economical, and has a high recognition accuracy, making it more suitable for identifying the flow pattern of gas-liquid two-phase flow in industrial processes.

[0014] 2) The genetic algorithm is used to optimize the variational mode decomposition parameters, which prevents over-decomposition and under-decomposition, improves the decomposition effect, reduces the noise factor and instability of the original signal, and improves the data quality. At the same time, the Pearson correlation coefficient and variance contribution rate are combined to screen components, which reduces the model training time and makes the recognition process more efficient.

[0015] 3) The RF-RFE algorithm is used to reduce the dimensionality of the feature vectors, optimize the analysis efficiency, reduce the subsequent model training and prediction time, and improve the speed and real-time performance of the algorithm in identifying manifolds;

[0016] 4) By optimizing the SVM parameters using an improved PSO algorithm, the optimization process better balances local and global search capabilities, resulting in higher search efficiency and avoiding premature convergence and getting trapped in local optima. This also improves the accuracy of the method in identifying flow patterns. The identification method of this invention is simple to operate, fast, has good real-time performance, is economical, has high accuracy, and is widely applicable. It provides a reference for identifying gas-liquid two-phase flow patterns using flow noise signals and for subsequent flow pattern identification methods. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the method of the present invention;

[0018] Figure 2 This refers to the flow noise signal collected by the method of this invention;

[0019] Figure 3 This is a flowchart of the genetic algorithm for optimizing variational mode decomposition parameters according to the present invention;

[0020] Figure 4 This is a flowchart of the improved particle swarm optimization algorithm for optimizing SVM parameters according to the method of this invention;

[0021] Figure 5 This is a classification result diagram of manifold recognition performed on a test set according to an embodiment of the method of the present invention. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0023] Reference Figure 1 The gas-liquid two-phase flow pattern identification method of the present invention is implemented according to the following steps based on the flow noise signal:

[0024] Step 1: Collect the flow noise signal of the gas-liquid two-phase flow in the pipeline system using a hydrophone to obtain the original time series. S(t) The collected flow noise signals under different flow patterns are classified to construct the original dataset. S i (t) ;

[0025] Specifically, the following steps were taken: Flow noise signals were collected using a hydrophone; the apparent velocities of the gas and liquid phases in the pipeline were continuously adjusted; and noise signals of the gas-liquid two-phase flow in the pipeline under different flow patterns were collected to obtain the original time series. S(t) ,See Figure 2 The collected flow noise signals under different flow patterns were classified to obtain the original dataset. S i (t) ;

[0026] Step 2, using the original dataset S i (t) Preprocessing is performed to obtain the IMF components of each flow noise signal after processing. Imf 1 (t) , Imf 2 (t) , Imf 3 (t) The specific process is as follows:

[0027] 2.1) Process the original dataset collected in step 1 using wavelet thresholding denoising method. S i (t) The original signal is decomposed into coefficients at each level using wavelet decomposition. A threshold is given, and each decomposed coefficient is compared to the threshold. Coefficients greater than the threshold are retained. The signal is then reconstructed using inverse transform to obtain the flowing noise signal. S’(t) ;

[0028] 2.2) Reference Figure 3 The flow noise signal after noise reduction in step 2.1) is processed by applying an optimization algorithm combined with variational mode decomposition. S’(t) Decompose;

[0029] To determine the parameters of variational mode decomposition (VMD) to prevent over- and under-decomposition, intelligent optimization algorithms are needed to find the optimal parameters. The number of decomposition levels K and the penalty factor α in the VMD process have a significant impact on the decomposition results. Therefore, a genetic algorithm is used to globally optimize the optimal combination of parameters [K, α] for VMD in order to obtain the best decomposition effect.

[0030] Information entropy can be used to characterize the amount of information contained in a variable when it occurs. The smaller the information entropy, the smaller the signal noise and the more effective information. The optimization is performed using the local minimum envelope entropy as the fitness function. The specific steps of the optimization are as follows:

[0031] 2.2.1) Initialize the genetic algorithm parameters and determine the range of values ​​for the number of decomposition layers K and the penalty factor α in the variational mode decomposition process;

[0032] 2.2.2) Perform variational mode decomposition on the flow noise signal and calculate the local minimum envelope entropy of the individual according to equation (1);

[0033] 2.2.3) Use the local minimum envelope entropy as the fitness function to perform a global search for the minimum value;

[0034] 2.2.4) Determine if the maximum number of generations has been reached. If it has, stop iterating and output the optimal solution; otherwise, continue the process.

[0035] 2.2.5) Perform selection, crossover, and mutation operations on the population to obtain the next generation population, and return to step 2.2.2) to calculate the envelope entropy of the individuals in the next generation population and continue the operation until the iteration stopping condition is met;

[0036] 2.2.6) Output the optimal parameter combination [K, α].

[0037] The formula for calculating the envelope entropy is as follows:

[0038] (1)

[0039] In the formula, a(t) It is the envelope signal obtained after variational mode decomposition and Hilbert demodulation of the modal components. x’(t) Compared with the original signal x(t) The sum of, N For signal length, E p Envelope entropy;

[0040] 2.3) After the flow noise signal is decomposed by variational mode decomposition algorithm, several IMF components are obtained. The signal characteristics of these IMF components are more obvious. The Pearson coefficients of each component with the original signal are calculated according to the cross-correlation method, and the IMF components are screened by combining the variance contribution rate. The three components with the highest similarity to the original signal are retained. Imf 1 (t) , Imf 2 (t) , Imf 3 (t) As a target for subsequent feature extraction, it reduces the noise and instability of the original signal, improves data quality, and reduces model training time;

[0041] The formula for calculating the Pearson correlation coefficient is:

[0042] (2)

[0043] In the formula, cov(X,Y) It is the covariance between the original signal and the IMF component. σ x and σ Y These are the standard deviations of the original signal and the IMF component, respectively.

[0044] The variance contribution rate is the ratio of the variance of the IMF component to the variance of the original signal, and is calculated as follows:

[0045] (3)

[0046] In the formula, X var 、Yvar These are the variance values ​​of the IMF components and the original signal sequence, respectively.

[0047] Step 3: Using signal analysis methods, extract the three components of the flow noise signal IMF obtained in Step 2. Imf 1 (t) , Imf 2 (t) , Imf 3 (t) The feature parameters are defined as follows:

[0048] 3.1) Extract features from the IMF components retained in step 2, including kurtosis, root mean square value, standard deviation, Hilbert marginal spectral energy, and permutation entropy, to form the original feature dataset;

[0049] The methods for extracting kurtosis, root mean square value, standard deviation, Hilbert marginal spectral energy, and permutation entropy are as follows:

[0050] First, the formula for calculating kurtosis is as follows:

[0051] (4)

[0052] In the formula n Represents the length of the signal sequence. μ Represents the average value of the signal sequence;

[0053] Secondly, the formula for calculating the root mean square value is as follows:

[0054] (5)

[0055] In the formula n Represents the length of the signal sequence;

[0056] Thirdly, the formula for calculating the standard deviation is as follows:

[0057] (6)

[0058] In the formula n Represents the length of the signal sequence. μ Represents the average value of the signal sequence;

[0059] Fourth, the solution process for Hilbert's marginal spectral energy is as follows:

[0060] A1) Perform a Hilbert transform on the noise signal IMF components obtained after variational mode decomposition and component filtering in step 2 to obtain the Hilbert spectrum, i.e.:

[0061] (7)

[0062] in,RC This indicates the operation of taking the real part. A i (t) Indicates instantaneous amplitude. F i (t) Indicates instantaneous frequency;

[0063] A2) Solve for the Hilbert marginal spectrum, using the following formula:

[0064] (8)

[0065] A3) Calculate the Hilbert marginal spectral energy using the following formula:

[0066] (9)

[0067] in, F1 , F2 These are the frequency ranges of Hilbert's marginal spectrum;

[0068] Fifth, the process of calculating the permutation entropy is as follows:

[0069] B1) For the selected IMF components (in... X(1), X(2)……X(n) (Representation), setting the embedding dimension m and the time delay t. delay ;

[0070] B2) By embedding dimension m and time delay t delay Phase space reconstruction is performed on each element of the original signal to obtain K =n-(m-1) t delay There are subsequences, each subsequence being named _ ... Y(i) express, Y(i) = X(i) , X(i + t delay )……X(i +(m-1) t delay ) ;

[0071] B3) For each Y(i) Sort in ascending order internally;

[0072] B4) Calculate the probability of each possibility occurring. P j ;

[0073] B5) Calculate the permutation entropy of these probabilities:

[0074] (10)

[0075] 3.2) Using the RF-RFE algorithm, the feature parameters of each component extracted in step 2.3) are ranked by feature importance to obtain the optimal feature subset. Feature vectors are then constructed, normalized, and further dimensionality reduction is performed to optimize analysis efficiency.

[0076] Using the RF-RFE algorithm for data dimensionality reduction mainly includes two steps: feature importance ranking and recursive feature elimination. The specific process is as follows:

[0077] C1) Extract features from the original feature dataset using the Bootstrap resampling method. X There are training samples, each sample has N The features that are not selected are called out-of-bounds (OOB) data.

[0078] C2) Initialization i =1, create a decision tree T i ;

[0079] C3) for the first i Training on samples T i Calculate the classification accuracy of the i-th OOB data. A i ;

[0080] C4) Features in the OOB dataset F j After adding random noise interference to obtain a new OOB dataset, the classification accuracy is calculated again. A ij ;

[0081] C5) For i =2,3..., repeat steps C2) to C4).

[0082] C6) Features F j The importance is calculated using the following formula:

[0083] (11)

[0084] C7) Rank the calculated feature importance and use cross-validation to obtain the classification accuracy of the initial feature subset;

[0085] C8) Remove the least important feature from the current feature subset, and calculate the feature importance of the new feature subset and the classification accuracy under cross-validation;

[0086] C9) Recursively repeat step C8) until the feature subset is empty, and take the subset with the highest classification accuracy as the optimal feature subset.

[0087] Step 4: After processing the feature parameters extracted in Step 3, use them as a sample set. The sample set is divided into a training set and a test set. Input the training set into the SVM training module to train the SVM flow pattern recognition model of gas-liquid two-phase flow. Input the test set as input parameters into the SVM flow pattern recognition model obtained by testing to obtain the flow pattern recognition result and verify the recognition accuracy of the trained model.

[0088] The specific process is as follows: the collected noise signal is processed by extracting feature parameters according to the optimal feature subset obtained after optimization in step 3, and a feature vector is constructed through normalization. At the same time, a one-dimensional element is added to the sample set as a label according to the flow pattern of the signal to indicate the attribute of the feature vector: 1 represents bubble flow, 2 represents slug flow, and 3 represents circumferential flow. The processed data is randomly divided into a training set and a test set. The data in the training set is used to train the SVM flow pattern recognition model. The data in the test set is used as input to verify the accuracy of the trained SVM flow pattern recognition model.

[0089] Reference Figure 4 In the process of training the SVM manifold recognition model, in order to determine the values ​​of the penalty parameter c and the kernel function parameter g during training, an improved particle swarm optimization algorithm is used to optimize the parameters. The specific process is as follows:

[0090] D1) Use the classification accuracy under cross-validation as the fitness function for algorithm optimization;

[0091] D2) Initialize the population and particle velocities, positions, and maximum number of iterations. Input the population particle information into the fitness function to calculate the initial fitness of the population, and use the initial particle fitness value as the individual extreme value. P best The individual's historical optimal fitness value is used as the global optimal extreme value. g best ;

[0092] D3) Update particle position and velocity, individual optimal values ​​and global optimal values ​​of the particle population using the following formula. Also, update the corresponding optimal values ​​when particle velocity and position exceed their original maximum or minimum values:

[0093] (12)

[0094] In the formula, w This represents the inertia weight, which changes with the number of evolutions and can better balance global search capability and local search capability. w max ,w min Take values ​​of 0.9 and 0.4 respectively; c 1 ,c 2 represents the acceleration coefficient, which is typically set to 2.0; r 1，r 2 is a random number between (0, 1), and its index is... i,j Indicates the first i The particle and the first j dimension, t Represents the number of iterations. t max Indicates the maximum number of generations;

[0095] D4) Perform adaptive mutation operation, using the mutation factor v calculated by the following formula as the probability to initialize the current position of the mutated particle. In the early stage of evolution, the mutation factor is larger, the global search capability is stronger, and the search is carried out in a larger area to prevent getting trapped in local optima. In the later stage of evolution, the mutation factor is smaller, and the search is concentrated near the excellent individuals to speed up the convergence speed of the algorithm.

[0096] (13)

[0097] In the formula, take v min ,v max They are 0.1 and 0.7 respectively. t Represents the number of iterations. t max Indicates the maximum number of generations;

[0098] D5) Calculate the mutated particle fitness value, compare the current particle population fitness value with the individual extreme value and the global optimal extreme value, and update the individual optimal and global optimal extreme values ​​of the particle population.

[0099] D6) Determine whether the iteration termination condition is met. If it is met, output the optimal parameter combination; otherwise, return to step D3) until the iteration termination condition is met.

[0100] Step 5: Extract feature parameters from the unknown manifold noise signal according to the optimal feature subset determined in Step 3, construct a feature vector, and input it into the trained SVM manifold recognition model to obtain the manifold recognition result.

[0101] Experimental verification:

[0102] The two-phase media used in the experiment were air and water. The apparent velocity of the gas phase was adjusted from 0.1 to 5 m / s, and the apparent velocity of the liquid phase was adjusted from 0.1 to 2 m / s. Hydrophones were used to collect the two-phase flow noise signal at a sampling frequency of 51.2 kHz. The flow pattern was adjusted by regulating the apparent velocities of the gas and liquid phases. A total of 150 samples were obtained, which were randomly divided into training and testing sets at a ratio of 4:1. 120 samples were used as the training set to train the SVM flow pattern recognition model, and the remaining 30 samples were used to verify the accuracy of the recognition model. Figure 5As shown, the trained SVM manifold recognition model achieved a recognition accuracy of 96.7% on the test set, meaning that only one sample was misidentified.

[0103] Although the sampling frequency of the original noise signal, which can reach several MHz, cannot be achieved in this example due to the hardware limitations of the data acquisition card, it is still much higher than the sampling frequency of commonly used methods such as pressure signals, conductivity probes, and phase density imaging. Moreover, the signal acquisition is simple and easy to implement.

[0104] Therefore, the method described in this invention can accurately identify the flow pattern of gas-liquid two-phase flow. Furthermore, the method is designed based on flow noise signals, has a high signal sampling frequency, low equipment requirements, is easy to implement, and can achieve rapid flow pattern identification, making it more suitable for industrial applications.

Claims

1.A method for flow pattern identification of gas-liquid two-phase flow based on flow noise signal, characterized in that, The following steps are implemented: Step 1, collect the gas-liquid two-phase flow noise signal in the pipeline system, classify the collected flow noise signals under different flow conditions, and construct the original data set, the specific process is as follows: The flow noise signal is collected by a hydrophone, the superficial velocities of the gas phase and the liquid phase are adjusted, the noise signals of the gas-liquid two-phase flow under different flow type conditions are collected, and original time series are obtained S(t) The collected flow noise signals under different flow type conditions are classified to obtain original data sets S i (t) ; Step 2, pre-process the original data set to obtain 3 components of the IMF component of each flow noise signal, the specific process is as follows: 2.1) Processing the original data set collected in step 1 by wavelet threshold denoising method S i (t) , decomposing the original signal into each layer coefficient by wavelet decomposition, comparing each decomposition coefficient with the threshold value, retaining the decomposition coefficient greater than the threshold value, and reconstructing the signal by inverse transformation to obtain the flow noise signal S'(t) ; 2.2) Apply optimization algorithm combined with variational mode decomposition method to the flow noise signal after denoising processing in step 2.1) S'(t) decomposition; The decomposition level K and the penalty factor a of the variational mode decomposition process have a great influence on the decomposition result, therefore the optimal combination parameters [K, a] of the variational mode decomposition are globally optimized by genetic algorithm, and the local minimum envelope entropy is used as the fitness function for optimization, the specific steps of optimization are as follows: 2.2.1) initialize the genetic algorithm parameters, determine the value range of the decomposition level K and the penalty factor a in the variational mode decomposition process; 2.2.2) perform variational mode decomposition on the flow noise signal, calculate the local minimum envelope entropy of each individual according to formula (1); 2.2.3) perform global search on the local minimum envelope entropy as the fitness function to find the minimum value; 2.2.4) determine whether the maximum evolution generation is reached, if yes, stop iteration and output the optimal solution; otherwise, continue operation; 2.2.5) select, cross and mutate the population to obtain the next generation population, return to step 2.2.2) to calculate the envelope entropy of the next generation population and continue operation; until the iteration stopping condition is met; 2.2.6) output the optimal parameter combination [K, a]; The calculation formula of envelope entropy is as follows: （1） In the formula, a(t) is the envelope signal obtained after the modal component after the Hilbert demodulation after the variational modal decomposition x’ (t) and the original signal x(t) , N is the signal length, E p is the envelope entropy; 2.3) After the flow noise signal is decomposed by the variational mode decomposition algorithm, a plurality of IMF components are obtained, the Pearson coefficients of each component and the original signal are calculated according to the cross-correlation method, and the IMF components are screened in combination with the variance contribution rate, and the three components with the highest similarity to the original signal are retained Imf 1 (t) 、 Imf 2 (t) 、 Imf 3 (t) , as the subsequent feature extraction object; The calculation formula of Pearson correlation coefficient is as follows: （2） wherein cov(X,Y) is the covariance of the original signal and the IMF component, σ x and σ Y are the standard deviations of the original signal and the IMF component, respectively. The variance contribution rate is the ratio of the variance of the IMF component to the variance of the original signal, and the calculation formula is as follows: （3） wherein X var 、Y var are the variance values of the IMF components and the original signal sequence, respectively; Step 3, use signal analysis method to extract the feature parameters of the 3 components of the flow noise signal IMF, the specific process is as follows: 3.1) extract the feature parameters of each component, including kurtosis, root mean square value, standard deviation, Hilbert marginal spectrum energy and permutation entropy, to form the original feature data set; the extraction methods of kurtosis, root mean square value, standard deviation, Hilbert marginal spectrum energy and permutation entropy are as follows: First, the calculation formula of kurtosis is as follows: （4） In the formula n represents the length of the signal sequence, μ represents the average value of the signal sequence; Second, the calculation formula of root mean square value is as follows: （5） In the formula n represents the length of the signal sequence; Third, the calculation formula of standard deviation is as follows: （6） In the formula n represents the length of the signal sequence, μ represents the average value of the signal sequence; Fourth, the solving process of Hilbert marginal spectrum energy is as follows: A1) perform Hilbert transform on the noise signal IMF component obtained after variational mode decomposition and screening in step 2 to obtain Hilbert spectrum, i.e.: （7） wherein, RC represents a real part operation, A i (t) represents an instantaneous amplitude, F i (t) represents an instantaneous frequency; A2) solve the Hilbert marginal spectrum, the calculation formula is as follows: （8） A3) calculate the Hilbert marginal spectrum energy, the calculation formula is as follows: （9） wherein F1 , F2 are the frequency bins of the Hilbert marginal spectrum, respectively. Fifth, the solving process of permutation entropy is as follows: B1) for the screened IMF components, set the embedding dimension m and the time delay t X(1), X(2)……X(n) representing, set the embedding dimension m and the time delay t delay ; B2) by embedding dimension m and time delay t delay The phase space reconstruction is performed for each element extension of the original signal, resulting in K = n - (m - 1) t delay subsequences, wherein each subsequence is represented by Y(i) Y(i) = t X(i) , X(i+ t delay )……X (i+(m-1) t delay ) ;​ B3) for each Y(i) ordered in ascending order. B4) calculating the probability of occurrence of each possibility P j ; B5) calculate the permutation entropy of these probabilities: （10） 3.2) Using RF-RFE algorithm, the feature parameters of each component extracted in step 2.3) are sorted according to feature importance to obtain the optimal feature subset, construct a feature vector, and perform normalization processing, The RF-RFE algorithm is used for data dimensionality reduction, including two steps of feature importance sorting and recursive feature elimination, and the process is as follows: C1) Bootstrap resampling method is used to extract X training samples from the original feature dataset, each sample has N features, and the unextracted data is called OOB data; C2) Initialization i = 1, create decision tree T i ; C3) training a first classifier on the first set of samples i T i , calculating a classification accuracy of the i-th OOB data A i ;​ C4) Features in the OOB dataset F j Random noise is added to the data set to get a new OOB dataset, and the classification accuracy is calculated again A ij ; C5) for i = 2, 3,..., repeat steps C2) to C4); C6) Feature F j The importance is calculated with the following formula: （11） C7) The calculated feature importance is sorted, and the classification accuracy of the initial feature subset is obtained by using the cross-validation method; C8) A feature with the smallest importance is deleted from the current feature subset, and the feature importance and classification accuracy of the new feature subset under cross-validation are calculated; C9) Step C8) is recursively repeated until the feature subset is empty, and the subset with the highest classification accuracy is taken as the optimal feature subset; Step 4: The feature parameters extracted in step 3 are processed as a sample set, the sample set is divided into a training set and a test set, the training set is input into the SVM training module, and the SVM flow pattern recognition model of the gas-liquid two-phase flow is trained, the test set is input into the trained flow pattern SVM flow pattern recognition model as an input parameter, and the flow pattern recognition result is obtained, and the recognition accuracy of the trained model is verified, and the specific process is as follows: The collected noise signals are subjected to feature parameter extraction according to the optimal feature subset obtained after optimization in step 3, a feature vector is constructed through normalization processing, and a one-dimensional element is added as a label in the sample set according to the flow pattern of the signal to indicate the attribute of the feature vector, 1 represents bubbly flow, 2 represents slug flow, and 3 represents annular flow; The processed data is randomly divided into a training set and a test set, the data of the training set is used to train the SVM flow pattern recognition model; the data of the test set is used as an input to verify the accuracy of the trained SVM flow pattern recognition model, In the process of training the SVM flow pattern recognition model, in order to determine the values of the penalty parameter c and the kernel function parameter g in the training process, the improved particle swarm algorithm is used to optimize the parameters, and the process is as follows: D1) The classification accuracy under cross-validation is used as the fitness function of the algorithm optimization; D2) initialize population and particle velocity, position and the maximum number of iterations, bring population particle information into fitness function to calculate initial fitness of population, take initial particle fitness value as individual extremum P best , individual historical optimal fitness value as global optimal extremum g best ; D3) The particle position and velocity, individual optimal and global optimal extreme value of the particle population are updated by the following formula, and when the particle velocity and position exceed the original maximum value, the corresponding maximum value is updated: （12） wherein w represents the inertia weight, which varies with the number of generations, w max ,w min are taken as 0.9 and 0.4, respectively; c 1 ,c 2 is an acceleration coefficient; r 1 ，r 2 is a random number between (0, 1), the subscript i,j denotes the i th particle and the j th dimension, t represents the number of iterations, t max denotes the maximum number of generations; D4) Perform adaptive mutation operation, calculate the mutation factor v according to the following formula, and use it as a probability to initialize the mutation of the current particle position to speed up the convergence speed of the algorithm; （13） wherein v min ,v max are 0.1 and 0.7, respectively, t is the number of iterations, t max is the maximum number of evolutionary generations; D5) Calculate the fitness value of the mutated particle, compare the fitness value of the current particle population with the individual optimal and global optimal extreme value, and update the individual optimal and global optimal extreme value of the particle population; D6) Determine whether the iteration termination condition is met, if yes, output the optimal parameter combination, otherwise return to step D3) until the iteration termination condition is met; Step 5: After the feature parameters of the unknown flow pattern noise signal are extracted according to the optimal feature subset determined in step 3, a feature vector is constructed and input into the trained SVM flow pattern recognition model to obtain the flow pattern recognition result.

Images