CN112114047B - GAs-liquid flow parameter detection method based on acoustic emission-GA-BP neural network - Google Patents

Abstract

The method for measuring the flow parameters of the GAs-liquid two-phase flow based on the acoustic emission-GA-BP neural network comprises the steps of setting a threshold according to environmental noise; collecting the sound signal and the pressure difference and the liquid holdup; calculating acoustic signal parameters under each flow condition; analyzing wavelet packets of acoustic signals at each gas-liquid flow rate; determining an initial structure of a BP neural network; determining the population size, the coding length and the fitness in a genetic algorithm; selecting a genetic operator; calculating an optimal weight and a threshold; checking the accuracy of the model; and calculating pressure drop and liquid holdup parameters of the marine oil-gas mixing pipeline system. The invention aims to solve the defects of the BP neural network, adopts a genetic algorithm to optimize the BP neural network, avoids the defects of slow iteration and sinking into a local optimal solution of the BP neural network, has wide application range and has the accuracy rate of more than 95 percent. Besides submarine gathering and transportation, the invention is also suitable for detecting the pressure drop and the liquid holdup flow parameters of the gas-liquid two-phase flow in the industrial fields of power engineering, nuclear energy utilization, chemical industry and the like.

Description

Gas-liquid flow parameter detection method based on acoustic emission-GA-BP neural network

Technical Field

The present invention relates to the field of multiphase flow detection, in particular to the field of gas-liquid two-phase flow parameter detection, and specifically to a gas-liquid flow parameter detection method based on acoustic emission-GA (genetic algorithm)-BP neural network, so as to perform real-time online detection of pipeline pressure drop and liquid holdup of gas-liquid two-phase flow parameters in an inclined (20-90°) or riser of an offshore oil and gas mixed transportation system.

Background technique

As terrestrial oil resources are increasingly depleted, people are increasingly focusing on submarine oil and gas resources. In the offshore oilfield production system, in order to save transportation costs, oil and gas are mostly transported in a mixed manner. In practical engineering applications, in order to monitor the flow state in the pipeline in real time, it is necessary to detect a variety of two-phase flow parameters in the pipeline. The two-phase flow in the riser gathering and transportation system is affected by many factors, and the real-time measurement of flow parameters is of great significance, such as liquid holdup, pressure drop, two-phase flow velocity, flow type, etc. Among them, pipeline pressure drop and liquid holdup are important operating parameters. Pressure drop reflects the resistance loss and pressure change along the oil and gas flow in the pipeline, and is one of the basic parameters of pipeline operation. The upstream wellhead back pressure caused by the pressure drop of the mixed pipeline has an important impact on the oil well production, and is an important parameter for the configuration and operation of the mixed pump. At the same time, the pressure drop also reflects the changes in wax deposition and hydrate formation in deepwater oil and gas pipelines. Liquid holdup reflects the ratio of liquid to the cross section of the pipeline, which is of great significance for understanding the distribution of the oil and gas interface in the pipeline and determining the pipeline cleaning cycle. Therefore, pressure drop and liquid holdup are important parameters in the design and operation of oil and gas mixed transportation pipelines.

Gamma rays are a commonly used method for measuring liquid holdup at oil and gas production sites. However, their application is limited by the strict and cumbersome supervision of ray sources and the inconvenience of equipment installation. Patent No. CN201910250057.X provides a method for detecting liquid holdup. Through a pressure vessel partial pressure close to the volume of the pipeline, the liquid holdup of the pipeline can be detected by the pressure coefficient before and after the pressure drop of the container in a specific range, but it does not have universality and applicability. Acoustic emission technology is a non-destructive testing technology. With the advantages of quick installation and simple operation, it has been successfully used in the fields of material defects, crack detection and pipeline leakage. Applying acoustic emission technology to the identification of two-phase flow parameters can meet the requirements of non-destructive and rapid online detection. Application No. 2020102334040 provides a method for identifying flow patterns of two-phase flow based on acoustic emission-BP neural network. The patent uses acoustic emission data of different gas-liquid velocities under inclined pipes, combined with BP neural network, to establish a correlation model between acoustic emission data and flow patterns of inclined risers, and realizes online identification of flow patterns. This patent application provides ideas for the feasibility of using acoustic emission data to predict flow parameters.

The biggest feature of BP neural network is that it can solve nonlinear problems well, but it also has the following disadvantages: First, BP neural network converges slowly. BP neural network randomly selects initial values and thresholds. If the selection is poor, it will cause the iteration of the neural network in this part to converge slowly and the calculation will take a long time. Secondly, the learning method of BP neural network is gradient descent, but nonlinear problems usually have multiple extreme points in the function interval, which causes the neural network to fall into the local optimal solution and cannot achieve accurate prediction of parameters.

Therefore, there is an urgent need for an online detection method for gas-liquid two-phase flow parameters, which not only needs to be effective in measuring pipeline pressure drop and liquid holdup, but also should overcome the above problems of the existing BP network.

Summary of the invention

The purpose of the present invention is to provide a gas-liquid flow parameter detection method based on acoustic emission-GA (genetic algorithm)-BP neural network, which can identify the pressure drop and liquid holdup of gas-liquid two-phase flow parameters in an inclined (20-90°) or vertical high-pressure pipeline of a marine oil and gas mixed transportation system online.

The present invention extracts and analyzes the acoustic emission signal characteristics under gas-liquid flow conditions with different gas-liquid converted velocity combinations, establishes a model of signal characteristics and flow parameters, and learns the sample signal characteristics under different flow conditions through a GA-BP neural network to achieve the goal of calculating the pressure drop and liquid holdup of the two-phase flow parameters.

The method for measuring flow parameters of gas-liquid two-phase flow based on acoustic emission-GA-BP neural network is characterized by comprising the following steps:

Step 1: Set the threshold according to the ambient noise

An acoustic emission sensor (1) is arranged on the outer wall of an inclined pipeline or a vertical high-pressure pipeline in the range of 20-90 degrees in a marine oil and gas mixed transmission system;

When the gas velocity and liquid velocity in the pipeline are both zero, empty pipe collection is performed, and the maximum value of the empty pipe signal is set as the threshold voltage.

Step 2: Collect acoustic signals, pressure difference, and liquid holdup

When the gas-liquid two-phase flow in the pipeline is in the working condition of different gas-liquid flow rate combinations, the two-phase flow acoustic signal is collected by an acoustic emission sensor (1); while collecting the acoustic signal, the pressure difference and liquid holdup are synchronously recorded by a pressure difference sensor and a dual parallel conductivity probe, which are used to train and verify the neural network model established in the subsequent steps.

Step 3: Calculate the acoustic signal parameters under various flow conditions

The collected original waveform data of the acoustic signal is statistically analyzed to calculate the acoustic signal amplitude, average voltage level, root mean square value, absolute energy value and ringing count.

Amplitude AMP, in dB, is defined as:

Wherein, V _max is the maximum value of the voltage data in the two-phase flow acoustic signal, in V.

The average voltage level ASL, in dB, is defined as:

Wherein, V _mean is the average value of the voltage data in the two-phase flow acoustic signal, in V.

The root mean square value RMS, in V, is defined as:

Wherein, V is the voltage signal of each data point in the two-phase flow acoustic signal, in V; n is the number of acoustic signal data points.

Absolute energy value ABS, in J, is defined as:

Wherein, V is the voltage signal of each data point in the two-phase flow acoustic signal, unit V; 10KΩ is the reference resistor; T is the sampling time, unit s;

Ringing count Counts, which indicates the number of oscillations that cross the threshold signal, that is, the number of effective peaks that exceed the threshold voltage;

When analyzing the correlation between flow parameters and various statistical parameters, it was found that the amplitude value AMP has strong randomness in a large time window and has basically no correlation with the flow parameters, so it was discarded.

Step 4: Analyze the wavelet packets of the acoustic signal at each gas-liquid flow rate

Wavelet packets are composed of a series of wavelet packet basis functions. Since different wavelet packet bases have different time-frequency characteristics, different wavelet packet bases will produce different results for the same signal. Therefore, choosing a suitable wavelet packet basis is very important for accurately extracting signal features.

Step 4.1: Select wavelet packet basis function

Select Symlets8 wavelet basis function to decompose wavelet packet;

Step 4.2: Decompose the wavelet packet

If the sampling frequency of the signal is _fs , according to the Nyquist sampling theorem, the measurable frequency range of the signal is [0, _fs /2]; because within the range of the signal frequency, the distribution range of the detail signal and the approximate signal is symmetrical, when the decomposition scale is 1, [0, _fs /4] and [ _fs /4, _fs /2] are the frequency ranges of the approximate signal and the detail signal. After performing J wavelet packet decompositions on the signal f(n) with a sampling frequency of _fs , the signal is decomposed into ^2J frequency segments. The calculation formulas for each frequency range are as follows:

If the sampling frequency is _fs kHz, the minimum recognition frequency required by the signal is _fmin . According to equation (5), the maximum decomposition scale J should satisfy:

Right now:

The decomposition scale is generally selected as J=3~4; after wavelet packet decomposition, the reconstructed waveform signal of each frequency band is obtained, and the norm square of the reconstructed waveform signal is calculated as the wavelet packet energy of each node.

Step 5: Determine the initial structure of the BP neural network

Generally, when designing a neural network, a three-layer network is given priority. The four statistical parameters of the acoustic signal, namely the average voltage level, root mean square value, absolute energy value, and ringing count, and the energy of the wavelet packet at 2 ^J frequency bands, i.e., the square norm of the reconstructed waveform, totaling 4+2 ^J eigenvalues, are used as the input of the BP neural network input layer. Therefore, the number of nodes in the input layer is the same as the number of statistical features of the input, which is 4+2 ^J.

The number of nodes in the output layer is 1; the number of nodes in the hidden layer is determined by the empirical formula for calculating the number of elements in the hidden layer. The commonly used empirical formula is as follows:

h＝2*m+1 (8)

Among them, h is the number of hidden layer nodes, and m is the number of input layer nodes.

In order to facilitate calculation and comparison, the input layer and output layer neurons need to be normalized. The normalization formula is:

Among them, α is the eigenvalue after normalization, _xi is the eigenvalue before normalization, _xmax and _xmin are the maximum and minimum values of the eigenvalue before normalization, respectively, i is the eigenvalue sequence number, indicating the i-th eigenvalue, i∈1,…,n; _ymax and _ymin are the expected maximum and minimum values after normalization, which are generally defaulted to 1 and -1.

Define the flow parameters - pipeline pressure drop or liquid holdup as the output layer, that is, the output layer has only one neuron.

Step 6: Determine the population size, encoding length and fitness in the genetic algorithm

The genetic algorithm starts with a randomly selected initial solution and generates a new solution through selection, crossover, and mutation operations based on individuals in the previous generation population.

Set the population size: Genetic algorithms focus on the population, which is a set of multiple given initial solutions. Each initial solution is called an individual, or chromosome. During iterative operations, the population size remains unchanged, and the chromosomes in the population are constantly changing in the problem solution space. Therefore, the population size will affect the iteration rate. In the process of solving actual problems, the size and position of the initial population should be determined according to the actual situation, generally between 20 and 100.

Encoding is to convert data into chromosomes, that is, genetic string structure data of fixed form and length in genetic space. The present invention is designed to use binary encoding, that is, 0 and 1 in the binary character set {0,1} are used to represent candidate solutions to the problem.

The number of coding bits is often given based on actual problems, usually between 5 and 20.

The fitness value of an individual in a genetic algorithm is determined by a fitness function, which describes the quality of the chromosome. Therefore, the fitness function calculates comparable non-negative results for the problem, which are then compared by subsequent selection operators.

The fitness function used in this method is a ranking function. By comparing the errors between the weights and thresholds obtained by training in the subsequent steps and the weights and thresholds of the previous generation, the errors are sorted from small to large, and the calculation of the next step is performed.

Step 7: Selection of genetic operators

It includes the selection of selection operator, crossover operator and mutation operator.

This method uses random traversal sampling, and the probability of each individual being selected is equal, which is the inverse of the total number of individuals.

The crossover operator is the most important component of the genetic algorithm and the core of the operation. Crossover can preserve the excellent chromosomes in the initial population and generate a new population by exchanging chromosomes between the two parties, which increases the complexity and diversity of the population. Commonly used crossover operators include single-point crossover operators and two-point crossover operators. This method uses a single-point crossover operator:

Suppose the two parent strings are x = [x ₁ , x ₂ , L, x _n ] and y = [y ₁ , y ₂ , L, y _n ], and the fork point is k, then the generated offspring is:

x'＝[x1,x2,L, _xk ,yk _{+ 1} ,yk _{+ 2} ,L, _yn ] (14)

y'＝[y ₁ ,y ₂ ,L,y _k ,x _k+1 ,x _k+2 ,L,x _n ] (15)

In a single-point crossover, there is only one breakpoint on the chromosome. If the length of the parent string is n, then the single-point crossover has (n-1) different crossover results.

The crossover probability is generally between 0.4 and 0.9.

The mutation operator of this method is a random selection of a chromosome with a certain probability to change the data, usually as an auxiliary means to generate new species, with local search capabilities, further expanding the diversity of the population;

The mutation probability generally ranges from 0.001 to 0.1.

Step 8: Calculate the optimal weights and thresholds

At the selected pipeline inclination angle, select no less than 150 groups of measured acoustic emission signals of different flow conditions with different superficial gas velocity and superficial liquid velocity combinations, and then use steps 3 and 4 to calculate 4+2 ^J characteristic parameters of the acoustic emission signal of each condition, select 60% of all conditions as training samples, and 40% as test samples; select the pressure drop or liquid holdup corresponding to the training samples as the training sample results, and select the pressure drop or liquid holdup corresponding to the test samples as the test sample results; normalize the training samples, test samples, training sample results, and test sample result parameters.

Bring the normalized training samples, training sample results, test samples, and test sample results into the BP neural network of step 5 for training to obtain initial weights and thresholds;

Then, the genetic algorithm optimization calculation is performed, and the fitness calculation of the weight and threshold is performed according to the population size, encoding form and fitness function set in step 6, and the genetic operator set in step 7 is used for iteration;

After the iteration of the genetic operator, the next generation population is obtained; the population is evaluated by the fitness function set in step 6, and the iteration stops when the number of iterations reaches a predetermined value, at which time the optimal weight and threshold are obtained;

The selection range of the number of iterations is less than 200 times; when the set number of iterations is reached, the optimal individual obtained is the optimal weight and threshold of the BP neural network.

Step 9: Model Accuracy Test

The optimal weights and thresholds are reassigned to the neural network, and the test samples are substituted into the BP neural network. The calculated output results are compared with the actual results. If the calculated error is below 5%, the model is successfully established. If the calculated error does not meet the requirements, steps 6 and 7 are repeated until the calculated error meets the design requirements. The BP neural network at this time is the optimal BP neural network optimized by the genetic algorithm.

Step 10: Calculation of pressure drop and liquid holdup parameters for offshore oil and gas mixed pipeline system

Then, the acoustic emission sensor (1) arranged on the outer wall of the pipeline in step 1 is used to collect the acoustic emission signal, the acoustic signal parameters are calculated in the manner of step 3, the wavelet packet energy of the acoustic signal is calculated in the manner of step 4, and the obtained parameters and wavelet packet energy are input into the optimal model obtained after training in the above steps, so that the pressure drop and liquid holdup parameters of the marine oil and gas mixed transmission pipeline system can be obtained.

In the step 4, J=3 is selected as the wavelet packet decomposition scale.

In step 5, a linear transfer function is used between the input layer, the hidden layer and the output layer, the training function adopts the trainlm function, that is, the LM back propagation algorithm, the number of training times is set to 1000, the training target is set to 10 ^-5 , and the learning rate is set to 0.1.

In step 6, 10 is used as the optimal binary coding bit number.

In step 6, 0.7 is selected as the optimal crossover probability.

In step 7, 0.01 is used as the optimal mutation probability.

In step 8, the maximum number of iterations, i.e., the maximum genetic generation, is 50.

Advantages of the invention

In order to solve the shortcomings of BP neural network, the genetic algorithm (GA) is used to optimize the BP neural network. The prediction model optimizes the weights and thresholds on the basis of keeping the structure of the BP prediction model unchanged. The genetic algorithm is an adaptive heuristic global search algorithm. It imitates the inheritance and evolution of organisms. Based on Darwin's evolutionary principle of natural selection and survival of the fittest, it focuses on overall parallel search and finally obtains the optimal solution to the problem. This method is oriented to overall parallel search and probability-oriented search process, avoiding the defects of slow iteration of BP neural network and falling into the local optimal solution.

Acoustic emission technology is a passive non-destructive testing technology. The sensor only needs to be attached to the outer wall of the pipe. It is simple to operate and has strong adaptability to the environment. It can detect two-phase flow signals without damaging the pipeline. Carbon steel metal oil and gas pipelines are conducive to the propagation of acoustic signals. The present invention is a method for calculating the flow parameters of gas-liquid two-phase flow based on acoustic emission measurement. It provides a real-time, fast, and online method, which is of great value for the operation status detection of high-pressure thick-walled oil and gas pipelines and the safety monitoring of deep-water oil and gas flow. The present invention can identify the pressure drop and liquid holdup flow parameters of gas-liquid two-phase flow in inclined and vertical pipelines, with a wide range of applicable flow rates and an accuracy rate of more than 95%. In addition to submarine gathering and transportation, the present invention is also suitable for the detection of gas-liquid two-phase flow pressure drop and liquid holdup flow parameters in industrial fields such as power engineering, nuclear energy utilization, and chemical industry.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 is a flow chart of two-phase flow parameter identification based on acoustic emission technology.

Figure 2 is a schematic diagram of acoustic emission signal acquisition in a two-phase mixed pipeline.

Among them, 1-acoustic emission sensor, 2-signal amplifier, 3-signal acquisition box, 4-computer.

Figure 3 is a three-scale wavelet packet decomposition tree diagram.

Figure 4 is a wavelet packet decomposition and reconstruction diagram of the original signal at an apparent water velocity of 5 m/s and an apparent air velocity of 1.6 m/s at an inclination angle of 45°.

Figure 5 is a flow chart of GA-BP neural network training.

Table 1 shows the frequency range of each node in wavelet packet decomposition.

Table 2 shows the error comparison of BP neural network and GA-BP neural network in predicting pipeline liquid holdup.

Table 3 shows the error comparison of pipeline pressure drop prediction by BP neural network and GA-BP neural network.

Table 4 is an example of the implementation of identifying the liquid holdup of a 45-degree inclined pipeline based on acoustic emission-GA-BP neural network.

Table 5 is an example of the implementation of identifying the pressure drop of a 45-degree inclined pipeline based on acoustic emission-GA-BP neural network.

Detailed ways

The method of the present invention is further described in detail below in conjunction with FIG1:

The method for measuring the flow parameters of gas-liquid two-phase flow based on acoustic emission-GA-BP neural network, the calculation process is shown in Figure 1, and the steps include:

Step 1: Set the threshold according to the ambient noise

An acoustic emission sensor 1 is arranged on the outer wall of an inclined pipeline or a vertical high-pressure pipeline in the range of 20-90 degrees in a marine oil and gas mixed transportation system. The acoustic emission sensor 1 is usually arranged in a fully developed flow section.

When the gas velocity and liquid velocity in the pipeline are both zero, empty pipe collection is performed, and the maximum value of the empty pipe signal is set as the threshold voltage.

Step 2: Collect acoustic signals, pressure difference, and liquid holdup

When the gas-liquid two-phase flow in the pipeline is in the working condition of different gas-liquid flow rate combinations, the two-phase flow acoustic signal is collected by the acoustic emission sensor 1; while collecting the acoustic signal, the pressure difference and liquid holdup are synchronously recorded by the pressure difference sensor and the dual parallel conductivity probe (or gamma ray phase content meter) for training and verification of the neural network model established in the subsequent steps.

If the test conditions are not available on site, a large multiphase flow test loop with the same size, inclination, operating conditions and medium can be used at the experimental base to test the acoustic emission signal, pressure drop and liquid holdup parameters.

Step 3: Calculate the acoustic signal parameters under various flow conditions

The collected original waveform data of the acoustic signal is statistically analyzed to calculate the acoustic signal amplitude, average voltage level, root mean square value, absolute energy value and ringing count.

Amplitude AMP, in dB, is defined as:

Wherein, V _max is the maximum value of the voltage data in the two-phase flow acoustic signal, in V.

The average voltage level ASL, in dB, is defined as:

Wherein, V _mean is the average value of the voltage data in the two-phase flow acoustic signal, in V.

The root mean square value RMS, in V, is defined as:

Wherein, V is the voltage signal of each data point in the two-phase flow acoustic signal, unit is V; n is the number of acoustic signal data points.

Absolute energy value ABS, in J, is defined as:

Wherein, V is the voltage signal of each data point in the two-phase flow acoustic signal, unit is V; 10KΩ is the reference resistor; T is the sampling time, unit is s.

The ringing count Counts indicates the number of oscillations that cross the threshold signal, that is, the number of effective peaks that exceed the threshold voltage.

When analyzing the correlation between flow parameters and various statistical parameters, it was found that the amplitude value AMP has strong randomness in a large time window and has basically no correlation with the flow parameters, so it was discarded.

Step 4: Analyze the wavelet packets of the acoustic signal at each gas-liquid flow rate

Wavelet packets are composed of a series of wavelet packet basis functions. Since different wavelet packet bases have different time-frequency characteristics, different wavelet packet bases will produce different results for the same signal. Therefore, choosing a suitable wavelet packet basis is very important for accurately extracting signal features.

Step 4.1: Select wavelet packet basis function

In actual application and processing of acoustic emission signals, Symlets8 or coif4 wavelets are generally selected. In the embodiment of the present invention, Symlets8 wavelet basis functions are selected to decompose wavelet packets.

Step 4.2: Decompose the wavelet packet

If the sampling frequency of the signal is _fs , according to the Nyquist sampling theorem, the measurable frequency range of the signal is [0, _fs /2]; because within the range of the signal frequency, the distribution range of the detail signal and the approximate signal is symmetrical, when the decomposition scale is 1, [0, _fs /4] and [ _fs /4, _fs /2] are the frequency ranges of the approximate signal and the detail signal. After performing J wavelet packet decompositions on the signal f(n) with a sampling frequency of _fs , the signal is decomposed into ^2J frequency segments. The calculation formulas for each frequency range are as follows:

If the sampling frequency is _fs kHz, the minimum recognition frequency required by the signal is _fmin . According to equation (5), the maximum decomposition scale J should satisfy:

Right now:

The signal sampling rate fs in the example of this application _is 2000kHz, so the range of the complete sound signal is 0-1000kHz. According to practical application experience, the decomposition scale is generally selected as J=3~4 to meet the requirements. The sound signal is decomposed by 3-scale wavelet packet, and the decomposition tree structure is shown in Figure 3. It can be seen that the third layer obtains 2 ³ = 8 frequency segments, and the frequency interval of each frequency segment is 1000/8=125kHz. The corresponding frequency segments of each interval are shown in Table 1.

After wavelet packet decomposition, the reconstructed waveform signals of each frequency band are obtained (Figure 4 shows the wavelet packet decomposition and reconstruction diagram of the original signal when the inclination angle is 45°, the apparent water velocity is 5m/s, and the apparent air velocity is 1.6m/s studied in the present invention), and the norm square of the reconstructed waveform signal is calculated as the wavelet packet energy of each node.

Step 5: Determine the initial structure of the BP neural network

Generally, when designing a neural network, three-layer networks (i.e., one hidden layer) should be given priority. Increasing the number of hidden layers can reduce network errors and improve accuracy, but it also complicates the network, thereby increasing the training time of the network and the tendency to "overfit".

The neural network designed and used in the present invention is a 3-layer network.

The four statistical parameters of the acoustic signal, namely the average voltage level, root mean square value, absolute energy value, and ringing count, and the wavelet packet energy (square norm of the reconstructed waveform) at 2 ^J frequency bands, totaling 4+2 ^J eigenvalues, are used as the input of the BP neural network input layer. When J is 3, this number is 8; therefore, the number of nodes in the input layer is the same as the number of input statistical features, which is 4+2 ^J.

The number of nodes in the output layer is 1; the number of nodes in the hidden layer is determined by the empirical formula for calculating the number of elements in the hidden layer. The commonly used empirical formula is as follows:

h＝2*m+1 (8)

Among them, h is the number of hidden layer nodes, m is the number of input layer nodes; according to the empirical formula, when J=3, the number of hidden layer nodes is calculated to be 25;

In order to facilitate calculation and comparison, the input layer and output layer neurons need to be normalized. The normalization formula is:

Among them, α is the eigenvalue after normalization, _xi is the eigenvalue before normalization, _xmax and _xmin are the maximum and minimum values of the eigenvalue before normalization, respectively, i is the eigenvalue sequence number, indicating the i-th eigenvalue, i∈1,…,n; _ymax and _ymin are the expected maximum and minimum values after normalization, which are generally defaulted to 1 and -1.

Define the flow parameters - pipeline pressure drop or liquid holdup as the output layer, that is, the output layer has only one neuron.

Considering the characteristics of liquid holdup and pressure drop, a linear transfer function is used between the input layer, hidden layer and output layer. This transfer function has the advantage of being able to be scaled within any numerical range, which is convenient for comparison with sample values. The training function uses the trainlm function, i.e., the LM back propagation algorithm. After many experiments, the number of trainings was set to 1000, the training target was set to 10 ^-5 , and the learning rate was set to 0.1.

This example targets pipeline pressure drop or liquid holdup. Tables 2 and 3 are examples of the implementation of flow parameter identification based on acoustic emission-GA-BP neural network.

Step 6: Determine the population size, encoding length and fitness in the genetic algorithm

The genetic algorithm starts with a randomly selected initial solution and generates a new solution through selection, crossover, and mutation operations based on the individuals (i.e., chromosomes) in the previous generation population.

Set the population size: The genetic algorithm focuses on the population, that is, a set of multiple given initial solutions, each of which is called an individual, or chromosome. During iterative operations, the population size remains unchanged, and the chromosomes in the population are constantly changing in the problem solution space. Therefore, the population size will affect the iteration rate. In the process of solving actual problems, the size and position of the initial population should be determined according to the actual situation. It is generally between 20 and 100. According to experience and repeated tests, the optimal population size of the embodiment of the present invention (J=3, i.e., 3-scale wavelet packet decomposition) is 40.

Encoding is to convert data into chromosomes, that is, genetic string structure data of fixed form and length in genetic space. The present invention is designed to use binary encoding, that is, 0 and 1 in the binary character set {0,1} are used to represent candidate solutions to the problem.

The number of coding bits is often given according to the actual problem, usually between 5 and 20. This example uses a 12-25-1 three-layer neural network, so there are 12×25×1=325 weights and 25+1=26 thresholds. After many experiments, the optimal number of binary coding bits in this example is 10, so the coding length of each individual is: (325+26)×10=3510.

The fitness value of an individual in a genetic algorithm is determined by a fitness function. The fitness describes the quality of the chromosome. Therefore, the fitness function calculates comparable non-negative results for the problem, which are then compared by subsequent selection operators. The fitness function uses the following methods:

1. The fitness function is directly transformed from the objective function

Find the maximum value: g(x) = f(x) (10)

Find the minimum value: g(x) = -f(x) (11)

2. Countdown method

Find the maximum value:

Find the minimum value:

Where c is a conservative estimate of the bounds of the objective function.

The fitness function used in the example application of this method is a ranking function. By comparing the errors between the weights and thresholds obtained by training in the subsequent steps and the weights and thresholds of the previous generation, the errors are sorted from small to large, and the calculation of the next step is performed.

Step 7: Selection of genetic operators

It includes the selection of selection operator, crossover operator and mutation operator.

The selection operator is to select excellent chromosomes from the current population and let these chromosomes reproduce the next generation population. The higher the fitness of the chromosome calculated by the fitness function, the higher the chance of being retained in the next generation or even multiple generations.

Common algorithms for selecting operators include roulette wheel selection, ranking-based selection, random traversal sampling, and tournament selection. This method uses random traversal sampling, where each individual has an equal chance of being selected, which is the inverse of the total number of individuals;

This method uses a single-point crossover operator:

Suppose the two parent strings are x = [x ₁ , x ₂ , L, x _n ] and y = [y ₁ , y ₂ , L, y _n ], and the cross point is k, then the generated offspring is:

x'＝[x1,x2,L, _xk ,yk _{+ 1} ,yk _{+ 2} ,L, _yn ] (14)

y'＝[y ₁ ,y ₂ ,L,y _k ,x _k+1 ,x _k+2 ,L,x _n ] (15)

In a single-point crossover, there is only one breakpoint on the chromosome. If the length of the parent string is n, then the single-point crossover has (n-1) different crossover results;

The crossover probability generally takes a value between 0.4 and 0.9. After multiple calculations, the optimal crossover probability in the embodiment of the present invention is selected as 0.7.

In this method, the mutation operator refers to a randomly selected chromosome with a certain probability to change the data. It is usually used as an auxiliary means to generate new species. It has local search capabilities and further expands the diversity of the population. The mutation operator determines the range of chromosomes in the population, determines the gene position, and mutates the gene at that position with a given mutation probability.

The value range of mutation probability is generally between 0.001 and 0.1. If it is greater than 0.5, the genetic algorithm degenerates into a random search. After multiple calculations, the optimal mutation probability of the embodiment of the present invention is 0.01. For the binary code used in the present invention, mutation refers to randomly swapping a certain bit of 0 or 1, for example:

As shown above, mutations occurred at positions 1 and 6;

Step 8: Calculate the optimal weights and thresholds

At the selected pipeline inclination angle, select no less than 150 groups of measured acoustic emission signals of different flow conditions with different superficial gas velocity and superficial liquid velocity combinations, then use steps 3 and 4 to calculate 4+2 ^J characteristic parameters of the acoustic emission signals of each condition, select 60% of all conditions as training samples, and 40% as test samples; select the pressure drop or liquid holdup corresponding to the training samples as the training sample results, and select the pressure drop or liquid holdup corresponding to the test samples as the test sample results; normalize the parameters of the training samples, test samples, training sample results (pressure drop or liquid holdup), and test sample results (pressure drop or liquid holdup);

Bring the normalized training samples, training sample results, test samples, and test sample results into the BP neural network of step 5 for training to obtain initial weights and thresholds;

Then, the genetic algorithm optimization calculation is performed, and the fitness calculation of the weight and threshold is performed according to the population size, encoding form and fitness function set in step 6, and the genetic operator set in step 7 is used for iteration;

After the iteration of the genetic operator, the next generation population is obtained; the population is evaluated by the fitness function set in step 6, and the iteration stops when the number of iterations reaches a predetermined value, at which time the optimal weight and threshold are obtained;

The range of the number of iterations is less than 200; when the set number of iterations is reached, the optimal individual is the optimal weight and threshold of the BP neural network. After multiple calculations and verifications, considering the iteration time, the maximum number of iterations given in this example, that is, the maximum genetic generation, is 50.

Step 9: Model Accuracy Test

The optimal weights and thresholds are reassigned to the neural network, and the test samples are substituted into the BP neural network. The calculated output results are compared with the actual results. If the calculated error is below 5%, the model is successfully established. If the calculated error does not meet the requirements, steps 6 and 7 are repeated until the calculated error meets the design requirements. The BP neural network at this time is the optimal BP neural network optimized by the genetic algorithm.

Step 10: Calculation of pressure drop and liquid holdup parameters for offshore oil and gas mixed pipeline system

Then, the acoustic emission sensor (1) arranged on the outer wall of the pipeline in step 1 is used to collect the acoustic emission signal, the acoustic signal parameters are calculated in the manner of step 3, the wavelet packet energy of the acoustic signal is calculated in the manner of step 4, and the obtained parameters and wavelet packet energy are input into the optimal model obtained after training in the above steps, so that the pressure drop and liquid holdup parameters of the marine oil and gas mixed transmission pipeline system can be obtained.

Example

When J=3, the method of the present invention is as follows: as shown in FIG1 , a flow chart of two-phase flow parameter identification based on acoustic emission technology.

Acoustic emission sensors are installed in the fully developed section of the flow in the pipe to collect acoustic emission signals. When the apparent gas velocity and apparent liquid velocity in the pipe are both zero, empty pipe collection is performed, and the maximum value of the empty pipe signal is set as the threshold voltage.

When the gas-liquid two-phase flow in the pipeline is in different flow states, the two-phase flow acoustic signal is collected by the acoustic emission sensor. While collecting the acoustic emission signal, the pressure drop and liquid holdup are measured by the pressure difference sensor and the dual parallel conductivity probes.

The average voltage level, root mean square value, absolute energy value and ring count of the acoustic emission signal collected in the previous step are calculated according to formulas 1 to 4 and the threshold voltage.

The acoustic emission signal is subjected to wavelet packet analysis to obtain the distribution of different flow states in each frequency band. The signal sampling rate in the example of this application is 2000kHz, so the range of the complete acoustic signal is 0-1000kHz. According to practical application experience, the decomposition scale is generally selected as J=3~4 to meet the requirements. The acoustic signal was decomposed into 3-scale wavelet packets, the wavelet basis function is Symlets8, and the decomposition tree structure is shown in Figure 3. It can be seen that the third layer obtains 2 ³ = 8 frequency bands, and the frequency interval of each frequency band is 1000/8 = 125kHz. The corresponding frequency bands of each interval are shown in Table 1. Figure 4 is a reconstruction diagram of the signal of each node obtained by wavelet packet decomposition of the original signal of the annular flow. By calculating the norm square of the reconstructed waveform signal, the wavelet packet parameters of each node are extracted.

The input layer of the BP neural network is 12 eigenvalues, namely, the average voltage level, RMS value, absolute energy value, ring count and the norm square of the reconstructed waveform signal at 8 nodes. The output layer of the BP neural network is the flow parameter, which is the pressure drop or liquid holdup in this example. According to Equation 8, the number of hidden layer nodes is calculated to be 25.

The training steps of the GA-BP neural network are shown in Figure 5. 200 groups of acoustic emission signals of experimental conditions with different superficial gas velocities and superficial liquid velocities were selected at the same inclination angle, and the characteristic parameters were calculated and extracted. 120 groups were selected as training samples and 80 groups were selected as test samples. The pressure drop or liquid holdup corresponding to the training samples was used as the training sample results, and the pressure drop or liquid holdup corresponding to the test samples was used as the test sample results. The training samples, training sample results, test samples, and test sample results were normalized according to formula 9 and input into the BP neural network established in step 5 to obtain the initial weights and thresholds. The initial weights and thresholds were optimized by genetic algorithm according to steps 7 and 8, and the specific settings were as follows: population size 40; binary coding; maximum genetic generation 50 generations; crossover was a single-point crossover operator with a probability of 0.7; mutation probability 0.01, and the method of selecting operators was random traversal sampling. When the offspring meets the fitness function conditions or the number of iterations reaches a predetermined value, the genetic algorithm calculation is stopped, and the optimal weights and thresholds are output.

The obtained optimal weights and thresholds are assigned to the BP neural network established in step 5, and the model accuracy is tested according to step 9. If the error requirements are met, the model is successfully established. If the error requirements are not met, steps 7 and 8 are repeated until the BP neural network with the optimal weights and thresholds meets the error requirements. The characteristic parameters of the acoustic emission signal to be identified are input into the BP neural network model with the optimal weights and thresholds to identify the pressure drop or liquid holdup flow parameters.

In order to compare the differences between the BP neural network and the GA-BP neural network, the present invention lists the error comparison of the BP neural network and the GA-BP neural network under the same pipeline operating conditions, as shown in Table 2 and Table 3. 100 sets of operating data at each inclination angle are used to verify the model. The comparison shows that the prediction effect of the improved model is significantly improved.

The present invention takes the pressure drop and liquid holdup of a pipeline with an inclination angle of 45 degrees as an example. After the model completes training and learning, a neural network recognition verification of 100 groups of working conditions is performed, as shown in Table 4 and Table 5 (due to space limitations, only the first 20 groups are listed). In this example, the average errors of the 100 groups of pressure drop and liquid holdup are 4.87% and 3.91% respectively, which can be seen to meet industrial applications.

Table 1 Frequency range of each node in wavelet packet decomposition

node (3,0) (3,1) (3,2) (3,3) Frequency band/kHz 0-125 125-250 250-375 375-500 node (3,4) (3,5) (3,6) (3,7) Frequency band/kHz 500-625 625-750 750-875 875-1000

Table 2 Comparison of the error of pipeline liquid holdup predicted by BP neural network and GA-BP neural network

Table 3 Comparison of pipeline pressure drop prediction errors between BP neural network and GA-BP neural network

Claims (7)

1. The method for measuring the flow parameters of the GAs-liquid two-phase flow based on the acoustic emission-GA-BP neural network is characterized by comprising the following steps of:

step 1, setting a threshold according to environmental noise

An acoustic emission sensor (1) is arranged on the outer wall of an inclined pipeline or a vertical high-pressure pipeline within the range of 20-90 degrees of the marine oil-gas mixing and conveying system;

When the gas speed and the liquid speed in the pipeline are zero, carrying out empty pipe acquisition, and setting the maximum value in an empty pipe signal as a threshold voltage;

Step 2, collecting the acoustic signals, the pressure difference and the liquid holdup

When the gas-liquid two-phase flow of the pipeline is in the working condition of different gas-liquid flow rate combinations, acquiring two-phase flow acoustic signals through an acoustic emission sensor (1); synchronously recording pressure difference and liquid holdup through a pressure difference sensor and a double parallel conductivity probe while collecting acoustic signals, and training and verifying a neural network model established in the subsequent step;

step 3, calculating acoustic signal parameters under each flow condition

Carrying out statistical analysis on the acquired original waveform data of the acoustic signal, and calculating to obtain the amplitude value, the average voltage level, the root mean square value, the absolute energy value and the ringing count of the acoustic signal;

Amplitude AMP, in dB, is defined as:

wherein V _max is the maximum value of voltage data in the two-phase flow acoustic signal, and the unit is V;

Average voltage level ASL, in dB, is defined as:

wherein V _mean is the average value of voltage data in the two-phase flow acoustic signal, and the unit is V;

root mean square RMS, unit V, defined as:

V is the voltage signal of each data point in the two-phase flow acoustic signal, and the unit is V; n is the number of acoustic signal data points;

Absolute energy value ABS, unit J, defined as:

v is the voltage signal of each data point in the two-phase flow acoustic signal, and the unit is V;10KΩ as a reference resistance; t is sampling time, unit s;

Ringing count Counts, which indicates the number of oscillations of the signal crossing the threshold, i.e., the number of effective peaks exceeding the threshold voltage;

When the correlation between the flow parameters and each statistical parameter is analyzed, the amplitude value AMP is found to have strong randomness in a large time window and is basically uncorrelated with the flow parameters, so that the amplitude value AMP is omitted;

step 4, analyzing wavelet packet of acoustic signal under each gas-liquid flow rate

The wavelet packet is composed of a series of wavelet packet basis functions, and as different wavelet packet bases have different time-frequency characteristics, different wavelet packet bases can obtain different results for the same signal, so that the selection of a proper wavelet packet base is very important to accurately extract the signal characteristics;

step 4.1, selecting wavelet packet basis function

Selecting Symlets8 wavelet basis function to decompose wavelet packet;

step 4.2, decomposing the wavelet packet

If the sampling frequency of the signal is f _s, according to the Nyquist sampling theorem, the measurable frequency range of the signal is [0, f _s/2 ]; since the distribution ranges of the detail signal and the approximate signal are symmetrical in the range of the signal frequency, when the decomposition scale is 1, [0, f _s/4 ] and [ f _s/4,f_s/2 ] are the frequency ranges of the approximate signal and the detail signal, after the signal f (n) with the sampling frequency f _s is subjected to J times of wavelet packet decomposition, the signal is decomposed into 2 ^J frequency segments, and the calculation formulas of the frequency ranges are as follows:

If the sampling frequency is f _s kHz, the minimum identification frequency required by the signal is f _min, and according to the formula (5), the maximum decomposition scale J should satisfy:

Namely:

the decomposition scale is selected from J=3 to 4; after the wavelet packet is decomposed, a reconstructed waveform signal of each frequency band is obtained, and the norm square of the reconstructed waveform signal is calculated to be used as the wavelet packet energy of each node;

step 5, determining the initial structure of the BP neural network

Designing a 3-layer neural network; the average voltage level, root mean square value, absolute energy value and ringing count of acoustic signals and the wavelet packet energy at2 ^J frequency segments, namely the norm square of the reconstruction waveform, are taken as the input quantity of the BP neural network input layer, wherein the total number of the characteristic values is 4+2 ^J; therefore, the number of nodes of the input layer is the same as the number of the input statistical features, and is 4+2 ^J;

The node number 1 of the output layer; the number of nodes of the hidden layer is determined by an empirical formula for calculating the number of hidden layer elements, and the conventional empirical formula is as follows:

h＝2*m+1 (8)

Wherein h is the number of hidden layer nodes, and m is the number of input layer nodes;

In order to facilitate calculation and comparison, the input layer neurons and the output layer neurons need to be normalized, and the normalized formula is as follows:

wherein, alpha is the characteristic value after normalization processing, x _i is the characteristic value before normalization processing, x _max、x_min is the maximum value and the minimum value of the characteristic value before normalization processing, i is the characteristic value serial number, and i is the i-th characteristic value, i is epsilon 1, … … and n; y _max、y_min is the maximum and minimum values expected after normalization, defaulting to 1 and-1;

Defining flow parameters-pipeline pressure drop or liquid holdup as an output layer, namely, the output layer has only one neuron;

Step 6, determining the population size, the coding length and the adaptability in the genetic algorithm

The genetic algorithm starts from an initial solution of random selection, and generates a new solution by selecting generations based on the selection, crossover and mutation operations of individuals in the previous generation population;

setting the population size: genetic algorithms look at populations, i.e., a collection of a plurality of given initial solutions, each of which is called an individual, i.e., a chromosome, with a population size between 20 and 100;

The coding is to convert the data into chromosome, namely, the genetic string structure data with fixed form and length in genetic space, and binary coding is used, namely, 0 and 1 in a binary character set {0,1} are used for representing candidate solutions of the problem;

the number of coding bits is between 5 and 20;

the fitness value of the individual in the genetic algorithm is judged by a fitness function, the fitness is used for illustrating the merits of the chromosome, so that the fitness function calculates a comparable non-negative result for the problem and then the non-negative result is compared by a subsequent selection operator;

The applied fitness function is a sorting function, and the weight and the threshold value obtained by training in the subsequent step are compared with the error between the weight and the threshold value of the previous generation, and the calculation of the next step is carried out according to the sorting from small error to large error;

step 7, selecting genetic operators

The method comprises the steps of selecting a selection operator, selecting a crossover operator and selecting a mutation operator;

Using a random traversal sampling method, wherein the probability of each individual being selected is equal and is the reciprocal of the total number of the individuals;

The crossover operator is the most important component in the genetic algorithm and is the core of operation, and common crossover operators comprise a single-point crossover operator and a two-point crossover operator, and the single-point crossover operator is used:

let two parent strings be x= [ x ₁,x₂,L,x_n ] and y= [ y ₁,y₂,L,y_n ], respectively, and the fork point be k, then the generated child is:

x'＝[x₁,x₂,L,x_k,y_k+1,y_k+2,L,y_n] (14)

y'＝[y₁,y₂,L,y_k,x_k+1,x_k+2,L,x_n] (15)

In the single-point crossing, the number of breakpoints of the chromosome is only one, and if the length of a father string is n, the single-point crossing has (n-1) different crossing results;

the value of the cross probability is between 0.4 and 0.9,

The mutation operator is a random selected chromosome with a certain probability to change data, and is usually used as an auxiliary means for generating new species, has local searching capability, and further expands the diversity of the population;

The value range of the variation probability is between 0.001 and 0.1;

Step 8, calculating the optimal weight and threshold value

Under the selected inclination angle of the pipeline, selecting at least 150 groups of measurement acoustic emission signals of different apparent gas speeds and apparent liquid speed combined flow working conditions, calculating 4+2 ^J characteristic parameters of the acoustic emission signals of each working condition by utilizing the step 3 and the step 4, and selecting 60% of all working conditions as training samples and 40% as test samples; selecting the pressure drop or the liquid holdup corresponding to the training sample as a training sample result, and selecting the pressure drop or the liquid holdup corresponding to the test sample as a test sample result; normalizing the training sample, the test sample, the training sample result and the test sample result parameters;

carrying out training on the normalized training samples, training sample results, test samples and test sample results in the BP neural network in the step 5 to obtain initial weights and thresholds;

then, genetic algorithm optimization calculation is carried out, fitness calculation is carried out on the weight and the threshold value according to the population size, the coding form and the fitness function set in the step 6, and iteration is carried out by utilizing the genetic operator set in the step 7;

After iteration of genetic operators, obtaining a next generation population; evaluating the population by the fitness function set in the step 6, and stopping iteration when the iteration times reach a preset value, so as to obtain an optimal weight and a threshold;

The selection range of the iteration times is below 200 times; when the set iteration times are reached, the obtained optimal individuals are the optimal weight and the threshold of the BP neural network;

step 9, model accuracy inspection

Re-assigning the obtained optimal weight and threshold value to the neural network, substituting the test sample into the BP neural network, comparing the calculated output result with the real result, if the calculated error is less than 5%, successfully establishing the model, and if the calculated error is not satisfied, repeating the steps 6 and 7 until the calculated error satisfies the design requirement; the BP neural network is the optimal BP neural network optimized by the genetic algorithm;

Step 10, calculating pressure drop and liquid holdup parameters of marine oil-gas mixed transportation pipeline system

And then, acquiring acoustic emission signals by using the acoustic emission sensors (1) arranged on the outer wall of the pipeline in the step 1, calculating acoustic signal parameters in a step 3 mode, calculating wavelet packet energy of the acoustic signals in a step 4 mode, and inputting the obtained parameters and the wavelet packet energy into an optimal model obtained after training in the steps, so that pressure drop and liquid holdup parameters of the marine oil-gas mixed transportation pipeline system can be obtained.

2. The method for measuring flow parameters of GAs-liquid two-phase flow based on acoustic emission-GA-BP neural network according to claim 1, wherein in step 4, j=3 is selected as the wavelet packet decomposition scale.

3. The method for measuring flow parameters of GAs-liquid two-phase flow based on acoustic emission-GA-BP neural network according to claim 1, wherein in step 5, a linear transfer function is used among the input layer, the hidden layer and the output layer, a trainlm function is used as a training function, i.e. an L-M back propagation algorithm, the training frequency is set to 1000, the training target is set to 10 ^-5, and the learning rate is set to 0.1.

4. The method for measuring flow parameters of GAs-liquid two-phase flow based on acoustic emission-GA-BP neural network according to claim 2, wherein in step 6, 10 is taken as the optimal binary coding bit number.

5. The method for measuring flow parameters of GAs-liquid two-phase flow based on acoustic emission-GA-BP neural network as set forth in claim 4, wherein in said step 6, 0.7 is selected as the optimal crossover probability.

6. The method for measuring flow parameters of GAs-liquid two-phase flow based on acoustic emission-GA-BP neural network as recited in claim 5, wherein in said step 7, 0.01 is used as the optimal variation probability.

7. The method for measuring flow parameters of GAs-liquid two-phase flow based on acoustic emission-GA-BP neural network of claim 6, wherein in said step 8, the maximum number of iterations, i.e., the maximum number of genetic algebra, is 50.