A method for calculating pipeline flow based on neural network and transducer topology setting

Abstract

The present invention relates to the field of ultrasonic velocity measurement, and in particular to a method for calculating pipeline flow based on neural network and transducer topology adjustment, which solves the problem of low measurement accuracy in the ultrasonic flow measurement process in the prior art. The present invention sets an ultrasonic transducer group on the outside of the pipeline and connects to a remote server through a network module. First, the velocity field and flow under input parameters are calculated by building a model through CFD; second, the real-time flow velocity of the pipeline is calculated; and third, the topology of the ultrasonic transducer group is feedback-adjusted through the error of the velocity field and the model field. The present invention improves the measurement accuracy of the transducer group by eliminating the influencing factors of ultrasonic waves in non-aqueous media; the flow under input parameters is calculated by a neural network model, and the variance is calculated with the flow measured by the transducer group, and the measurement accuracy is further improved by adjusting the topology of the transducer group. The remote server is connected to the transducer group to realize real-time monitoring of the measurement.

Description

Technical Field

[0001] The invention relates to the field of ultrasonic velocity measurement, and in particular to a method for calculating pipeline flow based on neural network and transducer topology structure setting. Background Art

[0002] The most widely used method in existing pipeline flow measurement is to calculate pipeline flow based on ultrasonic flow velocity measurement. However, the propagation speed of ultrasonic waves in the medium is affected by many factors, such as temperature, pressure and medium density. When these influencing factors change significantly, the calculation accuracy of the existing flow meter will be greatly reduced, and there are problems such as low robustness and complex structure of ultrasonic flow meters.

[0003] In the design of ultrasonic pipeline flowmeters based on the principle of "time difference method", the measurement accuracy of ultrasonic propagation time in the pipeline determines the measurement accuracy of average flow velocity, which ultimately affects the measurement of pipeline flow. However, in the actual measurement of propagation time, in addition to the propagation time in the pipeline, it also includes the propagation time in solid media such as the pipe wall and transducer. Since the propagation speed of ultrasonic waves in the pipeline will be affected by factors such as temperature, pressure, and liquid density, it is difficult to ensure the measurement accuracy.

[0004] There is an urgent need for a new flow calculation method that can solve the above problems. Summary of the invention

[0005] The present invention proposes a method for calculating pipeline flow based on neural network and transducer topology structure adjustment, which solves the problem of low measurement accuracy in the process of ultrasonic flow measurement in the prior art.

[0006] The technical solution of the present invention is implemented as follows: a method for calculating pipeline flow based on neural network and transducer topology structure adjustment, an ultrasonic transducer group is arranged on the outside of the pipeline, and the transducer group is connected to a remote server through a network module. First, a flow velocity field and flow rate under input parameters are calculated by building a model through CFD; second, the real-time flow velocity of the pipeline is calculated; the network module transmits the signal of the transducer group to the remote server through the network module, and calculates the flow velocity distribution in real time; third, the topology structure of the ultrasonic transducer group is feedback-adjusted through the error of the flow velocity field and the model field.

[0007] Furthermore, the step 2 includes ultrasonic transmission time calibration, including simulating an experimental platform and acquiring experimental data; the simulated experimental platform includes a temperature-controllable pipe and an ultrasonic transducer arranged on the outside of the pipe.

[0008] Furthermore, the ultrasonic transmission time calibration is specifically as follows: 1. in an empty pipe state and at a temperature T0, obtaining a result of τ0 at the temperature by changing the ultrasonic propagation path; Wherein, c is the speed of sound, L1 and L2 are the lengths of the two propagation paths respectively, t1 and t2 are the propagation times of the ultrasonic wave on the two propagation paths respectively; 2 Change the temperature and repeat the experimental calculation for many times to obtain the results of τ0 at different temperatures; 3 Fit the experimental calculation results to obtain the description model τ0=F(T) between the time delay of ultrasonic wave in non-aqueous medium and temperature.

[0009] The specific step three is sub-area division: adjusting the ultrasonic transducer topology: adjusting the ultrasonic transducer array topology based on the error distribution of step one and step two; dynamic adjustment based on feedback mechanism: taking the reconstruction error as the target variable, the inner circle radius R of the sub-area ir and radial angle θ as control variables to determine the optimal inner ring radius R ir and radial angle θ.

[0010] Furthermore, the step one is specifically as follows: a. calculating the pipeline flow rate under set parameters by computational fluid dynamics (CFD); b. increasing the pipeline flow data: changing the parameters and simulating to obtain CFD simulation results under different states; c. training the CFD simulation results by a neural network model: training the simulation results in step b by a neural network model; d. actual measured results: inputting the parameter data of the actual measured pipeline into the neural network model in step c to obtain the flow results of the pipeline.

[0011] Furthermore, the fitting in step 3 is a machine learning method that combines learning of more than two ultrasound transmission characteristics individual learners and individual learner strategies.

[0012] The method for calculating pipeline flow based on neural network and transducer topology setting disclosed in the present invention improves the measurement accuracy of the transducer group by eliminating the influence factors of ultrasonic wave in non-aqueous medium; calculates the flow under input parameters through the neural network model, performs variance calculation with the flow measured by the transducer group, and further improves the measurement accuracy by adjusting the topology of the transducer group. The remote server is connected to the transducer group to realize real-time monitoring of the measurement. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0014] Figure 1 : Simulation experiment platform;

[0015] Figure 2 : Schematic diagram of pipeline sub-area division;

[0016] Figure 3 :Modeling flow chart of reactor three-dimensional cylindrical pipe;

[0017] Figure 4 : Meshing diagram for reactor cylindrical pipe modeling;

[0018] Among them: 1. Ultrasonic transducer; 2. Temperature-controllable pipe. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] The present invention discloses a method for calculating pipeline flow based on neural network and transducer topology structure adjustment. An ultrasonic transducer group is arranged on the outside of the pipeline. The transducer group is connected to a remote server through a network module. First, a flow velocity field and flow rate under input parameters are calculated by building a model through CFD; second, the real-time flow velocity of the pipeline is calculated; the network module transmits the signal of the transducer group to the remote server through the network module, and calculates the flow velocity distribution in real time; third, the topology structure of the ultrasonic transducer group is feedback-adjusted through the error of the flow velocity field and the model field.

[0021] Furthermore, the step 2 includes ultrasonic transmission time calibration, including simulating an experimental platform and acquiring experimental data; the simulated experimental platform includes a temperature-controllable pipe 2 and an ultrasonic transducer 1 arranged on the outside of the pipe.

[0022] Furthermore, the ultrasonic transmission time calibration is specifically:

[0023] 1 In the empty pipe state and at temperature T0, the result of τ0 at this temperature is obtained by changing the ultrasonic propagation path; Among them, c is the speed of sound, L1 and L2 are the lengths of the two propagation paths, t1 and t2 are the propagation times of the ultrasonic wave on the two propagation paths;

[0024] 2. Change the temperature and repeat the experiment many times to obtain the results of τ0 at different temperatures;

[0025] 3 The experimental calculation results are fitted to obtain the description model τ0=F(T) between the time delay of ultrasound in non-aqueous media and temperature.

[0026] Furthermore, the step three is specifically sub-area division: adjusting the topological structure of the ultrasonic transducer 1: adjusting the array topological structure of the ultrasonic transducer 1 based on the error distribution of step one and step two; dynamic adjustment based on the feedback mechanism: taking the reconstruction error as the target variable, the inner circle radius R of the sub-area ir and radial angle θ as control variables to determine the optimal inner ring radius R ir and radial angle θ.

[0027] Ultrasonic wave propagation time from downstream to upstream t u and upstream to downstream propagation time t d Can be expressed as:

[0028]

[0029] In view of the complex flow patterns at different positions of the reactor pipeline, taking into account the existence of turbulent vortices in the pipeline flow field, the flow calculation method based on the three-dimensional cylindrical flow field distribution will be more accurate than the two-dimensional pipeline cross-section average flow velocity calculation scheme.

[0030] An ultrasonic detection array is installed in the reactor pipe, an appropriate ultrasonic signal emission period is selected, and the flight time of the sound wave path is measured. After the flight time is obtained, the flow velocity is calculated based on the difference in the upstream and downstream flight times to invert the three-dimensional flow velocity field.

[0031] The relationship between ultrasonic flight time and flow velocity can be obtained from formula (1):

[0032]

[0033] Therefore, after obtaining the flight time on each effective sound wave path, the three-dimensional flow velocity field is inverted using a radial basis based reconstruction algorithm using the relationship between the ultrasonic wave flight time and the flow velocity.

[0034] Several ultrasonic transducers 1 generate I effective ultrasonic paths between each other, which are divided into J sub-areas, such as Figure 2 The sub-areas are divided into inner circle radius R ir The relationship between the flight time and flow velocity of the ultrasonic wave along a specific propagation path in the test area can be expressed as the following line integral:

[0035]

[0036] Among them, a(x,y,z) represents the inverse of the flow velocity, l k is the ultrasonic propagation path. When using radial basis functions, a(x,y,z) needs to be expressed as a linear combination of J radial basis functions:

[0037]

[0038] Where: ε j is the unknown coefficient; (x j ,y j ,z j ) The coordinates of the center point of the jth sub-region; α is the radial basis function φ j The shape parameters of (x, y, z) need to be determined appropriately during the specific experiment.

[0039] make

[0040]

[0041] Then we can get

[0042]

[0043] Rewrite equation (6) into matrix form

[0044] Δt=F·E(8)

[0045] In the formula: Δt=[Δt1,Δt2,...,Δt I ] T , F=(f kj ) k＝1,2,...I；j＝1,2,...J ,E=[ε1,ε2,...,ε J ] T ;

[0046] Using the singular value decomposition of the matrix F and Tikhonov regularization technology, the regularization solution of equation (8) is:

[0047]

[0048] Where: σ is the non-zero singular value of the matrix F, J is the total number of non-zero singular values; u j 、v j are the left and right singular vectors of F respectively; μ is the regularization parameter. When the position of the ultrasonic transceiver array is determined, the shape parameter α of the radial basis function is given, and the matrix F and its singular values ​​can be obtained. Δt can also be determined by actually measuring the acoustic wave flight time and the sound channel angle θ, and the parameter vector ε can be determined according to formula (9). Substituting it into formula (4) can obtain the distribution function of the inverse of the flow velocity, and then the distribution function of the flow velocity can be obtained by the following formula, thereby realizing the reconstruction of the flow velocity field.

[0049]

[0050] According to the established reactor three-dimensional cylindrical pipeline position-time velocity distribution model, the error between the reconstructed field and the model field is calculated, the center coordinates in the divided sub-areas are brought in, and the velocity value of the center coordinates of each sub-area is obtained. The maximum absolute error, minimum absolute error, average relative error, and root mean square error between the velocity value of the center coordinate point of each sub-area of ​​the reconstructed velocity field and the velocity value calculated by the model are solved. The expression is as follows:

[0051] E max =max|v(x i ,y i ,z i )-v m (x i ,y i ,z i )|(11)

[0052] E min =min|v(x i ,y i ,z i )-v m (x i ,y i ,z i )|(12)

[0053]

[0054] In formulas (11)-(14):

[0055] n: the number of flow velocity values ​​of all center coordinates in the area to be measured

[0056] v a : Average flow velocity in the simulation field

[0057] v(x i ,y i ,z i ): Model field at coordinate (x i ,y i ,z i ) flow rate value

[0058] v m (x i ,y i ,z i ): velocity field at coordinate (x i ,y i ,z i ) flow rate value

[0059] Furthermore, the step one is specifically as follows: a. calculating the pipeline flow rate under set parameters by computational fluid dynamics (CFD); b. increasing the pipeline flow data: changing the parameters and simulating to obtain CFD simulation results under different states; c. training the CFD simulation results by a neural network model: training the simulation results in step b by a neural network model; d. actual measured results: inputting the parameter data of the actual measured pipeline into the neural network model in step c to obtain the flow results of the pipeline.

[0060] like Figure 3 The modeling flow chart of the reactor three-dimensional cylindrical pipeline is shown in detail: the space-time domain velocity distribution model of the reactor cylindrical pipeline is modeled. Since the fluid does not undergo phase change in the pipeline, it can be modeled as a single-phase turbulent flow. Therefore, the following time-averaged single-phase continuity equation, momentum equation and energy equation can be established to describe the fluid flow process in the pipeline.

[0061] Continuity equation:

[0062]

[0063] Momentum equation:

[0064]

[0065] where μ eff =μ+μ turb .

[0066] Energy equation:

[0067]

[0068] where k eff =k+Pr t (μ t / ρα).

[0069] Since turbulence affects both mass and energy transfer, it is necessary to establish a suitable turbulence model to quantify turbulent viscosity. The standard k-ε turbulence model can be used to assist in the analysis. The k-ε model solves the equation for the turbulent momentum k and its dissipation rate ε, which is expressed as follows:

[0070]

[0071] The turbulent viscosity is defined as μ t =ρC μ (k 2 / ε). The turbulent Prandtl number is also introduced in the above formulas (18) and (19) to quantify the turbulent energy dissipation coefficient of turbulent viscosity.

[0072] like Figure 4As shown in the mesh division diagram of the reactor cylindrical pipeline modeling, since the calculation of the above equations is to convert them into discrete equations, the CFD software is used to mesh the modeling area to achieve discretization and obtain the flow velocity distribution of the pipeline flow measurement section.

[0073] In order to establish the space-time velocity distribution model of the reactor's three-dimensional cylindrical pipeline flow measurement section, it is necessary to predict data based on existing data. The radial basis function RBF neural network is a feedforward neural network with excellent performance. It can approximate any nonlinear function with arbitrary precision and has global approximation capability. Its topological structure is compact, and its structural parameters can be separated for learning. It has fast convergence speed, simple learning rules, and is easy to implement on computers. The following method can be used to predict data using RBF neural networks: the velocity data obtained by CFD analysis is used as training samples. The training sample format is {X, V}, where X = [x1, x2, x3, x4] T , x1 is the horizontal coordinate of the model grid position, x2 is the vertical coordinate of the model grid position, x3 is the vertical coordinate of the model grid position, x4 is the relative movement time t of water in the reactor pipe; V is the one-dimensional output velocity obtained by CFD simulation; the RBF network output is Y = [y]; p is the number of hidden layer nodes (p>4), N is the total number of input samples; ε is the iteration termination accuracy. The output of the RBF neural network can be expressed as:

[0074]

[0075] The radial basis neuron structure is as follows Figure 4 As shown, the formula is:

[0076]

[0077] The purpose of model training is to obtain the three parameters W in the formula j , C j and D j , and obtain the velocity distribution model in the space-time domain. The training of the RBF neural network is mainly divided into two steps. First, unsupervised learning is required to determine the parameter C between the input layer and the hidden layer. j and D j , and then supervised learning is performed to determine the weight W between the hidden layer and the output layer j .

[0078] First, the connection weights from the hidden layer to the output layer are W = [w1, w2, ..., w p ] T , the central parameter C of each node in the hidden layer j =[c j1 ,c j2 ,c j3 ,c j4 ]T , the width vector D of each node in the hidden layer j =[d j1 ,d j2 ,d j3 ,d j4 ] T (j=1,2,…,p) for initialization, and the initialization formulas are:

[0079]

[0080] In formula (22), p is the number of hidden layer nodes (p>4), min(V k ) is the minimum value of all expected outputs of the output neurons in the training set, max(V k ) is the maximum value of all expected outputs of the output neurons in the training set.

[0081]

[0082] In formula (23), min(x i ) is the minimum value of all input information in the i-th input feature of the training set, max(x i ) is the maximum value of all input information in the i-th input feature of the training set.

[0083]

[0084] In formula (24), d f It is the width adjustment coefficient, and its value is generally less than 1. It mainly makes it easier for each hidden layer neuron to perceive local information and helps the RBF neural network improve its local response ability.

[0085] Then, the gradient descent method is used to adaptively adjust the center parameter, width vector, and adjustment weight. The iterative calculation formulas are:

[0086]

[0087] Among them, η is the learning factor and E is the RBF network evaluation function:

[0088]

[0089] Finally, determine whether the conditions for exiting the iterative operation are met, that is, the root mean square error RMS ≤ ε. If so, the training ends. If not, the weight iterative calculation will continue.

[0090] After the training, the values ​​of relevant parameters can be obtained, that is, the space-time domain velocity distribution model of the three-dimensional cylindrical pipeline of the reactor is established.

[0091] Furthermore, the fitting of step 3 is a machine learning method that combines the learning of more than two individual learners of ultrasonic transmission characteristics and the strategy of individual learners. The present invention divides the experimental data for ultrasonic transmission time calibration into a training set, a validation set and a test set. The training set is used for model training and optimization of parameters such as weights and biases; the validation set is used to optimize hyperparameters, learning rates, regularization coefficients, etc.; and the test set is used to evaluate the generalization ability of the model. The best data classification method, individual learner model and integration strategy are determined through experiments, and finally a mapping relationship description model is obtained.

[0092] The "average method" is used for integration, including the simple average method and the weighted average method. When the amount of data is large, the "learning method" can be used for integration, that is, a new learner is used to combine several individual learners to improve the learning accuracy. The model with the smallest root mean square error obtained by testing the test set is the optimal model.

[0093] The method for calculating pipeline flow based on neural network and transducer topology setting disclosed in the present invention improves the measurement accuracy of the transducer group by eliminating the influence factors of ultrasonic wave in non-aqueous medium; calculates the flow under input parameters through the neural network model, performs variance calculation with the flow measured by the transducer group, and further improves the measurement accuracy by adjusting the topology of the transducer group. The remote server is connected to the transducer group to realize real-time monitoring of the measurement.

[0094] Of course, without departing from the spirit and essence of the invention, technicians familiar with the field should be able to make various corresponding changes and modifications based on the invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the invention.

Claims

1. A method for calculating pipeline flow based on neural network and transducer topology setting. An ultrasonic transducer group is arranged outside the pipeline, and the transducer group is connected to a remote server through a network module. Features:

1. Velocity field reconstruction: Calculate the velocity field and flow rate under input parameters by building a CFD model; a Calculate the pipeline flow rate under set parameters through computational fluid dynamics (CFD); b Increase pipeline flow data: change parameters and simulate to obtain CFD simulation results under different states; c Neural network model training CFD simulation results: train the simulation results in step b through the neural network model; d Actual measurement results: input the parameter data of the actual measured pipeline into the neural network model in step c to obtain the pipeline flow results; 2. Model field velocity measurement: calculate the real-time flow velocity of the pipeline; the network module transmits the signal of the transducer group to the remote server through the network module, and calculates the velocity distribution in real time; including ultrasonic transmission time calibration, including simulation experiment platform and obtaining experimental data; the simulation experiment platform includes a temperature-controllable pipeline and an ultrasonic transducer arranged on the outside of the pipeline; 3. Feedback adjustment of the topological structure of the ultrasonic transducer group is performed through the error of the velocity field and the model field; specifically, the sub-area division: Adjusting the ultrasonic transducer topology: adjusting the ultrasonic transducer array topology based on the error distribution of the reconstructed field and the model field; Dynamic adjustment based on feedback mechanism: Taking reconstruction error as target variable, the inner circle radius R of sub-area ir and radial angle θ as control variables to determine the optimal inner ring radius R ir and radial angle θ.

2. The method for calculating pipeline flow based on neural network and transducer topology setting according to claim 1, Features: The ultrasonic transmission time calibration is specifically: (1) In the empty pipe state and temperature T 0 By changing the ultrasonic propagation path, we can obtain the τ at this temperature. 0 Result; Among them, L 1 and L 2 are the lengths of the two propagation paths, t 1 and t 2 are the propagation time of ultrasonic wave on two propagation paths respectively; (2) Change the temperature and repeat the experiment many times to calculate τ at different temperatures 0 Result; (3) Fit the experimental calculation results to obtain the description model τ between the time delay of ultrasound in non-aqueous media and temperature 0 =F(T).

3. The method for calculating pipeline flow based on neural network and transducer topology setting according to claim 2, Features: The ultrasonic transducer array topology is a broadcast type, specifically, the ultrasonic transducers are evenly arranged on the cross section of the pipeline to be tested.

4. The method for calculating pipeline flow based on neural network and transducer topology setting according to claim 3, Features: The fitting of step (3) is a machine learning method that combines the learning of more than two individual learners of ultrasonic transmission characteristics and the strategy of individual learners.

Images