A method for constructing a prediction model of a high-pressure fuel filter and a flow rate prediction method

Abstract

The application discloses a kind of high-pressure fuel filter prediction model construction method and flow prediction method, including two prediction models, method specifically includes: using computational fluid dynamics method carries out fluid dynamics simulation calculation to high-pressure fuel filter, constructs working condition sample data set;The size of filter hole, quantity, filter working pressure, fuel viscosity and filter pressure drop are used as input variable, and filter flow is used as output variable, two new high-pressure filter general nonlinear flow prediction models are established using artificial neural network and regression prediction method respectively, the flow characteristics of different models of filter in working interval can be accurately predicted using the neural network prediction method proposed in the application at any working condition point, and the error is low;The flow of different models of filter under different conditions can be quickly and accurately obtained, the accuracy meets the engineering needs, while greatly shortening the experimental time and calculation time, saving design cost.

Description

Technical Field

[0001] This invention belongs to the field of fuel pretreatment for internal combustion engines, specifically relating to a method for constructing a predictive model and a flow prediction method for a high-pressure fuel filter. Background Technology

[0002] During high-pressure fuel flow, fine metal particles are generated on the surface of metal pipes and other metal components due to friction and shear forces. The function of a high-pressure fuel filter is to filter out these coarse particles, preventing erosion of critical components such as high-pressure injectors and ensuring their performance and lifespan. In the design and research of high-pressure fuel filters, it is essential not only to ensure the filter's filtration efficiency for metal particles but also to guarantee good flow characteristics and minimize flow resistance during the flow process.

[0003] Existing common rail fuel injection systems for internal combustion engines use slit filters within the fuel injectors, which can only coarsely filter out impurities and cannot achieve fine filtration of fuel. The problem of particulate impurity filtration severely impacts the efficiency and lifespan of various components. A new type of high-pressure fuel filter (CN205868583U) can effectively filter fuel particles while ensuring minimal pressure drop at the filter inlet and outlet. However, the actual flow rate calculation and testing methods for the filter are cumbersome. Experimental methods are costly, and simulation methods are time-consuming. Therefore, a predictive model construction method for high-pressure fuel filters is needed to improve the efficiency of evaluating the relationship between pressure drop and flow rate in high-pressure fuel filters. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a method for constructing a predictive model for a high-pressure fuel filter. The model can not only calculate the flow rate of existing filter models under different operating conditions, but also predict the flow rate and pressure drop characteristics of newly designed filter models. It is applicable to, but not limited to, high-pressure common rail fuel filtration systems for fuel engines.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: a method for constructing a predictive model for a high-pressure fuel filter, comprising the following steps:

[0006] A geometric model of the fuel filter is established based on the type variables corresponding to the filter model and the variables that affect the flow rate of a fixed filter model in actual working conditions.

[0007] Based on the range of variation of filter boundary conditions, construct a working condition sample dataset;

[0008] Based on the operating condition sample dataset, the geometric model of the fuel filter is used as the calculation object, and the simulation is performed using computational fluid dynamics. The number of operating conditions is consistent with the size of the operating condition sample dataset. The flow data of all operating condition samples in the dataset is obtained, and the accuracy of the model is verified by experimental data at the selected operating point.

[0009] Based on the flow data of all working condition samples, the flow data of all working condition samples are divided into training set and validation set. The BP artificial neural network model is used to learn the simulation results of the working condition training set and to verify them using the validation set, so as to establish a neural network nonlinear prediction model between input and output.

[0010] A multivariate nonlinear regression prediction model was constructed based on the working condition sample dataset. The regression prediction inputs are the number of filter holes NUM and the filter hole diameter d. hole Actual work pressure P work The fuel viscosity μ and the inlet and outlet pressure difference ΔP are used to determine the filter flow rate FLOW, and the undetermined coefficients in the prediction function are determined by the lsqcurvefit method to obtain the filter flow rate regression prediction model.

[0011] Furthermore, the type variables corresponding to the filter model include the number of filter holes (NUM) and the filter hole diameter (d). hole In actual operating conditions, the variables that affect the flow rate of a fixed filter model include the actual operating pressure P of the filter. work The viscosity of the filtered fuel (μ) and the actual inlet and outlet pressure difference (ΔP) of the filter.

[0012] Furthermore, when establishing the geometric model of the fuel filter, four or more different filter models were identified, with the number of filter holes (NUM) ranging from 1,000 to 20,000. The filter holes of different models were all gradually expanding, with the inner wall hole diameter ranging from 10 μm to 100 μm and the expansion ratio ranging from 1.1 to 1.3. Based on the different filter hole data, the filter geometric model was constructed. High-pressure fuel flowed into the calculation domain from the filter pressure inlet, and after passing through the fine filter holes, the liquid phase fuel reached the filter outlet and flowed out.

[0013] Furthermore, the specific range of filter boundary conditions is as follows: the filter inlet working pressure range under actual operating conditions is 20MPa~200MPa; 3~10 different pressure differentials are set under different working pressures, with a pressure distribution range of 3bar~14bar; the corresponding fuel viscosity is calculated using a formula based on the different working pressures of the filter.

[0014]

[0015] μ0 = 3.2158·e [-0.0263(T-298)]

[0016] Where p is the working pressure of the filter and T is the actual working temperature of the fuel in the filter.

[0017] Furthermore, when constructing the working condition sample dataset, a sampling method is used to select at least 50 different working conditions within the computational domain boundary conditions of four or more types of filters to construct the working condition sample dataset.

[0018] Furthermore, the simulation using computational fluid dynamics methods specifically involves building and calculating a flow simulation model of the filter using the fluid simulation software FLUENT. Specifically, the fluid domain is calculated using the Realizable k-ε turbulence model from computational fluid dynamics, employing the SIMPLE algorithm. Pressure is discretized using a second-order method, momentum is calculated using a second-order upwind scheme, and turbulent kinetic energy and its dissipation rate are calculated using a first-order upwind scheme. The convergence criterion is set to a residual error below 10. -5 Furthermore, the filter flow rate remains stable, and the filter flow rate at all operating points in the operating condition sample dataset is calculated.

[0019] Furthermore, when dividing the traffic data of all working condition samples into training and validation sets, 10% to 30% of the working condition points in the working condition sample dataset are randomly selected as the validation set, and the remaining working condition points are used as the training set.

[0020] The input parameters for the neural network nonlinear prediction model include the number of filter holes (NUM) and the filter hole diameter (d). hole Actual work pressure P work The fuel viscosity μ and the inlet and outlet pressure difference ΔP were used to predict the filter flow rate FLOW. The backpropagation algorithm was used as the training method for the neural network to predict the flow rate of the high-pressure filter. The nonlinear prediction model of the neural network includes two hidden layers, with 6 to 9 neurons in each layer. The neurons are fully connected, and the learning rate is set in the range of 0.01 to 0.1.

[0021] Furthermore, a flow regression prediction model for fixed-type filters is constructed based on the operating condition sample dataset:

[0022]

[0023] After determining the basic form of the prediction function, it is extended for different filter models. For different filter models, the number of filter pores NUM and the pore diameter d are added. hole The correction coefficients are used to derive an extended flow regression prediction model, applicable to all filter models:

[0024]

[0025] Furthermore, the coefficients in the regression prediction model are:

[0026]

[0027] Based on the prediction model of the high-pressure fuel filter obtained by the above construction method, this invention provides a method for calculating and predicting the flow rate of a high-pressure fuel filter, which includes the number of filter holes NUM and the filter hole diameter d. hole Actual work pressure P work The fuel viscosity μ and the inlet / outlet pressure difference ΔP are used as inputs to the prediction model to obtain the filter flow rate FLOW.

[0028] Compared with the prior art, the present invention has at least the following beneficial effects:

[0029] The filter flow prediction model described in this invention can replace physical experiments, providing a systematic method for evaluating flow and pressure drop characteristics in the design and development of high-pressure fuel filters. Furthermore, due to the complexity of filter characteristics, full-domain simulation requires high computational costs and long computation times. The flow calculation method described in this invention can quickly and accurately obtain the flow and pressure drop curves of the filter under different operating pressures, significantly shortening the design and development time of high-pressure fuel filters. The neural network prediction method proposed in this invention can accurately predict the flow characteristics of different filter models at any operating point within the working range with low error, filling the gap in high-pressure fuel filter flow calculation and prediction methods.

[0030] This invention proposes a method for calculating and predicting fuel filter flow rate based on full-domain computational fluid dynamics simulation of a high-pressure fuel filter, utilizing artificial neural networks and multivariate nonlinear regression. The method employs FLUENT to perform full-domain computational simulation of the high-pressure fuel filter within its operating range, constructing a dataset of operating points. A nonlinear prediction model for filter flow rate is built using an artificial neural network, and its accuracy is verified. Furthermore, the multivariate nonlinear regression prediction relationship for the filter is obtained through flow rate analysis of the dataset, and its accuracy is also verified. Both models can accurately predict the flow rate of different filter models under specific operating conditions. The proposed method for calculating and predicting high-pressure fuel filter flow rate can quickly and accurately calculate or predict filter flow rate, significantly reducing design costs and shortening experimental and computational time, demonstrating significant scientific value and promising engineering applications. Attached Figure Description

[0031] The accompanying drawings, which form part of this specification, are provided to further illustrate embodiments of the invention and, together with the textual description, explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0032] Figure 1 Drawings and a full-range geometric model of a high-pressure fuel filter provided for embodiments of the present invention.

[0033] Figure 2 The values ​​of the inlet and outlet pressure difference of the filter constructed under different working pressures in the embodiments of the present invention are given.

[0034] Figure 3 The image shows a simulation cloud diagram of filter velocity and pressure obtained from an embodiment of the present invention.

[0035] Figure 4 The working condition sample dataset constructed for embodiments of the present invention.

[0036] Figure 5 This is a relative error diagram of the neural network nonlinear prediction model in an embodiment of the present invention.

[0037] Figure 6 This is a relative error diagram of the regression prediction model for multiple filter types in an embodiment of the present invention.

[0038] Figure 7 This is an absolute error graph of the regression prediction model for multiple filter types in an embodiment of the present invention. Detailed Implementation

[0039] Specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. In the following description, specific details are set forth for purposes of explanation and not limitation, in order to aid in a thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention may be practiced in other embodiments departing from these specific details.

[0040] It should be noted that, in order to avoid obscuring the invention with unnecessary details, only the device structure and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0041] This invention provides a predictive model construction method for high-pressure fuel filters. This design method not only satisfies the flow rate calculation requirements of existing filter models under different operating conditions, but can also be used to predict the flow rate and pressure drop characteristics of newly designed filter models, significantly reducing experimental and computational costs and shortening design time. The method includes the following steps:

[0042] S100 defines the type variables for high-pressure fuel filter: the number of filter holes NUM and the filter hole diameter d. hole .

[0043] The number of filter holes NUM and the filter hole diameter d holeThere is a potential connection: due to the limited fuel filtration area, the more filter holes (NUM) there are, the larger the filter hole diameter (d) will be. hole The lower.

[0044] S200 defines the variable that affects the flow rate of a fixed filter model under actual operating conditions: the actual working pressure P of the filter. work The viscosity of the filtered fuel (μ) and the actual inlet and outlet pressure difference (ΔP) of the filter.

[0045] S300, determine the number of filter holes (NUM) and the pore diameter (d) corresponding to the filter model. hole Establish a geometric model of the fuel filter.

[0046] As an example, this invention identifies four or more different filter models, with the number of filter pores (NUM) ranging from 1000 to 20000. All filter pores in each model are of a gradually expanding shape, with the inner wall pore diameter ranging from 10 μm to 100 μm and the expansion ratio ranging from 1.1 to 1.3. A filter geometric model is constructed based on the different pore sizes. High-pressure fuel flows into the computational flow domain from the filter pressure inlet, and after passing through the fine pores, the liquid fuel flows out at the filter outlet.

[0047] S400, determines the range of variation of the filter boundary conditions, including the actual operating pressure range P of the filter. work The viscosity μ of the filtered fuel and the corresponding pressure difference ΔP between the filter inlet and outlet under different working pressures.

[0048] The actual operating pressure range of the filter inlet is 20MPa to 200MPa; 3 to 10 different pressure differentials (ranging from 3 bar to 14 bar) are selected under different operating pressures; the corresponding fuel viscosity is calculated using a formula based on the different operating pressures of the filter.

[0049]

[0050] μ0 = 3.2158·e [-0.0263(T-298)]

[0051] Where p is the working pressure of the filter and T is the actual working temperature of the fuel in the filter.

[0052] S500: Based on the range of computational domain boundary conditions determined by S400, construct the working condition sample dataset;

[0053] As an example, this invention uses a sampling method to select 190 different working conditions within the computational domain boundary conditions of five types of filters, and constructs a working condition sample dataset.

[0054] S600: Based on the working condition sample dataset obtained from S500, the filter geometric model established in S300 is used as the calculation object. The simulation is performed using computational fluid dynamics. The number of working conditions is consistent with the size of the dataset in S500. The flow data of all working condition samples in the dataset is obtained, and specific working condition points are selected to verify the accuracy of the model using experimental data.

[0055] The flow model for the filter was calculated using the fluid simulation software FLUENT. Specifically, the fluid domain was calculated using the Realizable k-ε turbulence model in computational fluid dynamics. The K-equation and ε-equation in the Realizable k-ε model are as follows:

[0056]

[0057]

[0058] The calculation methods for each coefficient and source term are as follows:

[0059]

[0060]

[0061]

[0062]

[0063]

[0064] Among them, G K G represents the turbulent kinetic energy caused by the average velocity gradient. b The turbulent kinetic energy generated by buoyancy; Y M C represents the contribution of wave expansion in compressible turbulence to the total dissipation rate; ε1 C ε2 and C ε3 It is a constant; Pr K and Pr ε These are the Prandtl numbers for turbulence, K and ε, respectively; the calculation coefficients are set as follows:

[0065] C ε1 =1.44C2=1.9Pr K =1.0Pr ε =1.0A0=4.04

[0066] By solving the computational fluid dynamics equations, the distribution of fluid pressure and velocity across the entire flow domain of the filter can be obtained. Specifically, the SIMPLE algorithm is used for calculation. The pressure term employs a second-order discretization method, momentum uses a second-order upwind scheme, and turbulent kinetic energy and its dissipation rate use a first-order upwind scheme. The convergence criterion is set to a residual error of less than 10. -5 Furthermore, the filter flow rate remained stable. The filter flow rate at all operating points in the operating condition sample dataset was calculated, and the results were verified using experimental data and the corresponding operating conditions in the calculated operating condition sample dataset.

[0067] S700, based on the flow data of all operating condition samples calculated by S600, divides the flow data of all operating condition samples into a training set and a validation set. This is to validate the input quantities (including the number of filter holes NUM, the filter hole diameter d). hole Actual work pressure P work There is a strong nonlinear relationship between the fuel viscosity (μ) and the inlet / outlet pressure difference (ΔP) and the output (filter flow rate). A BP artificial neural network model is used to learn from the training set of the simulation results and to verify the model using the validation set, thus establishing a neural network nonlinear prediction model between the input and output quantities.

[0068] From the flow data of all operating conditions, 30 operating points were randomly selected as the validation set, and the remaining 160 operating points were used as the training set. The inputs to the constructed neural network prediction model included the number of filter orifices NUM and the orifice diameter d. hole Actual work pressure P work The fuel viscosity μ and the inlet / outlet pressure difference ΔP are used as the inputs, and the predicted quantity is the filter flow rate FLOW. An error backpropagation algorithm is used as the training method for the neural network to predict the flow rate of the high-pressure filter. This prediction model has two hidden layers (each layer has 6-9 neurons), fully connected neurons, and a learning rate ranging from 0.01 to 0.1. A nonlinear prediction model between the input and output quantities is established, and the coefficient of determination R is calculated using the validation set. 2 And the root mean square error (RMSE).

[0069] To simplify the flow prediction model, a multivariate nonlinear regression prediction model for the filter is constructed based on the calculated operating condition sample dataset. The regression prediction inputs are the number of filter holes NUM and the filter hole diameter d. hole Actual work pressure P work The fuel viscosity μ and the inlet / outlet pressure difference ΔP are used as the output of the regression prediction function, which is the filter flow rate FLOW. The undetermined coefficient a in the prediction function is determined using the lsqcurvefit function. 1～8 Thus, a filter flow regression prediction model was obtained.

[0070] To determine the basic form of the regression prediction function, a flow prediction function for a fixed-type filter is constructed, with the actual working pressure P as the input to the regression prediction function. work Given the fuel viscosity μ and the inlet / outlet pressure difference ΔP, the regression prediction function output is the filter flow rate FLOW. The regression prediction model for the flow rate of a fixed filter model is determined as follows:

[0071]

[0072] Using the lsqcurvefit function and the least squares method, calculate the undetermined coefficient α in the regression prediction function. 1～6 The objective function of the least squares method is set as the sum of the squares of the differences between the predicted and actual values. A lower objective function indicates a more accurate prediction. The undetermined coefficients 'a' are solved by ensuring that the partial derivatives of the objective function with respect to the undetermined coefficients are zero. 1～6 :

[0073]

[0074] Where f represents the defined regression function expression, i.e., the calculation formula for FLOW; a i Represents undetermined coefficients (here i = 1 to 6); y j This represents the flow rate obtained from the simulation.

[0075] After determining the basic form of the regression model, it was expanded for different filter models. For different filter models, the number of filter pores NUM and the pore diameter d were added. hole The correction coefficients are then used to obtain the extended flow regression prediction method through the same process, calculating the undetermined coefficients 'a' in the prediction function. 1～8 This yields a flow regression prediction method applicable to all filter models:

[0076]

[0077] Using the lsqcurvefit function and the least squares method, the coefficients in the flow regression prediction model are obtained as follows:

[0078] a1 = 4.4540 × 10 -6 a2=0.3515a3=1.2201a4=-0.0951

[0079] a5 = 0.7105a6 = -1.3628 × 10 -4 a7 = 0.6082a8 = 0.7537

[0080] To facilitate understanding of the above technical solution, this invention provides an application example of the predictive model construction method for a high-pressure fuel filter, specifically implemented through the following steps:

[0081] Step 1: Determine the filter geometry model and build a working condition sample dataset.

[0082] Step 1.1 Determine the filter geometry model.

[0083] In this example, five different filter models were identified, with the number of filter pores (NUM) being 1600, 1900, 2400, 4000, and 5000, respectively. All five filter models used diffuser nozzles with pore diameters of 50μm–55μm (inlet to outlet diameter), 45μm–50μm, 40μm–45μm, 30μm–35μm, and 30μm–35μm, respectively, and diffuser ratios of 1.1, 1.111, 1.125, 1.167, and 1.167, respectively. The geometric model of the 2400-pore filter is shown below. Figure 1 As shown, (a) is a cross-sectional view of the geometric model of the 2400-pore filter, and (b) is a perspective view of the geometric model of the 2400-pore filter. Geometric models of five different filter types were established based on the filter's geometric parameters.

[0084] Step 1.2 Construct the working condition sample dataset.

[0085] Based on the general operating pressure range of high-pressure fuel filters, the operating pressure is divided into eight levels, from low to high: 20MPa, 40MPa, 55MPa, 105MPa, 130MPa, 160MPa, 180MPa, and 200MPa. Different operating pressures correspond to different inlet and outlet pressure drops of the filter, such as... Figure 2 As shown.

[0086] A total of 38 different operating conditions were generated under 8 different working pressures, basically covering the operating range of the high-pressure fuel filter. A total of 190 (38×5) operating condition samples were generated for five different filter models, forming the operating condition sample dataset.

[0087] Step 2: Simulate each operating point in the dataset to obtain the filter flow data set.

[0088] Step 2.1 Perform simulation preprocessing on the filter model.

[0089] To study the flow state of high-pressure fuel within the filter, meshes were created for the geometric models of different filter types. The mesh counts for these different filter types were 8.98 million (1600 filter holes), 13.2 million (1900 filter holes), 14.94 million (2400 filter holes), 21.45 million (4000 filter holes), and 23.21 million (5000 filter holes), respectively. The more filter holes a filter has, the more complex its geometry becomes, and the higher the corresponding mesh count.

[0090] To determine the dynamic viscosity of fuel under different operating pressures, the fuel viscosity is calculated using the formula:

[0091]

[0092] μ0 = 3.2158·e [-0.0263(T-298)]

[0093] Where p is the working pressure of the filter and T is the actual working temperature of the fuel in the filter. The calculated fuel viscosities under different working pressures are 0.0022 Pa·s (20 MPa), 0.0029 Pa·s (40 MPa), 0.0034 Pa·s (55 MPa), 0.0061 Pa·s (105 MPa), 0.0081 Pa·s (130 MPa), 0.011 Pa·s (160 MPa), 0.014 Pa·s (180 MPa), and 0.018 Pa·s (200 MPa).

[0094] Step 2.2 Calculate the filter flow rate using computational fluid dynamics.

[0095] In FLUENT, set the boundary conditions for the high-pressure fuel filter calculation model according to the constructed operating condition sample dataset.

[0096] The K-equation and ε-equation in the Realizable k-ε turbulence method are as follows:

[0097]

[0098]

[0099] The calculation methods for each coefficient and source term are as follows:

[0100]

[0101]

[0102]

[0103]

[0104]

[0105] Among them, G K G represents the turbulent kinetic energy caused by the average velocity gradient. b The turbulent kinetic energy generated by buoyancy; Y M C represents the contribution of wave expansion in compressible turbulence to the total dissipation rate; ε1 C ε2 and C ε3 It is a constant; PrK and Pr ε These are the Prandtl numbers for turbulence, K and ε, respectively; the calculation coefficients are set as follows:

[0106] C ε1 =1.44C2=1.9Pr K =1.0Pr ε =1.0A0=4.04

[0107] By solving the above equations, the distribution of fluid pressure and velocity across the entire flow range of the filter can be obtained. The specific calculation method uses the SIMPLE algorithm, with pressure using a second-order discretization method, momentum using a second-order upwind scheme, and turbulent kinetic energy and its dissipation rate using a first-order upwind scheme. The convergence criterion is set to a residual error of less than 10. -5 Furthermore, the filter flow rate remains stable. The filter flow rate at all operating points in the operating condition sample dataset is calculated, and this can be verified using experimental data and the corresponding operating conditions in the calculated dataset. Taking a filter with 2400 pores as an example, the velocity and pressure distribution are as follows: Figure 3 As shown, the filter flow rate at different operating points is obtained. Finally, the filter flow rate dataset is calculated for all operating points, as shown in the figure. Figure 4 As shown.

[0108] Step 2.3 Validate the filter traffic dataset based on the experimental data.

[0109] Simulated flow rate data and experimental data were compared under six different operating conditions. These six conditions included: inlet and outlet pressure drops of 3 bar and 5 bar at an operating pressure of 20 MPa; inlet and outlet pressure drops of 4 bar and 6 bar at an operating pressure of 40 MPa; and inlet and outlet pressure drops of 5 bar and 7 bar at an operating pressure of 55 MPa. The error between the simulated and experimental flow rates was within 10%, indicating a good agreement between the simulation and experimental results.

[0110] Step 3: Construct a nonlinear prediction model for filter flow based on artificial neural networks.

[0111] To verify the input quantities (including the number of filter holes NUM, the filter hole diameter d) hole Actual work pressure P work A strong nonlinear relationship exists between fuel viscosity (μ) and inlet / outlet pressure difference (ΔP) and output (filter flow rate FLOW), leading to the construction of a neural network nonlinear prediction model. The backpropagation algorithm is used as the training method for the neural network to predict the flow rate of the high-pressure filter. This prediction model has two hidden layers, each containing six neurons, forming a 5×6×6×1 neural network topology. The neurons are fully connected, and the activation function f of the hidden layer neural network is the sigmoid function. The expression for calculating the value of each layer is as follows:

[0112]

[0113] Where x 1～5 This represents the input to the neural network; O1 and O2 represent the values ​​of the first and second hidden layers, respectively; z m represents the output predicted by the neural network, here representing the filter flow rate; w represents the weights of different layers; θ represents the thresholds of different layers. The mean squared error (MSE) is used as the loss function for the neural network, and its calculation expression is:

[0114]

[0115] Where E represents the mean square error; y m The actual value (filter flow rate obtained from simulation calculation); z m The value is obtained through neural network calculation. By calculating the partial derivative of the error with respect to the weights, the correction formula for the weights in the neural network can be obtained, as shown below:

[0116]

[0117] Where η 1 η 2 and η 3 , respectively, are the learning rates of the two hidden layers and the output layer, set to 0.02 here; t represents the number of iterations. Similarly, the threshold correction formula can be obtained through the partial derivative of the error with respect to the threshold, as shown below:

[0118]

[0119] A nonlinear prediction model between the input and output quantities is established using the above equations. The model is learned using a training set and validated using a validation set. The entire process is completed within 2 seconds. The validation results are as follows: Figure 5 As shown, the maximum relative error is less than 3%, the average relative error is 0.85%, and the RMSE is 0.0113, indicating that the neural network prediction model is relatively accurate; the coefficient of determination R... 2 The value of 0.9990 indicates a strong nonlinear relationship between the input and output. Using a neural network nonlinear prediction model can save time on simulating the entire flow domain of the filter, thus reducing design costs.

[0120] Step 4: Construct a multivariate nonlinear regression prediction model for filter flow.

[0121] Step 3.1 Construct a flow regression prediction model for fixed-type filters.

[0122] At this point, the input to the regression prediction function is the actual working pressure P. workThe inputs are fuel viscosity μ and inlet / outlet pressure difference ΔP. The output of the regression prediction function is the filter flow rate FLOW. Since fuel viscosity is related to the actual operating pressure, the inputs are simplified to fuel viscosity μ and inlet / outlet pressure difference ΔP. Therefore, the simplified regression function form is:

[0123]

[0124] However, the relative error in calculating filter flow using the above formula is relatively large, with an average relative error of 4.54%, and an R² of 0.966 and an RMSE of 0.0581 for the calculated dataset. After analyzing the relative error plot, a working pressure P was added to the above regression prediction function. work The more accurate regression prediction function, obtained by adding the correction term for the inlet and outlet pressure difference ΔP, is shown below:

[0125]

[0126] Using the sqcurvefit function and the least squares method, the coefficient 'a' in the flow regression prediction function for a fixed-type filter is determined. 1～6 Their values ​​are as follows:

[0127] a1=0.0368a2=0.5755a3=-0.0873

[0128] a4=0.0525a5=0.5340a6=-0.0046

[0129] The mean relative error was then calculated to be 2.14%, and the R-squared of the dataset was calculated. 2 The mean relative error of the dataset is 0.990, and the RMSE is 0.0242. This shows a significant reduction in the mean relative error of the dataset; the coefficient of determination R0 is [missing value]. 2 The slight increase indicates a stronger regression relationship between the input and output; the root mean square error (RMSE) has decreased to half of its original value, indicating an improvement in model accuracy.

[0130] Step 3.2 Construct flow regression prediction models for different filter models.

[0131] Step 3.1 yielded the flow regression prediction model for a fixed-type filter. For different filter models, it is necessary to add the number of filter holes (NUM) and the pore size (d). hole The correction coefficients yield a new model applicable to all filter types, and the final regression prediction function expression is:

[0132]

[0133] The coefficients in the new model were calculated using the same method as in step 3.1, as follows:

[0134] a1 = 4.4540 × 10 -6 a2=0.3515a3=1.2201a4=-0.0951

[0135] a5 = 0.7105a6 = -1.3628 × 10 -4 a7 = 0.6082a8 = 0.7537

[0136] The relative error of the dataset was then calculated as follows: Figure 6 As shown, the absolute error is as follows Figure 7 As shown, the maximum relative error is 10.1%, and the average relative error is 2.73%. The coefficient of determination R0 of the dataset is calculated. 2 The mean square error (RMSE) is 0.9990 and the root mean square error (RMSE) is 0.0330. Therefore, it can be considered that the regression prediction model meets the engineering requirements for filter flow calculation, and the regression prediction model can save the time of full-basin simulation and experimentation of the filter, thus saving design costs.

[0137] Many features and advantages of these embodiments are apparent from this detailed description, and therefore the appended claims are intended to cover all such features and advantages of these embodiments that fall within their true spirit and scope. Furthermore, since many modifications and alterations will readily occur to those skilled in the art, the embodiments of the invention are not intended to be limited to the precise structures and operations illustrated and described, but rather all suitable modifications and equivalents falling within their scope are to be covered. The parts of the invention not described in detail are well known to those skilled in the art.

Claims

1. A method for constructing a predictive model for a high-pressure fuel filter, characterized in that, It includes the following steps: A geometric model of the fuel filter is established based on the type variables corresponding to the filter model and the variables that affect the flow rate of a fixed filter model in actual working conditions. Based on the range of variation of filter boundary conditions, construct a working condition sample dataset; Based on the operating condition sample dataset, the geometric model of the fuel filter is used as the calculation object, and the simulation is performed using computational fluid dynamics. The number of operating conditions is consistent with the size of the operating condition sample dataset. The flow data of all operating condition samples in the dataset is obtained, and the accuracy of the model is verified by experimental data at the selected operating point. Based on the flow data of all working condition samples, the flow data of all working condition samples are divided into training set and validation set. The BP artificial neural network model is used to learn the simulation results of the working condition training set and to verify them using the validation set, so as to establish a neural network nonlinear prediction model between input and output. A multivariate nonlinear regression prediction model was constructed based on the working condition sample dataset, with the number of filter holes as the regression prediction input. NUM Filter pore size d hole Actual work pressure P work fuel viscosity μ and inlet / outlet pressure difference Δ P The output of the regression prediction function is the filter flow rate. FLOW The lsqcurvefit method is used to determine the undetermined coefficients in the prediction function, resulting in a filter flow regression prediction model. Based on the operating condition sample dataset, a flow regression prediction model for a fixed-model filter is constructed as follows: After determining the basic form of the prediction function, it is extended for different filter models. For different filter models, the number of filter holes is increased. NUM and filter pore size d hole The correction coefficients are used to derive an extended flow regression prediction model, applicable to all filter models: 。 2. The method for constructing a predictive model for a high-pressure fuel filter as described in claim 1, characterized in that, The type variable corresponding to the filter model includes the number of filter holes. NUM and filter pore size d hole In actual operating conditions, the variables that affect the flow rate of a fixed filter model include the actual operating pressure of the filter. P work Filter fuel viscosity μ and the actual inlet and outlet pressure difference Δ of the filter P .

3. The method for constructing a predictive model for a high-pressure fuel filter as described in claim 1, characterized in that, When creating the geometric model of the fuel filter, determine four or more different filter models and the number of filter holes. NUM The number of filters ranges from 1,000 to 20,000. The filter holes of different models are all gradually expanding, with the inner wall hole diameter ranging from 10 μm to 100 μm and the expansion ratio ranging from 1.1 to 1.

3. The filter geometric model is constructed based on the different filter hole data. High-pressure fuel flows into the calculation domain from the filter pressure inlet. After passing through the fine filter holes, the liquid phase fuel flows out at the filter outlet.

4. The method for constructing a predictive model for a high-pressure fuel filter as described in claim 1, characterized in that, The specific range of filter boundary conditions is as follows: Under actual operating conditions, the filter inlet working pressure ranges from 20MPa to 200MPa; 3 to 10 different pressure differentials are set at different working pressures, with a pressure distribution range of 3 bar to 14 bar; the corresponding fuel viscosity is calculated using a formula based on the different working pressures of the filter. in p The working pressure of the filter, T This is the actual operating temperature of the fuel in the filter.

5. The method for constructing a predictive model for a high-pressure fuel filter as described in claim 1, characterized in that, When constructing the working condition sample dataset, a sampling method is used to select at least 50 different working conditions within the computational domain boundary conditions of four or more types of filters to construct the working condition sample dataset.

6. The method for constructing a predictive model for a high-pressure fuel filter as described in claim 1, characterized in that, The simulation using computational fluid dynamics (CFD) methods specifically involves building and calculating a flow simulation model of the filter using the fluid simulation software FLUENT. The fluid domain is calculated using the Realizable k-α turbulence model from CFD, employing the SIMPLE algorithm. Pressure is discretized using a second-order discretization method, momentum using a second-order upwind scheme, and turbulent kinetic energy and its dissipation rate using a first-order upwind scheme. The convergence criterion is set to a residual error below 10. -5 Furthermore, the filter flow rate remains stable, and the filter flow rate at all operating points in the operating condition sample dataset is calculated.

7. The method for constructing a predictive model for a high-pressure fuel filter as described in claim 1, characterized in that, When dividing the traffic data of all working condition samples into training and validation sets, 10% to 30% of the working condition points in the working condition sample dataset are randomly selected as the validation set, and the remaining working condition points are used as the training set. The input to a neural network nonlinear prediction model includes the number of filter holes. NUM Filter pore size d hole Actual work pressure P work fuel viscosity μ and inlet / outlet pressure difference Δ P The predicted quantity is the filter flow rate. FLOW The backpropagation algorithm was used as the training method for the neural network to predict the flow rate of the high-pressure filter. The nonlinear prediction model of the neural network includes two hidden layers, with 6 to 9 neurons in each layer. The neurons are fully connected, and the learning rate is set to a range of 0.01 to 0.

1.

8. The method for constructing a predictive model for a high-pressure fuel filter as described in claim 1, characterized in that, The coefficients in the regression prediction model are: 。 9. A method for calculating and predicting the flow rate of a high-pressure fuel filter, characterized in that, The flow prediction model obtained based on the prediction model construction method of the high-pressure fuel filter according to any one of claims 1-7, includes the number of filter holes. NUM Filter pore size d hole Actual work pressure P work fuel viscosity μ and inlet / outlet pressure difference Δ P The filter flow rate is obtained as input to the prediction model. FLOW .

Images