CN111723513A - Method for inverting simulation parameters through machine learning neural network - Google Patents
Method for inverting simulation parameters through machine learning neural network Download PDFInfo
- Publication number
- CN111723513A CN111723513A CN202010338720.4A CN202010338720A CN111723513A CN 111723513 A CN111723513 A CN 111723513A CN 202010338720 A CN202010338720 A CN 202010338720A CN 111723513 A CN111723513 A CN 111723513A
- Authority
- CN
- China
- Prior art keywords
- neural network
- parameters
- weight set
- machine learning
- model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 88
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 84
- 238000010801 machine learning Methods 0.000 title claims abstract description 51
- 238000004088 simulation Methods 0.000 title claims abstract description 37
- 238000003062 neural network model Methods 0.000 claims abstract description 20
- 238000002347 injection Methods 0.000 claims abstract description 14
- 239000007924 injection Substances 0.000 claims abstract description 14
- 230000008569 process Effects 0.000 claims description 60
- 238000012549 training Methods 0.000 claims description 13
- 230000006870 function Effects 0.000 claims description 9
- 238000004364 calculation method Methods 0.000 claims description 7
- 230000001537 neural effect Effects 0.000 claims 1
- 238000001746 injection moulding Methods 0.000 abstract description 14
- 238000004422 calculation algorithm Methods 0.000 abstract description 7
- 239000004743 Polypropylene Substances 0.000 description 15
- 229920001155 polypropylene Polymers 0.000 description 15
- -1 polypropylene Polymers 0.000 description 13
- 239000004033 plastic Substances 0.000 description 12
- 238000004519 manufacturing process Methods 0.000 description 6
- 239000000463 material Substances 0.000 description 6
- 230000000694 effects Effects 0.000 description 5
- 238000005206 flow analysis Methods 0.000 description 5
- 238000012360 testing method Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 238000005259 measurement Methods 0.000 description 4
- 238000005094 computer simulation Methods 0.000 description 3
- 238000002844 melting Methods 0.000 description 3
- 230000008018 melting Effects 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 230000001105 regulatory effect Effects 0.000 description 2
- 239000000243 solution Substances 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000008602 contraction Effects 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 230000004907 flux Effects 0.000 description 1
- 238000009434 installation Methods 0.000 description 1
- 239000012764 mineral filler Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 239000002861 polymer material Substances 0.000 description 1
- 239000000843 powder Substances 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 239000002994 raw material Substances 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/17—Mechanical parametric or variational design
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2113/00—Details relating to the application field
- G06F2113/22—Moulding
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- Geometry (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Computer Hardware Design (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Computational Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Injection Moulding Of Plastics Or The Like (AREA)
Abstract
The invention discloses a method for inverting simulation parameters by a machine learning neural network, which comprises the following steps: establishing a machine learning neural network; and injection simulation software based on the physical model is used as a learning object, and the weight of the neural network is adjusted by using an inversion algorithm in a machine learning algorithm, so that the corresponding result calculated by the injection simulation software based on the physical model is calculated after the same input parameters are used. And then inverting input parameters which should be used in the injection molding simulation software according to the difference between the measured value in the injection molding system quantified by the sensor and the predicted value of the trained and learned neural network model, wherein the most important parameters are, but not limited to, rheological model coefficients. The input parameters after being adjusted by machine learning inversion can reduce the difference between the simulation predicted value and the actual measured value by replacing injection molding simulation software, and the digital twin of the simulation injection molding process is realized so as to facilitate intelligent control.
Description
Technical Field
The invention relates to the field of intelligent control of physical model simulation, in particular to a method for inverting simulation parameters through a machine learning neural network.
Background
The plastic part is manufactured by an injection molding process after being designed, and certain computer simulation operation is firstly carried out on various working conditions, so that a process condition predicted value is obtained by utilizing mold flow analysis and is used for testing and adjusting the machine, a sensor measured value installed in a mold is obtained after at least one time of testing and adjusting the machine, and the measured value and a mold flow software predicted value have differences, and the existence of the differences can cause that a reliable intelligent control system cannot be established.
And simulation software can quantitatively predict various possible results, provide a group of good reference values for manufacturing and processing, predict possible defects and provide analysis and comparison for various working conditions and design changes of plastic parts and molds. These simulation analyses may take hours or days, sometimes even weeks, to arrive at a DOE with multiple design of manufacture (DFM) options, providing a variety of possible optimal choices. However, the predicted results may be different from reality in the injection molding process, for example, the simulation environment and the real environment directly have different influence factors, so that errors are generated, or numerical model errors in simulation software or programming errors occur, which may cause the simulation results to have deviation, and the matching degree between simulation and reality is insufficient.
It is currently of utmost importance to find the cause of the difference and correct the input parameters, especially the rheological model parameters, accordingly to match the prediction with the measured data. Therefore, there is a need to re-fit the parameters of the rheological model and correct other process parameters that may be erroneous with an on-line sensor system in the actual injection molding of plastic parts.
Machine Learning (ML) technology has been widely used in many fields as an important component of Artificial Intelligence (AI), and has been rapidly developed in recent years. Because of the installation of sensors, the data volume in the manufacturing process system is large, and in order to realize an efficient Manufacturing Execution System (MES), it is necessary to propose a method for inversely training a neural network to adjust input parameters by using a machine learning technique to minimize the difference between a predicted value and an actual measured value.
Disclosure of Invention
In order to solve the defects in the prior art, the invention provides a method for inverting simulation parameters through a machine learning neural network, which utilizes a machine learning technology to invert and train the neural network to adjust input parameters so as to minimize the difference between a simulation predicted value and an actual measured value.
In order to achieve the purpose, the invention adopts the specific scheme that:
a method for inverting simulation parameters through a machine learning neural network, comprising the steps of:
s1, establishing a machine learning neural network;
s2, providing a physical model for simulation prediction;
s3, inputting the same original parameters in the machine learning neural network and the physical model, taking the physical model as a learning object, performing inversion adjustment on the operation of the machine learning neural network by using forward-backward propagation to match a third weight set, and minimizing the mean square sum L of the difference value between the neural network predicted value and the physical model predicted valuec-nTo obtain a trained third weight set of the neural network;
s4, minimizing the mean square sum L of the difference between the neural network predicted value after training and the measured value of the sensor under the same conditione-nAiming at the target, the operation of inverse adjustment of the machine learning neural network is matched with the first weight set and/or the second weight set by using forward-backward propagation so as to obtain input parameters after inverse adjustment of the neural network;
and S5, substituting the input parameters subjected to inversion adjustment in the S4 into the physical model for recalculation.
Preferably, the raw parameters input into the machine learning neural network and the physical model at least comprise process parameters and rheological model parameters.
Preferably, the third weight set is disposed in a hidden layer of the neural network, and the third weight set is used for matching the trained neural network model to perform the prediction calculation.
Preferably, the first weight set and the second weight set are arranged in the neural network input data layer, the first weight set at least comprises a process parameter weight set, the second weight set at least comprises a rheological model parameter weight set, the process parameter weight set is used for matching process parameters, and the rheological model parameter weight set is used for matching rheological model parameters.
Preferably, the step S3 further includes minimizing a sum of mean square values L of the difference valuesc-nAnd then adjusting the third weight value to train the machine learning neural network.
Preferably, the step S4 further includes minimizing a sum of mean square of the difference value Le-And n is a target, and the process parameter weight set and the rheological model parameter weight set are adjusted to inversely adjust the input parameters which accord with the measured values.
Preferably, in step S4, the process parameter weight set and the rheological model parameter weight set are adjusted, specifically, if the input of the process parameter is correct, the rheological model parameter weight set is adjusted by inverting the neural network system, and if the input of the process parameter is incorrect, the process parameter weight set is adjusted by inverting the process parameter, so that the mean square sum L of the difference values is obtainede-nAnd (4) minimizing.
Preferably, in step S4, the input parameters after being inverted and adjusted by the neural network are respectively a calculation value of the process parameter and the adjusted process parameter weight set, and a calculation value of the rheological model parameter and the adjusted rheological model parameter weight value.
Preferably, in the steps S3 and S4, the mean square sum L of the difference valuesc-nMean square sum of sum and difference values Le-nCalculated by a loss function.
Preferably, the physical model is established by simulation software, and the simulation software is injection mold flow analysis software.
The method disclosed by the invention integrates the predicted value of the simulation software and the measured value of the sensor together by utilizing the machine learning neural network to form intelligent control, can quickly realize the digital twin effect, enables the computer simulation to generate a result which is almost the same as that generated by actual manufacturing, utilizes the machine learning technology provided by the invention to close the difference between the predicted value and the measured value, can automatically identify the reason of the difference and correct the difference so as to minimize the difference, and reduces the simulation error.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the drawings without creative efforts.
FIG. 1 is a block diagram of the calculations of the present invention;
FIG. 2 is an algorithmic flow chart of a neural network model of the present invention;
FIG. 3 shows the conditions and test results of the injection molding process used in example 1;
FIG. 4 is a Carreau-WLF model parameter for virgin polypropylene and recycled polypropylene;
FIG. 5 shows the Carreau-WLF model parameters obtained in example 1;
FIG. 6 is a convergence curve of the third weight set W3 inverted by the Carreau-WLF model in example 1;
FIG. 7 is the Cross-WLF model parameters derived from the inversion in example 1;
FIG. 8 is a convergence curve of the rheological model parameter weight set W2 when inverting Cross-WLF model parameters in example 1;
FIG. 9 is the inversion of the viscosity model for recycled polypropylene in example 2;
FIG. 10 is the injection molding process condition parameters and test results used in example 3;
FIG. 11 is a die flow analysis of the spiral plastic part used in example 3;
FIG. 12 is the Cross-WLF model parameters inverted from the quadratic viscosity model in example 3.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Example 1
Referring to fig. 1 to 8, a method for performing inverse simulation parameters by using a machine learning neural network includes the following steps:
s1, establishing a machine learning neural network, which comprises an input layer and an output layer, wherein the input layer and the output layer are the same as the test conditions of the physical model;
s2, carrying out simulation prediction on the basis of a simple physical model, and simulating various possible working conditions of the injection molding process of the designed plastic part; the present example simulates 18 working conditions of the plastic part, and the input data is shown in fig. 3;
s3, inputting the same original parameters in the machine learning neural network and the physical model according to the injection molding process, taking the physical model as a learning object, and performing forward-backward propagation to inversely adjust the operation of the machine learning neural network to match a third weight set W3, and minimizing the mean square sum L of the difference value between the calculated value of the neural network and the predicted value of the physical modelc-nTo obtain a third weight set W3 of the trained neural network; in this embodiment, 18 working conditions of plastic parts are simulated, the original parameters input in the machine learning neural network and the physical model include 18 process parameters, the input data is shown in fig. 3, and the viscosity model parameters are input, as shown in fig. 4, in this embodiment, the original polypropylene plastic parts are adopted. The parameters of the Carreau-WLF model obtained are shown in FIG. 5, in which the uppermost curve represents a melting temperature of 200 ℃ and the lower curves represent 220 ℃ and 240 ℃ in sequence, and the description is omitted here for the sake of brevity.The convergence curve of the third weight set W3 was inverted using the Carreau-WLF model, as shown in fig. 6.
S4, minimizing the mean square sum L of the difference value between the neural network predicted value after training and the measured value of the sensor under the same conditione-nAiming at the goal, the operation of inversely regulating the machine learning neural network by using forward-backward propagation is matched with the second weight set W2 and/or the first weight set W1 to obtain the input parameters after being inversely regulated by the neural network; providing a calculated value of the trained neural network and a measured value of the sensor under the same condition, performing inversion adjustment on a second weight set W2 of the rheological model Cross-WLF according to the measured value of the sensor under the condition that the input process parameters are correct, and minimizing the mean square sum L of the difference value of the measured value of the sensor and the predicted value of the trained neural networke-nObtaining Cross-WLF viscosity model parameters after inversion adjustment of the neural network; the curves and parameters of the deduced Cross-WLF model are inverted, and the convergence curves of the second weight set W2 when the Cross-WLF model parameters are inverted are shown in fig. 7 and 8, respectively.
And S5, substituting the input parameters subjected to inversion adjustment in the S4 into the physical model for recalculation.
Specifically, the viscosity model parameters of the Cross-WLF inverted and adjusted in S4 were plotted and compared with the viscosity model Carreau-WLF and the parameters obtained, i.e., fig. 5 and 7.
In this embodiment, a method applied to inverting viscosity model parameters in a physical model in an injection molding process using a machine learning neural network is presented. In this embodiment, the measured value of the sensor is used to reflect the real situation, the physical model in this embodiment calculates the predicted value by using the physical model of the injection mold flow, and in other embodiments, other physical models may be used, or simulation software for analyzing other injection mold flows may be used.
At least the following input parameters may be used to calculate the difference between the predicted and measured values of the injection mold flow analysis software: namely, the process parameters at least include process setting parameters and boundary condition parameters, the process parameters include speed curve, melt temperature, V/P conversion percentage, pressure maintaining conditions and the like, and the boundary conditions include mold surface temperature/heat flux conditions, sliding mold walls, vent hole size/position, mold deformation in the injection process and the like. And the rheological model parameters include at least a contraction flow loss and a PVT model parameter. Of course, in other embodiments, the corresponding parameters, such as the measuring device and its precision parameters, and the solver precision parameters programmed into the module flow analysis software, may also be adjusted according to possible factors.
In this embodiment, a simulation predicted value R is obtained by inputting process parameters and rheological model parameters into injection mold flow analysis softwarecMeanwhile, the same process parameters and rheological model parameters are input into the established machine learning neural network model, and the neural network model takes injection mold flow analysis software as a learning object to carry out inversion adjustment on the operation of the machine learning neural network to match with a third weight set W3 so as to enable the trained neural network to calculate a value RnAnd the predicted value R of the physical modelcThe weight value is adjusted by an inversion algorithm in the process, and a third weight set W3 is arranged in a hidden layer of the neural network and is used for matching the trained neural network calculated value RnTraining the machine learning neural network model to calculate a value R of the neural networknMinimizing the difference value with the predicted value Rc of the physical model to obtain the final calculated value R of the neural networkn. The minimization of the difference is obtained by calculating a loss function, and the specific formula is as follows: min { Lc-n＝∑(Rc-Rn)2In which L isc-nRepresents the mean square sum of the difference between the calculated values Rn of the neural network and the predicted values Rc of the physical model, the minimum difference referred to in this implementation is such that Lc-nApproaching to 0, the third weight set W3 is continuously adjusted in the inversion algorithm process to make the neural network calculate the value RnAnd the predicted value R of the physical modelcMean square sum of difference between Lc-nGradually tends to stabilize, making its gradient zero. The inversion algorithm used in the present embodiment for the third weight set W3 is a random gradient descent (SGD) algorithm, or any other fast-running optimization model, so as to adjust the learning rate (iteration step) of the neural network model to achieve fast convergence.
Calculating the neural network by using the final trainingValue RnAnd providing the measured value R of the sensor under the same conditionseThe measured value of the sensor is inverted and the operation of the machine learning neural network is adjusted to match the second weight set W2 and/or the first weight set W1, and the measured value R of the sensor is calculatedeAnd neural network calculated value RnMean square sum of difference values of Le-nBy continuously inverting and adjusting the weight set to make the mean square sum L of the differencese-nAnd (4) minimizing. The second weight set W2 and the first weight set W1 are arranged in the neural network input data layer, the weight set arranged in the neural network input data layer at least comprises a process parameter weight set W1 and a rheological model parameter weight set W2, the process parameter weight set W1 is used for matching process parameters, the rheological model parameter weight set W2 is used for matching rheological model parameters, the mean square sum minimization of the difference is calculated through a loss function, and the specific formula is as follows: min { Le-n＝∑(Re-Rn)2In which L ise-nRepresenting sensor measurements ReAnd neural network calculated value RnThe sum of the mean square deviations of the differences of (a) and (b), likewise in the inversion process, minimizes L by continuously adjusting the weight valuese-nMaking it approach to 0 to obtain the input parameters which are inverted and adjusted by the neural network; specifically, a process parameter weight set W1 and a rheological model parameter weight set W2 are adjusted to train a machine learning neural network model, if the input of process parameters is correct, the rheological model parameter weight set W2 is adjusted in an inversion mode for a neural network system, and if the input of the process parameters is incorrect, the process parameter weight set W1 is adjusted in an inversion mode aiming at the process parameters; specifically, assuming that the process parameters are correctly input, the process parameter weight set W1 multiplied by the process condition parameters is fixed, and the rheological model parameter weight set W2 for adjusting the rheological model parameters is propagated in reverse. However, this may not be the case in practical applications of simulation models and equivalent neural network models, because the process parameters and the actual effects of each injection molding machine may be slightly different, and in any case, human input errors may occur. Therefore, if it is found that the adjustment of the process parameter weight set W1 does indeed apply to the loss function Le-nThe minimization contributes, the vector values of the process parameter weight set W1 may be adjusted.Firstly, the identification of the input correctness of the process parameters is realized through the back propagation of a rheological model parameter weight set W2, and when the rheological model parameter weight set W2 converges to a minimum loss value, a simulation model predicted value RcAnd the measured value ReStill with no small difference, the back propagation adjusted process parameter weight set W1 may be activated by fixing the rheological model parameter weight set W2. Likewise, when the adjustment of the process parameter weight set W1 using back propagation is not applied to the loss function L any moree-nWhen minimization is effected, activation of the adjustment to the set of rheological model parameter weights W2 continues to cause the loss function L to bee-nAnd (4) minimizing. In other words, the back-propagation adjustments to the process parameter weight set W1 and the rheological model parameter weight set W2 may be iteratively performed interactively. The gradient of the weight set W2 of the rheological model parameters in the embodiment can be derived by a neural network through a chain rule of differentiating complex relations.
Difference loss function Le-nAfter minimization, the input parameters subjected to the neural network inversion adjustment at this time are respectively the operation values of the process parameters and the adjusted process parameter weight set W1 and the operation values of the rheological model parameters and the adjusted rheological model parameter weight value W2, and the operation methods adopted in the embodiment are multiplication, that is, the parameters finally input into the physical model for simulation are the product of the process parameters and the adjusted process parameter weight set W1 and the product of the rheological model parameters and the adjusted rheological model parameter weight value W2.
The finite element method used in the simulation of the physical model is that the three-dimensional space uses the unit and the node to form the geometric grid to approximately represent the object, the discrete form of the physical equation model is realized on the unit and the node, and the overall matrix equation composed of the unit matrix is constructed, so that the problem to be solved is described by the numerical equation set. Such a system of equations is based on a geometric connection relationship in which each node is contributed by adjacent connected elements. The neural network model has similar relations, and each node contributes to other nodes in a weight value mode by utilizing the topological relation among the nodes of the neural network.
The principle and the implementation method of the neural network model based on the difference back propagation push algorithm are not only superior to the simulation based on the physical model in the calculation speed, but also can learn to different objects, not only can learn the physical model, but also can be trained according to the measured data, the more the data, the better the effect produced, and the smaller the error value. The method disclosed by the invention integrates the predicted value of the physical model and the measured value of the sensor by utilizing the machine learning neural network, can quickly realize the digital twin effect, enables the computer simulation to generate a result which is almost the same as that generated by actual manufacturing, utilizes the machine learning technology provided by the invention to close the difference between the predicted value and the measured value, and reduces the simulation error.
Example 2
The present embodiment specifically describes training of a machine learning neural network model, and is different from the above embodiments in that the rheological model parameters of the recycled polypropylene plastic corresponding to the original plastic in embodiment 1 are inverted, and the rheological model parameters in the present embodiment specifically refer to at least one set of shear viscosity model parameters and one set of additional model parameters for calculating pressure loss under a systolic flow condition. This additional model requires a pressure sensor to be placed before and after the systolic flow to measure the pressure loss of the systolic flow. Designing and using at least two or more different contracted flows may better pair the measured additional model parameters.
On the measurement data, two forms are accepted:
1) a constant value at a certain moment, for example the moment corresponding to the peak value measured by one sensor, and the measured values of the other sensors at the same moment.
2) Time series of all sensor measurements.
The Cross-WLF viscosity coefficient of the polymer material tested with the instrumental forming system was derived using the method proposed by the present invention using the process conditions listed in figure 3, which includes three melt temperatures, respectively 200 ℃, 220 ℃, 240 ℃, corresponding injection speeds, and the pressure values measured with two pressure sensors as indicated in the table, the material used for the experiment was polypropylene (PP) mixed with 10% mineral filler powder, Hostacom CR 1171G1A, the recovered material corresponding to the original material was tested.
The viscosity model used in examples 1 and 2 is a modified Carreau-WLF model with the following equation:
αTexpressed in the following way:
wherein η is the viscosity (Pa.s), KiIs the model coefficient;
is the shear rate (1/s); t is temperature (. degree.C.)
The following formula is used herein to calculate the wall shear stress:
Δ P is the difference between the pressures measured at the two sensor locations (Pa); a (═ 0.003m) is the height of the rectangular cross section of the measurement channel, and b (═ 0.02m) is the width; l (═ 0.1m) is the distance between the two sensor locations that measure the pressure drop.
The shear rate is calculated according to the following Weissenberg-Rabinowitsch equation:
wherein the apparent shear rate
Is to use a band with a shape factor FpIs calculated by the following formula:
where Q is the injection volume flow.
Viscosity is the ratio of shear stress to shear rate, and equation (3-5) is a physical model that is apparently an approximate model since thermal effects are not considered, i.e., the system is assumed to be isothermal. The coefficient d in equation (4) and Fp in equation (5) are correction coefficients, and Fp for the selected geometry is 0.919.
The viscosity model coefficients of the original material and the corresponding recycled polypropylene material are shown in fig. 4, and the Carreau-WLF model with the fitted coefficients in example 1 is shown in fig. 5, in which the uppermost curve represents a melting temperature of 200 ℃, and the lower curves are 220 ℃ and 240 ℃ in sequence, and the description is omitted. The Carreau-WLF model is used as the training model of the neural network model of the present invention to train the third weight set W3, the convergence curve of the loss function of the third weight set W3 is shown in fig. 6, and the inversion result of the recycled polypropylene plastic corresponding to the raw material in the present embodiment is shown in fig. 9. The neural network model with this third set of weights W3 is then used as a training model to fit the coefficients of the Cross-WLF model to train the rheological model parameters second set of weights W2. Cross-WLF model is as follows:
wherein the zero shear rate viscosity is:
wherein
And D1,D2,D3,n,A1,
τ*Are all model coefficients. It must be noted that the pressure-related coefficient D3Is an important parameter and can influence the pressure prediction in the injection molding simulation.
The curve diagrams of Cross-WLF model coefficients and viscosity models after fitting of the original polypropylene are shown in fig. 7, wherein the curve diagrams include a set of model coefficients of the original PP, a training model for training the neural network model of the present invention trains a rheological model parameter weight set W2, and a convergence curve diagram of the neural network model for the rheological model parameter weight set W2 is shown in fig. 8.
Similarly, the polypropylene recovered in this example also performs the machine learning neural network inversion training of the present invention in the same manner as the original polypropylene. The Carreau-WLF model and Cross-WLF model coefficients for the recycled polypropylene, as well as the trained third weight set W3 and rheological model parameter weight set W2 are shown in fig. 9.
Example 3
In this embodiment, a spiral mold is proposed, the physical model of this embodiment is injection mold flow analysis software, and in other embodiments, any injection mold flow analysis physical model is also possible, and three melt temperature conditions are also used, namely 220 ℃, 240 ℃ and 260 ℃ respectively, and have corresponding speed and pressure values at each melt temperature, and specific values are listed in fig. 10. The material used for the experiment was polypropylene Stamylan PHC 31, shown in FIG. 11 as a mold flow analysis of the spiral plastic part used, and simulated calculations were performed on the spiral mold to provide a prediction equation based on a second order viscosity model, as follows:
wherein η is the viscosity (Pa.s); AiA model coefficient;
shear rate (1/s); t temperature (. degree. C.).
Similarly, in this embodiment, a second-order viscosity model is used as the training model of the neural network model of the present invention to train the third weight set W3, then the neural network model with the third weight set W3 is used as the training model to fit to obtain a Cross-WLF model, and then the rheological model parameter weight set W2 is trained by the Cross-WLF model coefficients. Fig. 12 shows a graph of the second order viscosity model coefficients provided, the fitted Cross-WLF model coefficients based on the invention, and the convergence curves of the neural network model of the trained third weight set W3 and rheological model parameter weight set W2, where the uppermost curve in the graphs of the second order viscosity model coefficients and the Cross-WLF model coefficients represents a melting temperature of 220 ℃ followed by 240 ℃ and 260 ℃.
The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims (10)
1. A method for inverting simulation parameters through a machine learning neural network, comprising the steps of:
s1, establishing a machine learning neural network;
s2, providing a physical model for simulation prediction;
s3, inputting the same original parameters in the machine learning neural network and the physical model, taking the physical model as a learning object, performing inversion adjustment on the operation of the machine learning neural network by using forward-backward propagation to match a third weight set, and minimizing the mean square sum L of the difference value between the neural network predicted value and the physical model predicted valuec-nTo obtain a trained third weight set of the neural network;
s4, minimizing the mean square sum L of the difference between the neural network predicted value after training and the measured value of the sensor under the same conditione-nAiming at the target, the operation of inverse adjustment of the machine learning neural network is matched with the first weight set and/or the second weight set by using forward-backward propagation so as to obtain input parameters after inverse adjustment of the neural network;
and S5, substituting the input parameters subjected to inversion adjustment in the S4 into the physical model for recalculation.
2. The method of claim 1 for inverting simulation parameters through a machine learning neural network, wherein: the original parameters input in the machine learning neural network and the physical model at least comprise process parameters and rheological model parameters.
3. The method of claim 2 for inverting simulation parameters through a machine learning neural network, wherein: the third weight set is arranged in a hidden layer of the neural network, and the third weight set is used for matching the trained neural network model to perform prediction calculation.
4. The method of claim 2 for inverting simulation parameters through a machine learning neural network, wherein: the first weight set and the second weight set are arranged in a neural network input data layer, the first weight set at least comprises a process parameter weight set, the second weight set at least comprises a rheological model parameter weight set, the process parameter weight set is used for matching process parameters, and the rheological model parameter weight set is used for matching rheological model parameters.
5. The method of claim 3 for inverting simulation parameters through a machine learning neural network, wherein: in the step S3, the method further includes minimizing the mean square sum L of the differencec-nAnd then adjusting the third weight value to train the machine learning neural network.
6. The method of claim 4, wherein the neural network is a machine learning neural networkThe method for inverting the simulation parameters is characterized by comprising the following steps: the step S4 further comprises minimizing the sum of mean square of the difference values Le-nAnd adjusting the process parameter weight set and the rheological model parameter weight set to inversely adjust the input parameters which accord with the measured values.
7. The method of claim 6 for inverting simulation parameters through a machine learning neural network, wherein: in step S4, the process parameter weight set and the rheological model parameter weight set are adjusted, specifically, if the input of the process parameter is correct, the rheological model parameter weight set is adjusted by inverting the neural network system, and if the input of the process parameter is incorrect, the process parameter weight set is adjusted by inverting the process parameter, so that the mean square sum L of the difference values is obtainede-nAnd (4) minimizing.
8. The method for inverting simulation parameters through the neural network of the robotic system as claimed in claim 6, wherein: in step S4, the input parameters after being inverted and adjusted by the neural network are the calculated values of the process parameters and the adjusted process parameter weight set, and the calculated values of the rheological model parameters and the adjusted rheological model parameter weight value, respectively.
9. The method for inverting simulation parameters through a neural network of a robotic system as claimed in claim 1, wherein: the mean square sum L of the difference values in the steps S3 and S4c-nMean square sum of sum and difference values Le-nCalculated by a loss function.
10. The method of claim 1 for inverting simulation parameters through a machine learning neural network, wherein: the physical model is established through simulation software, and the simulation software is injection mold flow analysis software.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010338720.4A CN111723513B (en) | 2020-04-26 | 2020-04-26 | Method for inverting simulation parameters through machine learning neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010338720.4A CN111723513B (en) | 2020-04-26 | 2020-04-26 | Method for inverting simulation parameters through machine learning neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111723513A true CN111723513A (en) | 2020-09-29 |
| CN111723513B CN111723513B (en) | 2024-04-26 |
Family
ID=72563678
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010338720.4A Active CN111723513B (en) | 2020-04-26 | 2020-04-26 | Method for inverting simulation parameters through machine learning neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111723513B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112487715A (en) * | 2020-11-27 | 2021-03-12 | 江苏科技大学 | Method for optimizing reliability of process parameters of key hole system of marine diesel engine body |
| CN113537354A (en) * | 2021-07-19 | 2021-10-22 | 吉林大学 | Aquifer structure stage type stochastic inversion identification method based on deep learning |
| CN114596919A (en) * | 2022-05-10 | 2022-06-07 | 安徽瑞邦数科科技服务有限公司 | Index prediction method and system and application thereof in phosphoric acid production |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20030050728A1 (en) * | 2001-04-24 | 2003-03-13 | Bahman Sarabi | Hybrid model and method for determining mechanical properties and processing properties of an injection-molded part |
| KR20070025111A (en) * | 2005-08-31 | 2007-03-08 | 동명대학교산학협력단 | A neural network-based warpage estimation method of an injection-molded thin plastic part |
| US20120209580A1 (en) * | 2011-02-11 | 2012-08-16 | Tisne Severine | Method and system employing flow simulation for improving material delivery in lens manufacturing |
| CN103914581A (en) * | 2013-12-27 | 2014-07-09 | 华中科技大学 | Optimization method for plastic injection molding technological parameter |
| CN110377948A (en) * | 2019-06-12 | 2019-10-25 | 江苏师范大学 | A kind of injection parameters Multipurpose Optimal Method |
-
2020
- 2020-04-26 CN CN202010338720.4A patent/CN111723513B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20030050728A1 (en) * | 2001-04-24 | 2003-03-13 | Bahman Sarabi | Hybrid model and method for determining mechanical properties and processing properties of an injection-molded part |
| KR20070025111A (en) * | 2005-08-31 | 2007-03-08 | 동명대학교산학협력단 | A neural network-based warpage estimation method of an injection-molded thin plastic part |
| US20120209580A1 (en) * | 2011-02-11 | 2012-08-16 | Tisne Severine | Method and system employing flow simulation for improving material delivery in lens manufacturing |
| CN103914581A (en) * | 2013-12-27 | 2014-07-09 | 华中科技大学 | Optimization method for plastic injection molding technological parameter |
| CN110377948A (en) * | 2019-06-12 | 2019-10-25 | 江苏师范大学 | A kind of injection parameters Multipurpose Optimal Method |
Non-Patent Citations (1)
| Title |
|---|
| 麻向军: "考虑压力对黏度影响的注塑填充过程数值模拟", 模具技术, no. 2010, pages 11 - 13 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112487715A (en) * | 2020-11-27 | 2021-03-12 | 江苏科技大学 | Method for optimizing reliability of process parameters of key hole system of marine diesel engine body |
| CN113537354A (en) * | 2021-07-19 | 2021-10-22 | 吉林大学 | Aquifer structure stage type stochastic inversion identification method based on deep learning |
| CN113537354B (en) * | 2021-07-19 | 2022-07-12 | 吉林大学 | Aquifer structure staged stochastic inversion identification method based on deep learning |
| CN114596919A (en) * | 2022-05-10 | 2022-06-07 | 安徽瑞邦数科科技服务有限公司 | Index prediction method and system and application thereof in phosphoric acid production |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111723513B (en) | 2024-04-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111723513A (en) | Method for inverting simulation parameters through machine learning neural network | |
| CN100507460C (en) | Dynamic soft measuring and form establishing method base pulse response formwork and parameter optumization | |
| JP2002505194A (en) | Method and apparatus for forming curved plate | |
| Fratczak et al. | Virtual commissioning for the control of the continuous industrial processes—Case study | |
| CN105116726A (en) | Parameter design method for nonlinear predictive controller based on mechanism model | |
| Byrski et al. | An adaptive identification method based on the modulating functions technique and exact state observers for modeling and simulation of a nonlinear MISO glass melting process | |
| CN114912364A (en) | Natural gas well flow prediction method, device, equipment and computer readable medium | |
| CN111931307A (en) | Part quality prediction and evaluation optimization method and system | |
| Ke et al. | Enhancement of multilayer perceptron model training accuracy through the optimization of hyperparameters: a case study of the quality prediction of injection-molded parts | |
| US20200290257A1 (en) | Moulding-parameters processing method for an injection press | |
| Sika et al. | Synergy of modeling processes in the area of soft and hard modeling | |
| Zobeiry et al. | Theory-guided machine learning composites processing modelling for manufacturability assessment in preliminary design | |
| CN101996272B (en) | Method for simulating vulcanization process of rubber product | |
| Toukhtarian et al. | Parameter identification of transport PDE/nonlinear ODE cascade model for polymer extrusion with varying die gap | |
| Ma et al. | Intelligent recommendation system of the injection molding process parameters based on CAE simulation, process window, and machine learning | |
| Yu et al. | Offline prediction of process windows for robust injection molding | |
| CN111210877B (en) | Method and device for deducing physical parameters | |
| CN111625995B (en) | Online time-space modeling method integrating forgetting mechanism and double ultralimit learning machines | |
| CN114880930A (en) | Entity equipment digital twin modeling simulation prediction method based on online learning | |
| KR102091815B1 (en) | Intelligent super precision plastic mold design apparatus | |
| Iacono et al. | Improvement of manufacturing technologies through a modelling approach: an air-steam sterilization case-study | |
| CN114297940B (en) | Method and device for determining unsteady state reservoir parameters | |
| EP4025406B1 (en) | Computer implemented method of designing a molding process | |
| CN113552438B (en) | Method for predicting load capacity of transformer | |
| TWI829076B (en) | Virtual and real correction and learning method of simulated pressure values assisted by machine learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |