CN114021414B - Finite element iteration process optimization method and device based on deep learning - Google Patents
Finite element iteration process optimization method and device based on deep learning Download PDFInfo
- Publication number
- CN114021414B CN114021414B CN202210002593.XA CN202210002593A CN114021414B CN 114021414 B CN114021414 B CN 114021414B CN 202210002593 A CN202210002593 A CN 202210002593A CN 114021414 B CN114021414 B CN 114021414B
- Authority
- CN
- China
- Prior art keywords
- finite element
- iteration
- deep learning
- data set
- 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.)
- Active
Links
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
- G06F30/23—Design optimisation, verification or simulation using finite element methods [FEM] or finite difference methods [FDM]
-
- 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
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/14—Force analysis or force optimisation, e.g. static or dynamic forces
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Geometry (AREA)
- Computer Hardware Design (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention provides a finite element iteration process optimizing method and device based on deep learning, and relates to the technical field of civil structural engineering and computers. The method comprises the following steps: establishing a single-unit finite element model based on the constitutive relation; carrying out iterative computation solution on the finite element model of the single unit, and summarizing iterative process data to form a data set; deep learning is carried out on the hard convergence data set of the finite element model of the single unit; and establishing a real structure finite element model according to the constitutive relation and the deep learning model, defining a loading condition and a loading and unloading path according to the actual stress condition of the structure, obtaining a finite element simulation result of the real structure, and completing an optimized finite element iteration process based on the deep learning. The method realizes the prediction of the state point of the next step in the iteration by using the deep learning algorithm, reduces the iteration steps, optimizes the calculation efficiency and the convergence of the finite element model on the basis of ensuring the accuracy of the simulation result, and has higher universality.
Description
Technical Field
The invention relates to the technical field of civil structure engineering and computers, in particular to a finite element iterative process optimizing method and device based on deep learning.
Background
In the civil structure engineering calculation analysis, the finite element method carries out modeling analysis on the structure based on the discretization thought, and becomes an important technical means due to shallow and clear physical concept, convenience, practicability and wide application range. Finite element calculation needs to solve a nonlinear equation set, and since the constitutive relation of materials or components is often complex, implicit equations are often involved in the solution, and an iterative method needs to be adopted for the solution.
At present, iteration methods commonly used in finite element calculation in the prior art include secant stiffness iteration method, newton-raphson iteration method and the like, but the methods have some defects: (1) the finite element estimates the state of the next step according to the gradient of the state of the previous step when iterative computation is carried out, and when the rigidity of the material or the component has sudden change, namely the gradient is discontinuous, the situation that iteration cannot be converged easily occurs; (2) when the rigidity matrix of the structure is close to the odd difference, the inversion operation of the rigidity matrix becomes very difficult, the calculation time is very long and the convergence is difficult; (3) the secant stiffness array is not in the steepest descent direction, the iteration speed is low, and the tangential stiffness array adopted by the Newton-Raphson iteration method is only in the local steepest descent direction and is not necessarily in the global steepest descent direction, so that the iteration speed can be influenced, and the finite element calculation efficiency is further influenced.
Disclosure of Invention
The invention provides a finite element iteration process optimizing method and device based on deep learning, aiming at the problem that the iteration speed is low in the prior art, and the finite element calculation efficiency is further influenced.
In order to solve the technical problems, the invention provides the following technical scheme:
in one aspect, a method for optimizing a finite element iterative process based on deep learning is provided, which includes:
s1: presetting a constitutive relation of components or materials, and establishing a single-unit finite element model adopting the constitutive relation;
s2: randomly generating a load working condition and a loading and unloading path in the single-unit finite element model, and performing multiple expansion on the iteration times of the single-unit finite element model; with multiple expansion being a preset maximum number of iterations in the conventional solution processMultiple, set the number of iterations after expansion to;
S3: carrying out iterative computation solution on the finite element model of the single unit, and summarizing iterative process data to form a data set; dividing the data set into an easy convergence data set, a difficult convergence data set and an unconverged data set from small iteration times to large iteration times;
s4: deep learning is carried out on the data set which is difficult to converge of the finite element model of the single unit, and a deep learning model is obtained;
s5: and establishing a real structure finite element model according to the constitutive relation and the deep learning model, defining a loading condition and a loading and unloading path according to the actual stress condition of the structure, obtaining a finite element simulation result of the real structure, and completing an optimized finite element iteration process based on the deep learning.
Optionally, in step S1, the constitutive relation includes: a macroscopic constitutive relation of a component level or a microscopic constitutive relation of a material level.
Optionally, in step S2, the load condition includes a single degree of freedom directional loading or a multiple degree of freedom directional coupling loading.
Optionally, in step S3, the data set includes status points of each iteration step, and the status points include: generalized force vectors, generalized displacement vectors and other variables of the constitutive state of each iteration step.
Optionally, in step S3, the data set is divided into an easy-to-converge data set, a hard-to-converge data set, and an unconverged data set according to the number of iterations, which includes:
dividing the data set into an easy convergence data set, a difficult convergence data set and an unconverged data set from small iteration times to large iteration times; the converged data are divided into data which are difficult to converge and data which are easy to converge according to the ratio of 1:4, and the maximum iteration number of the data which are easy to converge is M.
Optionally, in step S4, deep learning is performed on the finite element model of a single element through the hard convergence data set, and a deep learning model is obtained, including:
dividing a data set difficult to converge into a training set, a verification set and a test set according to the ratio of 6:2: 2; according to the divided training set, verification set and test set, the first step isThe sub-iteration generalized force vector, the generalized displacement vector and other variables of the constitutive state are taken as inputAnd taking the sub-iteration generalized force vector as output, and carrying out deep learning neural network training based on loading and unloading time sequence data to obtain a deep learning model.
Optionally, in step S5, a real structure finite element model is established according to the constitutive relation and the deep learning model, a load condition and a loading and unloading path are defined according to an actual stress condition of the structure, a finite element simulation result of the real structure is obtained, and an optimized finite element iterative process based on the deep learning is completed, including:
s51: establishing a real structure finite element model adopting a constitutive relation, and defining a loading condition and an loading and unloading path according to the actual stress condition of the structure;
s52: the initial iterative calculation adopts a traditional numerical method, when the iterative step number is larger than M, the first M iterative processes are used as time sequence input to predict the first time sequence based on the trained deep learning modelGeneralized force vectors of state points after the secondary iteration; according to the conventional numerical methodOther variables of the structure state of the secondary iteration and the residual error of the iteration objective function;
s53: using the predicted generalized force vector and the calculated other variables of the constitutive state in the second stepA second iteration of obtainingThe prediction result of the step state point is continuously iteratedAnd completing the process of deep learning optimization finite element iteration until the iteration target function residual meets the requirement, and obtaining a finite element simulation result of the real structure.
In one aspect, an optimized finite element iterative process device based on deep learning is provided, and the device includes:
the finite element model establishing module is used for setting a constitutive relation and establishing a single-unit finite element model adopting the constitutive relation;
the iteration expansion module is used for randomly generating a load working condition and a loading and unloading path in the single unit finite element model and performing multiple expansion on the preset maximum iteration times of the single unit finite element model; with expansion times being preset maximum number of iterations in the solution processMultiple, set the number of iterations after expansion to;
The finite element model calculation module is used for calculating and solving the finite element model of a single unit and summarizing model iteration process data which is calculated and converged to form a data set; dividing the data set into an easy convergence data set, a difficult convergence data set and an unconverged data set from small iteration times to large iteration times;
the finite element model training module is used for carrying out deep learning on the hard convergence data set of the finite element model of the single unit to obtain a deep learning model;
and the real structure model optimization module is used for establishing a real structure finite element model according to the constitutive relation and the deep learning model, defining the loading working condition and the loading and unloading path according to the actual stress condition of the structure, obtaining a finite element simulation result of the real structure and completing the finite element optimization iterative process based on the deep learning.
Optionally, the constitutive relation includes: a macroscopic constitutive relation of a component level or a microscopic constitutive relation of a material level.
Optionally, the load condition includes single degree of freedom directional loading or multiple degrees of freedom directional coupling loading.
In one aspect, an electronic device is provided, which includes a processor and a memory, where at least one instruction is stored in the memory, and the at least one instruction is loaded and executed by the processor to implement the method for optimizing a finite element iteration process based on deep learning.
In one aspect, a computer-readable storage medium is provided, in which at least one instruction is stored, and the at least one instruction is loaded and executed by a processor to implement the method for optimizing finite element iteration process based on deep learning.
The technical scheme of the embodiment of the invention at least has the following beneficial effects:
in the scheme, the invention provides a method for optimizing a finite element iteration process based on deep learning, which is characterized in that the prediction of the state point of the next step in iteration is realized by using a deep learning algorithm in the finite element model iteration calculation process, the iteration step number is reduced, and the calculation efficiency and the convergence of the finite element model are optimized. Meanwhile, the method is based on the constitutive relation of the components or materials and finite element analysis software, the mechanical concept is clear, and the authenticity and the accuracy of a simulation result are ensured. The method can be applied to the finite element model iterative solving process adopting various materials or component constitutive relations, has no specific requirements on the constitutive relations and the structure types, and has higher universality.
Drawings
In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed to be used in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.
FIG. 1 is a flowchart of an optimized finite element iterative process method based on deep learning according to an embodiment of the present invention;
FIG. 2 is a flowchart of a finite element iterative process optimization method based on deep learning according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a node unit to be analyzed of an optimized finite element iterative process method based on deep learning according to an embodiment of the present invention;
FIG. 4 is a schematic diagram of a constitutive model of a bolt sliding equivalent spring of an optimized finite element iterative process method based on deep learning according to an embodiment of the present invention;
FIG. 5 is a schematic diagram of a constitutive model of a lower flange contact compression equivalent spring of an optimized finite element iterative process method based on deep learning according to an embodiment of the present invention;
FIG. 6 is a statistical diagram of iterable records of a deep learning-based optimized finite element iterative process method according to an embodiment of the present invention;
FIG. 7 is a graph comparing iteration count of a conventional algorithm and iteration count of an optimization finite element iterative process method based on deep learning according to an embodiment of the present invention;
FIG. 8 is a block diagram of an apparatus for optimizing finite element iteration process based on deep learning according to an embodiment of the present invention;
fig. 9 is a schematic structural diagram of an electronic device according to an embodiment of the present invention.
Detailed Description
In order to make the technical problems, technical solutions and advantages of the present invention more apparent, the following detailed description is given with reference to the accompanying drawings and specific embodiments.
The embodiment of the invention provides a method for optimizing a finite element iterative process based on deep learning, which can be realized by electronic equipment, wherein the electronic equipment can be a terminal or a server. The method for optimizing finite element iterative process based on deep learning as shown in fig. 1 is a flow chart, and the processing flow of the method may include the following steps:
s101: presetting a constitutive relation of components or materials, and establishing a single-unit finite element model adopting the constitutive relation;
s102: randomly generating load working conditions and loading and unloading paths in the finite element model of the single unit, and carrying out finite element analysis on the single unitThe iteration times of the model are multiplied; with multiple expansion being a preset maximum number of iterations in the conventional solution processMultiple, set the number of iterations after expansion to;
S103: carrying out iterative computation solution on the finite element model of the single unit, and summarizing iterative process data to form a data set; dividing the data set into an easy convergence data set, a difficult convergence data set and an unconverged data set from small iteration times to large iteration times;
s104: deep learning is carried out on the data set which is difficult to converge of the finite element model of the single unit, and a deep learning model is obtained;
s105: and establishing a real structure finite element model according to the constitutive relation and the deep learning model, defining a loading condition and a loading and unloading path according to the actual stress condition of the structure, obtaining a finite element simulation result of the real structure, and completing an optimized finite element iteration process based on the deep learning.
Optionally, in step S101, the constitutive relation includes: a macroscopic constitutive relation of a component level or a microscopic constitutive relation of a material level.
Optionally, in step S102, the load condition includes a single degree of freedom directional loading or a multiple degree of freedom directional coupling loading.
Optionally, in step S103, the data set includes state points of each iteration step, where the state points include: generalized force vectors, generalized displacement vectors and other variables of the constitutive state of each iteration step.
Optionally, in step S103, the data set is divided into an easy-to-converge data set, a hard-to-converge data set, and an unconverged data set according to the iteration number from small to large, and the method includes:
dividing the data set into an easy convergence data set, a difficult convergence data set and an unconverged data set from small iteration times to large iteration times; the converged data are divided into data which are difficult to converge and data which are easy to converge according to the ratio of 1:4, and the maximum iteration number of the data which are easy to converge is M.
Optionally, in step S104, performing deep learning on the hard-to-converge data set of the single-element finite element model to obtain a deep learning model, including:
dividing a data set difficult to converge into a training set, a verification set and a test set according to the ratio of 6:2: 2; according to the divided training set, verification set and test set, the first step isThe sub-iteration generalized force vector, the generalized displacement vector and other variable vectors of the constitutive state are taken as inputAnd taking the sub-iteration generalized force vector as output, and carrying out deep learning neural network training based on loading and unloading time sequence data to obtain a deep learning model.
Optionally, in step S105, a real structure finite element model is established according to the constitutive relation and the deep learning model, a load condition and an unloading path are defined according to an actual stress condition of the structure, a finite element simulation result of the real structure is obtained, and an optimized finite element iterative process based on the deep learning is completed, including:
s151: establishing a real structure finite element model adopting a constitutive relation, and defining a loading condition and an loading and unloading path according to the actual stress condition of the structure;
s152: the initial iterative calculation adopts a traditional numerical method, when the iterative step number is larger than M, the first M iterative processes are used as time sequence input to predict the first time sequence based on the trained deep learning modelGeneralized force vectors of state points after the secondary iteration; according to the conventional numerical methodOther variables of the structure state of the secondary iteration and the residual error of the iteration objective function;
s153: using the predicted generalized force vector and the calculated other variables of the constitutive state in the second stepA second iteration of obtainingAnd (5) continuously iterating the prediction result of the step state point until the residual error of the iteration target function meets the requirement, and completing the process of deep learning optimization finite element iteration to obtain a finite element simulation result of the real structure.
In the embodiment of the invention, the invention provides a method for optimizing a finite element iteration process based on deep learning, which is used for realizing the prediction of the state point of the next step in iteration by using a deep learning algorithm in the finite element model iteration calculation process, reducing the iteration step number and optimizing the calculation efficiency and the convergence of the finite element model. Meanwhile, the method is based on the constitutive relation of the components or materials and finite element analysis software, the mechanical concept is clear, and the authenticity and the accuracy of a simulation result are ensured. The method can be applied to the finite element model iterative solving process adopting various materials or component constitutive relations, has no specific requirements on the constitutive relations and the structure types, and has higher universality.
The embodiment of the invention provides a method for optimizing a finite element iterative process based on deep learning, which can be realized by electronic equipment, wherein the electronic equipment can be a terminal or a server. As shown in fig. 2, a flow chart of a method for optimizing finite element iterative process based on deep learning, a processing flow of the method may include the following steps:
s201: presetting the constitutive relation of the components or materials, and establishing a single-unit finite element model adopting the constitutive relation.
In a feasible implementation mode, a constitutive relation is set, and a single-unit finite element model adopting the constitutive relation is established on the basis of commercial finite element analysis software or self-written software; the given constitutive relation may be a microscopic constitutive relation (stress-strain relation) of a material layer or a macroscopic constitutive relation (force-displacement relation) of a member layer.
In the embodiment of the invention, the establishment of the single-element finite element model instead of the multi-element model takes into account: (1) the modeling is simple, the calculation is quick, and the modeling and calculation time is saved; (2) the model of a single unit is simple in structure, the model calculation convergence process is not affected by the common problems of contact nonlinearity, geometric instability and the like of a multi-unit complex structure model, the convergence iteration process is only determined by the loading condition and the constitutive relation, and the inaccuracy of the training model caused by the fact that a data set of a subsequent training deep learning neural network contains other interference factors is avoided.
S202: randomly generating a load working condition and a loading and unloading path in the single-unit finite element model, and performing multiple expansion on the iteration times of the single-unit finite element model; with multiple expansion being a preset maximum number of iterations in the conventional solution processMultiple, set the number of iterations after expansion to。
In a feasible implementation mode, the load working condition can be single-degree-of-freedom direction loading or multi-degree-of-freedom direction coupling loading, and various characteristics of the finite element model corresponding to the finite element model can be fully covered so as to accumulate corresponding data.
In a practical implementation, the iteration number of the invention needs to be enlarged on the basis of the maximum iteration number in the traditional solving processAnd the iteration process which is not easy to converge can be fully calculated.
S203: carrying out iterative computation solution on the finite element model of the single unit, and summarizing iterative process data to form a data set; and dividing the data set into an easy convergence data set, a hard convergence data set and an unconverged data set from small iteration times to large iteration times.
In one possible embodiment, the data set includes state points for each iteration, and the state points include: generalized force vectors, generalized displacement vectors and other variables of the constitutive state of each iteration step.
In a possible implementation, as shown in fig. 3, the node unit to be analyzed according to the embodiment of the present invention is equivalent to a plurality of layers of springs connected in parallel, and the springs are formed by connecting specific sub-springs in series. In the case of the example shown in the figure,showing the tension and compression of the equivalent spring,Showing the equivalent spring of concrete tension and compression,Showing equivalent springs of the web of the steel beam with the hole,Showing equivalent springs of the perforated gusset plate,Showing the equivalent spring of the bolt in shear,showing the equivalent spring of the bolt slippage,showing the lower flange contacting the compressed equivalent spring. Combining these springs results in an equivalent single element finite element model of the node.Representing the bending moment at the node point,which represents the axial force at the node point,indicating the thickness of the concrete floor at the node,representing the height of the steel beam at the node,representing the distance of the nodal centroid from the bottom of the steel beam. Equivalent spring due to bolt slippageLower flange contact compression equivalent springThe rigidity mutation points exist in the constitutive model, so that the equivalent unit finite element model of the node is easy to converge at the rigidity mutation position.
In a possible implementation, fig. 4 shows an equivalent spring of bolt sliding in the embodiment of the present inventionSchematic diagram of a constitutive model of (1), wherein,showing the displacement of the bolt in the sliding way,the maximum amount of slippage of the bolt is indicated,indicating the horizontal force caused by the bolt slippage,representing the bolt slip critical load.
In one possible implementation, FIG. 5 shows an equivalent spring with a lower flange contacting a compression spring according to an embodiment of the present inventionSchematic diagram of a constitutive model of (1), wherein,the contact and compression of the lower flange is shown to be equivalent to the deformation of the spring,the equivalent spring counterforce of the contact and compression of the lower flange is shown,the stiffness of the lower flange contacting the compressed equivalent spring is shown,representing the initial spacing of the lower flanges at the node.
In one possible embodiment, the data set is divided into an easy-to-converge data set, a hard-to-converge data set, and an unconverged data set according to the number of iterations, which includes: dividing the data set into an easy convergence data set, a difficult convergence data set and an unconverged data set from small iteration times to large iteration times; wherein, the converged data is divided into the data which is difficult to converge and the data which is easy to converge according to the ratio of 1:4, and the maximum iteration number of the data which is easy to converge is. FIG. 6 is a diagram illustrating an example of an iterative record statistic of a single unit under random loading conditions and loading/unloading paths, where in the conventional calculation, the maximum number of iterations is 10 and the number of iterations is increased by a multipleTaking 50, dividing the calculation result of convergence into data which is difficult to converge and data which is easy to converge according to the ratio of 1:4, and then obtaining the maximum iteration times of the data which is easy to convergeIs 11.
S204: deep learning is carried out on the data set which is difficult to converge of the finite element model of the single unit, and a deep learning model is obtained;
in one possible implementation, the data set which is difficult to converge is divided into a training set, a verification set and a test set according to the ratio of 6:2: 2; according to the divided training set, verification set and test set, the first step isThe sub-iteration generalized force vector, the generalized displacement vector and other variables of the constitutive state are taken as inputAnd taking the sub-iteration generalized force vector as output, and carrying out deep learning neural network training based on loading and unloading time sequence data to obtain a deep learning model. Constructing the current state point and the current state point in the finite element calculation iterative process under the state of difficult convergenceAnd (5) association relation between state points after the secondary iteration.
In one possible implementation, the deep learning neural network based on loading and unloading time series data may use a multi-dimensional multi-layer LSTM (Long Short-Term Memory network). The stack LSTM model can be used for training, the first layer is a long-short term hidden layer LSTM, the input dimension is the sum of the length of a generalized displacement vector, the length of a generalized force vector and other variable lengths of the constitutive state, the second layer is an overfitting prevention layer Dropout which is inactivated randomly, the third layer is another long-short term hidden layer LSTM, the fourth layer is another overfitting prevention layer Dropout, the fifth layer is a full connection layer, and the output dimension is the length of the generalized force vector.
S205: establishing a real structure finite element model adopting a constitutive relation, and defining a loading condition and an loading and unloading path according to the actual stress condition of the structure;
s206: the initial iterative calculation adopts the traditional numerical method, and when the number of iterative steps is more than that of the iterative stepsWhen it is at restDeep learning model after training, will be beforeStep iteration process as time sequence input, predictGeneralized force vectors of state points after the secondary iteration; according to the conventional numerical methodOther variables of the structure state of the secondary iteration and the residual error of the iteration objective function;
in one possible embodiment, each iteration is first performed using a conventional numerical method, such as Newton-Raphson iteration.
S207: using the predicted generalized force vector and the calculated other variables of the constitutive state in the second stepA second iteration of obtainingAnd (5) continuously iterating the prediction result of the step state point until the residual error of the iteration target function meets the requirement, and completing the process of deep learning optimization finite element iteration to obtain a finite element simulation result of the real structure.
In the embodiment of the present invention, fig. 7 shows that, in the test set of the embodiment of the present invention, for a scene difficult to converge, that is, the iteration number is greater than 11, the iteration number of the conventional algorithm is compared with the iteration number of the optimization algorithm, the data distribution of the optimization algorithm is more concentrated in a section with less iteration number than that of the conventional algorithm, and the average iteration number of the optimization algorithm is shortened by 31.2% than that of the conventional algorithm through statistics.
In the embodiment of the invention, the invention provides a method for optimizing a finite element iteration process based on deep learning, which is used for realizing the prediction of the state point of the next step in iteration by using a deep learning algorithm in the finite element model iteration calculation process, reducing the iteration step number and optimizing the calculation efficiency and the convergence of the finite element model. Meanwhile, the method is based on the constitutive relation of the components or materials and finite element analysis software, the mechanical concept is clear, and the authenticity and the accuracy of a simulation result are ensured. The method can be applied to the finite element model iterative solving process adopting various materials or component constitutive relations, has no specific requirements on the constitutive relations and the structure types, and has higher universality.
FIG. 8 is an apparatus block diagram illustrating an iterative process for optimizing finite elements based on deep learning in accordance with an exemplary embodiment. Referring to fig. 8, the apparatus 300 includes:
a finite element model establishing module 310, configured to set a constitutive relation, and establish a single-element finite element model using the constitutive relation;
the iteration expansion module 320 is used for randomly generating a load working condition and a loading and unloading path in the single-unit finite element model and performing multiple expansion on the preset maximum iteration times of the single-unit finite element model; with expansion times being preset maximum number of iterations in the solution processMultiple, set the number of iterations after expansion to;
The finite element model calculation module 330 is used for calculating and solving the finite element model of a single unit, and summarizing the model iteration process data of calculation convergence to form a data set; dividing the data set into an easy convergence data set, a difficult convergence data set and an unconverged data set from small iteration times to large iteration times;
the finite element model training module 340 is used for deep learning the hard convergence data set of the finite element model of the single unit to obtain a deep learning model;
and the real structure model optimization module 350 is used for establishing a real structure finite element model according to the constitutive relation and the deep learning model, defining a loading condition and a loading and unloading path according to the actual stress condition of the structure, obtaining a finite element simulation result of the real structure, and completing the finite element optimization iterative process based on the deep learning.
Optionally, the constitutive relation includes: a macroscopic constitutive relation of a component level or a microscopic constitutive relation of a material level.
Optionally, the load condition includes single degree of freedom directional loading or multiple degrees of freedom directional coupling loading.
Optionally, the data set includes state points of each iteration step, and the state points include: generalized force vectors, generalized displacement vectors and other variables of the constitutive state of each iteration step.
Optionally, the finite element model calculating module 330 is further configured to divide the data set into an easy-to-converge data set, a hard-to-converge data set, and a non-converge data set according to the iteration number from small to large; wherein, the converged data is divided into the data which is difficult to converge and the data which is easy to converge according to the ratio of 1:4, and the maximum iteration number of the data which is easy to converge is。
Optionally, the finite element model training module 340 is further configured to divide the data set difficult to converge into a training set, a verification set, and a test set according to a ratio of 6:2: 2; according to the divided training set, verification set and test set, the first step isThe sub-iteration generalized force vector, the generalized displacement vector and other variables of the constitutive state are taken as inputAnd taking the sub-iteration generalized force vector as output, and carrying out deep learning neural network training based on loading and unloading time sequence data to obtain a deep learning model.
Optionally, the real structure model optimization module 350 is further configured to establish a real structure finite element model using a constitutive relation, and define a loading condition and a loading and unloading path according to an actual stress condition of the structure;
the initial iterative calculation adopts the traditional numerical method, and when the number of iterative steps is more than that of the iterative stepsThen, based on the deep learning model after training, will be the frontStep iteration process as time sequence input, predictGeneralized force vectors of state points after the secondary iteration; according to the conventional numerical methodOther variables of the structure state of the secondary iteration and the residual error of the iteration objective function;
using the predicted generalized force vector and the calculated other variables of the constitutive state in the second stepA second iteration of obtainingAnd (5) continuously iterating the prediction result of the step state point until the residual error of the iteration target function meets the requirement, and completing the process of deep learning optimization finite element iteration to obtain a finite element simulation result of the real structure.
In the embodiment of the invention, the invention provides a method for optimizing a finite element iteration process based on deep learning, which is used for realizing the prediction of the state point of the next step in iteration by using a deep learning algorithm in the finite element model iteration calculation process, reducing the iteration step number and optimizing the calculation efficiency and the convergence of the finite element model. Meanwhile, the method is based on the constitutive relation of the components or materials and finite element analysis software, the mechanical concept is clear, and the authenticity and the accuracy of a simulation result are ensured. The method can be applied to the finite element model iterative solving process adopting various materials or component constitutive relations, has no specific requirements on the constitutive relations and the structure types, and has higher universality.
Fig. 9 is a schematic structural diagram of an electronic device 400 according to an embodiment of the present invention, where the electronic device 400 may generate relatively large differences due to different configurations or performances, and may include one or more processors (CPUs) 401 and one or more memories 402, where the memory 402 stores at least one instruction, and the at least one instruction is loaded and executed by the processor 401 to implement the following steps of the method for optimizing a finite element iterative process based on deep learning:
s1: presetting a constitutive relation of components or materials, and establishing a single-unit finite element model adopting the constitutive relation;
s2: randomly generating a load working condition and a loading and unloading path in the single-element finite element model, and performing multiple expansion on the iteration times of the single-element finite element model; with multiple expansion being a preset maximum number of iterations in the conventional solution processMultiple, set the number of iterations after expansion to;
S3: performing iterative computation solution on the finite element model of the single unit, and summarizing iterative process data to form a data set; dividing the data set into an easy convergence data set, a difficult convergence data set and an unconverged data set from small iteration times to large iteration times;
s4: deep learning is carried out on the data set which is difficult to converge of the finite element model of the single unit, and a deep learning model is obtained;
s5: and establishing a real structure finite element model according to the constitutive relation and the deep learning model, defining a loading working condition and a loading and unloading path according to the actual stress condition of the structure, obtaining a finite element simulation result of the real structure, and completing an optimized finite element iteration process based on deep learning.
In an exemplary embodiment, a computer-readable storage medium, such as a memory, is also provided that includes instructions executable by a processor in a terminal to perform the above-described method for optimizing a finite element iteration process based on deep learning. For example, the computer readable storage medium may be a ROM, a Random Access Memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, and the like.
It will be understood by those skilled in the art that all or part of the steps for implementing the above embodiments may be implemented by hardware, or may be implemented by a program instructing relevant hardware, where the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a magnetic disk or an optical disk, etc.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.
Claims (6)
1. An optimized finite element iterative process method based on deep learning is characterized by comprising the following steps:
s1: presetting a constitutive relation of components or materials, and establishing a single-unit finite element model adopting the constitutive relation;
s2: randomly generating a load working condition and a loading and unloading path in the single-element finite element model, and performing multiple expansion on the iteration times of the single-element finite element model; with multiple expansion being a preset maximum number of iterations in the conventional solution processxMultiple, set the number of iterations after expansion toN;
S3: performing iterative computation solution on the finite element model of the single unit, and summarizing iterative process data to form a data set; dividing the data set into an easy convergence data set, a difficult convergence data set and an unconverged data set from small iteration times to large iteration times;
in step S3, the data set includes state points of each iteration step, where the state points include: generalized force vectors, generalized displacement vectors and other variables of the constitutive state of each iteration step;
in step S3, the data set is divided into an easy-to-converge data set, a hard-to-converge data set, and an unconverged data set according to the number of iterations, which includes:
dividing the data set into an easy convergence data set, a difficult convergence data set and an unconverged data set from small iteration times to large iteration times; dividing the converged data into data which is difficult to converge and data which is easy to converge according to a ratio of 1:4, wherein the maximum iteration number of the data which is easy to converge is M;
s4: deep learning is carried out on the data set which is difficult to converge of the finite element model of the single unit, and a deep learning model is obtained;
in step S4, the deep learning of the hard-to-converge data set of the single-element finite element model to obtain a deep learning model includes:
dividing the data set difficult to converge into a training set, a verification set and a test set according to the ratio of 6:2: 2; according to the divided training set, verification set and test set, the first step isiThe sub-iteration generalized force vector, the generalized displacement vector and other variables of the constitutive state are taken as inputi+xTaking the sub-iteration generalized force vector as output, and performing deep learning neural network training based on loading and unloading time sequence data to obtain a deep learning model;
s5: establishing a real structure finite element model according to the constitutive relation and the deep learning model, defining a loading working condition and a loading and unloading path according to the actual stress condition of the structure, obtaining a finite element simulation result of the real structure, and completing an optimized finite element iteration process based on deep learning;
in step S5, establishing a real structure finite element model according to the constitutive relation and the deep learning model, defining a loading condition and an unloading path according to an actual stress condition of the structure, obtaining a finite element simulation result of the real structure, and completing an optimized finite element iteration process based on deep learning, including:
s51: establishing a real structure finite element model adopting the constitutive relation, and defining a loading condition and an loading and unloading path according to the actual stress condition of the structure;
s52: the initial iterative calculation adopts the traditional numerical methodWhen the number of iteration steps is larger than M, the first M steps of iteration process is used as time sequence input to predict the first step based on the trained deep learning modeliGeneralized force vectors of state points after the secondary iteration; according to the conventional numerical methodiOther variables of the structure state of the secondary iteration and the residual error of the iteration objective function;
s53: using the predicted generalized force vector and the calculated other variables of the constitutive state in the second stepi+1 iteration to getiAnd (6) continuously iterating the prediction result of the state point in the step +1 until the residual error of the iteration target function meets the requirement, and completing the process of deep learning optimization finite element iteration to obtain a finite element simulation result of the real structure.
2. The optimized finite element iterative process method based on deep learning of claim 1, wherein in the step S1, the constitutive relation comprises: a macroscopic constitutive relation of a component level or a microscopic constitutive relation of a material level.
3. The optimized finite element iterative process method based on deep learning of claim 2, wherein in step S2, the loading condition comprises one degree of freedom directional loading or multiple degrees of freedom directional coupling loading.
4. An optimized finite element iterative process device based on deep learning, which is characterized by comprising:
the finite element model establishing module is used for setting a constitutive relation and establishing a single-unit finite element model adopting the constitutive relation;
the iteration expansion module is used for randomly generating a load working condition and a loading and unloading path in the single unit finite element model and performing multiple expansion on the preset maximum iteration times of the single unit finite element model; with expansion times being preset maximum number of iterations in the solution processxMultiple, set the number of iterations after expansion toN;
The finite element model calculation module is used for calculating and solving the finite element model of the single unit and summarizing model iteration process data which is calculated and converged to form a data set; dividing the data set into an easy convergence data set, a difficult convergence data set and an unconverged data set from small iteration times to large iteration times;
the data set includes state points for each iteration step, the state points including: generalized force vectors, generalized displacement vectors and other variables of the constitutive state of each iteration step;
the finite element model calculation module is also used for dividing the data set into an easy convergence data set, a difficult convergence data set and an unconverged data set from small to large according to the iteration times; dividing the converged data into data which is difficult to converge and data which is easy to converge according to a ratio of 1:4, wherein the maximum iteration number of the data which is easy to converge is M;
the finite element model training module is used for carrying out deep learning on the data set which is difficult to converge of the single unit finite element model to obtain a deep learning model;
the finite element model training module is also used for dividing the data set which is difficult to converge into a training set, a verification set and a test set according to the proportion of 6:2: 2; according to the divided training set, verification set and test set, the first step isiThe sub-iteration generalized force vector, the generalized displacement vector and other variables of the constitutive state are taken as inputi+xTaking the sub-iteration generalized force vector as output, and performing deep learning neural network training based on loading and unloading time sequence data to obtain a deep learning model;
the real structure model optimization module is used for establishing a real structure finite element model according to the constitutive relation and the deep learning model, defining a loading working condition and a loading and unloading path according to the actual stress condition of the structure, obtaining a finite element simulation result of the real structure and completing an optimized finite element iteration process based on the deep learning;
the real structure model optimization module is also used for establishing a real structure finite element model adopting a constitutive relation and defining a loading working condition and a loading and unloading path according to the actual stress condition of the structure;
the initial iterative calculation adopts a traditional numerical method, and when the iterative step number is larger than M, the previous M-step iterative process is used as the basis of the trained deep learning modelFor time-series input, predictiGeneralized force vectors of state points after the secondary iteration; according to the conventional numerical methodiOther variables of the structure state of the secondary iteration and the residual error of the iteration objective function;
using the predicted generalized force vector and the calculated other variables of the constitutive state in the second stepi+1 iteration to getiAnd (6) continuously iterating the prediction result of the state point in the step +1 until the residual error of the iteration target function meets the requirement, and completing the process of deep learning optimization finite element iteration to obtain a finite element simulation result of the real structure.
5. The deep learning based optimized finite element iterative process device according to claim 4, wherein the constitutive relation comprises: a macroscopic constitutive relation of a component level or a microscopic constitutive relation of a material level.
6. The optimized finite element iterative process device based on deep learning of claim 5, wherein the loading condition comprises single degree of freedom directional loading or multiple degree of freedom directional coupling loading.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210002593.XA CN114021414B (en) | 2022-01-05 | 2022-01-05 | Finite element iteration process optimization method and device based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210002593.XA CN114021414B (en) | 2022-01-05 | 2022-01-05 | Finite element iteration process optimization method and device based on deep learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN114021414A CN114021414A (en) | 2022-02-08 |
CN114021414B true CN114021414B (en) | 2022-04-08 |
Family
ID=80069331
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210002593.XA Active CN114021414B (en) | 2022-01-05 | 2022-01-05 | Finite element iteration process optimization method and device based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114021414B (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116975989B (en) * | 2023-09-25 | 2024-02-27 | 深圳市城市交通规划设计研究中心股份有限公司 | Method for evaluating bridge running state based on finite element guided deep learning agent model |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111881601A (en) * | 2020-07-01 | 2020-11-03 | 中国电建集团昆明勘测设计研究院有限公司 | Earth and rockfill dam material constitutive model construction method based on deep learning and finite element unit method |
CN112163328A (en) * | 2020-09-18 | 2021-01-01 | 武汉大学 | Geotechnical particle material constitutive modeling method based on deep learning and data driving |
CN112651153A (en) * | 2020-12-09 | 2021-04-13 | 中南大学 | Method for determining material parameters of crystal plastic finite element model |
Family Cites Families (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10740513B2 (en) * | 2013-12-20 | 2020-08-11 | Dassault Systèmes Simulia Corp. | Intra-increment adjustments of implicit finite element simulation |
US10489533B2 (en) * | 2015-07-31 | 2019-11-26 | Autodesk, Inc. | Techniques for warm starting finite element analyses with deep neural networks |
US20200397283A1 (en) * | 2018-02-28 | 2020-12-24 | Narayana Nethralaya Foundation | A method to quantify the corneal parameters to improve biomechanical modeling |
CN110362860B (en) * | 2019-06-06 | 2023-05-09 | 国网江西省电力有限公司电力科学研究院 | Electric field measuring instrument bracket optimization method based on finite element simulation and differential evolution algorithm |
CN111046609A (en) * | 2019-12-24 | 2020-04-21 | 上海智殷自动化科技有限公司 | BP neural network based welding process parameter optimization method |
US11093831B1 (en) * | 2020-02-08 | 2021-08-17 | Suntracker Technologies Ltd. | System and method for neural network radiosity calculations |
CN112784331B (en) * | 2020-09-25 | 2022-09-13 | 汕头大学 | Soil stress-strain relation determination method based on improved LSTM deep learning method |
-
2022
- 2022-01-05 CN CN202210002593.XA patent/CN114021414B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111881601A (en) * | 2020-07-01 | 2020-11-03 | 中国电建集团昆明勘测设计研究院有限公司 | Earth and rockfill dam material constitutive model construction method based on deep learning and finite element unit method |
CN112163328A (en) * | 2020-09-18 | 2021-01-01 | 武汉大学 | Geotechnical particle material constitutive modeling method based on deep learning and data driving |
CN112651153A (en) * | 2020-12-09 | 2021-04-13 | 中南大学 | Method for determining material parameters of crystal plastic finite element model |
Non-Patent Citations (1)
Title |
---|
板带轧制过程刚塑性有限元求解的初速度场设定;陈伟 等;《江苏技术师范学院学报(自然科学版)》;20081231;第14卷(第4期);全文 * |
Also Published As
Publication number | Publication date |
---|---|
CN114021414A (en) | 2022-02-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Jensen et al. | Implementation of an adaptive meta-model for Bayesian finite element model updating in time domain | |
CN111970163B (en) | Network flow prediction method of LSTM model based on attention mechanism | |
Jensen et al. | An adaptive scheme for reliability-based global design optimization: A Markov chain Monte Carlo approach | |
Jensen et al. | Reliability sensitivity analysis of stochastic finite element models | |
Wei et al. | Extended Monte Carlo simulation for parametric global sensitivity analysis and optimization | |
Möller et al. | Structural optimization for performance-based design in earthquake engineering: Applications of neural networks | |
WO2021164250A1 (en) | Turbulence field update method and apparatus, and related device | |
Chen et al. | Sequential design strategies for mean response surface metamodeling via stochastic kriging with adaptive exploration and exploitation | |
CN106372278A (en) | Sensitivity analysis method jointly considering input parameter uncertainty and proxy model uncertainty | |
JP2023510922A (en) | Optimizing High-Cost Functions for Complex Multidimensional Constraints | |
Chang | Dissipative, noniterative integration algorithms with unconditional stability for mildly nonlinear structural dynamic problems | |
Jensen et al. | Model-reduction techniques for reliability-based design problems of complex structural systems | |
CN114021414B (en) | Finite element iteration process optimization method and device based on deep learning | |
CN110990135A (en) | Spark operation time prediction method and device based on deep migration learning | |
Tamura et al. | Machine learning for combinatorial optimization of brace placement of steel frames | |
Ryan et al. | A gaussian process modeling approach for fast robust design with uncertain inputs | |
Yang et al. | Structural design optimization under dynamic reliability constraints based on the probability density evolution method and highly-efficient sensitivity analysis | |
Luo et al. | Active Kriging-based conjugate first-order reliability method for highly efficient structural reliability analysis using resample strategy | |
CN113537614A (en) | Construction method, system, equipment and medium of power grid engineering cost prediction model | |
CN116011071A (en) | Method and system for analyzing structural reliability of air building machine based on active learning | |
Amrane et al. | On the use of ensembles of metamodels for estimation of the failure probability | |
Liu et al. | Modified sequential bifurcation for simulation factor screening under skew-normal response model | |
Ramu et al. | Design Optimization of Complex Structures Using Metamodels. | |
Abdollahi et al. | Soft Monte Carlo Simulation for imprecise probability estimation: A dimension reduction-based approach | |
Kowalczyk et al. | Numpress–integrated computer system for analysis and optimization of industrial sheet metal forming processes |
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 |