Stress prediction method and device, equipment and medium
Through the physical information neural network model based on elastic-plastic constitutive relationship, the stress in the product manufacturing process is predicted, which solves the problem of difficulty in collecting stress data in the prior art, and effectively predicts and collects stresses.
Patent Information
- Application Number
- CN202411990307.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-31
- Publication Date
- 2025-04-29
AI Technical Summary
In the prior art, stress data is difficult to collect during product manufacturing, resulting in the inability to effectively control the influence of multiple stresses.
By obtaining the spatial coordinates of the target object to be pressed and the loading time during the pressing process, it is input to the physical information neural network model trained based on the elastic-plastic constitutive relationship, stress prediction is performed, and the prediction results of stress and strain are obtained.
It realizes the prediction of stress during product manufacturing process, and collects stress data to help ensure the safety of engineering structures and optimize manufacturing processes.
Smart Images
Figure CN119903743A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of stress prediction technology, and in particular to a stress prediction method, device, equipment and medium. Background Art
[0002] With the development of science and technology, the functions and structures of products are becoming increasingly complex. Stress analysis is of great significance in ensuring the safety of engineering structures and optimizing manufacturing processes. When a product is affected by multiple stresses at the same time, each stress needs to be strictly controlled. However, it is currently difficult to collect stress data for the product manufacturing process. Summary of the invention
[0003] The purpose of the present invention is to provide a stress prediction method, device, equipment and medium for solving the problem that stress data in the product manufacturing process is difficult to collect.
[0004] In order to achieve the above object, in a first aspect, the present invention provides a stress prediction method, comprising:
[0005] Obtaining the spatial coordinates of the target object to be pressed;
[0006] Inputting the spatial coordinates of the target object and the loading time corresponding to the pressing process of the target object into a stress prediction model to perform stress prediction, and obtaining a prediction result, wherein the prediction result includes a stress prediction result and a strain prediction result;
[0007] The stress prediction model is obtained by training a physical information neural network based on the elastic-plastic constitutive relationship of each training object during the pressing process.
[0008] In some embodiments, the method further comprises:
[0009] Acquire a training data set, wherein the training data set includes the spatial coordinates of each training object among a plurality of training objects and a corresponding loading time during a pressing process of the training object;
[0010] Based on the training data set, a physical information artifact network is trained using a partial differential equation loss function to obtain a pre-trained model, wherein the partial differential equation loss function is related to a loss term of an elastic-plastic constitutive equation of the training object during a pressing process, and the elastic-plastic constitutive equation is used to characterize an elastic-plastic constitutive relationship;
[0011] Based on the training data set, the pre-training model is trained using a first loss function to obtain the stress prediction model, wherein the first loss function is related to the loss term of the boundary conditions and the loss term of the initial conditions of the training object during the pressing process.
[0012] In some embodiments, the partial differential equation loss function is the sum of the loss function of the elastic-plastic constitutive equation of the training object during the pressing process, the loss function of the stress balance equation of the training object during the pressing process, and the loss function of the geometric equation that characterizes the geometric relationship between stress and displacement of the training object during the pressing process; wherein, the loss function of the elastic-plastic constitutive equation is the product of the loss term of the elastic-plastic constitutive equation and the first weight coefficient, the loss function of the stress balance equation is the product of the loss term of the stress balance equation and the second weight coefficient, and the loss function of the geometric equation is the product of the loss term of the geometric equation and the third weight coefficient.
[0013] In some embodiments, the pre-training model is trained based on the training data set using a first loss function to obtain the stress prediction model, including:
[0014] Based on the training data set, in the i-th training process, for each loss term constituting the first loss function before weight adjustment, a new weight coefficient corresponding to the loss term is determined according to the average gradient of the loss term and the size of the loss term itself, i=1, 2, ...;
[0015] Determine a weight-adjusted first loss function according to each loss term constituting the first loss function and a new weight coefficient corresponding to each loss term;
[0016] Using the weight-adjusted first loss function, adjusting the network parameters of the pre-trained model for the (i+1)th training;
[0017] After assigning i+1 to i, return to execute the above steps until the training meets the conditions, and obtain the stress prediction model.
[0018] In some embodiments, determining a new weight coefficient corresponding to the loss term according to the average gradient of the loss term and the size of the loss term itself includes:
[0019] Calculate a fourth weight coefficient according to the average gradient of the loss term and the size of the loss term itself;
[0020] A new weight coefficient corresponding to the loss term is calculated based on the weight coefficient corresponding to the loss term before weight adjustment and the fourth weight coefficient.
[0021] In some embodiments, adjusting the network parameters of the pre-trained model using the weight-adjusted first loss function includes:
[0022] Fixing the weights and parameters of the preset layers of the pre-trained model;
[0023] The weights and parameters of the remaining layers of the pre-trained model except the preset layer are adjusted using a first learning rate and the first loss function after the weight adjustment, wherein the first learning rate is less than a preset threshold.
[0024] In a second aspect, the present invention further provides a stress prediction device, comprising:
[0025] A first acquisition module is used to acquire the spatial coordinates of the target object to be pressed;
[0026] A stress prediction module, used for inputting the spatial coordinates of the target object and the loading time corresponding to the pressing process of the target object into a stress prediction model to perform stress prediction and obtain a prediction result, wherein the prediction result includes a stress prediction result and a strain prediction result;
[0027] The stress prediction model is obtained by training a physical information neural network based on the elastic-plastic constitutive relationship of each training object during the pressing process.
[0028] In a third aspect, the present invention further provides a stress prediction device, comprising a processor and a transceiver, wherein the transceiver receives and sends data under the control of the processor, and the processor is used to perform the following operations:
[0029] Obtaining the spatial coordinates of the target object to be pressed;
[0030] Inputting the spatial coordinates of the target object and the loading time corresponding to the pressing process of the target object into a stress prediction model to perform stress prediction, and obtaining a prediction result, wherein the prediction result includes a stress prediction result and a strain prediction result;
[0031] The stress prediction model is obtained by training a physical information neural network based on the elastic-plastic constitutive relationship of each training object during the pressing process.
[0032] In a fourth aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the stress prediction method as described in the first aspect.
[0033] In a fifth aspect, the present invention further provides a computer program product, comprising computer instructions, which, when executed by a processor, implement the steps in the stress prediction method as described in the first aspect.
[0034] The above technical solution of the present invention has at least the following beneficial effects:
[0035] In an embodiment of the present invention, the spatial coordinates of the target object to be pressed and formed are obtained; the spatial coordinates of the target object and the corresponding loading time during the pressing process of the target object are input into a stress prediction model to perform stress prediction and obtain prediction results, and the prediction results include stress prediction results and strain prediction results; wherein the stress prediction model is based on the elastic-plastic constitutive relationship of each training object during the pressing process and is obtained by training a physical information neural network. In this way, the stress prediction model obtained by training a physical information neural network using the elastic-plastic deformation theory can realize stress prediction of the product during the manufacturing process and collect stress data. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 A schematic diagram showing a flow chart of a stress prediction method according to an embodiment of the present invention;
[0037] Figure 2 A schematic diagram showing the training process of the pre-training model according to an embodiment of the present invention;
[0038] Figure 3 A schematic diagram showing a flow chart of a model training process using a transfer learning solution according to an embodiment of the present invention;
[0039] Figure 4 A schematic diagram showing the training process of the stress prediction model according to an embodiment of the present invention;
[0040] Figure 5 A schematic diagram showing a module of a stress prediction device according to an embodiment of the present invention;
[0041] Figure 6 A schematic diagram showing the hardware structure of a stress prediction device according to an embodiment of the present invention. DETAILED DESCRIPTION
[0042] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.
[0043] like Figure 1 FIG. 1 is a flow chart of a stress prediction method provided by an embodiment of the present invention. The method may specifically include:
[0044] Step 101, obtaining the spatial coordinates of the target object to be pressed;
[0045] Here, the target object to be pressed and formed may be a certain product to be pressed and formed. It should be understood that the spatial coordinates of the target object refer to the spatial coordinate points of the target object in three-dimensional space.
[0046] Step 102, inputting the spatial coordinates of the target object and the loading time corresponding to the pressing process of the target object into a stress prediction model to perform stress prediction and obtain a prediction result, wherein the prediction result includes a stress prediction result and a strain prediction result; wherein the stress prediction model is based on the elastic-plastic constitutive relationship of each training object during the pressing process and is obtained by training a physical information neural network.
[0047] The above method of the embodiment of the present invention utilizes the elastic-plastic deformation theory and the stress prediction model obtained by training the physical information neural network, so as to realize the stress prediction of the product during the manufacturing process and collect stress data.
[0048] It should be noted that the stress prediction model is pre-trained. How to train the stress prediction model can be seen in the following embodiment.
[0049] As an optional embodiment, the method of the present invention further includes:
[0050] 1) obtaining a training data set, wherein the training data set includes the spatial coordinates of each training object among a plurality of training objects and the corresponding loading time during the pressing process of the training object;
[0051] 2) based on the training data set, using a partial differential equation loss function to train a physical information artifact network to obtain a pre-trained model, wherein the partial differential equation loss function is related to a loss term of an elastic-plastic constitutive equation of the training object during a pressing process, and the elastic-plastic constitutive equation is used to characterize an elastic-plastic constitutive relationship;
[0052] Optionally, the elastic-plastic constitutive equation is constructed based on consistency conditions, fluidity laws and elastic stress-strain relationships, and the elastic-plastic constitutive equation is used to describe the incremental relationship between stress and strain of the training object during the pressing process.
[0053] Here, the stress prediction model is obtained by training the physical information neural network (Physics-Informed Neural Networks, referred to as PINN). Among them, in order to construct the physical information module in the physical information neural network architecture, it is necessary to describe the elastic-plastic constitutive equation in the product manufacturing process. Taking the Planger hardening criterion as its subsequent yield function after entering the injection molding deformation, the incremental theory is used to construct its stress-strain relationship. Specifically, during the press-fitting process, when the stress point in the product is already on the yield surface and continues to be elastically-plastically loaded, the incremental method is required to construct the stress-strain relationship of the elastic-plastic behavior of the product. The above-mentioned elastic-plastic constitutive equation is used to describe the incremental relationship between stress and strain of the training object during the pressing process.
[0054] Among them, the elastic-plastic constitutive equation is constructed based on the consistency condition, fluidity law and elastic stress-strain relationship, that is, the construction of the elastic-plastic constitutive equation must follow the following principles.
[0055] a.Consistency conditions
[0056] When the product is subjected to external loading during the press-fitting process, the new stress point (σ ij +dσ ij ) is still on the yield surface.
[0057] b. Liquidity Law
[0058] New elastic-plastic strain increment The direction of the external normal divided by the original stress point on the yield surface.
[0059] c. Elastic stress-strain relationship
[0060] Strain increment dε ij The elastic part and stress increment dσ ij It still obeys the generalized Hooke's law of elasticity.
[0061] Based on the above three principles, the stress-strain relationship of the product manufacturing process is constructed. According to the elastic-plastic increment theory, the strain increment dε ij It can be regarded as the elastic strain increment and plastic strain increment The sum of
[0062]
[0063] From the above basic principles, it can be seen that the elastic strain increment conforms to the generalized Hooke's law, that is,
[0064]
[0065] in, is the stiffness matrix, which is a fourth-order tensor. According to the flow law, the plastic strain increment It can be expressed as:
[0066]
[0067] Among them, dλ is the size of the plastic increment, which can be determined according to the consistency condition. f is related to the subsequent yield function of the Planck kinematic hardening criterion, which is:
[0068] F=fK=0 (4)
[0069]
[0070] Among them, α and K are experimental constants related to the friction angle and cohesion, obtained from uniaxial compression tests. I1 is the first invariant of stress, and its specific form is:
[0071] I1=σ ii =σ1+σ2+σ3 (6)
[0072] J2 is the second invariant of stress deviator, and its specific expression is:
[0073]
[0074] At the same time, the strain and displacement of the product manufacturing process also satisfy the geometric relationship, that is,
[0075]
[0076] Given the conditions for the product material to undergo elastic-plastic deformation, determine whether it is in the elastic loading or plastic loading stage. In the elastic zone,
[0077] F<0 (9)
[0078] In the plastic zone,
[0079]
[0080] Finally, the elastic-plastic constitutive equation can be described by the following form:
[0081]
[0082] in, It can be calculated by formula (3).
[0083] It should be noted that the physical information neural network is based on a fully connected neural network and has the basic structure of a fully connected network, including an input layer, an output layer, and a hidden layer in the middle.
[0084] Taking three-dimensional space as an example, the output layer contains the spatial coordinate points X=(X1, X2, X3) in three-dimensional space. These coordinate points are configured according to specific boundary conditions and model training requirements. Generally speaking, the coordinate points inside the plane should be kept as uniform as possible. The elastic-plastic changes in the product manufacturing process are related to the loading path and loading history. Here, the product manufacturing process is regarded as a quasi-static process. The time t is set, and t is made to be the parameter t of the loading displacement. In this way, the loading displacement can be associated with time t, thereby adjusting the loading path. Therefore, the input of the physical information neural network includes spatial coordinates (X1, X2, X3) and time t.
[0085] The output layer of the physical information neural network depends on the object to be predicted. In order to predict the stress and strain of the product manufacturing process, the unknown stress field tensor σ(X,t)=(σ11 ,σ 22 ,σ 33 ,σ 12 ,σ 13 ,σ 23 ) and the corresponding strain tensor ε(X,t)=(ε 11 ,ε 22 ,ε 33 ,ε 12 ,ε 13 ,ε 23 ) and displacement tensor u(X,t)=(u1,u2,u3) and other 8 variables are used as the output of the physical information neural network, so that the output of the physical information neural network can fit the stress and strain in the product manufacturing process. Let θ represent the internal parameters of the physical information neural network, then the physical information neural network can be expressed as follows:
[0086] [σ,ε,u]=N(X,t;θ)
[0087] Physical information neural network establishes the mapping relationship between input and output by taking the error term of physical information as part of the loss function.
[0088] Among them, the loss function of the physical information neural network is generally composed of data loss terms, partial differential equation (PDE) loss terms, boundary condition loss terms, initial condition loss terms and other loss terms.
[0089] Optionally, the partial differential equation loss function is the sum of the loss function of the elastic-plastic constitutive equation of the training object during the pressing process, the loss function of the stress balance equation of the training object during the pressing process, and the loss function of the geometric equation that characterizes the geometric relationship between stress and displacement of the training object during the pressing process; wherein, the loss function of the elastic-plastic constitutive equation is the product of the loss term of the elastic-plastic constitutive equation and the first weight coefficient, the loss function of the stress balance equation is the product of the loss term of the stress balance equation and the second weight coefficient, and the loss function of the geometric equation is the product of the loss term of the geometric equation and the third weight coefficient.
[0090] Among them, the loss term of the partial differential equation is constructed based on the stress balance equation, the elastic-plastic constitutive equation and the geometric equation. In the stress balance equation, there is the following relationship:
[0091] ▽σ(X N )+f=0, X N ∈Ω (12)
[0092] Among them, Ω is the problem domain in the product manufacturing process, f is the body force, and σ(X) is the unknown stress field in the domain. When the body force is 0, the stress balance equation can be described as:
[0093]
[0094] When using the elastic-plastic constitutive equation to construct the loss term, since the physical information neural network outputs stress and strain, and the elastic-plastic constitutive equation describes the incremental relationship between stress and strain, the loss function uses the derivative of the stress variable with respect to time t to replace the increment of stress and strain. Therefore, the elastic-plastic constitutive equation can be described by the above formula (11).
[0095] Optionally, the loss function is expressed as mean square error loss, so the partial differential equation loss function can be described as:
[0096]
[0097] Where, L T is the loss term of the stress balance equation, L F is the loss term of the elastoplastic constitutive equation, L G is the loss term of the geometric equation, ω T is the second weight coefficient, ω F is the first weight coefficient, ω G is the third weight coefficient.
[0098] Here, the physical information artifact network is pre-trained using PDE loss, so that the pre-trained model (its model structure see Figure 2 As shown in the figure, the loss function (i.e., the first loss function) during subsequent training under new boundary conditions converges faster, thereby improving the model training speed.
[0099] 3) Based on the training data set, the pre-trained model is trained using a first loss function to obtain the stress prediction model, wherein the first loss function is related to the loss term of the boundary conditions and the loss term of the initial conditions of the training object during the pressing process.
[0100] In order to make the stress and strain output by the stress prediction model meet the boundary conditions and initial conditions, it is necessary to construct the loss function of the boundary conditions and initial conditions in the loss term, that is, the first loss function. Among them, the boundary conditions of the product manufacturing process are set as:
[0101] u(t,X U )=f(t,X U ), X U ∈Γ D1 (15)
[0102] Among them, Γ D1is the boundary range, u(t,X) represents the displacement of the boundary, which changes with time, and f(t,X) is a custom function. By adjusting this function, the boundary conditions of the displacement can be changed. The initial displacement conditions of the product manufacturing process are set as:
[0103]
[0104] In addition to setting the initial displacement conditions, the initial stress conditions must also be set. Since the gaps between particles are large in the initial stage of the product manufacturing process, the stress and strain at this stage are difficult to predict, so a certain stress state in the middle of the manufacturing process is selected as the initial condition, and the initial condition is set as follows:
[0105]
[0106] Therefore, the mathematical expression of the loss term of the boundary condition is:
[0107]
[0108] The mathematical expression of the loss term of the initial condition is:
[0109]
[0110] Therefore, the first loss function can be described as:
[0111] L1=ω BC L BC +ω IC L IC (20)
[0112] In some embodiments, the above 3) based on the training data set, using the first loss function to train the pre-trained model to obtain the stress prediction model may include:
[0113] Step 31, based on the training data set, in the i-th training process, for each loss term constituting the first loss function before weight adjustment, according to the average gradient of the loss term and the size of the loss term itself, determine a new weight coefficient corresponding to the loss term, i=1, 2, ...;
[0114] Step 32, determining a weight-adjusted first loss function according to each loss term constituting the first loss function and a new weight coefficient corresponding to each loss term;
[0115] It should be noted that the first loss function contains multiple loss terms. When training the model, unbalanced training of the loss terms will lead to reduced accuracy of the prediction results. Through the weight optimization processing of the above steps 31 and 32, that is, dynamically weighting the loss terms of the loss function based on the weighting strategy of the average gradient value of the loss term and the size of the loss term itself, it is possible to avoid the gradient of a loss term of the loss function being 0 during the training process and to avoid imbalance in training caused by a loss term being too large, thereby improving the accuracy of the prediction results.
[0116] In an optional embodiment, step 32, determining a new weight coefficient corresponding to the loss term according to the average gradient of the loss term and the size of the loss term itself, may include:
[0117] ① Calculate the fourth weight coefficient according to the average gradient of the loss term and the size of the loss term itself;
[0118] Here, the first loss function can also be written as:
[0119]
[0120] Among them, L1(θ) represents the loss function of the pre-trained model when the internal parameters of the network are θ, L i (θ) represents the loss terms that make up the first loss function, ω i is the weight coefficient of the corresponding loss term, and the weighted expression based on the average gradient value of the loss term is:
[0121]
[0122] The weighted average gradient value of the loss term measures the reciprocal of the average gradient of each loss term of the pre-trained model at the τth iteration, and its denominator is is the average gradient of the i-th loss term. In order to keep the numerator and denominator the same in order of magnitude, the maximum gradient value of the first loss term is selected As a molecule.
[0123] The weighted expression based on the size of the loss term itself is:
[0124]
[0125] Therefore, based on the average gradient value of the loss term and the size of the loss term itself, the fourth weight coefficient is calculated It is expressed as:
[0126]
[0127] Fourth weight coefficient The value of is affected by both the average gradient value of the loss term and the size of the loss term itself.
[0128] ② According to the weight coefficient corresponding to the loss item before the weight adjustment and the fourth weight coefficient, a new weight coefficient corresponding to the loss item is calculated.
[0129] Specifically, according to the weight coefficient corresponding to the loss term before the weight adjustment and the fourth weight coefficient, a new weight coefficient ω corresponding to the loss term is calculated. i (τ+1) is expressed as:
[0130]
[0131] It should be understood that i (τ) is the weight coefficient corresponding to the i-th loss term before weight adjustment.
[0132] When the gradient of a loss term is too small or the value of the loss term is too large, the weight of the loss term will be increased in the next iteration, so that the weights of each loss term of the first loss function are dynamically adjusted during the model training process.
[0133] To explain the weight optimization algorithm in detail, the pseudo code of the algorithm is shown below.
[0134] Algorithm 1: Weight Optimization Algorithm
[0135] 1: Define a loss function with multiple loss terms, where M is the total number of loss terms
[0136]
[0137] 2: Assume the number of pre-training model iterations is S:
[0138] for τ=1,…,S executes
[0139] 3: Calculation
[0140] 4: Update weight ω i (τ),
[0141] α is the coefficient of the moving average, which is a fixed value and is set to 0.2 here.
[0142] 5: Update the internal parameters θ of the pre-trained model through the loss function:
[0143]
[0144] β is the learning rate parameter of the pre-trained model, which is set to 10 here. -2
[0145] 6: end
[0146] Step 33, using the first loss function after the weight adjustment, adjusting the network parameters of the pre-trained model for the (i+1)th training;
[0147] Step 34, after assigning i+1 to i, return to execute the above steps until the training meets the conditions, and obtain the stress prediction model.
[0148] Here, the training condition can be that the number of training (iterations) reaches a preset number, or the model is trained until convergence (that is, the loss value of the loss function no longer decreases).
[0149] It should be noted that since product manufacturing often requires different manufacturing conditions, and each manufacturing condition requires training a new stress prediction model to predict stress, this will greatly increase the training cost of the stress prediction model and reduce the training efficiency. In order to solve this problem, the above 2) and 3) are adopted, that is, firstly, the physical information neural network is pre-trained using the elastic-plastic constitutive relationship, and secondly, based on the pre-trained model, the pre-trained model is fine-tuned according to the specific boundary conditions.
[0150] In some embodiments, step 33 may include:
[0151] Fixing the weights and parameters of the preset layers of the pre-trained model;
[0152] The weights and parameters of the remaining layers of the pre-trained model except the preset layer are adjusted using a first learning rate and the first loss function after the weight adjustment, wherein the first learning rate is less than a preset threshold.
[0153] Here, based on the pre-trained model, the transfer learning method is used to fix (freeze) the weights and parameters of the preset layers of the pre-trained model, and the weights and parameters of the remaining layers of the pre-trained model except the preset layers are fine-tuned at a smaller learning rate. Transfer learning is used to improve the generalization of the stress prediction model to speed up the training of the stress prediction model under different boundary conditions.
[0154] It should be noted that when fine-tuning the pre-trained model, the loss function needs to be changed to a loss term consisting of boundary conditions, that is, the first loss function, so that the fine-tuned pre-trained model can adapt to the new boundary conditions. Figure 3 shown.
[0155] The total loss function of the stress prediction model is the sum of the loss term of the partial differential equation PDE, the loss term of the boundary condition, and the loss term of the initial condition. Its expression is:
[0156] L=ω PDE L PDE +ω BC L BC +ω ICL IC
[0157] The stress prediction model performs back propagation through the loss function, so that the stress and strain output by the stress prediction model satisfies the physical equation (i.e., PDE) in the loss function as well as the boundary conditions and initial conditions, and updates the internal parameters of the network. Figure 4 , describing the overall network architecture and back-propagation process of the stress prediction model.
[0158] The stress prediction method of the present invention dynamically weights each loss term of the loss function through a weighting strategy based on the size of the loss term itself and the average gradient of the loss term, so as to avoid the gradient of a certain loss term of the loss function falling to 0 during the training process and avoid the imbalance of training caused by a certain loss term being too large. Compared with the stress prediction model with fixed weights, the stress prediction model using the dynamic weighting strategy can further improve the accuracy of stress prediction; by pre-training the stress prediction model using the PDE loss, the loss function of the pre-trained model converges faster when trained under new boundary conditions, and the model training speed is improved, which all help to improve the feasibility of the stress prediction model in application, and improve the prediction effect and training speed of the stress prediction model. In terms of grid generation, the physical information neural network provides a gridless method for predicting stress and strain. In terms of improving the prediction effect, in addition to the dynamic weighting method, there are also methods such as domain decomposition and parallel computing of the physical information neural network, which is also an aspect that can further improve the prediction effect of the stress prediction model in the future.
[0159] like Figure 5 As shown, an embodiment of the present invention further provides a stress prediction device, the device comprising:
[0160] A first acquisition module 501 is used to acquire the spatial coordinates of the target object to be pressed;
[0161] A stress prediction module 502 is used to input the spatial coordinates of the target object and the loading time corresponding to the pressing process of the target object into a stress prediction model to perform stress prediction and obtain a prediction result, wherein the prediction result includes a stress prediction result and a strain prediction result;
[0162] The stress prediction model is obtained by training a physical information neural network based on the elastic-plastic constitutive relationship of each training object during the pressing process.
[0163] Optionally, the device of the present invention further comprises:
[0164] A second acquisition module is used to acquire a training data set, wherein the training data set includes the spatial coordinates of each training object in a plurality of training objects and a corresponding loading time during a pressing process of the training object;
[0165] A first model training module is used to train the physical information artifact network based on the training data set using a partial differential equation loss function to obtain a pre-trained model, wherein the partial differential equation loss function is related to a loss term of an elastic-plastic constitutive equation of the training object during a pressing process, and the elastic-plastic constitutive equation is used to characterize an elastic-plastic constitutive relationship;
[0166] The second model training module is used to train the pre-trained model based on the training data set using a first loss function to obtain the stress prediction model. The first loss function is related to the loss terms of the boundary conditions and the initial conditions of the training object during the pressing process.
[0167] Optionally, the partial differential equation loss function is the sum of the loss function of the elastic-plastic constitutive equation of the training object during the pressing process, the loss function of the stress balance equation of the training object during the pressing process, and the loss function of the geometric equation that characterizes the geometric relationship between stress and displacement of the training object during the pressing process; wherein, the loss function of the elastic-plastic constitutive equation is the product of the loss term of the elastic-plastic constitutive equation and the first weight coefficient, the loss function of the stress balance equation is the product of the loss term of the stress balance equation and the second weight coefficient, and the loss function of the geometric equation is the product of the loss term of the geometric equation and the third weight coefficient.
[0168] Optionally, the second model training module includes:
[0169] A first processing unit is configured to determine, based on the training data set, in an i-th training process, for each loss term constituting the first loss function before weight adjustment, according to an average gradient of the loss term and a size of the loss term itself, a new weight coefficient corresponding to the loss term, i=1, 2, ...;
[0170] A weight optimization unit, used to determine a weight-adjusted first loss function according to each loss term constituting the first loss function and a new weight coefficient corresponding to each loss term;
[0171] A network parameter adjustment unit, used to adjust the network parameters of the pre-trained model for the (i+1)th training by using the first loss function after the weight adjustment;
[0172] After assigning i+1 to i, return and execute the steps executed by the above units in sequence until the training meets the conditions, and the stress prediction model is obtained.
[0173] Optionally, the first processing unit is specifically configured to:
[0174] Calculate a fourth weight coefficient according to the average gradient of the loss term and the size of the loss term itself;
[0175] A new weight coefficient corresponding to the loss term is calculated based on the weight coefficient corresponding to the loss term before weight adjustment and the fourth weight coefficient.
[0176] Optionally, the network parameter adjustment unit is specifically used for:
[0177] Fixing the weights and parameters of the preset layers of the pre-trained model;
[0178] The weights and parameters of the remaining layers of the pre-trained model except the preset layer are adjusted using a first learning rate and the first loss function after the weight adjustment, wherein the first learning rate is less than a preset threshold.
[0179] The stress prediction device of the embodiment of the present invention obtains the spatial coordinates of the target object to be pressed and formed; inputs the spatial coordinates of the target object and the corresponding loading time during the pressing process of the target object into a stress prediction model to perform stress prediction and obtain prediction results, and the prediction results include stress prediction results and strain prediction results; wherein the stress prediction model is based on the elastic-plastic constitutive relationship of each training object during the pressing process and is obtained by training a physical information neural network. In this way, the stress prediction model obtained by training a physical information neural network using the elastic-plastic deformation theory can realize stress prediction of the product during the manufacturing process and collect stress data.
[0180] In order to better achieve the above goals, Figure 6 As shown, an embodiment of the present invention further provides a stress prediction device, including a processor 600 and a transceiver 610, wherein the transceiver 610 receives and sends data under the control of the processor 600, and the processor 600 is used to perform the following process:
[0181] Obtaining the spatial coordinates of the target object to be pressed;
[0182] Inputting the spatial coordinates of the target object and the loading time corresponding to the pressing process of the target object into a stress prediction model to perform stress prediction, and obtaining a prediction result, wherein the prediction result includes a stress prediction result and a strain prediction result;
[0183] The stress prediction model is obtained by training a physical information neural network based on the elastic-plastic constitutive relationship of each training object during the pressing process.
[0184] Optionally, the processor 600 is further configured to:
[0185] Acquire a training data set, wherein the training data set includes the spatial coordinates of each training object among a plurality of training objects and a corresponding loading time during a pressing process of the training object;
[0186] Based on the training data set, a physical information artifact network is trained using a partial differential equation loss function to obtain a pre-trained model, wherein the partial differential equation loss function is related to a loss term of an elastic-plastic constitutive equation of the training object during a pressing process, and the elastic-plastic constitutive equation is used to characterize an elastic-plastic constitutive relationship;
[0187] Based on the training data set, the pre-training model is trained using a first loss function to obtain the stress prediction model, wherein the first loss function is related to the loss term of the boundary conditions and the loss term of the initial conditions of the training object during the pressing process.
[0188] Optionally, the partial differential equation loss function is the sum of the loss function of the elastic-plastic constitutive equation of the training object during the pressing process, the loss function of the stress balance equation of the training object during the pressing process, and the loss function of the geometric equation that characterizes the geometric relationship between stress and displacement of the training object during the pressing process; wherein, the loss function of the elastic-plastic constitutive equation is the product of the loss term of the elastic-plastic constitutive equation and the first weight coefficient, the loss function of the stress balance equation is the product of the loss term of the stress balance equation and the second weight coefficient, and the loss function of the geometric equation is the product of the loss term of the geometric equation and the third weight coefficient.
[0189] Optionally, the processor 600 is further configured to:
[0190] Based on the training data set, in the i-th training process, for each loss term constituting the first loss function before weight adjustment, a new weight coefficient corresponding to the loss term is determined according to the average gradient of the loss term and the size of the loss term itself, i=1, 2, ...;
[0191] Determine a weight-adjusted first loss function according to each loss term constituting the first loss function and a new weight coefficient corresponding to each loss term;
[0192] Using the weight-adjusted first loss function, adjusting the network parameters of the pre-trained model for the (i+1)th training;
[0193] After assigning i+1 to i, return to execute the above steps until the training meets the conditions to obtain the stress prediction model.
[0194] Optionally, the processor 600 is further configured to:
[0195] Calculate a fourth weight coefficient according to the average gradient of the loss term and the size of the loss term itself;
[0196] A new weight coefficient corresponding to the loss term is calculated based on the weight coefficient corresponding to the loss term before weight adjustment and the fourth weight coefficient.
[0197] Optionally, the processor 600 is further configured to:
[0198] Fixing the weights and parameters of the preset layers of the pre-trained model;
[0199] The weights and parameters of the remaining layers of the pre-trained model except the preset layer are adjusted using a first learning rate and the first loss function after the weight adjustment, wherein the first learning rate is less than a preset threshold.
[0200] The stress prediction device of the embodiment of the present invention obtains the spatial coordinates of the target object to be pressed and formed; inputs the spatial coordinates of the target object and the corresponding loading time during the pressing process of the target object into a stress prediction model to perform stress prediction and obtain prediction results, and the prediction results include stress prediction results and strain prediction results; wherein the stress prediction model is based on the elastic-plastic constitutive relationship of each training object during the pressing process and is obtained by training a physical information neural network. In this way, by using the elastic-plastic deformation theory and the stress prediction model obtained by training the physical information neural network, it is possible to realize stress prediction of the product during the manufacturing process and collect stress data.
[0201] An embodiment of the present invention further provides a stress prediction device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, each process in the stress prediction method embodiment described above is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be described here.
[0202] The embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, each process in the stress prediction method embodiment described above is implemented, and the same technical effect can be achieved. To avoid repetition, it is not repeated here. The computer-readable storage medium is, for example, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.
[0203] The embodiment of the present invention also provides a computer program product, including computer instructions, which, when executed by a processor, implement the following Figure 1 The various processes of the method embodiment shown can achieve the same technical effect, and will not be described again here to avoid repetition.
[0204] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage and optical storage, etc.) containing computer-usable program code.
[0205] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A device that specifies functions in one or more processes and / or one or more blocks.
[0206] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable storage medium produce a paper product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0207] These computer program instructions may also be loaded onto a computer or other programmable data processing device so that the computer or other programmable device executes a series of operating steps to produce a computer-implemented process, thereby providing instructions executed on the computer or other programmable device for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0208] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.
Claims
1. A stress prediction method, characterized in that: include: Obtaining the spatial coordinates of the target object to be pressed; Inputting the spatial coordinates of the target object and the loading time corresponding to the pressing process of the target object into a stress prediction model to perform stress prediction, and obtaining a prediction result, wherein the prediction result includes a stress prediction result and a strain prediction result; The stress prediction model is obtained by training a physical information neural network based on the elastic-plastic constitutive relationship of each training object during the pressing process.
2. The method according to claim 1, characterized in that The method further comprises: Acquire a training data set, wherein the training data set includes the spatial coordinates of each training object among a plurality of training objects and a corresponding loading time during a pressing process of the training object; Based on the training data set, a physical information artifact network is trained using a partial differential equation loss function to obtain a pre-trained model, wherein the partial differential equation loss function is related to a loss term of an elastic-plastic constitutive equation of the training object during a pressing process, and the elastic-plastic constitutive equation is used to characterize an elastic-plastic constitutive relationship; Based on the training data set, the pre-training model is trained using a first loss function to obtain the stress prediction model, wherein the first loss function is related to the loss term of the boundary conditions and the loss term of the initial conditions of the training object during the pressing process.
3. The method according to claim 2, characterized in that The partial differential equation loss function is the sum of the loss function of the elastic-plastic constitutive equation of the training object during the pressing process, the loss function of the stress balance equation of the training object during the pressing process, and the loss function of the geometric equation that characterizes the geometric relationship between stress and displacement of the training object during the pressing process; wherein, the loss function of the elastic-plastic constitutive equation is the product of the loss term of the elastic-plastic constitutive equation and the first weight coefficient, the loss function of the stress balance equation is the product of the loss term of the stress balance equation and the second weight coefficient, and the loss function of the geometric equation is the product of the loss term of the geometric equation and the third weight coefficient.
4. The method according to claim 2, characterized in that: The method of training the pre-trained model based on the training data set using a first loss function to obtain the stress prediction model includes: Based on the training data set, in the i-th training process, for each loss term constituting the first loss function before weight adjustment, a new weight coefficient corresponding to the loss term is determined according to the average gradient of the loss term and the size of the loss term itself, i=1, 2, ...; Determine a weight-adjusted first loss function according to each loss term constituting the first loss function and a new weight coefficient corresponding to each loss term; Using the weight-adjusted first loss function, adjusting the network parameters of the pre-trained model for the (i+1)th training; After assigning i+1 to i, return to execute the above steps until the training meets the conditions, and obtain the stress prediction model.
5. The method according to claim 4, characterized in that Determining a new weight coefficient corresponding to the loss term according to the average gradient of the loss term and the size of the loss term itself includes: Calculate a fourth weight coefficient according to the average gradient of the loss term and the size of the loss term itself; A new weight coefficient corresponding to the loss term is calculated based on the weight coefficient corresponding to the loss term before weight adjustment and the fourth weight coefficient.
6. The method according to claim 4, characterized in that The adjusting the network parameters of the pre-trained model by using the first loss function after the weight adjustment includes: Fixing the weights and parameters of the preset layers of the pre-trained model; The weights and parameters of the remaining layers of the pre-trained model except the preset layer are adjusted using a first learning rate and the first loss function after the weight adjustment, wherein the first learning rate is less than a preset threshold.
7. A stress prediction device, characterized in that: include: A first acquisition module is used to acquire the spatial coordinates of the target object to be pressed; A stress prediction module, used for inputting the spatial coordinates of the target object and the loading time corresponding to the pressing process of the target object into a stress prediction model to perform stress prediction and obtain a prediction result, wherein the prediction result includes a stress prediction result and a strain prediction result; The stress prediction model is obtained by training a physical information neural network based on the elastic-plastic constitutive relationship of each training object during the pressing process.
8. A stress prediction device, comprising a processor and a transceiver, wherein the transceiver receives and sends data under the control of the processor, characterized in that: The processor is configured to perform the following operations: Obtaining the spatial coordinates of the target object to be pressed; Inputting the spatial coordinates of the target object and the loading time corresponding to the pressing process of the target object into a stress prediction model to perform stress prediction, and obtaining a prediction result, wherein the prediction result includes a stress prediction result and a strain prediction result; The stress prediction model is obtained by training a physical information neural network based on the elastic-plastic constitutive relationship of each training object during the pressing process.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the stress prediction method according to any one of claims 1 to 6 are implemented.
10. A computer program product, characterized in that The method comprises computer instructions, which, when executed by a processor, implement the steps in the stress prediction method according to any one of claims 1 to 6.