Assembly joint surface contact deformation prediction method based on enhanced depth operator network
By building an enhanced depth operator network, the image generator is used to predict contact deformation of the assembly bonding surface of mechanical product parts, the problem of high computing costs in the prior art is solved, and efficient contact deformation prediction and optimized design are achieved.
Patent Information
- Application Number
- CN202510378270.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-28
- Publication Date
- 2025-08-01
AI Technical Summary
The prior art has high calculation cost when calculating contact deformation of the assembly bonding surface of mechanical product parts, resulting in low optimization design efficiency and fewer proxy models based on isometric analysis methods.
A method for predicting the assembly bonding surface contact deformation based on an enhanced depth operator network is constructed. By establishing an equal geometric analysis model, the rough morphology and loading capacity of the bonding surface are changed, and the depth operator network of the image generator is used as the backbone network for training to predict the bonding surface contact deformation.
It significantly reduces the calculation cost, improves the prediction accuracy and efficiency of the contact deformation of the assembly bonding surface, and enhances the universality and learning ability of the deep operator network.
Smart Images
Figure CN120409190A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for predicting the contact deformation of an assembly joint surface in the field of mechanical products, and more particularly to a method for predicting the contact deformation of an assembly joint surface based on an enhanced deep operator network. Background Art
[0002] In order to ensure the performance of mechanical products such as heat conduction, electrical conductivity, and sealing effects, it is necessary to optimize the surface rough morphology of the assembly joint surface of product parts during the R & D process. During the optimization design process, it is often necessary to calculate the contact deformation of the joint surface generated by many different-shaped rough morphologies under the action of an external loading force, so as to obtain the distribution and area of the contact region of the joint surface corresponding to each rough morphology, and then evaluate the product performance corresponding to different-shaped rough morphologies, and screen out the rough morphology with the best performance. Therefore, calculating the contact deformation of the joint surface affected by the rough morphology is the basis for ensuring product performance.
[0003] Isogeometric analysis method can be used to calculate the contact deformation of the joint surface, but during its calculation process, it is necessary to continuously assemble the stiffness matrix and solve the linear equations. When the number of elements is large, its calculation cost will increase significantly, resulting in a large amount of time consumed in the optimization design process, which hinders the improvement of the R & D design efficiency of products. Therefore, improving the calculation efficiency of the contact deformation of the joint surface has important application value. By constructing a surrogate model, the calculation results of the isogeometric analysis method can be predicted, thereby significantly reducing the calculation cost and improving the R & D design efficiency of products. With the continuous development of deep learning technology, the method for constructing a surrogate model based on neural networks has gradually been applied. However, most of the existing methods are based on the finite element method to construct a surrogate model, and there are relatively few surrogate models constructed based on the isogeometric analysis method and used to predict the contact deformation of the joint surface. Summary of the Invention
[0004] In order to solve the problems existing in the background art, the present invention proposes a method for predicting the contact deformation of an assembly joint surface based on an enhanced deep operator network. Aiming at the problem of the contact deformation of the assembly joint surface affected by the surface rough morphology of product parts, the present invention first establishes an isogeometric analysis model of the assembly joint surface, changes the rough morphology of the joint surface, and uses isogeometric analysis to solve for the contact deformation of the joint surface under different rough morphologies and loading forces; then represents the joint surface deformation in the form of an image, and constructs a training set based on the rough morphology of the joint surface, the loading force, and the joint surface deformation data; finally, constructs an enhanced deep operator network of an image generator with the image generator as the backbone network and the fully connected neural network as the branch network, and uses the training set to train the deep operator network.
[0005] The technical solution of the present invention is as follows:
[0006] 1. A method for predicting the contact deformation of an assembly joint surface based on an enhanced depth operator network
[0007] Step 1: Establish an isogeometric analysis model of the assembly joint surface, and construct a joint surface deformation dataset by changing the rough morphology of the joint surface of the isogeometric analysis model and applying different loading forces;
[0008] Step 2: Construct an enhanced depth operator network, and use the joint surface deformation dataset to train the enhanced depth operator network to obtain a prediction model for the contact deformation of the assembly joint surface;
[0009] Step 3: Input the rough morphology and loading force of the assembly joint surface to be measured into the prediction model for the contact deformation of the assembly joint surface, and the model outputs the contact deformation result of the assembly joint surface.
[0010] In the above Step 1, the joint surface deformation dataset is constructed by changing the rough morphology of the joint surface of the isogeometric analysis model and applying different loading forces, specifically as follows:
[0011] Use isogeometric analysis to solve for the contact deformation data of the joint surface under each rough morphology of the joint surface and the corresponding loading force; taking the index of the control point parameter coordinates of the joint surface of the isogeometric analysis model as the pixel point index, and the deformation amount of the physical coordinates of the control point as the pixel value of the pixel point, convert each contact deformation data of the joint surface into a deformation true value map, take each rough morphology of the joint surface and the corresponding loading force as the input and the corresponding deformation true value map as the true value, so as to form a joint surface deformation sample; change the rough morphology of the joint surface of the isogeometric analysis model and apply different loading forces, obtain the corresponding contact deformation data of the joint surface and convert it into a deformation true value map, so as to obtain different joint surface deformation samples, and finally obtain the joint surface deformation dataset.
[0012] The enhanced depth operator network includes a backbone network and a branch network. The combined surface roughness profile is used as the input of the backbone network, and the loading force is used as the input of the branch network. The backbone network includes a number of fully connected layers and a number of deconvolution layers, and the deconvolution layers are connected in sequence. The branch network includes a number of fully connected layers. The fully connected layers in the branch network except the first and second fully connected layers are connected in sequence, and the output of the last fully connected layer is used as the output of the branch network. The first and second fully connected layers of the backbone network are connected, and the first and second fully connected layers of the branch network are connected. The output of the second fully connected layer of the backbone network and the output of the second fully connected layer of the branch network are concatenated and then used as the input of the third fully connected layer of the backbone network and the input of the third fully connected layer of the branch network. The third fully connected layer of the backbone network is connected to the first deconvolution layer of the backbone network, and the output of the last deconvolution layer of the backbone network is used as the output of the backbone network. Finally, using the vector elements in the output of the branch network as weights, the channels of the basic image output by the backbone network are weighted and summed to obtain an image representing the physical coordinate deformation of the combined surface control points, which is used as the output of the enhanced depth operator network.
[0013] During the training process of the enhanced depth operator network, the root mean square error is used as the loss function.
[0014] II. An assembly joint surface contact deformation prediction system based on an enhanced depth operator network
[0015] A combined surface deformation dataset generation unit is used to establish an isogeometric analysis model based on the assembly joint surface, and generate a combined surface deformation dataset by changing the combined surface roughness profile of the isogeometric analysis model and changing different loading forces.
[0016] A network storage unit is used to store the enhanced depth operator network.
[0017] A model training unit is used to train the enhanced depth operator network using the combined surface deformation dataset until the training is completed.
[0018] An assembly joint surface contact deformation prediction unit is used to input the roughness profile and loading force of the to-be-tested assembly joint surface into the trained enhanced depth operator network to obtain the assembly joint surface contact deformation result.
[0019] III. A computer device
[0020] The device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the method for predicting the assembly joint surface contact deformation based on the enhanced depth operator network are implemented.
[0021] IV. A computer-readable storage medium
[0022] A computer program is stored on the medium. When the computer program is executed by a processor, the steps of the method for predicting the contact deformation of an assembly joint surface based on an enhanced depth operator network are implemented.
[0023] V. A computer program product
[0024] The product includes computer programs / instructions. When the computer programs / instructions are executed by a processor, the steps of the method for predicting the contact deformation of an assembly joint surface based on an enhanced depth operator network are implemented.
[0025] The beneficial effects of the present invention are as follows:
[0026] The present invention uses the contact deformation of the assembly joint surface under different rough morphologies and loading forces as training data. The trained depth operator network can quickly calculate the contact deformation of the joint surface under rough morphologies and loading forces that have not been trained.
[0027] The present invention represents the joint surface as NURBS and uses the deformation amount of the NURBS control points as the output of the neural network, so that various types of results such as the contact force and contact area of the joint surface can be calculated through post-processing, enhancing the versatility of the prediction results of the depth operator network.
[0028] The present invention represents the contact deformation of the joint surface calculated by isogeometric analysis in the form of an image and uses an image generator as the backbone network of the depth operator network, enhancing the ability of the depth operator network to learn the correlation relationship between the contact deformations at different positions of the joint surface and improving the prediction accuracy of the contact deformation. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] Figure 1 It is a flow block diagram of the method of the present invention.
[0030] Figure 2 It is a schematic diagram of the contact problem in the embodiment of the present invention.
[0031] Figure 3 It is a schematic diagram of representing the deformation amount of the control point coordinates in the form of an image in the embodiment of the present invention.
[0032] Figure 4 It is a schematic diagram of the depth operator network enhanced by the image generator in Embodiment 1 of the present invention.
[0033] Figure 5 It is a schematic diagram of the prediction effect of the depth operator network enhanced by the image generator on the contact deformation in Embodiment 1 of the present invention.
[0034] Figure 6 It is a schematic diagram of the prediction effect of the depth operator network without being enhanced by the image generator on the contact deformation in Embodiment 1 of the present invention.
[0035] Figure 7 Schematic diagram of the enhanced depth operator network of the image generator in Embodiment 2 of the present invention.
[0036] Figure 8 Schematic diagram of the prediction effect of the enhanced depth operator network of the image generator in Embodiment 2 of the present invention on contact deformation.
[0037] Figure 9 Schematic diagram of the prediction effect of the contact deformation without using the enhanced depth operator network of the image generator in Embodiment 2 of the present invention. Detailed implementation manners
[0038] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.
[0039] The method for predicting the contact deformation of the assembly joint surface based on the enhanced depth operator network proposed by the present invention includes: establishing an isogeometric analysis model for the contact problem of the assembly joint surface, changing the rough morphology of the joint surface, and using isogeometric analysis to solve for the contact deformation of the joint surface under different rough morphologies and loading forces; then representing the joint surface deformation in the form of an image, and constructing a training set and a test set based on the rough morphology of the joint surface, the loading force, and the joint surface deformation data; finally, constructing an enhanced depth operator network of the image generator with the image generator as the backbone network and the fully connected neural network as the branch network, training the depth operator network using the training set, predicting the contact deformation of the joint surface corresponding to the rough morphology not trained in the test set based on the trained depth operator network, and evaluating the prediction accuracy at the same time.
[0040] Embodiment 1
[0041] As Figure 1 shown, Figure 1 the block diagram of is the flow block diagram of the method for predicting the contact deformation of the assembly joint surface based on the enhanced depth operator network proposed by the present invention. In this embodiment, the contact problem between a hyperelastic cube with an ideally smooth surface and a rigid cuboid with a rough morphology as shown in Figure 2 is used to demonstrate the implementation process of the method for predicting the contact deformation of the assembly joint surface based on the enhanced depth operator network proposed by the present invention, including the following steps:
[0042] Step 1: Establish an isogeometric analysis model for the assembly joint surface, and construct a joint surface deformation data set by changing the rough morphology of the joint surface of the isogeometric analysis model and changing different loading forces;
[0043] The shape parameter of the hyperelastic cube takes the value of a = 1 mm. First, based on the NURBS parameters shown in Table 1, the initial isogeometric model of the hyperelastic cube is constructed. Then, based on the knot vectors shown in Table 2, h-refinement and p-refinement are performed on the initial isogeometric model of the hyperelastic cube. Finally, the obtained isogeometric model has n = 30, m = 30, and l = 6. The shape parameters of the rigid cuboid take the values of L = 2 mm and H = 0.3 mm. Similarly, first, based on the NURBS parameters shown in Table 3, the initial isogeometric model of the rigid cuboid with an ideally smooth surface is constructed, and based on the knot vectors shown in Table 4, h-refinement and p-refinement are performed on the initial isogeometric model of the rigid cuboid. Finally, the obtained isogeometric model has n = 100, m = 100, and l = 6. On the upper surface of the hyperelastic cube, a loading force P in the negative z-axis direction is applied, and the displacements of the two surfaces of the hyperelastic cube parallel to the yz plane in the x and y directions are constrained to be zero. The lower surface of the rigid cuboid is constrained to be fixed, and the lower surface of the hyperelastic cube and the upper surface of the rigid cuboid are defined as a contact pair in contact with each other. The Neo-Hookean model is used to describe the material properties of the hyperelastic cube, with a Young's modulus of 1 MPa and a Poisson's ratio of 0.3.
[0044] Table 1 shows the NURBS order, knot vector, physical coordinates of control points, and weights of the initial isogeometric model of the hyperelastic cube used in this implementation in each parameter direction.
[0045]
[0046] Table 2 shows the NURBS order and knot vector of the final isogeometric model of the hyperelastic cube used in this implementation in each parameter direction.
[0047]
[0048] Table 3 shows the NURBS order, knot vector, physical coordinates of control points, and weights of the initial isogeometric model of the rigid cuboid used in this implementation in each parameter direction.
[0049]
[0050]
[0051] Table 4 shows the NURBS order and knot vector of the final isogeometric model of the rigid cuboid used in this implementation in each parameter direction.
[0052]
[0053] By changing the rough morphology of the joint surface of the isogeometric analysis model and applying different loading forces, a joint surface deformation dataset is constructed, specifically:
[0054] The isogeometric analysis model is represented by NURBS, and its expression is as follows:
[0055]
[0056] Among them, is the isogeometric representation of the geometric body, ξ is the parametric coordinate along the x direction, η is the parametric coordinate along the y direction, is the parametric coordinate along the z direction, n is the number of control points along the ξ direction, m is the number of control points along the η direction, l is along the direction of the control points, is the NURBS basis function, p is the order of the basis function along the ξ direction, q is the order of the basis function along the η direction, r is the order of the basis function along the direction, B i,j,k is the coordinate of the control point corresponding to the NURBS basis function in the physical space.
[0057] The bonded surface is the boundary surface of the isogeometric analysis model, and its expression is:
[0058]
[0059] Among them, S(ξ,η) is the isogeometric representation of the boundary surface of the geometric body; is the NURBS basis function of the isogeometric representation of the boundary surface; B i,j is the coordinate of the control point of the NURBS basis function of the isogeometric representation of the boundary surface in the physical space.
[0060] Set a discrete point sequence evenly distributed along the x and y directions on the upper surface of the rigid cuboid Keep the x and y coordinate values of these discrete points unchanged, and adjust their z coordinates to satisfy the Weibull distribution shown in the following formula under different given parameters:
[0061]
[0062] Among them, the shape parameter b of the Weibull distribution takes 100 values uniformly distributed within the range of [1, 3], and the scale parameter corresponding to each shape parameter b in λ, Γ(·) is the gamma function, and h is the value of the z coordinate of the discrete point.
[0063] Using the discrete point sequence that satisfies the Weibull distribution in the rough morphology of the bonded surface as the shape value points, the coordinate B of the control points on the upper surface of the rigid cuboid is calculated by NURBS interpolation i,j The value is as shown in the following formula:
[0064]
[0065] Among them, Ds,t are the physical coordinates of the shape value points, are the parametric coordinates of the shape value points in the ξ and η directions respectively, and The values of are shown in Table 5.
[0066] Table 5 shows the parametric coordinates of the shape value points of the NURBS interpolation of the rough surface morphology on the upper surface of the rigid cuboid used in this implementation in each parametric direction
[0067]
[0068] Using isogeometric analysis to solve, the contact deformation data of each joint surface rough morphology and the corresponding loading force are obtained; taking the index of the parametric coordinates of the joint surface control points of the isogeometric analysis model as the pixel point index, and the deformation amount of the control point physical coordinates as the pixel value of the pixel point, each joint surface contact deformation data is converted into a deformation true value map, and each joint surface rough morphology and the corresponding loading force are used as inputs and the corresponding deformation true value map is used as the true value, thus forming a joint surface deformation sample; the parametric coordinates of the control points of the joint surface include two different directions, that is, they are respectively used as the horizontal direction and the vertical direction of the image pixel points, and the numbers of the parametric coordinates of the control points in different directions are respectively used as the horizontal and vertical indexes of the image pixel points, as Figure 3 shown. By changing the rough morphology of the joint surface of the isogeometric analysis model (that is, changing the value of to obtain different rough morphologies of the joint surface) and applying different loading forces, the corresponding contact deformation data of the joint surface are obtained and converted into a deformation true value map, thus obtaining different joint surface deformation samples, and finally obtaining a joint surface deformation data set.
[0069] Step 2: Construct an enhanced depth operator network, and use the joint surface deformation data set to train the enhanced depth operator network to obtain an assembly joint surface contact deformation prediction model;
[0070] As Figure 4 shown, the enhanced depth operator network includes a backbone network and a branch network, and the rough morphology of the joint surface As the input of the backbone network, the loading force is used as the input of the branch network. The backbone network consists of 3 fully connected layers and 5 deconvolution layers. The activation function of each layer is leaky Relu. The number of neurons in the 3 fully connected layers is 10000, 512, and 1024 respectively. The output of the third layer is shape-transformed to obtain an image with a size of 2×2 and 256 channels. Several deconvolution layers are connected in sequence; the branch network consists of 7 fully connected layers. The fully connected layers in the branch network except the first and second fully connected layers are connected in sequence. The output of the last fully connected layer is used as the output of the branch network, which is a one-dimensional vector; the first and second fully connected layers of the backbone network are connected, the first and second fully connected layers of the branch network are connected. The output of the second fully connected layer of the backbone network and the output of the second fully connected layer of the branch network are concatenated and then used as the input of the third fully connected layer of the backbone network and the input of the third fully connected layer of the branch network. The third fully connected layer of the backbone network is connected to the first deconvolution layer of the backbone network. The output of the last deconvolution layer of the backbone network is used as the output of the backbone network; since the number of channels of the basic image output by the backbone network is equal to the number of elements of the vector output by the branch network, finally, the elements of the vector in the output of the branch network are used as weights to perform weighted summation on each channel of the basic image output by the backbone network, so as to obtain an image representing the physical coordinate deformation of the joint surface control points and use it as the output of the enhanced depth operator network, that is, an image representing the physical coordinate z-direction deformation of the control points on the lower surface of the hyperelastic cube where θ is the parameter of the trained depth operator network, and i and j are the indices of the image pixels..
[0071] The convolution kernel size of the first four deconvolution layers is 3×3, the padding value is 2, the dilation rate is 1, and the stride is 1. The convolution kernel size of the fifth deconvolution layer is 3×3, the padding value is 2, the dilation rate is 3, and the stride is 1.
[0072] During the training process of the enhanced depth operator network, the root mean square error is used as the loss function loss, and the formula is as follows:
[0073]
[0074] where n and m are the numbers of control points on the lower surface of the hyperelastic cube along the ξ and η directions respectively.
[0075] Predict the rough surface contact deformation in the test set that has not been trained, and at the same time use the L2 average relative error to evaluate the prediction accuracy. The formula for the L2 average relative error (MRL2E) is as follows:
[0076]
[0077] where, is the deformation of the physical coordinates of the control points of the joint surface with index numbers i and j calculated based on isogeometric analysis. is the input of the loading force P k and the discrete point sequence of the rough surface topography is the deformation of the physical coordinates of the control points of the joint surface with index numbers i and j predicted by the deep operator network with parameter θ based on the training, where N train is the number of training data in each training batch, and N test is the number of test data in the test set.
[0078] Step 3: Input the rough surface topography and the loading force of the assembly joint surface to be measured into the assembly joint surface contact deformation prediction model. The model outputs the assembly joint surface contact deformation result. According to the assembly joint surface contact deformation result, screen the rough surface topography of the assembly joint surface, and select the rough surface topography with the best performance.
[0079] The average value of the L2 average relative error of the method of the present invention in the test set is 0.0238. Randomly select a set of data in the test set. The isogeometric analysis solution, the deep operator network solution, and the point-by-point error between the two are as Figure 5 shown; the prediction result of the deep operator network that does not use the image generator as the backbone network but uses the fully connected neural network as the backbone network is as Figure 6 shown, and the average value of its L2 average relative error in the test set is 0.2534; therefore, the prediction accuracy of the method of the present invention for the assembly joint surface contact deformation is significantly better than the existing methods.
[0080] Embodiment 2
[0081] This embodiment uses the contact problem between an elastically deformable cube with an ideally smooth surface and an elastically deformable cuboid with a rough surface topography as shown in Figure 2 to demonstrate the implementation process of the assembly joint surface contact deformation prediction method based on the enhanced deep operator network proposed by the present invention, including the following steps:
[0082] Step 1: Establish an isogeometric analysis model of the assembly joint surface, and construct a joint surface deformation dataset by changing the rough surface topography of the isogeometric analysis model and changing different loading forces.
[0083] The shape parameter of the linear elastic cube takes the value of a = 1 mm. First, based on the NURBS parameters shown in Table 6, the initial isogeometric model of the linear elastic cube is constructed. Then, based on the knot vectors shown in Table 7, h-refinement and p-refinement are performed on the initial isogeometric model of the linear elastic cube. Finally, the obtained isogeometric model has n = 30, m = 30, and l = 6. The shape parameters of the linear elastic cuboid take the values of L = 2 mm and H = 0.3 mm. Similarly, first, based on the NURBS parameters shown in Table 8, the initial isogeometric model of the linear elastic cuboid with an ideally smooth surface is constructed, and based on the knot vectors shown in Table 9, h-refinement and p-refinement are performed on the initial isogeometric model of the linear elastic cuboid. Finally, the obtained isogeometric model has n = 100, m = 100, and l = 6. On the upper surface of the linear elastic cube, a loading force P in the negative z-axis direction is applied, and the displacements of the two surfaces of the linear elastic cube parallel to the yz plane in the x and y directions are constrained to be zero. The lower surface of the linear elastic cuboid is constrained to be fixed. The lower surface of the linear elastic cube and the upper surface of the linear elastic cuboid are defined as a contact pair in contact with each other. The Young's modulus of both the linear elastic cube and the cuboid is defined as 1 MPa, and the Poisson's ratio is 0.3. Table 6 shows the NURBS orders, knot vectors, physical coordinates of control points, and weights of the initial isogeometric model of the linear elastic cube used in this implementation.
[0084]
[0085] Table 7 shows the NURBS orders and knot vectors of the final isogeometric model of the linear elastic cube used in this implementation in each parameter direction.
[0086]
[0087]
[0088] Table 8 shows the NURBS orders, knot vectors, physical coordinates of control points, and weights of the initial isogeometric model of the linear elastic cuboid used in this implementation in each parameter direction.
[0089]
[0090] Table 9 shows the NURBS orders and knot vectors of the final isogeometric model of the linear elastic cuboid used in this implementation in each parameter direction.
[0091]
[0092] By changing the rough morphology of the joint surface of the isogeometric analysis model and applying different loading forces, a joint surface deformation data set is constructed. Specifically:
[0093] The isogeometric analysis model is represented by NURBS, and its expression is as follows:
[0094]
[0095] wherein, is the isogeometric representation of the geometric body, ξ is the parametric coordinate in the x direction, η is the parametric coordinate in the y direction, is the parametric coordinate in the z direction, n is the number of control points in the ξ direction, m is the number of control points in the η direction, l is the number of control points in the direction, is the NURBS basis function, p is the order of the basis function in the ξ direction, q is the order of the basis function in the η direction, r is the order of the basis function in the direction, B i,j,k is the coordinate of the control point corresponding to the NURBS basis function in the physical space.
[0096] The bonding surface is the boundary surface of the isogeometric analysis model, and its expression is:
[0097]
[0098] wherein, S(ξ,η) is the isogeometric representation of the boundary surface of the geometric body; is the NURBS basis function of the isogeometric representation of the boundary surface; B i,j is the coordinate of the control point of the NURBS basis function of the isogeometric representation of the boundary surface in the physical space.
[0099] Set a discrete point sequence uniformly distributed in the x and y directions on the upper surface of the linearly elastic cuboid Keep the x and y coordinate values of these discrete points unchanged, and adjust their z coordinates to satisfy the Weibull distribution shown in the following formula under different given parameters:
[0100]
[0101] wherein, the shape parameter b of the Weibull distribution takes 100 values uniformly distributed within the range of [1, 3], and the scale parameter λ in which Γ(·) is the gamma function, and h is the value of the z coordinate of the discrete point.
[0102] Using the discrete point sequence that satisfies the Weibull distribution in the rough topography of the bonding surface as the control points, the coordinate B of the control points on the upper surface of the linearly elastic cuboid is calculated through NURBS interpolation i,j as shown in the following formula:
[0103]
[0104] wherein, D s,t is the physical coordinate of the control point, are the parametric coordinates of the shape value points in the ξ and η directions respectively, and The values of are shown in Table 10.
[0105] Table 10 shows the parametric coordinates of the shape value points of the NURBS interpolation of the rough surface morphology on the upper surface of the linearly elastic cuboid used in this implementation in each parametric direction
[0106]
[0107] Using isogeometric analysis to solve, the contact deformation data of the joint surface under each joint surface rough morphology and the corresponding loading force are obtained; taking the index of the parametric coordinates of the control points on the joint surface of the isogeometric analysis model as the pixel point index, and taking the deformation amount of the physical coordinates of the control points as the pixel value of the pixel point, each joint surface contact deformation data is converted into a deformation true value map, taking each joint surface rough morphology and the corresponding loading force as the input and the corresponding deformation true value map as the true value, thus forming a joint surface deformation sample; the parametric coordinates of the control points on the joint surface include two different directions, that is, taking them as the horizontal direction and the vertical direction of the image pixel points respectively, and taking the numbers of the parametric coordinates of the control points in different directions as the horizontal and vertical indexes of the image pixel points respectively, as Figure 3 shown. By changing the rough morphology of the joint surface of the isogeometric analysis model (that is, changing the values of to obtain different rough morphologies of the joint surface) and applying different loading forces, the corresponding contact deformation data of the joint surface are obtained and converted into a deformation true value map, thus obtaining different joint surface deformation samples, and finally obtaining a joint surface deformation data set.
[0108] Step 2: Construct an enhanced depth operator network, and use the joint surface deformation data set to train the enhanced depth operator network to obtain an assembly joint surface contact deformation prediction model;
[0109] As Figure 7 shown, the enhanced depth operator network includes a backbone network and a branch network, and the rough morphology of the joint surface As the input of the backbone network, the loading force is used as the input of the branch network. The backbone network includes 3 fully connected layers and 5 deconvolution layers. The activation function of each layer is leaky Relu. The number of neurons in the 3 fully connected layers is 900, 512, and 1024 respectively. The output of the third layer is shape-transformed to obtain an image with a size of 2×2 and 256 channels. Several deconvolution layers are connected in sequence; the branch network includes 7 fully connected layers. The fully connected layers in the branch network except the first and second fully connected layers are connected in sequence. The output of the last fully connected layer is used as the output of the branch network, which is a one-dimensional vector; the first and second fully connected layers of the backbone network are connected, the first and second fully connected layers of the branch network are connected. The output of the second fully connected layer of the backbone network and the output of the second fully connected layer of the branch network are concatenated and then used as the input of the third fully connected layer of the backbone network and the input of the third fully connected layer of the branch network. The third fully connected layer of the backbone network is connected to the first deconvolution layer of the backbone network. The output of the last deconvolution layer of the backbone network is used as the output of the backbone network; since the number of channels of the basic image output by the backbone network is equal to the number of elements of the vector output by the branch network, finally, the channels of the basic image output by the backbone network are weighted and summed with the vector elements in the output of the branch network as weights, so as to obtain an image representing the physical coordinate deformation of the joint surface control points and use it as the output of the enhanced depth operator network, that is, an image representing the physical coordinate z-direction deformation of the control points on the lower surface of the linear elastic cube where θ is the parameter of the trained depth operator network, and i and j are the indices of the image pixels..
[0110] The convolution kernel size of the first four deconvolution layers is 3×3, the padding value is 2, the dilation rate is 1, and the stride is 1. The convolution kernel size of the fifth deconvolution layer is 3×3, the padding value is 2, the dilation rate is 3, and the stride is 1.
[0111] During the training process of the enhanced depth operator network, the root mean square error is used as the loss function loss, and the formula is as follows:
[0112]
[0113] where n and m are the numbers of control points on the lower surface of the linear elastic cube along the ξ and η directions respectively.
[0114] Predict the rough surface contact deformation in the test set that has not been trained, and at the same time use the L2 average relative error to evaluate the prediction accuracy. The formula of the L2 average relative error (MRL2E) is as follows:
[0115]
[0116] where, is the deformation of the physical coordinates of the joint surface control points with index numbers i and j calculated based on isogeometric analysis. is the applied load P k and the discrete point sequence of the rough surface topography as the input, and based on the deep operator network with parameter θ obtained through training, the predicted deformation of the physical coordinates of the joint surface control points with index numbers i and j, N train is the number of training data in each training batch, N test is the number of test data in the test set.
[0117] Step 3: Input the rough surface topography and the applied load of the assembly joint surface to be measured into the assembly joint surface contact deformation prediction model. The model outputs the assembly joint surface contact deformation result, and based on the assembly joint surface contact deformation result, screen the rough surface topography of the assembly joint surface to select the rough surface topography with the best performance.
[0118] The mean value of the L2 average relative error of the method of the present invention in the test set is 0.0015. Randomly select a set of data in the test set. The isogeometric analysis solution, the deep operator network solution, and the point-by-point error between the two are as Figure 8 shown; the prediction result of the deep operator network that does not use the image generator as the backbone network but uses the fully connected neural network as the backbone network is as Figure 9 shown, and its mean value of the L2 average relative error in the test set is 0.0362; therefore, the prediction accuracy of the method of the present invention for the assembly joint surface contact deformation is significantly better than the existing methods.
[0119] The above embodiments are only specific embodiments of the present invention, which are used to illustrate the technical solutions of the present invention and are not intended to limit it. The protection scope of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: Any person skilled in the art within the technical scope disclosed by the present invention can still modify the technical solutions recorded in the foregoing embodiments or can easily think of changes, or perform equivalent replacements on some of the technical features; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the present invention and should all be covered by the protection scope of the present invention.
Claims
1. A method for predicting the contact deformation of an assembly joint surface based on an enhanced depth operator network, characterized in that It includes the following steps: Step 1: Establish an isogeometric analysis model of the assembly joint surface, and construct a joint surface deformation dataset by changing the rough morphology of the joint surface of the isogeometric analysis model and changing different loading forces; Step 2: Construct an enhanced depth operator network, and use the joint surface deformation dataset to train the enhanced depth operator network to obtain an assembly joint surface contact deformation prediction model; Step 3: Input the rough morphology and loading force of the assembly joint surface to be measured into the assembly joint surface contact deformation prediction model, and the model outputs the assembly joint surface contact deformation result.
2. The method for predicting the contact deformation of an assembly joint surface based on an enhanced depth operator network according to claim 1, wherein In the above Step 1, by changing the rough morphology of the joint surface of the isogeometric analysis model and applying different loading forces to construct the joint surface deformation dataset, specifically: Use isogeometric analysis to solve and obtain the joint surface contact deformation data under each joint surface rough morphology and corresponding loading force; Taking the index of the joint surface control point parameter coordinates of the isogeometric analysis model as the pixel point index, and the deformation amount of the control point physical coordinates as the pixel value of the pixel point, convert each joint surface contact deformation data into a deformation true value map, take each joint surface rough morphology and the corresponding loading force as the input and the corresponding deformation true value map as the true value, so as to form a joint surface deformation sample; Change the rough morphology of the joint surface of the isogeometric analysis model and apply different loading forces, obtain the corresponding joint surface contact deformation data and convert it into a deformation true value map, so as to obtain different joint surface deformation samples, and finally obtain the joint surface deformation dataset.
3. A method for predicting the contact deformation of an assembly joint surface based on an enhanced depth operator network according to claim 1, characterized in that, The enhanced depth operator network includes a backbone network and a branch network. The rough morphology of the joint surface is used as the input of the backbone network, and the loading force is used as the input of the branch network. The backbone network includes several fully connected layers and several deconvolution layers, and the several deconvolution layers are connected in sequence; the branch network includes several fully connected layers, and the fully connected layers in the branch network except the first fully connected layer and the second fully connected layer are connected in sequence, and the output of the last fully connected layer is used as the output of the branch network; the first fully connected layer and the second fully connected layer of the backbone network are connected, the first fully connected layer and the second fully connected layer of the branch network are connected, the output of the second fully connected layer of the backbone network and the output of the second fully connected layer of the branch network are spliced and then used as the input of the third fully connected layer of the backbone network and the input of the third fully connected layer of the branch network, and the third fully connected layer of the backbone network is connected to the first deconvolution layer of the backbone network, and the output of the last deconvolution layer of the backbone network is used as the output of the backbone network; finally, using the vector elements in the output of the branch network as weights, perform weighted summation on each channel of the basic image output by the backbone network, so as to obtain an image representing the deformation amount of the joint surface control point physical coordinates and use it as the output of the enhanced depth operator network.
4. A method for predicting the contact deformation of an assembly joint surface based on an enhanced depth operator network according to claim 3, characterized in that During the training process of the enhanced depth operator network, the root mean square error is used as the loss function.
5. An assembly joint surface contact deformation prediction system based on an enhanced depth operator network, characterized in that, It includes: A joint surface deformation dataset generation unit, which is used to establish an isogeometric analysis model based on the assembly joint surface, and generate a joint surface deformation dataset by changing the rough morphology of the joint surface of the isogeometric analysis model and changing different loading forces; A network storage unit, which is used to store the enhanced depth operator network; A model training unit, configured to train an enhanced depth operator network using a dataset of joint surface deformations until the training is completed; An assembled joint surface contact deformation prediction unit, configured to input the rough morphology and loading force of a to-be-tested assembled joint surface into the trained enhanced depth operator network to obtain an assembled joint surface contact deformation result.
6. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method for predicting the contact deformation of an assembled joint surface based on an enhanced depth operator network according to any one of claims 1 to 4.
7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for predicting the contact deformation of an assembled joint surface based on an enhanced depth operator network according to any one of claims 1 to 4.
8. A computer program product comprising computer programs / instructions, characterized in that, When the computer program / instructions are executed by the processor, it implements the steps of the method for predicting the contact deformation of an assembled joint surface based on an enhanced depth operator network according to any one of claims 1 to 4.