Machine learning based laser additive manufacturing stress distortion prediction method and system

Abstract

The application belongs to the technical field of metal additive manufacturing, and discloses a laser additive manufacturing stress deformation prediction method and system based on machine learning, which comprises the following steps: constructing a machine learning agent model, establishing a nonlinear mapping relationship between simplified temperature field parameters, structural characteristics and inherent strain; constructing a residual stress prediction network, structurally coding the inherent strain tensor through a dynamic self-adjusting tensor coding layer, establishing a mapping relationship from inherent strain to residual stress based on the coded inherent strain tensor; adaptively partitioning a large component in geometry topology, predicting the inherent strain of each sub-region, and continuously fusing the strain fields of each sub-region through a strain field gradient fusion algorithm; predicting the residual stress of the overall component by using the fused strain field, and verifying and correcting the prediction result through experimental data. The application solves the problems of low calculation efficiency, insufficient prediction accuracy and lack of universality of the traditional numerical simulation method.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of metal additive manufacturing, in particular to a laser additive manufacturing stress deformation prediction method and system based on machine learning. BACKGROUND

[0002] The laser powder directed energy deposition (LP-DED) technology uses a high-energy laser beam to melt metal powder layer by layer and deposit a shaped metal part, has the advantages of high technical flexibility, high material utilization rate, short manufacturing cycle, etc., and has a significant advantage in forming large complex structures compared to other additive manufacturing and traditional processing technologies. Because of its unique advantages in manufacturing large complex geometric components, it has been widely used in the fields of aerospace, etc.

[0003] However, the LP-DED layer-by-layer deposition process of large components experiences periodic, severe, non-steady-state, cyclic heating and cooling and its short-time non-equilibrium cyclic solid-state phase change, which generates high horizontal, evolving and interacting thermal stress, phase change microstructure stress and constraint stress and their strong nonlinear and strongly coupled interactions in the component, which can easily lead to serious deformation problems. Relying solely on a large number of experiments and empirical exploration of forming process parameters has the problems of high cost and low efficiency, so there is an urgent need to develop an efficient numerical simulation method.

[0004] Current numerical calculation methods for metal additive manufacturing mainly include "thermal elastic-plastic mechanics model" and "inherent strain theory model". However, these two methods have obvious limitations. The thermal elastic-plastic model can accurately predict the distribution of residual stress and deformation, but the current research objects are small-sized structures, and the simulation calculation of medium and large-sized structures and components has not been realized. Small-sized structures still require a long calculation time of several days, and the simulation of residual stress and deformation of larger structures will face more severe challenges in terms of calculation amount and efficiency. The inherent strain method couples the material thermal-mechanical coupling behavior and the inherent strain evolution law, which can effectively capture the residual stress distribution induced by multi-stage thermal cycling and phase change in the additive manufacturing process while simplifying the calculation complexity, but it still has the problems of relatively low solution accuracy, the need for strain tensor calibration based on specific geometric features, and sensitivity to changes in component geometry, resulting in low strain extraction efficiency, which makes it difficult to be applied to efficient prediction and regulation of stress and deformation of large-sized complex structure additive components.

[0005] Therefore, the traditional numerical simulation method has the problem of being difficult to realize efficient prediction of large components while ensuring accuracy. SUMMARY

[0006] In order to solve the above problems, the present application provides a laser additive manufacturing stress deformation prediction method and system based on machine learning, to solve the problems of low calculation efficiency, insufficient prediction accuracy and lack of universality of traditional numerical simulation methods.

[0007] To achieve the above object, the present application adopts the following technical solutions:

[0008] In a first aspect, the present application provides a laser additive manufacturing stress deformation prediction method based on machine learning, comprising the following steps:

[0009] A machine learning agent model is constructed to establish a nonlinear mapping relationship between simplified temperature field parameters, structural features and inherent strain;

[0010] A residual stress prediction network based on inherent strain tensor coding is constructed, the inherent strain tensor is structured and coded through a dynamic self-adjusting tensor coding layer, and a mapping relationship from inherent strain to residual stress is established based on the coded inherent strain tensor;

[0011] The large component is adaptively partitioned in geometry and topology, the inherent strain of each sub-region is predicted, and the continuous fusion of the strain fields of each sub-region is performed through a strain field gradient fusion algorithm;

[0012] The residual stress of the overall component is predicted using the fused strain field, and the prediction results are verified and corrected through experimental data to form a closed-loop optimization system.

[0013] As an optional implementation, the simplified temperature field parameters include an equivalent peak temperature, a transverse heat affected zone radius and a longitudinal heat affected zone depth.

[0014] As an optional implementation, the machine learning agent model adopts a multilayer perceptron regression model, the input is temperature field parameters and structural type coding, and the output is the inherent strain value of each deposition layer.

[0015] As an optional implementation, the dynamic self-adjusting tensor coding layer realizes bilinear coding of the strain tensor through symmetric weight matrices and anti-symmetric weight matrices, and introduces a symmetry regularization term to maintain physical constraints.

[0016] As an optional implementation, the strain field gradient fusion algorithm adopts a bilinear interpolation based on distance weight at the junction of the sub-regions to ensure the first-order continuity of the strain field.

[0017] As an optional implementation, the geometry and topology adaptive partitioning adopts a graph cut algorithm to automatically partition based on the curvature and structural connectivity of the three-dimensional geometric model.

[0018] In a second aspect, the present application provides a laser additive manufacturing stress deformation prediction system based on machine learning, comprising:

[0019] The temperature field modeling module is configured to construct a machine learning agent model to establish a nonlinear mapping relationship between simplified temperature field parameters, structural features and inherent strain;

[0020] The machine learning module is configured to: construct a residual stress prediction network based on intrinsic strain tensor coding, structureally code the intrinsic strain tensor through a dynamic self-adjusting tensor coding layer, and establish a mapping relationship from intrinsic strain to residual stress based on the coded intrinsic strain tensor;

[0021] The partition and fusion module is configured to: adaptively partition a large component in geometry topology, predict the intrinsic strain of each sub-region, and continuously fuse the strain fields of each sub-region through a strain field gradient fusion algorithm.

[0022] The verification and correction module is configured to: predict the residual stress of the whole component using the fused strain field, verify and correct the prediction result through experimental data, and form a closed-loop optimization system.

[0023] In a third aspect, the present application provides an electronic device, comprising a memory and a processor, and computer instructions stored in the memory and running on the processor, when the computer instructions are run by the processor, the method of the first aspect is completed.

[0024] In a fourth aspect, the present application provides a computer readable storage medium for storing computer instructions, when the computer instructions are executed by the processor, the method of the first aspect is completed.

[0025] In a fifth aspect, the present application provides a computer program product comprising a computer program, when the computer program is executed by the processor, the method of the first aspect is completed.

[0026] Compared with the prior art, the present application has the following beneficial effects:

[0027] The laser additive manufacturing stress deformation prediction method based on machine learning of the present application replaces the traditional repetitive thermal-mechanical coupling simulation with a machine learning agent model, realizes instantaneous prediction, and improves the calculation efficiency by several orders of magnitude, which is suitable for engineering application of large and complex components.

[0028] The laser additive manufacturing stress deformation prediction method based on machine learning of the present application introduces a dynamic self-adjusting tensor coding layer, maintains the physical symmetry and directionality of the strain tensor, significantly improves the generalization ability and prediction stability of the model; the strain field gradient fusion algorithm eliminates the non-physical discontinuity at the junction of the sub-regions, ensures the continuity of the whole field prediction, and significantly improves the prediction accuracy and robustness.

[0029] The machine learning-based laser additive manufacturing stress deformation prediction method of the application has high automation degree in the whole process, reduces the dependence on simulation experience of operators, provides reliable technical support for process optimization and deformation control of metal additive manufacturing, and has remarkable engineering practical value.

[0030] The machine learning-based laser additive manufacturing stress deformation prediction method of the application has high automation degree in the whole process, reduces the dependence on simulation experience of operators, provides reliable technical support for process optimization and deformation control of metal additive manufacturing, and has remarkable engineering practical value.

[0031] In summary, the application realizes efficient and high-precision prediction of residual stress and deformation in the additive manufacturing process by constructing a closed-loop prediction framework.

[0032] The advantages of the additional aspects of the application will be partially given in the following description, partially become obvious from the following description, or be known by the practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0033] The drawings accompanying the specification of the application form a part of the application and serve to provide further understanding of the application, the illustrative embodiments of the application and their descriptions serve to explain the application, and do not constitute improper limitations on the application.

[0034] Figure 1 The flowchart of the machine learning-based laser additive manufacturing stress deformation prediction method of the application is shown in the figure.

[0035] Figure 2 The flowchart of the machine learning-based laser additive manufacturing stress deformation prediction method of the application is shown in the figure.

[0036] Figure 3 The grid diagram of the typical feature structure of the application is shown in the figure.

[0037] Figure 4 The framework diagram of the residual stress prediction network based on the inherent strain tensor coding of the application is shown in the figure.

[0038] Figure 5 The flowchart of the large component partition and fusion prediction of the application is shown in the figure. DETAILED DESCRIPTION

[0039] The application will be further described below in combination with the drawings and embodiments.

[0040] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0041] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments of the application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and "comprising", when used in this specification, specify the presence of stated features, integers, steps, or components, but do not preclude the presence or addition of one or more other features, integers, steps, components, or groups thereof.

[0042] The embodiments in the application and the features in the embodiments can be combined with each other without conflict.

[0043] Embodiment 1

[0044] As shown in Figures 1 to 5 The embodiment provides a laser additive manufacturing stress deformation prediction method based on machine learning, which comprises the following steps:

[0045] S1: Constructing a machine learning agent model, establishing a nonlinear mapping relationship between simplified temperature field parameters, structural characteristics and inherent strain;

[0046] S2: Constructing a residual stress prediction network based on inherent strain tensor coding, structurally coding the inherent strain tensor through a dynamic self-adjusting tensor coding layer, and establishing a mapping relationship from inherent strain to residual stress based on the coded inherent strain tensor;

[0047] S3: Geometric topology adaptive partitioning of large components, respectively predicting the inherent strain of each sub-region, and continuously fusing the strain fields of each sub-region through a strain field gradient fusion algorithm;

[0048] S4: Using the fused strain field to predict the residual stress of the overall component, and verifying and correcting the prediction results through experimental data to form a closed-loop optimization system.

[0049] The specific scheme of the application is as follows:

[0050] The materials, equipment and basic process parameters of the application are as follows.

[0051] Metal material: titanium alloy TC11 metal powder.

[0052] Additive manufacturing equipment: LP-DED equipment system developed independently or commercialized, including high-power laser, coaxial powder feeding nozzle, inert atmosphere protection cabin and motion control system.

[0053] Basic process parameter range: Through preliminary process test, a set of stable forming core process parameters are determined as the benchmark: laser power P=3000W, scanning speed v=700mm / min, powder feeding rate 15g / min, layer thickness 1mm, spot diameter 2mm.

[0054] S1: Construct machine learning agent model, establish nonlinear mapping relationship between simplified temperature field parameters, structural features and inherent strain. Flow chart as shown in Figure 2 .

[0055] S1.1: Establish simplified parameter temperature field model: propose a simplified model to characterize the heat effect of laser heat source on the molten pool and surrounding area by three key parameters: equivalent peak temperature T P , transverse heat affected zone radius R x and longitudinal heat affected zone depth R z . There is a certain quantitative conversion relationship between the model parameters and process parameters such as laser power and scanning speed to replace the complex traditional heat source model.

[0056] Equivalent peak temperature TP(℃): represents the highest temperature at the center of the molten pool, which is related to laser power and scanning speed, and the calculation formula is:

[0057] ;

[0058] Transverse heat affected zone radius Rx(mm): represents the width of the molten pool in the scanning plane, which is measured by metallographic dissection and regression:

[0059] ;

[0060] Longitudinal heat affected zone radius Rz(mm): represents the penetration depth of the molten pool, which is measured by metallographic dissection and regression:

[0061] .

[0062] S1.2: Obtain inherent strain training data set: for a variety of typical feature structures (such as single-layer wall, block, cantilever, ring, cross structure, etc.), use the simplified temperature field model for finite element simulation. Extract the inherent strain tensor generated after each deposition layer cools down by calculation to form the initial data set.

[0063] Typical feature structure selection: five kinds of geometric feature structures widely existing in additive manufacturing and representative are selected as the research object, and the grid model is as shown in Figure 3Dimensions are as follows:

[0064] A: Single-wall multi-layer structure (Dimensions: 100 mm long x 6 mm wide x 10 mm high);

[0065] B: Solid cube (Dimensions: 10 mm x 10 mm x 10 mm);

[0066] C: Cantilever structure (Dimensions: base plate 15 mm x 10 mm x 5 mm, cantilever length 10 mm);

[0067] D: Thin-walled circular ring (Outer diameter 20 mm, wall thickness 1 mm, height 10 mm);

[0068] E: Cross structure (Two 10 mm long walls intersected);

[0069] Finite element simulation: A thermal-mechanical sequential coupling model was established for each structure in the commercial software ABAQUS. In the thermal analysis step, the simplified model defined in S1.1 was used as a thermal input in the form of a user subroutine DFLUX. For each structure, 5 different (TP, Rx, Rz) parameter combinations were input, a total of 25 transient thermal analyses were performed by fluctuating up and down from the baseline process parameters. Subsequently, the calculated temperature field history was imported as a predefined field into the mechanical analysis module for elastoplastic mechanical analysis considering material nonlinearity;

[0070] Data extraction: After each mechanical calculation is completed, the longitudinal equivalent intrinsic strain e of each deposited layer is extracted under the condition that the model is completely cooled to room temperature 25°C and the constraint at the bottom of the substrate is removed. Finally, a dataset containing 5(structures) x 5(parameters) = 25 samples is obtained. The input features of each sample are (TP, Rx, Rz, structure type), where the structure type is one-hot encoded; the output is the intrinsic strain value of all deposited layers of the structure.

[0071] S1.3: Build and train the first machine learning model: Use the key parameter set of the simplified temperature field model and the structure type encoding as input features, and the corresponding calculated intrinsic strain of each layer as output, to build and train the first machine learning model, to establish a high-dimensional nonlinear mapping relationship from "temperature field parameters and structure type" to "intrinsic strain interlayer distribution", as a proxy model for fast prediction of intrinsic strain.

[0072] Environment and tools: Python's Scikit-learn and TensorFlow libraries are used for machine learning model development.

[0073] Data preprocessing: The dataset of 25 samples was randomly divided into a training set (20 samples) and a test set (5 samples) in an 8:2 ratio. The input features were standardized, and the output was normalized.

[0074] Model construction: A multi-layer perceptron regression model was constructed with the following structure:

[0075] Input layer: 8 neurons corresponding to 4 input features (TP, Rx, Rz) and 4-dimensional structure one-hot encoding.

[0076] Hidden layer: Two fully connected layers. The first hidden layer contains 32 neurons using ReLU activation function; the second hidden layer contains 16 neurons also using ReLU activation function.

[0077] Output layer: The number of neurons corresponds to the maximum number of deposited layers of the component, which is 20 neurons in this embodiment, corresponding to 20 layers, using a linear activation function, outputting the inherent strain value of each layer.

[0078] Model training: The training set was used to train the model. The optimizer used Adam, and the loss function was mean squared error. The training period was set to 1000, and the batch size was 4. During training, the test set was used for validation, and the early stopping method was used to prevent overfitting. Finally, the determination coefficient R² of the prediction results of the model on the test set and the finite element calculation results was >0.95, indicating that it successfully established a high-precision nonlinear mapping relationship from the temperature field parameters to the inherent strain, which can be used as an efficient proxy model.

[0079] S2: As shown in Figure 4 , a residual stress prediction network based on inherent strain tensor encoding is constructed. The inherent strain tensor is structured and encoded through a dynamic self-adjusting tensor encoding layer. Based on the encoded inherent strain tensor, a mapping relationship from inherent strain to residual stress is established.

[0080] Existing deep learning-based residual stress prediction methods usually input the inherent strain (εxx, εyy, εzz, εxy, εyz, εxz) directly as independent scalars into the residual stress prediction network. This approach has the following problems:

[0081] Tensor structure is destroyed: The symmetry between components (such as εxy = εyx) is not maintained, and the network cannot understand the physical coupling relationship.

[0082] Spatial direction information is lost: Scalar input ignores the directional properties of strain in the spatial coordinate system.

[0083] Difficult to generalize: Under different loading directions, geometric rotations, or coordinate transformations, the stress mapping learned by the network is no longer consistent.

[0084] Thus, the model causes the prediction difference of different printing directions and geometric postures, high training data requirement, and poor physical consistency.

[0085] Therefore, the present application introduces a new "tensor feature coding structure" (i.e. tensor coding unit TEU) between the input layer and the hidden layer of the residual stress prediction network, solves the problem that the existing prediction network cannot efficiently process anisotropic strain tensor. The residual stress prediction network based on intrinsic strain tensor coding of the present application includes an input layer, a tensor coding unit, a hidden layer and an output layer, the input layer receives the intrinsic strain tensor obtained in the laser additive manufacturing process; the tensor coding unit realizes the structural coding of the intrinsic strain tensor through the bilinear tensor operation of the symmetric weight matrix and the antisymmetric weight matrix, outputs the high-dimensional embedded features that maintain the symmetry and directional information of the strain tensor; the hidden layer predicts the residual stress tensor according to the embedded features; and the output layer outputs the residual stress field distribution result.

[0086] The core of the residual stress prediction network based on intrinsic strain tensor coding is to insert a dynamic self-adjusting tensor coding layer (D-TEL) between the network input layer and the hidden layer, and the core of the dynamic self-adjusting tensor coding layer (D-TEL) is a tensor coding unit TEU, which is used to map the strain tensor into a high-dimensional tensor embedding vector in a way that maintains symmetry and directionality, so that the network can automatically capture the physical coupling relationship inside the tensor.

[0087] The structural principle of the tensor coding unit TEU: the unit includes two learnable matrix parameters: a symmetric weight matrix W s ∈R3×3 and an antisymmetric weight matrix W a ∈R3×3.

[0088] Given the intrinsic strain tensor of a node or element:

[0089] which physically satisfies εxy=εyx, εyz=εzy, εzx=εxz.

[0090] The coding process of the TEU is as follows: where ":" denotes the bilinear tensor inner product; "x" denotes tensor multiplication; n is the node normal direction (derived from local grid geometric information); and z is the coding vector used as the input of the subsequent network.

[0091] The TEU learns the symmetric coupling between components through W s and captures the antisymmetric information related to rotation through W a . Finally, the coding vector z satisfies the equivariance condition under any coordinate rotation: .

[0092] The overall training objective of the network is to minimize the loss function: .

[0093] The second term in the loss function is a symmetry regularization term, which ensures that the weight matrix remains physically symmetric.

[0094] D-TEL is a new type of neural network structure unit, whose weights are generated by local physical field self-adaption, and the structured coding of inherent strain is realized through dynamic tensor operation.

[0095] The input of D-TEL is the inherent strain tensor ε(x), the normal vector n(x), the temperature gradient , the historical stress residual h(x). Dynamic weight generation is realized by lightweight HyperNet, which is mapped as: ; Where, W s (x): symmetric tensor weight (maintain energy consistency); W a (x): antisymmetric tensor weight (capture rotation-related components).

[0096] Tensor coding performs the following mapping: ; Where, “:” represents the bilinear tensor inner product; “x” represents the directional projection; vec() represents the flattening to vector. This structure maintains the physical symmetry and direction dependence. This structure dynamically generates tensor mapping weights through HyperNet, realizes the physical self-adaptive coding and local dynamic adjustment of strain tensor, and significantly improves the precision, generalization and physical consistency of the model.

[0097] The coded inherent strain tensor is input into the main prediction network to predict the residual stress.

[0098] The specific implementation steps of S2 are as follows:

[0099] S2.1: Obtain the residual stress training data set: use the inherent strain results of the feature structure obtained by finite element simulation in step S1.2 as the initial mechanical load, apply it to each deposition layer of the corresponding feature structure finite element model using the inherent strain method, and quickly obtain the residual stress field data of the entire feature structure through one elastic-plastic statics calculation.

[0100] S2.2: Dynamic self-adjusting tensor encoding layer (D-TEL): To improve the expressive ability and physical consistency of the prediction model for the inherent strain tensor, the invention introduces a dynamic self-adjusting tensor encoding layer (D-TEL). This layer generates tensor mapping weights in real time according to local geometric information, temperature gradient and strain state through a dynamic weight generation module, realizing regional adaptive coding of inherent strain. The features processed by D-TEL are input into the main network for residual stress prediction, which can significantly improve the prediction accuracy, generalization and stability. The inherent strain distribution data obtained in step S1.2 is used as input, and the corresponding residual stress field calculated in step S2.1 is used as output to construct a data set. Using this data set, a residual stress prediction network based on inherent strain tensor encoding is trained. To achieve residual stress prediction for different types of structural components.

[0101] S3: Geometric topology adaptive partitioning of large components, predicting the inherent strain of each sub-region, and continuously fusing the strain fields of each sub-region through a strain field gradient fusion algorithm.

[0102] S3.1: Geometric topology adaptive feature partitioning: For large complex components, based on the topology structure of their three-dimensional geometric model and stress concentration risk analysis, the graph cut algorithm is used to automatically decompose them into several sub-regions similar to the typical feature structures described in step S1.

[0103] Component to be predicted: A complex large-size complex component, whose geometry includes upright support walls, outwardly extending reinforcing ribs (similar to cantilevers), bosses connected to other parts (similar to blocks), and arch-shaped support structures.

[0104] Partitioning algorithm: Use the pyvista library of Python to read its STL model, based on curvature and structural connectivity, use the graph cut algorithm to automatically segment it into several sub-regions. The segmentation results are as follows: 4 "wall" regions, 2 "cantilever" regions and 1 "block" region; the arch-shaped support can be approximated as a combination of multiple features.

[0105] S3.2: Machine learning to predict the inherent strain of sub-regions: For each sub-region, according to its corresponding laser process parameters, convert the key parameters of the simplified parameter temperature field model, call the "temperature field parameter-inherent strain" machine learning model trained in step S1.3, and quickly predict the inherent strain value of the deposited layer.

[0106] S3.3: Strain field gradient fusion and overall stress prediction: The inherent strain field of each sub-region corrected in step S2.3 is applied as a load to the finite element model of the overall large component.

[0107] Overall modeling: A fine finite element mesh model of the entire large-size complex component is established in ABAQUS;

[0108] Load application and fusion: the intrinsic strain field of each sub-region predicted by S3.2 and corrected by the method described in S2.3 is applied to the corresponding region element as a predefined field;

[0109] Key fusion step: at the junction of any two adjacent sub-regions, a transition zone with a width of 2 times the element size is set. For each integration point P in the transition zone, the final intrinsic strain value ε no longer belongs to a certain region, but is calculated by a bilinear interpolation algorithm based on distance weight.

[0110] Wherein, the bilinear interpolation fusion algorithm based on distance weight is used at the junction of sub-regions: a transition zone is set at the junction of adjacent sub-regions A and B, and for any point P in the zone, the fusion strain is:

[0111] ;

[0112] Wherein, εA and εB are the strain values of the point predicted according to the A and B sub-region models respectively. WA and WB are distance weights:

[0113] ;

[0114] ;

[0115] dA and dB are the nearest distances from point P to the boundaries of A and B sub-regions respectively. This operation ensures the first-order continuity of the strain field in the transition zone.

[0116] S4: using the fused strain field to predict the residual stress of the whole component, and verifying and correcting the prediction results by experimental data to form a closed-loop optimization system.

[0117] Based on the fused strain field, a first static mechanical analysis is performed using the intrinsic strain method to obtain the residual stress and deformation prediction results of the whole large complex component.

[0118] The titanium alloy component is actually printed using an LP-DED device. After manufacturing, the residual stress is measured at the root section of the stiffener with the highest risk of stress concentration using the profile method. The measured stress distribution curve is compared with the stress results on the same section predicted by the finite element method in S3.3, ensuring that the maximum error is less than 15%, proving that the method of the present application is significantly better than the traditional method without using the integrated correction and fusion strategy, fully verifying the high precision and engineering practicability of the present application.

[0119] The specific embodiment fully demonstrates the whole process of the present application from basic data preparation, model construction, correction optimization to final application in large complex component prediction. By deeply integrating physical mechanism, data driving and experimental verification, the present application successfully overcomes the bottleneck of traditional numerical simulation method in efficiency, accuracy and universality, and provides an efficient, accurate and reliable solution for process optimization and deformation control of metal additive manufacturing.

[0120] Embodiment 2

[0121] The embodiment provides a laser additive manufacturing stress deformation prediction system based on machine learning, comprising:

[0122] The temperature field modeling module is configured to construct a machine learning agent model, and establish a nonlinear mapping relationship between simplified temperature field parameters, structure characteristics and inherent strain;

[0123] The machine learning module is configured to construct a residual stress prediction network based on inherent strain tensor coding, to structure the inherent strain tensor through a dynamic self-tuning tensor coding layer, and to establish a mapping relationship from inherent strain to residual stress based on the coded inherent strain tensor;

[0124] The partitioning and fusion module is configured to adaptively partition the large component in terms of geometry topology, to predict the inherent strain of each sub-region, and to continuously fuse the strain fields of each sub-region through a strain field gradient fusion algorithm;

[0125] The verification and correction module is configured to predict the residual stress of the whole component by using the fused strain field, and to verify and correct the prediction result by using experimental data, to form a closed-loop optimization system.

[0126] It should be noted that the above modules correspond to the steps in Embodiment 1, and the above modules and the examples and application scenarios realized by the steps corresponding thereto are the same, but are not limited to the content disclosed in Embodiment 1. It should be noted that the above modules as part of the system can be executed in a computer system.

[0127] In more embodiments, there are also provided:

[0128] An electronic device comprising a memory and a processor, and computer instructions stored on the memory and running on the processor, when the computer instructions are run by the processor, the method in Embodiment 1 is completed. For brevity, it will not be repeated here.

[0129] It should be understood that, in this embodiment, the processor can be a central processing unit (CPU), and the processor can also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0130] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the method in embodiment 1.

[0131] The method in embodiment 1 can be directly executed by a hardware processor, or executed by a combination of hardware and software modules in the processor. The software modules can be located in a storage medium mature in the art, such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in a memory, and the processor reads information in the memory to complete the steps of the above method in combination with the hardware thereof. To avoid repetition, no further description is given here.

[0132] A computer program product comprising a computer program, which, when executed by a processor, implements the method in embodiment 1.

[0133] The present application also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as those included in program modules, executed by devices at a destination real or virtual processor to perform processes / methods as described above. Generally, program modules include routines, programs, libraries, objects, classes, components, data structures, etc. that perform particular tasks or implement particular abstract data types. In various embodiments, the functionality of program modules can be combined or split between program modules as desired. Machine-executable instructions for program modules can be executed within a local or distributed device. In a distributed device, program modules can be located in local and remote storage media.

[0134] Computer program code for carrying out operations of the present application can be written in one or more programming languages. These computer program code can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus, so that the program code, when executed by the computer or other programmable data processing apparatus, causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on a computer, partially on a computer, as a standalone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0135] In the context of the present application, the computer program code or related data can be carried by any suitable carrier to enable the device, apparatus or processor to perform the various processes and operations described above. Examples of carriers include signals, computer readable media, and the like. Examples of signals can include electrical, optical, radio, sound or other forms of propagated signals, such as carrier waves, infrared signals, and the like.

[0136] Those skilled in the art can understand that the units and algorithm steps of the examples described in combination with the embodiments can be realized by electronic hardware or a combination of electronic hardware and computer software. Whether the functions are realized in hardware or software manner depends on the specific application and design constraints of the technical solutions. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0137] Although the specific embodiments of the present application are described above in combination with the drawings, it is not a limitation on the scope of protection of the present application. Those skilled in the art should understand that various modifications or variations made by those skilled in the art without creative labor on the basis of the technical solutions of the present application are still within the scope of protection of the present application.

Claims

1. A method for predicting stress distortion in laser additive manufacturing based on machine learning, characterized in that, The method comprises the following steps: a machine learning agent model is constructed to establish a nonlinear mapping relationship between simplified temperature field parameters, structure characteristics and inherent strain; a residual stress prediction network based on inherent strain tensor coding is constructed, the inherent strain tensor is structured and coded through a dynamic self-adjusting tensor coding layer, and a mapping relationship from inherent strain to residual stress is established based on the coded inherent strain tensor; a geometric topology adaptive partitioning is performed on the large component, the inherent strain of each sub-region is predicted, and the continuous fusion of the strain fields of each sub-region is performed through a strain field gradient fusion algorithm; the residual stress of the overall component is predicted using the fused strain field, and the prediction result is verified and corrected through experimental data to form a closed-loop optimization system; the machine learning agent model adopts a multilayer perceptron regression model, the input is temperature field parameters and structure type coding, and the output is the inherent strain value of each deposition layer; the dynamic self-adjusting tensor coding layer realizes the bilinear coding of the strain tensor through symmetric weight matrices and antisymmetric weight matrices, and introduces a symmetry regularization term to maintain physical constraints.

2. The machine learning based laser additive manufacturing stress distortion prediction method of claim 1, wherein, The simplified temperature field parameters include equivalent peak temperature, transverse heat affected zone radius and longitudinal heat affected zone depth.

3. The machine learning based laser additive manufacturing stress distortion prediction method of claim 1, wherein, The strain field gradient fusion algorithm adopts a bilinear interpolation based on distance weight at the junction of the sub-regions to ensure the first-order continuity of the strain field.

4. The machine learning based laser additive manufacturing stress distortion prediction method of claim 1, wherein, The geometric topology adaptive partitioning adopts a graph cut algorithm to automatically partition based on the curvature and structure connectivity of the three-dimensional geometric model.

5. A machine learning based laser additive manufacturing stress distortion prediction system characterized by, The method for predicting stress deformation in laser additive manufacturing based on machine learning according to any one of claims 1-4 comprises: a temperature field modeling module configured to construct a machine learning agent model to establish a nonlinear mapping relationship between simplified temperature field parameters, structure characteristics and inherent strain; a machine learning module configured to construct a residual stress prediction network based on inherent strain tensor coding, to structure and code the inherent strain tensor through a dynamic self-adjusting tensor coding layer, and to establish a mapping relationship from inherent strain to residual stress based on the coded inherent strain tensor; a partitioning and fusion module configured to perform geometric topology adaptive partitioning on the large component, to predict the inherent strain of each sub-region, and to perform continuous fusion of the strain fields of each sub-region through a strain field gradient fusion algorithm; a verification and correction module configured to predict the residual stress of the overall component using the fused strain field, and to verify and correct the prediction result through experimental data to form a closed-loop optimization system.

6. An electronic device, comprising: A computer program product comprising a memory and a processor, and computer instructions stored on the memory and executed on the processor, when executed by the processor, complete the method of any one of claims 1-4.

7. A computer readable storage medium characterized in that, A computer program product for storing computer instructions, when executed by a processor, completes the method of any one of claims 1-4.

8. A computer program product, characterised in that, A computer program product for storing computer instructions, when executed by a processor, completes the method of any one of claims 1-4.

Images