Training method and training device of machine learning model, prediction system

Abstract

The application discloses a machine learning model training method, a machine learning model training device, a prediction system, a computer device and a computer readable storage medium. In the input data of the training data of the machine learning model, the printing features of the object three-dimensional model include voxel features and model geometric features. By obtaining the characteristics of each basic unit of the object three-dimensional model and the correlation between the basic units, the target machine learning model obtained by training can calculate the temperature history or / and stress history data changing over time in the printing of complex structures based on the voxel features and model geometric features obtained by discretizing the complex structure when predicting the printing process of different printing components. The machine learning model obtained by the machine learning model training method of the application is not only limited to the printing prediction of simple geometric bodies such as filled cubes, but also applicable to the printing prediction of various printing components.

Description

Technical Field

[0001] This application relates to the field of computer data processing, and more specifically to a training method for a machine learning model, a training device, a prediction system, a computer device, and a computer-readable storage medium. Background Technology

[0002] Existing fused deposition modeling (FDM) slicing software uses purely geometric methods to determine the printing path and parameters, which cannot effectively guarantee printing speed and the performance of the final printed product, such as shape distortion, interlayer adhesion, elastic modulus, and strength. Analyzing the printing process and evaluating the performance of the obtained object based on existing slicing technology to adjust the printing path and parameters is difficult to track changes in the material's temperature and stress fields during the printing process, and it also incurs high economic and time costs.

[0003] To address the aforementioned issues, several multiphysics simulation methods based on the finite element method (FEM) exist, which obtain temperature and stress data of printed materials during the material forming process. However, the FEM simulation process often suffers from low computational efficiency, long computation time, and high computational resource consumption, making it difficult to deploy the FEM simulation method to user-operable applications. Summary of the Invention

[0004] In view of the shortcomings of the above-mentioned related technologies, this application provides a training method for a machine learning model, a training device for a machine learning model, a prediction system, a computer device, and a computer-readable storage medium to solve the problems of low computational efficiency and high computational resource consumption in the simulation printing process of 3D printing in the prior art.

[0005] To achieve the above and other related objectives, the first aspect of this application discloses a method for training a machine learning model. The machine learning model is used to predict the temperature history and / or stress history data of printing materials over time in 3D printing. The training method includes the following steps: acquiring printing features of multiple sets of three-dimensional object models, and acquiring the temperature history and / or stress history data of the printing process over time corresponding to the multiple sets of three-dimensional object models; wherein the printing features include model geometric features and voxel features; wherein the voxel features include one or more of the following: type of basic unit in the three-dimensional object model, distance from the basic unit to the free surface, printing speed of the basic unit, printing time of the layer where the basic unit is located, position coordinates of the basic unit, and density of the position of the basic unit; using the printing features of the multiple sets of three-dimensional object models as input data and the temperature history and / or stress history data of the multiple sets of three-dimensional object models as output data, and performing supervised learning based on the input data and output data to obtain the machine learning model.

[0006] The second aspect of this application discloses a training apparatus for a machine learning model. The machine learning model is used to predict the temperature history and / or stress history data of printing materials over time in 3D printing. The training apparatus includes: a training sample acquisition module, used to acquire printing features of multiple sets of three-dimensional object models, and to acquire the temperature history and / or stress history data of the printing process over time corresponding to the multiple sets of three-dimensional object models; wherein the printing features include model geometric features and voxel features; wherein the voxel features include one or more of the following: type of basic unit in the three-dimensional object model, distance from the basic unit to the free surface, printing speed of the basic unit, printing time of the layer where the basic unit is located, position coordinates of the basic unit, and density of the position of the basic unit; and a training module, used to take the printing features of the multiple sets of three-dimensional object models as input data and the temperature history and / or stress history data of the multiple sets of three-dimensional object models as output data, and to perform supervised learning based on the input data and output data to obtain the machine learning model.

[0007] The third aspect of this application discloses a prediction system for predicting the temperature history and / or stress history data of printing materials over time in 3D printing, comprising: a receiving module for receiving printing features of the object's three-dimensional model; and a prediction module for calling a machine learning model generated by the training method of the machine learning model as described in any embodiment of the first aspect of this application to predict the printing features of the object's three-dimensional model, and outputting the temperature history and / or stress history data of the printing materials of the object's three-dimensional model over time during printing.

[0008] The fourth aspect of this application discloses a computer device comprising: a storage device for storing at least one program and a machine learning model trained by a training method for a machine learning model as described in any embodiment of the first aspect of this application; and a processing device connected to the storage device for executing the at least one program to invoke the machine learning model in the storage device to predict the printing features of the three-dimensional model of the object, and output temperature history and / or stress history data of the printing material of the three-dimensional model of the object over time during printing.

[0009] The fifth aspect of this application discloses a computer device network-connected to a service system, comprising: a communication device for acquiring a machine learning model trained by a training method of a machine learning model as described in any embodiment of the first aspect of this application from the service system; a storage device for storing at least one program; and a processing device connected to the storage device for executing the at least one program to invoke the machine learning model acquired from the service system to predict the printing features of the three-dimensional model of the object, and outputting temperature history and / or stress history data of the printing material of the three-dimensional model of the object over time during printing.

[0010] The sixth aspect of this application discloses a computer-readable storage medium, characterized in that it stores at least one program, which, when executed by a processor, implements a training method for a machine learning model as described in any embodiment of the first aspect of this application.

[0011] In summary, the training method of the machine learning model of this application has the following beneficial effects in one embodiment: This application provides a training method for a machine learning model to predict the temperature history and / or stress history of printing materials over time in 3D printing. The target machine learning model obtained thereby can quickly predict the temperature history and / or stress history of the printing material during the printing process after obtaining the printing features of the three-dimensional model of the object. This can solve the problems of low computational efficiency, long computation time, and high computational resource consumption in actual printing or finite element simulation printing. The target machine learning model obtained by the training method of this application can also be deployed to an application that can be directly operated by the user, which is beneficial to the application in actual production.

[0012] Furthermore, the training method of the machine learning model in this application provides an embodiment for designing (determining) and acquiring (generating) input and output data as training data. The input data, namely the printing features of the three-dimensional model of the object, includes voxel features and model geometric features. By acquiring the characteristics of each basic unit of the three-dimensional model of the object and the correlation between the basic units, the trained target machine learning model can calculate and predict the temperature history and / or stress history data that changes over time during the printing of complex structures based on the voxel features and model geometric features obtained by discretizing the complex structure when predicting the printing process of different printed components, including complex structures. Therefore, the machine learning model obtained by the training method of the machine learning model in this application is not limited to the printing prediction of simple geometries such as infilled cubes, but can also be adapted to the printing prediction of various printed components.

[0013] Other aspects and advantages of this application will readily be apparent to those skilled in the art from the detailed description below. Only exemplary embodiments of this application are shown and described in the following detailed description. As will be appreciated by those skilled in the art, the content of this application enables them to make modifications to the disclosed specific embodiments without departing from the spirit and scope of the invention to which this application pertains. Accordingly, the descriptions in the accompanying drawings and specification of this application are merely exemplary and not restrictive. Attached Figure Description

[0014] The specific features of the invention involved in this application are shown in the appended claims. The features and advantages of the invention can be better understood by referring to the exemplary embodiments and drawings described in detail below. A brief description of the drawings is as follows:

[0015] Figure 1 The flowchart shown is an embodiment of a method for training the machine learning model of this application.

[0016] Figure 2 The flowchart shown is a process for determining a feature quantity in one embodiment of the training method for the machine learning model of this application.

[0017] Figure 3 The flowchart shown is a process for determining another feature quantity in one embodiment of the training method for the machine learning model of this application.

[0018] Figures 4a-4b The training methods of the machine learning model of this application are shown respectively as the basic units within the search range of different target units in the three-dimensional model of the object in one embodiment.

[0019] Figure 5 The diagram shown is a flowchart illustrating the process of determining the density of basic unit locations in one embodiment of the training method for the machine learning model of this application.

[0020] Figure 6 The diagram shown is a simplified schematic of some basic units in a three-dimensional model of an object, illustrating the training method of the machine learning model of this application in one embodiment.

[0021] Figure 7 The diagram shown in another embodiment is a simplified schematic of some basic units in a three-dimensional model of an object, illustrating the training method of the machine learning model of this application.

[0022] Figure 8 The diagram shown is a block diagram of the module composition of a training apparatus for the machine learning model of this application in one embodiment.

[0023] Figure 9 The diagram shown is a block diagram of the module composition of the prediction system of this application in one embodiment.

[0024] Figure 10 The diagram shown is a block diagram of the module composition of a computer device according to one embodiment of the present application.

[0025] Figure 11 The diagram shown is a block diagram of the module composition of a computer device according to one embodiment of the present application. Detailed Implementation

[0026] The following specific embodiments illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification.

[0027] In the following description, reference is made to the accompanying drawings, which illustrate several embodiments of the present application. It should be understood that other embodiments may also be used, and changes in module or unit composition, electrical and operational aspects may be made without departing from the spirit and scope of this disclosure. The following detailed description should not be considered limiting, and the scope of the embodiments of the present application is defined solely by the claims of the published patents. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the present application.

[0028] While the terms first, second, etc., are used in some instances herein to describe various elements, information, or parameters, these elements or parameters should not be limited by these terms. These terms are used only to distinguish one element or parameter from another. For example, a first type of basic unit may be referred to as a second type of basic unit, and similarly, a second type of basic unit may be referred to as a first type of basic unit, without departing from the scope of the various described embodiments. Both the first type of basic unit and the second type of basic unit describe a class of basic units, but they are not the same class of basic units unless the context otherwise explicitly indicates otherwise. Depending on the context, the word "if," as used herein, may be interpreted as "when" or "when...".

[0029] Furthermore, as used herein, the singular forms “a,” “an,” and “the” are intended to include the plural forms as well, unless the context indicates otherwise. It should be further understood that the terms “comprising,” “including,” indicate the presence of the stated feature, step, operation, element, component, item, kind, and / or group, but do not preclude the presence, occurrence, or addition of one or more other features, steps, operations, elements, components, items, kinds, and / or groups. The terms “or” and “and / or” as used herein are to be interpreted inclusively, or mean any one or any combination thereof. Thus, “A, B, or C” or “A, B, and / or C” means “any one of: A; B; C; A and B; A and C; B and C; A, B, and C.” Exceptions to this definition occur only when combinations of elements, functions, steps, or operations are inherently mutually exclusive in some way.

[0030] In 3D printing technology, taking fused deposition modeling (FDM) as an example, the material is added in a high-temperature fluid state and then extruded and cooled. During the printing process, the material is gradually added in a molten state, or kept in a molten state by a moving heat source (such as a heating element), and then cooled on the continuously evolving surface.

[0031] Typically, the thermo-mechanical coupling relaxation behavior of the printed polymer material during printing, as well as the stress history during molding, have a significant impact on the mechanical properties of the printed object. Therefore, determining the temperature history and stress history of the printing material over time is an effective way to judge or control the final molding accuracy, strength, and other properties of the printed object. Controlling the temperature history and stress history of the printing material during molding must be determined by setting printing parameters such as extrusion speed, printing speed, layer height, material heating temperature, and printing chamber temperature.

[0032] For example, in FDM printing, the bonding strength of interlayer adhesion caused by thermal expansion and extrusion between different printed layers has a significant impact on the mechanical properties of the printed object. The temperature of the filament contact interface and the diffusion time of polymer molecules in the printing material affect the bonding quality of the filament. Therefore, it is necessary to conduct thermal analysis on the filament deposition process and study the variation of filament temperature with forming parameters. Studying the temperature and stress history during the printing process, and using this to form thermal and stress history data from printing to cooling and forming, can provide data support for improving the performance of printed objects in production.

[0033] However, as described in the background section, existing multiphysics simulations of the printing process suffer from high computational costs, data storage requirements, and computation time, all of which pose challenges in practical applications. Although simulations can reproduce the printing process and environment multiple times to obtain theoretical data, due to the aforementioned problems, simulated printing is difficult to reduce time and economic costs compared to actual printing experiments. Furthermore, finite element simulation methods are difficult to deploy on applications that users can directly operate, making them inconvenient for actual production.

[0034] Therefore, this application provides a method for training a machine learning model. For ease of explanation, the coordinate system used in the embodiments provided is a Cartesian three-dimensional coordinate system, with the directions of the three-dimensional coordinates being X, Y, and Z. (x, y, z) can refer to a coordinate point in the defined three-dimensional space. In actual or simulated printing, the Z direction is the normal direction of the printing surface (usually perpendicular to the horizontal plane). Taking fused deposition modeling as an example, the height direction of each printing layer is the Z direction.

[0035] The machine learning model is used to predict the temperature history and / or stress history data of the printing material over time in 3D printing. The training method includes the following steps: acquiring printing features of multiple sets of three-dimensional object models, and acquiring temperature history and / or stress history data of the printing process over time corresponding to multiple sets of three-dimensional object models; wherein, the printing features include model geometric features and voxel features; wherein, the voxel features include one or more of the following: type of basic unit in the three-dimensional object model, distance from the basic unit to the free surface, printing speed of the basic unit, printing time of the layer in which the basic unit is located, position coordinates of the basic unit, and density of the position of the basic unit; using the printing features of the multiple sets of three-dimensional object models as input data and the temperature history and / or stress history data of the multiple sets of three-dimensional object models as output data, and performing supervised learning based on the input data and output data to obtain the machine learning model.

[0036] The machine learning model training method of this application is applied to the 3D printing industry. Here, 3D printing is a type of rapid prototyping technology, specifically a technology that constructs objects by printing layer by layer using powdered metal, plastic, resin, and other bondable or curable materials based on digital model files. To facilitate the explanation of the machine learning model training method described in this application, the following embodiments mainly use its application in FDM printing (also known as FFF printing) as an example; of course, it should be understood that the machine learning model training method of this application can also be used in FLM printing, as well as adhesive jetting, micro-droplet jetting printing, etc., and is not limited here.

[0037] In the examples provided in this application, the training method of the machine learning model uses printing features as input data and temperature history and / or stress history data of the three-dimensional model of the object changing over time during the printing process as output data to form training data for the machine learning model. The machine learning model obtained thereby can predict its temperature history and / or stress history data after receiving the printing features of the printed component. Thus, this application provides a method for training and learning a model by obtaining temperature history and / or stress history data of the printing material during the molding process, which can improve computational efficiency, reduce computational power and computation time.

[0038] Please see Figure 1 The diagram shows a flowchart of the training method for the machine learning model of this application in one embodiment.

[0039] The training method for the machine learning model is the process of obtaining a machine learning model that can predict the target output data of the samples by analyzing the samples in the training dataset.

[0040] In the embodiments of this application, the training data is a set of training samples used to train a machine learning model in a supervised manner. The training samples include labeled input data and output data. The input data can be features, specifically the printing features in this application. The output data is the target output data of the machine learning model to be trained. By using labeled printing features and output data, the machine learning model learns the correlation (pattern) between sample features and target output data during training, thereby obtaining a machine learning model that can predict the temperature history and / or stress history data of the object's 3D model during the printing process based on printing features.

[0041] In step S10, the printing features of multiple sets of object 3D models are obtained, as well as the temperature history and / or stress history data of the printing process corresponding to the multiple sets of object 3D models over time are obtained.

[0042] In step S10, the object's three-dimensional model is a three-dimensional model corresponding to a preset printed component, such as common printed components used in 3D printing manufacturing, such as molds, medical devices, customized products such as shoe soles, jewelry models, or dental molds.

[0043] In some embodiments of this application, the three-dimensional model of the object includes a regular geometric structure and actual product components. The geometric structure includes one or more of the following: vertical thin wall, horizontal thin plate, inclined thin plate, cube, cylinder, cylindrical body, square cylinder, ring, square ring, cone, inclined column, solid block, and grid-filled structure. The actual product components include one or more of the following: connector, transmission component, and shell.

[0044] The mesh filling structure is, for example, a structure in which a defined printing area (or a three-dimensional contour area of ​​a 3D printed component) is filled by a mesh with gaps, such as a triangle, quadrilateral, or hexagon. In one example, the mesh filling structure can be determined based on the area of ​​the region and the selected mesh shape.

[0045] The actual product components mentioned herein can be representative structures such as connectors, transmission components, housings, and other standard parts commonly used in appliances, equipment, and devices. It should be understood that the actual product components mentioned herein do not necessarily have to be complete products or independent sales units, but can also be components of a product.

[0046] Training the machine learning model requires training data, and the object's 3D model can be used to form the training data for the machine learning model. In the examples above, the selected object's 3D model geometry includes common printing geometries and representative product components, so that the 3D model predicted by the trained machine learning model is more universal, enabling the prediction of temperature history and / or stress history for various 3D models.

[0047] The data of the three-dimensional model of the object can be in any known format, including but not limited to Standard Tessellation Language (STL) or Stereo Lithography Contour (SLC) format, Virtual Reality Modeling Language (VRML), Additive Manufacturing File (AMF) format, Drawing Exchange Format (DXF), Polygon File Format (PLY) format, or any other format suitable for Computer-Aided Design (CAD).

[0048] The printing features serve as the input data for the machine learning model. Correspondingly, output data that corresponds to the input data is also needed to form a training dataset. The output data refers to the temperature history and / or stress history data over time corresponding to the multiple sets of 3D object models in step S10. Training is performed based on the training dataset to generate temperature history and / or stress history data that can be used to predict the temperature history and / or stress history data of the printing material over time during printing.

[0049] The training dataset contains multiple sets of input data and multiple sets of output data that have corresponding relationships, and labels or markers are formed in the training dataset based on these relationships.

[0050] To illustrate the correspondence, for example, for a given 3D object model A1, a set of printing parameter information is assigned to the model to form printing feature B. 11 (B i1 B j1 B k1 B l1 B m1 B n1 The stress history data S is obtained by conducting printing experiments or finite element simulation printing. 11 or / and historical temperature data T 11 The stress history data S mentioned here11 or / and temperature history data T 11 With printing feature B 11 Label them to correspond to each other, forming a set of training data Q. 11 (B 11 T 11 S 11 During the training process, the stress history data S is based on the aforementioned correspondence. 11 or / and temperature history data T 11 Can be identified as printing feature B 11 The expected data. Among them, B i1 B j1 B k1 B l1 B m1 B n1 This can refer to different printing parameters, such as the print head's print path, layer height, print head movement speed, filament output speed, filament heating temperature, and extrusion temperature. Here, for the same 3D printed component model A1, another set of printing parameter information is configured to form the printed feature B. 12 (B i2 B j2 B k2 B l2 B m2 B n2 The stress history data S is obtained by conducting printing experiments or finite element simulation printing. 12 Historical temperature data T 12 The printed feature B 12 With stress history data S 12 Historical temperature data T 12 By labeling them to correspond to each other, another set of training data Q12(B) can be formed. 12 S 12 T 12 During the training process, the stress history data S is based on the aforementioned correspondence. 12 Historical temperature data T 12 Can be identified as printing feature B 12 Expected data.

[0051] Multiple sets of training data can be obtained by changing the slice pattern or printing parameters during printing experiments. It should be understood that the values ​​of at least one parameter in the printing parameters and / or the layered (sliced) patterns of the 3D object model may differ between any two sets of training data.

[0052] The layered (sliced) graphics are obtained in advance based on cross-sectional slices of the 3D object model along the Z-axis. Specifically, a sliced ​​graphic outlined by the contour of the 3D object model is formed on each cross-sectional layer created by adjacent cross-sections. If the cross-sectional layer is sufficiently thin, the contour lines of the upper and lower cross-sectional surfaces of the cross-sectional layer can be considered to be consistent. For 3D printing equipment based on surface projection, each sliced ​​graphic is also called a layered image (pattern) or sliced ​​image (pattern).

[0053] In the embodiments provided in this application, the layered (sliced) graphics are all the layered graphics that constitute the 3D printed component model. The corresponding slice data includes each layered (sliced) graphics and its configured layer height, as well as one or more of the following: the printing path of the layered graphics and the printing material output speed.

[0054] The printing parameters include printing material properties such as filament type, filament diameter, maximum material heating temperature, and material thermal parameters, as well as printing equipment parameters such as printing material heating temperature, extrusion temperature, component plate heating temperature of the printing equipment, and initial printing temperature field.

[0055] It should also be noted that in the embodiments of this application, the input data and output data used to train the machine learning model have a corresponding relationship. In the implementation, the order in which the input data and output data are obtained is not limited, only the correspondence between the two needs to be ensured; for example, the same layered graphic and printing parameter information are simulated to obtain output data, and the corresponding printing features are calculated based on the layered graphic and printing parameter information.

[0056] Typically, in existing methods, prediction of printed components based on neural network models is often limited to a few specific simple geometric structures, such as solid cuboids, thus severely restricting practical applications. In contrast, the machine learning model trained using the method described in this application can predict a variety of printed components. This is due to two factors: firstly, the type of the object's 3D model structure determined in this application; and secondly, the printing features determined from the object's 3D model. These printing features characterize the model structure and the relationships between different voxels, enabling the trained machine learning model to predict more complex geometric structures. To illustrate the method for determining these printing features, the following embodiments are also provided.

[0057] Generally speaking, a voxel can represent the smallest unit of digital data in three-dimensional space, and can also be referred to as a basic unit in the various embodiments provided in this application. For example, a three-dimensional model of an object can be divided into multiple basic units, each of which is the smallest indivisible unit, and the basic units used to compose the three-dimensional model of the object are in a state of being completely filled with printing material.

[0058] In the embodiments provided in this application, the printing features are feature quantities formed by various printing information from actual printing or simulated printing. These feature quantities can be vectors, matrices, numbers, coordinates, digits, codes, coded characters, images, etc., and their specific forms are not limited. It should be noted that the feature quantities can be used to represent or can be translated into information used to form the feature quantities. For example, the heating temperature of the printing material can be represented by degrees Celsius or a numerical value of thermodynamic temperature; the printing path can be represented by a series of sequential coordinate points; and the single-layer printing time can be represented by duration. By acquiring these feature quantities, printing information can be used to form the feature quantities.

[0059] The model geometric features can be used to characterize the geometric shape of the three-dimensional model of the object. At the same time, during the 3D printing process, the printing material is in a state of gradual accumulation. In some examples, the model geometric features can also characterize the gradually forming geometric shape of the dynamically printed object.

[0060] In some embodiments, the model geometric features include feature quantities formed by the G-Code data of the three-dimensional model of the object, the G-Code data including one or more information such as printing path, print head movement speed, layer height, printing material output speed, and single-layer printing time.

[0061] The G-Code data is primarily a numerical control programming language, and in practice, it can have multiple representations.

[0062] In practice, G-Code data includes a series of spatial coordinate points with a sequential execution order. The G-Code data of the object's three-dimensional model is a three-dimensional model represented by coordinates and time series or sequential order. For example, when the G-Code data is input into a computing device with processing capabilities, the path formed by following the order of each spatial coordinate point of the G-Code data constitutes the object's three-dimensional model.

[0063] Here, based on the spatial coordinates with execution order in the G-Code data, the printing path information of the 3D object in actual printing can be obtained. Taking FDM printing as an example, in FDM printing, filamentous thermoplastic material is typically fed into the thermoplastic nozzle via a filament feeding mechanism (usually a roller). Inside the nozzle, the filamentous material is heated and melted, while the nozzle moves along the contour and infill trajectory of the part layers, extruding the molten material and depositing it at a designated location where it solidifies and bonds with the previously formed layer, layer by layer, ultimately forming the product model. The printing path information can be used to indicate the movement path of the print head of the FDM printing equipment, and the movement path also includes the temporal sequence of movement. For example, by inputting the G-Code data into the printing equipment, the print head moves along the path formed by the sequence of each spatial coordinate point in the G-Code data, thus forming the 3D printed component.

[0064] It should be understood that the print head is a heating head (also known as a heating nozzle, spray nozzle, or nozzle) in an FDM type 3D printing device, used to heat and melt the filament, which is the raw material for 3D printing, into a liquid material and apply it to the component plate; the component plate (also known as a component platform or printing platform) is a platform for attaching the target 3D component, and it can move along the vertical Z-axis according to the signal provided by the computer-operated controller.

[0065] In some embodiments of the training method of this application, the training dataset for the machine learning model can be obtained based on printing experiments or simulated printing. Therefore, the print head can also be a virtual print head, such as a print head virtually modeled in a finite element simulation environment or a functionally equivalent print head. For example, in a finite element simulation environment, the print head can serve as the starting point for conveying printing material, causing the printing material to increase according to the print head's movement path (which can be represented as an activation sequence in the simulation). The functionally equivalent print head achieves this by influencing the print head on the printing material, such as influencing the printing material to increase along a predetermined path of the print head, or influencing the temperature of the extruded printing material. In actual finite element simulations, it is not necessary to establish a print head model.

[0066] The layer height refers to the slice thickness of the printed component model. In some examples, it is the spacing between adjacent layer printing paths in the vertical direction (Z-axis direction) in the G-Code data. The layer height can be used to indicate the height of the print head movement in the Z-direction. Of course, in terms of data representation, the layer height can be represented by the printed layer thickness, or it can be represented by the height of each printed layer from a reference surface, such as the surface of the component board (e.g., the Z coordinate in the G-Code data). After obtaining the height of each printed layer, the difference between the height of the current printed layer and the adjacent printed layers can also be used to determine the layer thickness of the printed layer.

[0067] Parameters defined by the print head, such as the printing path, print head moving speed, and heating temperature using the print head as a moving heat source, can be set based on actual print head data during printing or can be set manually. The parameter information of the print head can be read by the 3D printing equipment or used for finite element simulation.

[0068] The output speed of the printing material is the output speed of the printing material relative to the print head, that is, the speed at which material is added in the printing environment.

[0069] The single-layer printing time is the total printing time of each printing layer. In the G-Code data of the object's 3D model, the total printing time of each printing layer can be determined by the time interval from the printing start point to the printing end point of that printing layer.

[0070] In some embodiments, the printing feature further includes feature quantities formed by printing equipment information, wherein the printing equipment information includes one or more of the following: printing material heating temperature, component plate heating temperature of the printing equipment, initial printing temperature field information, and printhead diameter.

[0071] The printing temperature field information is the temperature field information inside the printing device or printing chamber at the initial moment of printing. In one example, the method of obtaining the initial printing temperature field information includes: obtaining the printing chamber temperature distribution information by measuring the temperature distribution of the printing chamber before printing based on a thermal imager or thermocouple during the printing experiment.

[0072] It should be understood that in actual printing, the initial printing temperature field is determined by the printing equipment and the environment in which the printing equipment is located, so there is no need to input the initial printing temperature field into the printing equipment; while in simulated printing, finite element simulation usually uses mathematical approximation methods to simulate real physical systems (such as geometry and load conditions), that is, it is usually necessary to ensure that the simulated environment is compared with the actual printing environment to form a simulation of the real physical environment.

[0073] Here, temperature distribution data within the printing chamber before printing begins is obtained using a thermal imager or thermocouple. This temperature distribution data, for example, is a temperature visualization image converted from a thermal imager. Based on this temperature distribution data, the temperature values ​​at different locations within the printing chamber at the initial printing moment are determined, forming initial printing temperature field information that can be used in the finite element simulation environment of the component. It should be understood that the temperature detection instrument or device used to obtain the temperature field distribution at the initial printing moment can also be a radiation thermometer, an electronic temperature sensor, etc., and this application does not impose any limitations.

[0074] In some examples, the initial printing temperature field is obtained based on multiple measurements of the printing device, and the measured temperature field is corrected via data processing.

[0075] The initial printing time can be set to the moment printing begins or the moment immediately preceding the start of printing. The temperature field in the printing environment is a function of time. Here, the initial time is typically a small time interval, such as 0.5s, 1s, 1.5s, etc. Within this time interval, the temperature field that changes with time can be equivalent to a constant temperature field. By determining the temperature distribution of the printing cavity through actual printing experiments to form the initial temperature field information in the finite element simulation, the accuracy of the training data obtained from the simulated printing can be determined.

[0076] The heating temperature of the printed component board can be a temperature directly set on the component board in an actual printing environment. In a finite element simulation environment, the heating temperature of the printed component board can be a constant temperature contact surface or a contact surface that heats up or dissipates heat according to a preset pattern. The contact surface is the receiving surface of the first layer of printing material, i.e., the bottom surface of the model. Typically, the temperature of the printed substrate is a preset constant value. In some examples of finite element simulations, the boundary condition at the bottom of the model is a preset fixed constant temperature to simulate the temperature of the printed substrate in the actual printing environment; or, based on the temperature change function of the printed substrate in the actual printing environment, the temperature at the bottom of the model is set to the same temperature change function in the simulation.

[0077] The heating temperature of the printing material is the temperature value or temperature range at which the printing material is heated to a molten state inside the printing device; the extrusion temperature is the temperature of the printing material when it is extruded from the print head. In one example, the extrusion temperature is determined by the heating temperature of the printing material and the temperature set at the print head.

[0078] The feature quantities corresponding to the printhead shape are used to characterize the cross-sectional information of the printhead. Generally, the printhead defaults to a circular cross-section or a polygonal cross-section (such as a square, rectangle, rhombus, pentagon, or hexagonal cross-section). The cross-section for delivering printing material to the printing chamber for cooling and shaping can be determined through the printhead shape information. Of course, the printhead cross-section is not limited to the above examples; geometric parameters can be defined according to the preset printhead shape in the simulation.

[0079] In some embodiments, the printing feature further includes feature quantities formed from material property information of the three-dimensional model of the object, wherein the material property information includes one or more of the following: filament type, filament diameter, filament cross-sectional shape, maximum heating temperature of the material, material thermal parameters, and initial residual stress of the material.

[0080] During the printing process, for a given 3D model of an object and equipment parameters, the temperature and stress histories of the printing materials may differ depending on the properties of the printing materials. Furthermore, the printing quality and the mechanical properties of the resulting printed components may also vary. Therefore, this application also provides an embodiment that uses the property information of the printing materials to form feature quantities, so that the trained machine learning model can predict the temperature and / or stress histories of 3D models of objects made of different printing materials.

[0081] The attribute information of the printing material is the material type and its related characteristic information, which can be parameters such as physical properties such as specific heat capacity or thermal conductivity, density, melting point, glass transition temperature, mechanical properties, and chemical properties, which usually correspond to different values ​​or characteristics for different materials.

[0082] The type of filament material is also the type of printing material. In some examples, the printing material includes PLA (Polylactic Acid), ABS (Acrylonitrile Butadiene Styrene), TPU, TPE (Thermoplastic Elastomer), Nylon, carbon fiber (e.g., Carbon Fiber), semi-crystalline thermoplastics, Metal PLA / Metal ABS, PEEK, FDM conductive filaments, Glow-in-the-Dark materials (e.g., adding different colored fluorescent agents to PLA or ABS), and Wood-like materials (by mixing a certain amount of wood fibers into PLA), etc.

[0083] The filament diameter is the diameter of the printing material extruded from the print head. In some embodiments, the filament diameter is the same as the diameter of the print head, or the cross-sectional shape of the filament extruded from the print head is the same as the cross-sectional shape of the print head.

[0084] The cross-sectional shape of the filament is the cross-sectional shape of the forming structure used to stack and form the component during the cooling process of the filament as it is extruded from the print head and continuously evolves on the surface. Ideally, the cross-sectional shape of the filament is the same at the print head position, but may evolve into different shapes after printing based on the heat transfer and mechanical states of different regions.

[0085] The maximum heating temperature of the printing material is the upper limit of the temperature allowed to maintain the performance of the printed object during the process of melting, extruding, and cooling the specific printing material. For example, excessively high temperatures can cause warping and shrinkage of the printing material. This printing information is used as a temperature limit to avoid the results obtained from the simulation analysis of the printing process by the finite element method from being impaired in actual printing due to temperature.

[0086] The material thermal parameters include the specific heat, thermal conductivity, convective heat transfer coefficient, and thermal radiation coefficient (emissivity) of the printing material or printing environment. In some examples, the material thermal parameters are obtained based on measurements or calculations of the actual printing environment. In other examples, the material thermal parameters determined based on actual printing can also be used to construct a comparative finite element simulation environment.

[0087] In some embodiments, the determination of the material thermal parameters includes at least one of the following: measuring the thermal radiation coefficient of the printing material using a thermal radiation coefficient tester; recording the temperature change of the printing filament after it is heated and placed in the printing chamber based on a filament printing experiment, thereby calculating the equivalent convective heat transfer coefficient; conducting a printing experiment on a regularly shaped geometric structure model, recording the temperature field distribution of the printing material over time during the printing process, and setting different convective heat transfer coefficients for the geometric structure model to perform finite element simulation, outputting the simulated temperature field of the simulated printing process, and using the convective heat transfer coefficient corresponding to the simulated temperature field that coincides with the temperature field of the printing experiment as the equivalent convective heat transfer coefficient.

[0088] The thermal radiation coefficient can be measured using a thermal radiation coefficient tester. In one example, the thermal radiation coefficient is obtained by testing the thermal radiation coefficient of the printed component during actual printing, thus forming the characteristic quantity corresponding to the thermal radiation coefficient.

[0089] In some examples, the convective heat transfer coefficient can be calculated using the convective heat transfer coefficient formula, i.e., Newton's law of cooling. For example, based on a defined material type, a filament printing experiment is conducted. The filament is uniformly heated to different temperatures and then placed in a printing cavity. Temperature detection devices, such as infrared thermal imagers or thermocouples, are used to record the temperature change of the filament over time. The convective heat transfer coefficient of the printing material is calculated based on the temperature of different areas within the printing cavity, i.e., the temperature field as a function of time, and the defined actual printing environment, such as the heat exchange area where convective heat transfer occurs.

[0090] In another embodiment, a printing experiment is conducted on a geometrically structured model with a regular shape. The temperature field distribution of the printing material changes over time during the printing process is recorded. Furthermore, a finite element simulation is performed on the geometrically structured model with different convective heat transfer coefficients to output the simulated temperature field of the printing process. The convective heat transfer coefficient corresponding to the simulated temperature field that coincides with the temperature field of the printing experiment is taken as the equivalent convective heat transfer coefficient.

[0091] Here, the geometric model with regular shape is, for example, a simple axisymmetric structure such as a horizontal thin plate or a thin-walled cylinder. In the printing experiment, the temperature change of the surface of the geometric structure is recorded by a temperature detection device such as an infrared thermal imager. For the same geometric structure model, multiple sets of simulated printing are performed with control variables for the printing parameter information. The variable is the convective heat transfer coefficient set in the finite element simulation environment. In the finite element environment, the other variables that affect the temperature change of the printing material during printing are set to be consistent with the printing experiment. The simulated printing is repeated until the temperature change law of the geometric structure surface output by the simulated printing (i.e., the temperature field of the simulated printing) coincides with the temperature field of the printing experiment. The convective heat transfer coefficient set in the finite element environment that obtains the simulated printing temperature field that coincides with the actual printing temperature field is taken as the equivalent convective heat transfer coefficient, and the convective heat transfer coefficient of the corresponding material type can be obtained.

[0092] The initial residual stress is a residual stress that arises from the moment the printing material is extruded from the print head and exists in the printed component until the printing is completed, influenced by factors such as the speed and temperature of the extrusion. This initial residual stress has a certain impact on the formed printed object, such as causing material deformation and cracking. In simulated printing, to ensure that the finite element simulation environment is consistent with the actual printing environment, the initial residual stress can be determined through printing experiments to form the input parameters for the finite element simulation environment, thus making the finite element simulation environment approximate the real printing environment.

[0093] Please see Figure 2 The diagram shows a flowchart of a method for determining the initial residual stress in one embodiment.

[0094] In step S201, multiple sets of monofilament printing experiments are conducted using different printing parameter information for the printing material.

[0095] Typically, the initial residual stress during printing is related to printing parameters, such as extrusion temperature and printing material type (filament type). Here, for different types of printing materials, printing experiments are conducted with different printing parameters to obtain multiple sets of monofilament structures that can be used for initial residual stress measurement. In actual printing, the initial residual stress is formed in the initial stage of printing. During this time, the extruded structure is usually a monofilament structure. Therefore, the residual stress of the monofilament can be used to characterize the initial residual stress of the printed component.

[0096] In step S202, the residual stress of the monofilament components obtained from multiple monofilament printing experiments is calculated or measured.

[0097] In some examples, step S202 may involve processing the monofilament structure to release residual strain, measuring the monofilament deformation to calculate residual stress, or determining the residual stress of the monofilament component based on physical testing methods.

[0098] In some specific examples, the printed monofilament structure is heated to above the glass transition temperature, for example, to 10°C above the glass transition temperature. The lengths of the monofilament structure before and after heating are recorded as L0 and L, respectively. During this process, the residual stress in the monofilament is released, and the residual stress can be determined by the deformation of the monofilament, i.e. (L0-L) / L.

[0099] Here, mechanical methods such as local separation, segmentation, drilling, and grooving can be used to release residual stress in the monofilament structure, and the residual stress can be calculated based on the resulting monofilament deformation. Alternatively, non-destructive physical testing methods such as X-ray diffraction, neutron diffraction, magnetic methods, ultrasonic methods, and indentation strain methods can be used to measure the residual stress distribution of the monofilament structure. By configuring the printing material with different printing parameters and conducting monofilament printing experiments, the residual stress of the monofilament can be measured to obtain the multiple sets of residual stresses.

[0100] In step S203, a residual stress database is obtained, which includes the different printing parameter information and the residual stress of the monofilament component obtained from multiple printing experiments; wherein the residual stress of the monofilament component obtained from the printing experiments has a corresponding relationship with the printing parameter information of the monofilament component.

[0101] The residual stress obtained from single-filament measurements is labeled with the printing parameters of the single filament, so that each residual stress data corresponds to a printing parameter. Single-filament printing experiments are conducted with different printing parameters for different printing materials, thus forming the residual strain database. In one example, the residual strain database is input into a finite element simulation system. During the finite element simulation, the system can match the corresponding residual stress, i.e., the initial residual stress, in the residual strain database based on the received printing parameter information.

[0102] The printing features of the object's three-dimensional model also include voxel features, wherein the voxel features include one or more of the following: the type of basic unit in the object's three-dimensional model, the distance of the basic unit to the free surface, the printing speed of the basic unit, the printing time of the layer where the basic unit is located, the position coordinates of the basic unit, and the density of the position of the basic unit.

[0103] The basic unit is the smallest indivisible unit, also known as a monolithic element. In the embodiments provided in this application, the basic unit can be obtained by discretizing the three-dimensional model of the object.

[0104] For example, the three-dimensional model of the object can be read by a device with processing capabilities. The device can be any computing device with mathematical and logical operations and data processing capabilities, including but not limited to: personal computer devices, single servers, server clusters, distributed servers, cloud servers, etc.

[0105] In some embodiments, the device is, for example, an electronic device loaded with an APP application or capable of accessing a webpage / website. The electronic device includes components such as memory, memory controller, one or more processing units (CPU), peripheral interfaces, RF circuitry, audio circuitry, speakers, microphones, input / output (I / O) subsystems, displays, other output or control devices, and external ports. These components communicate via one or more communication buses or signal lines. The electronic device includes, but is not limited to, personal computers such as desktop computers, laptops, tablets, smartphones, and smart TVs. The electronic device can also be an electronic device consisting of a host with multiple virtual machines and corresponding human-computer interaction devices (such as touchscreens, keyboards, and mice) for each virtual machine.

[0106] In one implementation, the device uses an optional preset density of any type of mesh to represent a real-world object as a discretization of multiple finite elements, where the finite element is a description of the geometric part of the real-world object, which is also the basic unit of this application.

[0107] The geometry and size of the basic unit can be determined by setting the mesh shape and density. In one embodiment, the basic unit used is a regular quadrilateral mesh on a surface, or a cubic mesh on a volume; in some embodiments, a hybrid mesh is used to discretize the three-dimensional model of the object, thereby obtaining basic units of different shapes.

[0108] In some embodiments of this application, the training method of the machine learning model further includes the step of performing finite element discretization on the three-dimensional model of the object to obtain multiple basic units, including: establishing a simulation domain based on the contour of the three-dimensional model of the object; and discretizing the simulation domain into multiple basic units according to a preset mesh type; determining the basic units to be activated in the simulation domain, wherein the basic units to be activated are determined to be non-empty units.

[0109] The simulation domain can be used to determine the computational domain for temperature and stress calculations during subsequent simulated printing. Generally, the larger the simulation domain, the closer it is to the actual printing conditions in production. In finite element simulation, to reduce computational resource consumption and computation time, the simulation domain can include the three-dimensional model of the object, the equivalent printing chamber boundary, and the component plate, with the indoor environment of the printing equipment being equivalent to the simulation domain boundary.

[0110] In some implementations, the step of determining the basic unit to be activated in the simulation domain includes: establishing an envelope range along the printing path in the simulation domain based on the property information of the printing material; determining whether the basic unit is within the envelope range; if it is within the envelope range, activating the unit; if it is not within the envelope range, not activating the unit.

[0111] The envelope range is a three-dimensional spatial region. In some embodiments, the G-Code data of the object's three-dimensional model can be discretized according to time steps. For a certain time step, a vertical circular cross-section is established with the starting and ending coordinates of the time step as the center coordinates and the diameter of the print head or the filament as the diameter. The circular cross-sections of the starting and ending points are connected with a smooth arc surface to form the envelope range within the time step. Alternatively, a vertical filament cross-section is established with the starting and ending coordinates of the time step as the geometric center of the filament cross-section, and the starting and ending cross-sections are connected to form the corresponding envelope range.

[0112] After determining the envelope range, the basic unit can be determined by the outline of the envelope range to determine whether it is within the envelope range. For basic units that intersect with the outline of the envelope range, they can be judged by preset rules. For example, if more than 50% of the volume of a basic unit is within the envelope range, it is determined to be a basic unit that needs to be activated. The preset rules can be customized, or they can be the rules preset in the processing device used to train the machine learning model.

[0113] In some implementations, the basic unit types include: a first type of basic unit containing a free surface and located on the surface of the model; a second type of basic unit containing a free surface and located inside the model; and a third type of basic unit not containing a free surface.

[0114] It should be understood that the type of basic unit needs to be determined according to certain classification criteria. In order to evaluate and associate the position of basic units and the relationship between basic units and the free surface of the model, this application defines a first type of basic unit as a basic unit that includes a free surface and is located on the model surface, i.e., the model surface is also the outer contour of the model; a second type of basic unit is defined as a basic unit that includes a free surface and is located inside the model, i.e., the second type of basic unit is located at the inner contour of the model; and a third type of basic unit is a basic unit that does not contain any free surface.

[0115] The outer and inner contours are determined by the geometry of the object's 3D model. The outer contour is part of the outer surface of the object's 3D model, and the inner contour is part of the internal structural surfaces of the object's 3D model, such as hole walls or groove walls. In some scenarios, the inner contour is formed by printing and filling. In the implementation, the type information of the basic unit of the object's 3D model can be determined by identifying the model's outer and inner contours.

[0116] In some implementations, the feature quantity formed by the type of the basic unit is a multi-dimensional vector formed by encoding the type of the basic unit using categorical variables.

[0117] The category variable is a variable that takes only a finite number of values. In the embodiments of this application, the type of the basic unit can be used as the category variable, so that each value describes the type attribute of a basic unit. For example, the basic unit type can be encoded using methods such as one-hot encoding, label encoding, feature hashing, target encoding, etc.

[0118] This application provides an embodiment for determining the type of basic units and aligning them using categorical variable encoding, so as to form the type features of the basic units selected in the design into numerical type features, making it easier for the machine learning model to obtain basic type information during training, increasing prediction accuracy and improving training efficiency.

[0119] In some implementations, the multidimensional vector is a multidimensional vector formed by one-hot encoding. Encoding the type information of the basic units into feature values ​​in Euclidean space facilitates the calculation of distances or similarities between features in algorithms for training machine learning models. In some scenarios, using one-hot encoding to form multidimensional vectors can meet the training needs of machine learning models that are sensitive to numerical values.

[0120] In some implementations, the multidimensional vector is a three-dimensional vector, the first type of basic unit is represented as (1,0,0), the second type of basic unit is represented as (0,1,0), and the third type of basic unit is (0,0,1).

[0121] In some implementations, the feature quantity formed by the distance of the basic unit to the free surface in the voxel feature is determined based on the distance of the basic unit to a first type of basic unit or a second type of basic unit within the printed layer.

[0122] The distance from the basic unit to the free surface can be used to describe the heat conduction and radiation of the basic unit at its location. The characteristic quantity of the distance from the basic unit to the free surface can be used to characterize the relationship between each voxel in the model and the model's geometry. Thus, the machine learning model obtained using the machine learning model training method of this application can correlate the voxel positions in the model, the relationship between voxels and the model's geometry, and other features with the temperature and / or stress history of the printing material during printing, thereby improving the accuracy of the prediction results.

[0123] Please see Figure 3The diagram shows a flowchart of the training method of this application for determining the distance of a basic unit to a first-class or second-class basic unit within its printed layer.

[0124] In step S211, a basic unit is determined as the target unit, and the first type of basic unit and the second type of basic unit in the printing layer are searched step by step within a preset range using the target unit as the search center. The step-by-step search method increases the search range step by step from the search center outward within the preset range.

[0125] It should be understood that the distance between each voxel in the three-dimensional object model and the first or second type of basic unit can be determined, and the target unit is the basic unit whose distance to the first or second type of basic unit needs to be determined. In the implementation, the determination methods for different basic units are similar; the following description uses the determination process for any one basic unit as an example.

[0126] A basic unit is identified as the target unit, and a search center is determined based on the target unit's location. Within a certain range from the search center and within a preset range, the system searches step-by-step within the printed layer containing the target unit to determine if it contains either a first-type or second-type basic unit. It should be noted that in actual processing, when determining the distance from each basic unit to the first-type or second-type basic unit in the discretized 3D model of the object, Figure 3 The illustrated process can simultaneously process multiple basic units, each with a different location and corresponding to multiple search centers and preset ranges defined by those search centers. Therefore, Figure 3 The embodiments shown are merely illustrative of the steps for determining the feature quantity of any basic unit, and do not limit the determination of the distance of the basic unit to the first or second type of basic unit within the printing layer to each basic unit in the three-dimensional model of the object.

[0127] The step-by-step search method involves progressively increasing the search range outward from the search center within a preset range. The preset range can also be set by the search center and the type of the basic unit. Since a basic unit is known to be an indivisible smallest unit, in all embodiments provided in this application, the preset range can be set to include an integer number of basic units. In one embodiment, when the basic unit is a hexahedron, the preset range is, for example, 5×5 basic units centered on the target unit, or 7×7 basic units centered on the target unit, etc.

[0128] As the search scope expands outward from the search center, the search scope for each level can be manually set. For example, in one example, the search might look for the existence of a first-type basic unit or basic unit within a 3×3 basic unit centered on the target unit. Then, in the next level, the search might look for the existence of a first-type basic unit or basic unit within a 5×5 basic unit centered on the target unit, and so on, increasing the search scope further in each level.

[0129] In step S212, the search stops after the first type of basic unit or the second type of basic unit is found, or after the first type of basic unit or the second type of basic unit is not found within the preset range.

[0130] When the search range is gradually increased from the search center outwards, the search can be stopped when the first type of basic unit or the second type of basic unit is found, or the search can be stopped when no basic unit is found within the preset range. That is, the preset range is the maximum search range.

[0131] For example, please refer to the following: Figure 4a and Figure 4b , Figure 4a and Figure 4b These are the basic units within the search range determined by different target units in the 3D model of the object. The search can be stopped when, in the first-level search, at least one of the first-type and second-type basic units is found among the 3×3 basic units centered on the target unit. For example... Figure 4a The target element D1 shown exists within the search range of the first-level search if it contains either a first-type or second-type basic element with a free surface (i.e., ...). Figure 4a The basic unit D shown f When the first-level search does not contain either the first or second type of basic unit, for example... Figure 4b If no type I or type II basic unit with a free surface exists among the basic units within the search range of the target unit D2 shown in the first-level search, then a second-level search is performed on 5×5 (or 4×4, 6×6, etc.) basic units centered on the target unit. The search can be stopped when at least one of type I or type II basic units exists in the basic units of the second-level search. If no target unit that is a type I or type II basic unit exists in each level of search, the search continues until the search range is the preset range centered on the target unit, and then the search can be stopped.

[0132] In step S213, the distance between the first type of basic unit or the second type of basic unit and the search center when the search stops is determined as the distance between the target unit and the first type of basic unit or the second type of basic unit within the printing layer.

[0133] When the search range is gradually increased from the search center outwards, the search stops when the first or second type of basic unit is found. That is, the first or second type of basic unit closest to the target unit is determined, and feature quantities are formed based on the first or second type of basic unit closest to the target unit.

[0134] In some scenarios, when no first-type basic unit or second-type basic unit exists within the preset range and the search stops, the distance from the target unit to the first-type basic unit or second-type basic unit within its printing layer is determined to be greater than the preset range.

[0135] It should be noted that in the examples provided in this application, the execution order of steps S211, S212, and S213 is not necessarily limited; the order here is merely an illustrative example for ease of understanding. For instance, steps S211, S212, and S213 are only used as conditions for determining distance and have no explicit order. For example, step S212 may be a search condition for step S211, and steps S212 and S213 may be implemented simultaneously.

[0136] In some implementations, the feature quantity formed by the distance from the basic unit to the free surface is a multidimensional vector formed by encoding the distance of the basic unit to the first type of basic unit or the second type of basic unit within the printed layer using categorical variables, wherein the vector dimension of the multidimensional vector is equal to the number of levels in the preset range.

[0137] By determining the distance classification of the search range at the end of the search as the category variable for the distance between the first or second type of basic unit and the target unit, a finite number of feature values ​​are formed for the distance between the first or second type of basic unit and the target unit. The dimension of the multidimensional vector is equal to the number of levels of the preset range classification. The number of levels of the preset range classification includes taking the area outside the preset range as a distance classification. For example, the preset range is defined as 5×5 basic units centered on the target unit within the same printing layer. The first level search within this preset range is 3×3 basic units centered on the target unit, the second level search is 5×5 basic units centered on the target unit, and the range outside the 5×5 basic units centered on the target unit is the third level distance. Correspondingly, in this example, the dimension of the multidimensional vector formed by encoding the distance category variable from the basic unit to the first or second type of basic unit within its printing layer is 3.

[0138] In one example, the categorical variable is encoded using one-hot encoding. In another example, target units that contain either a first-class basic unit or a second-class basic unit in the first-level search are marked as (1,0,0), target units that contain either a first-class basic unit or a second-class basic unit in the second-level search (the corresponding search range is the preset range) are marked as (0,1,0), and target units that do not contain either a first-class basic unit or a second-class basic unit in the preset range are marked as (0,0,1).

[0139] The feature quantity formed by the position coordinates of the basic unit in the voxel feature can be, for example, obtained by taking the coordinate value of the basic unit. In some implementations, the coordinate value of the basic unit can be represented by the coordinates of the geometric center of the basic unit, that is, the position coordinate information of each basic unit is indicated by the coordinates of its geometric center.

[0140] The geometric center of the basic unit is related to its geometric shape. In one example, the basic unit is a regular hexahedron, and its position coordinates are its centroid coordinates. The basic unit can be considered as being uniformly filled with printing material, and the centroid coordinates of the corresponding regular hexahedron basic unit are also the coordinates of its geometric center. In the printing environment (including actual printing environment and simulated printing environment) using a three-dimensional Cartesian coordinate system as defined in this application, the centroid coordinates can be represented as (x, y, z).

[0141] Please refer to the diagram for further information. Figure 5 and Figure 6 ,in, Figure 5 The diagram shows a flowchart of determining the density of the location of the basic unit in one embodiment. Figure 6 The diagram shown is a simplified schematic of some basic units in the three-dimensional model of the object.

[0142] In some implementations, the method of forming a characteristic quantity from the density at the location of the basic unit includes the following steps:

[0143] In step S221, a basic unit is determined as the target unit, and the number of non-empty units within N preset ranges centered on the target unit is determined; where N is a positive integer;

[0144] Here, the N preset ranges can be one or more preset ranges, and any two preset ranges do not completely overlap. In one implementation, the preset range is a preset range within the printing layer where the target unit is located. For example, for a 3D model of an object with a regular hexahedron as the basic unit, the multiple preset ranges may be, for example, 3×3, 5×5, or 7×7 basic units centered on the target unit in the printing layer where the target unit is located.

[0145] The number of non-empty units within N preset ranges centered on the target unit is determined. These non-empty units correspond to the solid parts of the printed product. In some embodiments, when the three-dimensional model of the object is discretized into multiple basic units by finite element analysis, the basic units identified as active in the simulation domain by identifying the model surface (e.g., outer and inner contours) are the non-empty units. Correspondingly, the basic units that do not need to be activated are empty units, such as pores in the internal filling structure of the printed object. It should be understood that any basic unit can be considered as being uniformly filled with printing material. When the printing material density is uniform, the number of non-empty units next to the target unit can be used to characterize the filling density at the location of the target unit.

[0146] In step S222, the number of non-empty units within each preset range and the total number of basic units within the preset range are determined to form a feature quantity with dimension N of the density at the location of the target unit.

[0147] Within N preset ranges, the number of non-empty cells in each range is determined, and the total number of basic unit boards corresponding to each range is also determined. The density feature number representing the location of the target cell is formed by the number of non-empty cells in each range and the total number of basic units within that range. Thus, a feature quantity of dimension N representing the location of the target cell can be formed using N preset ranges. The density feature number can be the ratio of the number of non-empty cells in the preset range to the total number of basic units within that range. For example, in one embodiment, for any target cell, the number of non-empty cells in three preset ranges centered on the target cell within the printed layer is determined as a1, b1, and c1, and the total number of basic units in the three preset ranges is A, B, and C, respectively. The density feature quantity formed by the location of the target cell can then be expressed as...

[0148] by Figure 6 The illustrated embodiment is an example. Figure 6 Displayed as 9 basic cells within the same printing layer, the basic cell E5 at the center is identified as the target cell. A preset range centered on the target cell is 3×3 basic cells as shown in the figure. Within this preset range, there are 4 non-empty cells (e.g., ...). Figure 6 The basic units E1, E3, E8, and E9 shown are used. When the target unit E5 is also a non-empty unit, the density of the target unit E5 within this preset range is 5 / 9.

[0149] In some implementations, the printing speed of the basic unit and / or the printing time of the layer in which the basic unit is located are calculated based on the G-code data of the three-dimensional model of the object to form the corresponding feature quantity.

[0150] The G-code data includes a series of spatial coordinates in temporal order, from which the printing speed at different basic units along the printing path can be calculated. In some embodiments, the G-code data also includes printhead movement speed information in the printing device; the printing speed of the basic unit can be obtained from the printhead movement speed information to form corresponding feature quantities. In some practical scenarios, the printing speed of each basic unit can be obtained by discretizing the G-code data of the object's 3D model using a finite element mesh, and the numerical value of the printing speed can be used as a feature quantity of the basic unit's printing speed.

[0151] The method for obtaining the printing time of the printing layer where the basic unit is located can refer to the single-layer printing time included in the G-code data in the aforementioned embodiment, and will not be repeated here.

[0152] In the training method of the machine learning model provided in this application, the three-dimensional model of the object is discretized into multiple basic units, and a method for defining and acquiring multiple feature quantities in the voxel features is provided. Based on the voxel features and model geometric features provided in this application as training data for the machine learning model, the machine learning model can learn the relationship between the basic units of the three-dimensional model of the object, as well as the relationship between the basic units and the model geometry, to predict the temperature history and / or stress history in the object forming (i.e., printing process). With the feature description method provided in this application (i.e. how to select and acquire printing features), the machine learning model trained can make printing predictions for complex geometric structures based on the voxel features of its basic units and the model geometric features, that is, it has the applicability to predict various printed components.

[0153] In some embodiments, the printing feature further includes an interlayer extension feature, which is the voxel feature of the basic unit at the same projection point in adjacent layers of the basic unit; for the basic unit in the Nth layer, its interlayer extension feature is the voxel feature of the basic unit at the same projection point in layers NK to N-1 and layers N+1 to N+K, where N is a positive integer and K is a positive integer less than N.

[0154] It should be understood that, generally speaking, the printing process is a process of stacking and accumulating printing material layer by layer. In the coordinate system defined in this application, the accumulation direction of the printing layers (the direction of the printing layer height) is the z-direction. After discretizing the three-dimensional model of the object into multiple basic units, it can also be regarded as the three-dimensional model of the object being formed by stacking multiple layers of basic units along the z-direction. For any basic unit, its corresponding interlayer expansion feature is the voxel feature of the basic unit at the same projection point in the adjacent layers of that basic unit. The projection direction mentioned here is the z-direction, and the basic units at the same projection point are the basic units with the same position in the XOY plane (O is the origin of the coordinate system). The adjacent layers can be customized. For example, for a basic unit in the Nth layer, its adjacent layers are layers NK to N-1 and layers N+1 to N+K. The sequence number of the layer where the basic unit is located can be determined by its coordinate size. For example, the printing layer where the basic unit is located, which is close to the component plate, i.e. the lower boundary of the model, is defined as the 1st layer. The 2nd, 3rd, 4th... Nth layers of basic units are defined in accordance with the cumulative height of the printing, up to the upper boundary of the model.

[0155] Please see Figure 7 The diagram shown is a simplified schematic of the basic unit in the three-dimensional model of the object in one embodiment of the training method of the machine learning model of this application.

[0156] As shown in the figure, a basic unit is defined as the target unit (e.g., Figure 7 The basic unit E shown N ), defining the adjacent layers of the target unit as the two layers below the target unit (e.g. Figure 7 The basic unit E shown N-1 Basic Unit E N-2 ) and the two layers above (such as Figure 7 The basic unit E shown N+1 Basic Unit E N+2 The voxel features of the basic units in the adjacent layers that share the same projection point as the target unit are used as the interlayer extension features of the target unit; wherein, "above" and "below" refer to the Z-direction. Of course, the adjacent layers can also be three layers above and three layers below the target unit, one layer above and one layer below, etc. Figure 7 This is just one example.

[0157] In some implementations, the method of obtaining the interlayer augmentation feature includes at least one of the following:

[0158] In one scenario, when a basic unit located inside the model in an adjacent layer is identified as the target unit, the voxel features of the basic units at the same projection point within the adjacent layers are added to the print features of the target unit. That is, when the adjacent layers of the target unit are located inside the model, or when the basic units at the same projection position within the adjacent layers of the target unit are all located inside the model (i.e., non-empty units), the voxel features of each basic unit are added to the print features of the current target unit.

[0159] In another scenario, when a basic unit in an adjacent layer located outside the model boundary is identified as the target unit, the voxel features of the basic units at the same projection point in the adjacent layers outside the model boundary are supplemented based on a preset constant value, and the supplemented voxel features are added to the print features of the target unit. This applies when at least one of the adjacent layers of the target unit is located outside the model, for example... Figure 7 In the illustrated embodiment, the first layer above the target unit, i.e., the N+1th layer, is located at the model boundary, and the N+2th layer among the adjacent layers of the target unit is located outside the boundary. In this example, for basic units at the same projection point in adjacent layers located inside the model, their voxel features are added to the printing features of the target unit. For basic units at the same projection point in adjacent layers located outside the model boundary, their voxel features are supplemented based on preset constant values ​​and then added to the printing features of the target unit.

[0160] The preset constant values ​​can be determined based on the definition of each feature quantity in the voxel features. For example, the feature quantity corresponding to the printing speed of the basic unit located outside the model boundary is a constant value of 0; the position coordinates of the basic unit can be taken as the geometric center coordinates of the basic unit in the coordinate system determined by the printing system; the printing time of the layer where the basic unit is located is a constant value of 0; and the density of the position where the basic unit is located is a constant value of 0. That is, based on the description method of the voxel features determined for non-empty units, the basic units outside the model boundary are formed into feature quantities in the same description method to supplement the interlayer expansion features of the corresponding target units.

[0161] In another scenario, when a basic unit with an empty cell at the same projection point in adjacent layers is identified as the target unit, the voxel features of the empty cells at the same projection point in adjacent layers are supplemented based on a preset constant value, and the supplemented voxel features are added to the printing features of the target unit. In this scenario, there are basic units within the model boundary that are empty cells at the same projection point in adjacent layers, such as basic units at the positions corresponding to the pores inside the model of an object filled by internal printing. Similarly, the feature quantity corresponding to the basic unit is determined based on the description method of the voxel features determined for non-empty cells; for example, if there are non-empty cells in the printing layer where the empty cell is located, the feature quantity formed by the density of its location is determined by the number of non-empty cells within N preset ranges centered on the basic unit and the total number of basic units within the N preset ranges. The corresponding feature quantity is formed based on the printing time of the layer where the empty cell is located, and the printing speed of the empty cell is supplemented with a constant value of 0, etc.

[0162] like Figure 7 In the embodiment shown, the basic unit E N The target element is identified as the basic element E, which is a non-empty element at the same projection position within its adjacent layer. N-1 Basic Unit E N-2 and basic unit E N+1 The voxel features of non-empty units are added to the printing features of the target unit; simultaneously, for the basic unit E of empty units in its adjacent layers... N+2 The feature quantities corresponding to the empty cell are determined based on the description method of the voxel features determined by the non-empty cell, for example, by supplementing the basic cell E with constant values. N+2 The voxel features and the printing features of adding the supplemented voxel features to the target cell.

[0163] In the aforementioned examples provided in this application, the information described by the voxel features includes the association between basic units and other basic units in the printing layer they are in, such as the distance of the basic unit to the first or second type of basic unit in the printing layer, the density of the basic unit's location, etc. The interlayer expansion features are the voxel features of basic units at the same projection point in adjacent layers of the basic unit. Thus, the interlayer expansion features of basic units can also describe the association between different printing layers. Predicting the temperature history and stress history of the printing process while considering interlayer association is beneficial to improving the accuracy of the prediction.

[0164] In some embodiments, the training method of the machine learning model described in this application further includes a step of distributing at least one feature among the printed features, and forming the input data of the machine learning model based on the feature that satisfies a Gaussian distribution after distribution processing.

[0165] By statistically distributing at least one feature in the input data of the training method—for example, by exponentiation of features that are random variables—to ensure that the training data received by the machine learning model conforms to a Gaussian distribution, it is beneficial to improve the speed and accuracy of model generation during the training process. For instance, in the voxel features described in the previous example, the basic unit printing speed, the basic unit position coordinates, and the density of the basic unit positions are the features that need to be distributed, thereby obtaining training data received by the machine learning model that conforms to a Gaussian distribution.

[0166] Please continue reading. Figure 1 In step S10, it is also necessary to obtain the temperature history and / or stress history data of multiple sets of object 3D models corresponding to the printing process over time.

[0167] The multiple sets of temperature history and / or stress history data of the three-dimensional models of the object are obtained in different finite element simulation environments, wherein the different finite element simulation environments are formed by setting different printing parameter information for the three-dimensional models of the object; or, the multiple sets of temperature history and / or stress history data of the three-dimensional models of the object are obtained in different printing experimental environments, wherein the different printing experimental environments are formed by setting different printing parameter information for the three-dimensional models of the object.

[0168] The printing parameter information described in this application includes the layered (sliced) graphics of the three-dimensional object model, as well as printing material property information and printing equipment parameter information. Different printing parameter information may involve different values ​​for at least one material parameter or equipment parameter, or different layered (sliced) graphics of the three-dimensional object model.

[0169] It should be understood that the temperature history and / or stress history data of the multiple sets of object 3D models correspond to the printing features of the multiple sets of object 3D models. After setting certain printing parameter information for an object 3D model, the printing features under this printing information are obtained, and the corresponding temperature history and / or stress history data are obtained by simulating printing or actual printing under this printing parameter information.

[0170] In some embodiments, the step of obtaining the temperature history and / or stress history data of the multiple sets of three-dimensional models of the object in different finite element simulation environments includes: performing coupled simulation calculations on the three-dimensional models of the object to obtain the temperature history and / or stress history data, wherein the coupled simulation calculation model includes a linear viscoelastic model describing the mechanical deformation of the printing material, and / or a transversely isotropic thermal conduction model or an orthotropic thermal conduction model used to describe the thermal conduction behavior of the printing material.

[0171] In one example, the linear elastoviscous model is a multi-branch thermoviscoelastic model, which considers the temperature-dependent relaxation behavior and flow shear phenomena of the printing material. The mechanical behavior of the material is temperature-dependent, and the total strain during printing consists of thermal strain and elastic strain. The thermo-mechanical-chemical coupling model describes the material's changes over time. The simulation results simultaneously consider material properties, temperature and mechanical state, as well as the interaction between temperature and stress-strain. The resulting deformation data more closely matches the actual printing conditions, reducing simulation errors.

[0172] In some implementations, during the simulation calculation, one of the following can be selected to output: a stress cloud map corresponding to the stress field, a displacement cloud map of the deformation, or a temperature cloud map corresponding to the temperature field. It should be understood that during the calculation, the temperature field and the stress field are fully coupled, and the transient temperature field and stress field can be obtained simultaneously in the calculation at each moment. During the simulated printing process, dynamic temperature history data and stress history data that change over time can be obtained.

[0173] Here, in step S10, the stress history data and temperature history data of the printed component can be obtained by outputting the calculation results of the simulated printing through finite element simulation, or by observing the actual printing process. Typically, the stress history data and temperature history data in the simulated printing calculation results are the stress history data and temperature history data at various spatial coordinates within the printed component. In some instances, the density of these spatial coordinate points is determined based on the mesh density set in the finite element environment. For example, the higher the mesh density and the smaller each mesh, the greater the data density corresponding to the calculated stress history data and temperature history data.

[0174] In some embodiments, the step of obtaining the temperature history and / or stress history data of the multiple sets of three-dimensional models of the object includes: determining the time starting point of the temperature history and / or stress history of the basic unit based on the G-code data of the three-dimensional model of the object, and obtaining the temperature and / or stress data of the basic unit from the time starting point at a preset sampling time interval.

[0175] During the printing process, the total time experienced by the printing material at different locations in the 3D model of the object varies. Therefore, this application also provides a method for unifying the format of the temperature and stress history data corresponding to the printing material at different locations in the 3D model of the object. For any printing material at any location in the 3D model of the object, or for any basic unit, starting from the time the material is extruded from the print head, temperature and / or stress data are sampled at intervals within a preset time length after extrusion. The preset time length should be selected as the duration for the printed object to recover to a stable state, for example, the duration for the overall temperature of the printed object to stabilize to room temperature. For example, in one scenario, sampling is performed at a preset interval of 1 second, and temperature and stress data for each basic unit are collected for a total duration of 500 seconds starting from its material extrusion time (i.e., activation time).

[0176] In an embodiment that uses finite element simulation to obtain temperature history and / or stress history data, the simulated printing process is calculated according to a preset time interval, and temperature and stress data at preset time sampling points are obtained.

[0177] In an embodiment where the temperature history of the printing process is obtained through printing experiments, the temperature distribution inside the printing cavity that changes over time during the printing process can be measured.

[0178] In one example, during the printing process, temperature field distribution data of the printing cavity inside the printing chamber, which changes over time, is detected using a temperature measuring device such as an infrared thermal imager. In this implementation, the infrared thermal imager is set up at the same height as the printed component during the curing process. It records the history of temperature changes on the projection surface of the material as the height of the printed component increases during the layer-by-layer accumulation of printing material; this is the temperature field distribution data. The recorded temperature change history can be in the form of visual images (photographs) converted from temperature distribution at different printing moments, or temperature field videos, dynamic graphs, etc. In another example, when the printed structure is non-axisymmetric, infrared thermal imagers can be set up in different directions of the printed component, or multiple printing operations can be performed by changing the relative orientation of the printed component and the thermal imager, such as rotating the printed component by a certain angle and repeating this process, thereby obtaining the temperature change history of different areas of the printed component.

[0179] In some examples, such as when an FDM device is actually printing, the infrared thermal imager is usually placed outside the printing cavity. In order to reduce the attenuation of infrared radiation collected by the infrared thermal imager by the cavity structure, such as a transparent glass plate, the temperature field can also be corrected. This correction process can be based on temperature field distribution data obtained from multiple printing experiments, and the attenuation value caused by the cavity structure can be calculated through data processing.

[0180] Here, the device used to record the temperature evolution of the printing material during actual printing can also be a thermocouple, a radiation thermometer, an electronic temperature sensor, or other instruments or devices that can be used for temperature detection; this application does not impose any limitations. It should be understood that the relative position of the temperature detector and the printed component is determined based on the different types of temperature detectors used, such as contact or non-contact detectors, to obtain the temperature change history of the printed component from the start of printing to its cooling and forming. For example, when the temperature detector is a thermocouple, the thermocouple can be arranged at different key locations in the cavity to record the temperature changes during printing.

[0181] In some implementations, the training method for the machine learning model further includes a step of dimensionality reduction of the historical temperature data, so that the dimensionality-reduced historical temperature data is used as the output data of the machine learning model. By reducing the dimensionality of the historical temperature data, the resulting output data can reduce the computational load of training the machine learning model.

[0182] The methods for dimensionality reduction processing of historical temperature data include, but are not limited to, Linear Decision Analysis (LDA), LASSO (least absolute shrinkage and selection operator), Laplacian eigenmaps, and Local Linear Embedding.

[0183] In some embodiments, dimensionality reduction processing of the temperature history data includes at least one of the following: converting the temperature history data obtained from finite element simulation or printing experiments into equivalent relaxation time data; performing principal component analysis processing on the temperature history data obtained from finite element simulation or printing experiments to achieve dimensionality reduction by retaining the principal components of the temperature history data.

[0184] Relaxation time refers to the time required for an object to return to its normal state after being deformed by force and the external force is removed. In the embodiments provided in this application, the equivalent relaxation time data can be used to characterize the stress relaxation experienced by each spatial coordinate point or local area in the printed component from the start of printing to stable formation. Here, the printing material is usually a polymer material. The mechanical relaxation behavior of polymer materials is the sum of relaxation behaviors throughout their entire history. The mechanical relaxation behavior is related to the material temperature. For example, the sum of mechanical relaxation behaviors generated by polymer materials at high temperature for a short time can be equivalent to the sum of mechanical relaxation behaviors generated at low temperature for a long time. The equivalent relaxation time data is the sum of relaxation times at the same preset temperature value, which is the equivalent of the temperature history experienced by each spatial coordinate point or local area in the printed component during printing. It should be understood that when multiple sets of equivalent relaxation time data are obtained using the same preset temperature value, the mechanical relaxation behavior of materials corresponding to multiple printing experiments or simulated printing can be compared based on the values ​​of the multiple sets of equivalent relaxation time data; or, for a set of equivalent relaxation time data, the mechanical relaxation behavior of materials experienced by different spatial coordinate points or areas in the model can be compared based on its data distribution.

[0185] In some implementations, the equivalent relaxation time data is obtained by equating the dynamic temperature field data of the printing material over time to the same preset temperature value based on the WLF equation (Williams-Landel-Ferry equation) and / or the Arrhenius equation during the actual printing process or the simulated printing process.

[0186] Based on the relationship between the relaxation behavior of polymer materials and temperature, the relaxation behavior at low temperature for a long time can be equivalent to the relaxation behavior at high temperature for a short time, as expressed by the following equation (1):

[0187]

[0188] Where τ(T1) and τ(T2) are the characteristic relaxation times of the material at different temperatures T1 and T2, respectively. The characteristic relaxation time can represent the stress relaxation ability of the material. a(T1) and a(T2) are temperature-related conversion factors (also called shift factors), that is, the conversion factor a is a function of temperature. Here, when the temperature in the material is higher than the glass transition temperature, the conversion factor a follows the WLF equation, as shown in the following equation (2):

[0189]

[0190] Where T is the actual temperature during printing, T M For the reference temperature, C1 and C2 are empirical parameters, derived from the reference temperature T. MThe value of is determined here, for any spatial coordinate or local area of ​​the printed component, if the actual temperature T is higher than the glass transition temperature of the printing material, a preset reference temperature T is determined. M Then, from equation (2), we can obtain the current actual temperature corresponding to the set reference temperature T. M The conversion factor 'a', for the mechanical relaxation experienced by the material at the current temperature T over time t, can be equivalent to the mechanical relaxation experienced by the material at the reference temperature T. M Mechanical relaxation experienced within the next time period t / a.

[0191] Furthermore, when the material temperature in the printed component is lower than the glass transition temperature, the conversion factor a follows the Arrhenius equation in the low-temperature state, as shown in equation (3) below:

[0192]

[0193] Where T is the actual temperature during printing, T g Here is the reference temperature, A is the material constant, and F is... C k is the configurational energy. B Boltzmann's constant is typically considered a constant when the temperature variation is small. For any spatial coordinate or local region within the printed component, if the actual temperature T is lower than the glass transition temperature of the printing material, a preset reference temperature T is determined. g Then, from equation (3), the current actual temperature corresponding to the set reference temperature T can be obtained. g The conversion factor 'a', for the mechanical relaxation experienced by the material at the current temperature T over time t, can be equivalent to the mechanical relaxation experienced by the material at the reference temperature T. g Mechanical relaxation experienced within the next time period t / a.

[0194] It should be understood that during actual or simulated printing, the temperature field inside the printed component changes with time. At the same time, the temperature history at different locations or spatial coordinate points in the printed component is different. Here, the printing process is divided into multiple small time intervals dt. Here, the temperature of the printing material within dt can be considered as a constant value. Then, the stress relaxation behavior within dt can be transformed into the stress relaxation behavior within dt / a time at the reference temperature by equations (2) and (3). The relaxation behavior experienced by the printing material at different locations during the entire printing process, i.e., from the start of printing, for example, time 0, to the time of cooling and forming, for example, time t1, can be integrated as shown in the following equation (4):

[0195]

[0196] Where (x, y, z) are spatial coordinates, and t r(x,y,z) represents the equivalent relaxation time of this point during the printing process from time 0 to time t1.

[0197] For example, the temperature data for each voxel's spatial coordinates consists of multiple temperature data points (e.g., 500) determined by a preset acquisition duration (e.g., 500 seconds) and data acquisition interval (e.g., 1 second). In this case, the temperature data can be considered to have a dimension corresponding to the number of temperature data points (e.g., 500 dimensions). After performing dimensionality reduction processing on the historical temperature data using equivalent relaxation time, the temperature data at each voxel location can be converted into relaxation time data at the same temperature value. This means that the multidimensional temperature data is equivalent to a single time parameter, which can be considered as converting the historical temperature data into a scalar value.

[0198] In some implementations, principal component analysis (PCA) is performed on the temperature history data obtained from finite element simulation or printed experiments to achieve dimensionality reduction by retaining the principal components of the temperature history data.

[0199] In one example, PCA can be performed on n-dimensional temperature data to identify and retain m principal components, thereby reducing the dimensionality of the temperature data from n-dimensional to m-dimensional (n > m). The n-dimensional temperature data can be, for example, temperature data obtained from n data samplings during simulated printing or actual printing.

[0200] In the embodiments of dimensionality reduction of historical temperature data provided in this application, the resulting training data can reduce the feature dimension and reduce the risk of overfitting. In some scenarios, data dimensionality reduction can effectively reduce the computational load of the training process and improve the training efficiency of machine learning models.

[0201] Regarding the input data, in some embodiments, the multiple sets of temperature history and / or stress history data of the three-dimensional model of the object are the temperature history and / or stress history data at the position coordinates of each basic unit in the three-dimensional model of the object.

[0202] In some implementations, the temperature and stress data can be in a data array format, where the data array includes a multidimensional matrix, such as a two-dimensional matrix or a three-dimensional matrix. In one example, each value in the data array can be the average of the stress or temperature data at the coordinates corresponding to multiple basic units of the object's three-dimensional model. In another example, the temperature and / or stress data of each basic unit can be taken from the coordinates corresponding to the geometric center of the basic unit, and the temperature and / or stress data at multiple coordinate points form the input data for the machine learning model.

[0203] In some embodiments, the multiple sets of temperature history and / or stress history data of the object's three-dimensional model are the average stress data and / or temperature data of each sub-region in the object's three-dimensional model; wherein, the sub-region is obtained by dividing the object's three-dimensional model based on a preset size or a preset number of partitions.

[0204] In some embodiments, the simulation domain of the three-dimensional model of the object in the finite element simulation environment is divided into multiple sub-regions according to a preset size or a preset number of partitions. The multiple sets of residual stress data and / or multiple sets of equivalent relaxation time data are the average residual stress and / or the average equivalent relaxation time in each sub-region of the simulation domain. The sub-regions can be obtained by dividing the simulation domain based on a preset sub-region size, or based on a preset number of divisions of the simulation domain. For example, the simulation domain can be divided into equal parts in the length direction l, the width direction m, and the height direction n, where l, m, and n are all natural numbers greater than or equal to 1. Alternatively, within a defined simulation domain, the preset sub-region size can be set to a×b×c, and the simulation domain can be divided into multiple sub-regions of size a×b×c.

[0205] Here, the stress and temperature data at the coordinates of each basic unit in each sub-region can be converted into equivalent values ​​within the sub-region, such as the average value, root mean square, and median of stress or temperature. This simplifies the input data used for training, reducing the amount of stress and temperature data from grid density to the number of sub-regions.

[0206] In some examples, the stress and / or temperature data of the multiple sets of object 3D models are the average values ​​of stress and / or temperature data in each sub-region of the object 3D model; wherein, the sub-region is obtained by dividing the object 3D model based on a preset size or a preset number of partitions. For example, when the stress and / or temperature data is obtained by measuring the printing process in an actual printing environment, the sub-region can be obtained by dividing the 3D contour of the printed component, for example, dividing the printed component model into multiple sub-regions based on a preset number, such as dividing the printed component model into l×m×n sub-regions according to the length, width, and height of l, m, and n respectively, or dividing the printed component model into multiple sub-regions according to the size of the set sub-regions; correspondingly, the stress and temperature data at different locations in each sub-region are converted into equivalent values ​​representing the whole sub-region, such as the average value, root mean square, median, etc., to reduce the amount of stress and temperature data in the training data, thereby reducing the computational load of training the machine learning model.

[0207] In some embodiments, the temperature history and / or stress history data of the multiple sets of three-dimensional models of the object are processed to be represented by equivalent values ​​of sub-regions, and the resulting data is three-dimensional data (three-dimensional array). Here, the three-dimensional data (three-dimensional array) can also be transformed into two-dimensional data (two-dimensional array), and the two-dimensional data (two-dimensional array) can be used as training data to train a machine learning model.

[0208] In step S11, the printing features of the multiple sets of object 3D models are used as input data and the temperature history and / or stress history data of the multiple sets of object 3D models are used as output data. Supervised learning is performed based on the input data and output data to obtain the machine learning model.

[0209] It should be understood that, in the embodiments provided in this application, using the temperature history and / or stress history data of the three-dimensional model of the object as output data means using the temperature history and / or stress history data as the target output data. In other words, the output data in the training data is the desired output of the target machine learning model to be obtained.

[0210] This application provides various implementation methods for obtaining input and output data. It should be understood that the input data, the feature values ​​of the printing features, and the corresponding data format described in step S10 determine the input data required for the trained machine learning model to predict the printing process. Similarly, the specific form of the output data described in step S10 determines the specific form of the temperature history and / or stress history data of the printing material over time predicted by the trained machine learning model. For example, in an embodiment where the average stress data and average temperature data of a sub-region of the object's 3D model are used as the output data for training, the trained machine learning model can predict the stress history data of each sub-region of the object's 3D model. And historical temperature data; for example, when only historical temperature data is used as input data in the training data, the corresponding machine learning model will only output the historical temperature data of the printing material in the printing prediction of the object's 3D model; or, the training data of the machine learning model includes the geometric features of the model and the types of basic units in the object's 3D model, the distance from the basic units to the free surface, and the voxel features formed by the position coordinates of the basic units. The target machine learning model obtained through training obtains the types of basic units, the distance from the basic units to the free surface, and the voxel features and geometric features formed by the position coordinates of the basic units based on the discrete object's 3D model in actual use, so as to predict the printing process of the object's 3D model.

[0211] In other words, different input or output data used as training data for the machine learning model in step S10 result in different target machine learning models. In real-world scenarios, the input data and predicted output data required by the target machine learning model are different.

[0212] The machine learning model is an algorithm model built based on supervised machine learning. It is trained on printing features and temperature history data and / or stress history data with corresponding relationships to generate a target machine learning model that can predict the temperature history and / or stress history data of printing material changes based on the printing features of the object's 3D model. The functions in the algorithm model include, for example, SVM (support vector machines), AdaBoost (adaptive boosting), linear regression algorithm, linear discriminant analysis (LDA), decision tree algorithm (e.g., classification and regression tree), random forest algorithm, Naive Bayes, K-nearest neighbor algorithm, learning vector quantization (LVQ), etc.

[0213] In some implementations, the machine learning model is a neural network model. For example, the preset neural network structure is a feedforward neural network, a feedback neural network, a deep neural network, a convolutional neural network, a self-organizing neural network, etc. The preset neural network is trained in supervised learning based on the input data and output data, so that the preset neural network structure is adjusted under the corresponding algorithm during the training process, thereby forming the expected neural network.

[0214] In some examples, step S11 further includes a step of evaluating the error based on the predicted data and the output data to adjust the machine learning model; wherein the predicted data is the performance evaluation result of the neural network making predictions based on the input data. For example, the preset model of the machine learning model is a neural network structure using a supervised learning algorithm, such as a BP neural network. During training, a sample data, i.e., a set of training data, is selected from the training dataset. The preset neural network calculates the corresponding temperature history and / or stress history data of the printing material changing over time during the printing process based on the input data in the training data, i.e., the printing features of the object's 3D model. The temperature history and / or stress history data predicted by the neural network is compared with the output data in the sample data, i.e., the temperature history and / or stress history data in the simulated printing or actual printing environment, to obtain a deviation value. The connection weights between different neurons in the neural network are adjusted according to the deviation value. This process is repeated until the deviation value meets the specified error range.

[0215] In one example, the preset model of the machine learning model is, for example, a deep neural network. A corresponding loss function, such as the mean squared error loss function, can be set to measure the output loss of the training data to determine the accuracy of the prediction results. This loss function represents the degree of inconsistency between the machine learning model's prediction results and the actual output data; it is a non-negative real-valued function. The smaller the loss function, the better the robustness of the machine learning model. Types of loss functions include logistic loss, squared loss, and exponential loss. During training, the neural network iteratively adjusts the parameters of the prediction algorithm—that is, the connection weights of neurons in different layers—based on multiple sets of training data until the loss function reaches its optimal extreme value. This means that the loss between the prediction results obtained by the neural network based on the input data in the training data and the output data stably reaches a set threshold, thus obtaining the target neural network.

[0216] Of course, the models and functions used in the training process in the above embodiments are only illustrative examples. Those skilled in the art can also use other preset machine learning models or preset algorithms to train after obtaining the temperature history and / or stress history data of the printing features and the three-dimensional model of the object provided in this application, so as to obtain a target machine learning model that can predict the thermal history and / or stress history of the printing process. This application does not limit this.

[0217] The first aspect of this application provides a training method for a machine learning model to predict the temperature history and / or stress history of printing materials over time in 3D printing. The resulting target machine learning model can quickly predict the temperature history and / or stress history of the printing material during the printing process after obtaining the printing features of the object's 3D model. This solves the problems of low computational efficiency, long computation time, and high computational resource consumption in actual printing or finite element simulation printing. The target machine learning model obtained by the training method of this application can also be deployed to user-operable applications, which is beneficial for practical production applications. Furthermore, the training method of the machine learning model of this application provides the design (determination) and acquisition (generation) of training data. According to the embodiment of the input and output data, the input data, namely the printing features of the three-dimensional model of the object, includes voxel features and model geometric features. By obtaining the characteristics of each basic unit of the three-dimensional model of the object and the relationship between the basic units, the trained target machine learning model can calculate and predict the temperature history and / or stress history data that changes over time during the printing of complex structures based on the voxel features and model geometric features obtained by discretizing the complex structure when predicting the printing process of different printed components, including complex structures. Therefore, the machine learning model obtained by the training method of the machine learning model of this application is not limited to the printing prediction of simple geometries such as infilled cubes, but can also be adapted to the printing prediction of various printed components.

[0218] In a second aspect, this application also provides a training apparatus for a machine learning model; please refer to [link / reference needed]. Figure 8 The diagram shown is a simplified schematic of a training apparatus for the machine learning model of this application in one embodiment.

[0219] In some embodiments, the modules in the training device for the machine learning model can be software modules, which can also be configured in a programming language-based software system. The software modules can be provided by the system of an electronic device. In some embodiments, the electronic device is, for example, an electronic device loaded with an APP application or capable of accessing a webpage / website. The electronic device includes components such as memory, memory controller, one or more processing units (CPUs), peripheral interfaces, RF circuitry, audio circuitry, speakers, microphones, input / output (I / O) subsystems, displays, other output or control devices, and external ports. These components communicate through one or more communication buses or signal lines. The electronic device includes, but is not limited to, personal computers such as desktop computers, laptops, tablets, smartphones, and smart TVs. The electronic device can also be an electronic device consisting of a host with multiple virtual machines and corresponding human-computer interaction devices (such as touch screens, keyboards, and mice) for each virtual machine.

[0220] In some implementations, the functional modules of the training device for acquiring training data, training based on the training data (including generating intermediate results), and obtaining the target machine learning model can be implemented collaboratively by various types of devices (such as terminal devices, servers, server clusters, or cloud server systems), or by computing resources such as processors and communication resources (such as those used to support various communication methods such as optical cables and cellular networks). The training data, the machine learning model obtained by the training device, and any of the prediction results can be stored in an electronic device that configures the training device, and can also be transmitted to other terminal devices, servers, server clusters, cloud server systems, etc. that communicate with the electronic device via a network.

[0221] In some embodiments of this application, the cloud server system can be deployed on one or more physical servers based on factors such as functionality and load. When distributed across multiple physical servers, the server can consist of servers based on a cloud architecture. For example, cloud-based servers include public cloud servers and private cloud servers, where public or private cloud servers include Software-as-a-Service (SaaS), Platform-as-a-Service (PaaS), and Infrastructure-as-a-Service (IaaS), etc. Examples of private cloud servers include Alibaba Cloud Computing Service Platform, Amazon Cloud Computing Service Platform, Baidu Cloud Computing Platform, Tencent Cloud Computing Platform, etc. The server can also consist of distributed or centralized server clusters. For example, the server cluster consists of at least one physical server. Each physical server is configured with multiple virtual servers, and each virtual server runs at least one functional module of the catering merchant information management server. The virtual servers communicate with each other via a network.

[0222] The network may be the Internet, mobile network, local area network (LAN), wide area network (WLAN), storage area network (SAN), or one or more intranets, or appropriate combinations thereof. This application does not limit the types of clients or servers, or the types or protocols of communication networks between publisher terminals and servers, or between responder terminals and servers.

[0223] The machine learning model is used to predict the temperature history and / or stress history data of the printing material over time in 3D printing, such as... Figure 8As shown, the training device includes: a training sample acquisition module 31, used to acquire printing features of multiple sets of object 3D models, and to acquire temperature history and / or stress history data that change over time during printing corresponding to multiple sets of object 3D models; wherein, the printing features include model geometric features and voxel features; wherein, the voxel features include one or more of the following: type of basic unit in the object 3D model, distance from basic unit to free surface, printing speed of basic unit, printing time of layer where basic unit is located, position coordinates of basic unit, and density of position of basic unit; and a training module 32, used to take the printing features of the multiple sets of object 3D models as input data and the temperature history and / or stress history data of the multiple sets of object 3D models as output data, and to obtain the machine learning model through supervised learning based on the input data and output data.

[0224] In some embodiments, the training sample acquisition module 31 is used to receive printing features of multiple sets of object 3D models that can be directly used as training data for machine learning models, and temperature history and / or stress history data that change over time during printing of multiple sets of object 3D models. For example, the printing features received by the sample training acquisition module 31 as input data are preset feature quantities in preset representation formats such as vectors, matrices, numbers, coordinates, etc.

[0225] In other embodiments, the training sample acquisition module 31 can also be used to perform a process of extracting printing features based on the object's 3D model to generate training samples. For example, the training sample acquisition module 31 can acquire the G-Code data (or layered graphics) and printing parameter information of the object's 3D model, convert the G-Code data (or layered graphics) into discrete units to calculate the feature quantities of each basic unit to form voxel features, and calculate the model's geometric features based on the printing parameter information and G-Code data. In some scenarios, the training sample acquisition module 31 can also calculate the feature quantities formed by material property information and printing equipment information, thereby forming preset printing features as input data for the machine learning model.

[0226] The training sample acquisition module 31 obtains the training data based on the printing features and the temperature history and / or stress history data over time during the printing of the 3D model of the object that corresponds to the printing features. The correspondence refers to the fact that the printing features, temperature history, and / or stress history data in each set of training samples are obtained based on the same layered graphics and printing parameter information. In some embodiments, the training sample acquisition module 31 is also used to label the data of each set of training samples, that is, to assign target attributes to the temperature history and / or stress history data that correspond to the printing features. The training sample acquisition module 31 obtains multiple sets of training samples by acquiring multiple different sets of layered graphics or printing parameter information to expand the training dataset.

[0227] Here, the method and data format for determining the printing features, temperature history and / or stress history data acquired by the training sample acquisition module 31 can refer to the implementation method provided in the first aspect of this application, and will not be repeated here.

[0228] After forming multiple sets of training samples, the training module 32 uses the printed feature set as input data and the temperature history and / or stress history data of the three-dimensional model of the object as output data for supervised learning to obtain the machine learning model.

[0229] The training process executed by the training module 32 and the target machine learning model obtained by training can be referred to the implementation method provided in the first aspect of this application, and will not be repeated here.

[0230] Regarding the presentation of the training device for the machine learning model in practical application scenarios, this application also provides the following exemplary description:

[0231] In scenario A, each module of the training device can be embedded in an electronic device's application (APP). The APP can obtain the training data from the electronic device's storage medium or from other devices or servers communicating with the electronic device via a network to complete the training process of the target machine learning model. During the training process, in some examples, the training device for the machine learning model can be implemented using computing resources provided by the electronic device used to configure the training device. In other examples, the computing resources required to execute the training process can also be allocated to terminal devices, servers, cloud server systems, or processors communicating with the device via a network. Simultaneously, the prediction results generated by the training module during the training process can be stored locally on the device, transmitted to terminal devices, servers, cloud server systems, or processors communicating with the device via a network, or provided to other applications or modules.

[0232] In scenario B, the training device is a software module running on a server. The server can also be a distributed, parallel computing platform composed of multiple servers. In this scenario, the required training data can be uploaded to the platform to execute the training process. The server can perform the training process of a machine learning model based on data stored in its storage medium or data from other devices communicating with it. The temperature history and / or stress history data obtained by the machine learning model trained to predict the printing of a 3D object model can be stored on the server and can also be provided to other applications or modules.

[0233] In scenario C, the modules of the training device can be APIs (Application Programming Interfaces), plugins, or software development kits (SDKs) provided to servers (including cloud servers) or electronic devices. The APIs, plugins, or SDKs can implement the functions of each module in the training device, such as acquiring training data and performing training based on the training data to generate a target machine learning model. In some embodiments, the training device presented in this form can be called by other servers or electronic devices to be embedded in various applications.

[0234] This application also provides a prediction system in a third aspect for predicting the temperature history and / or stress history data of the printing material over time in 3D printing. Please refer to [link to relevant documentation]. Figure 9 The figure shows a simplified schematic diagram of the prediction system of this application in one embodiment. As shown, the prediction system includes:

[0235] Receiving module 41 is used to receive the printing features of the three-dimensional model of the object;

[0236] The prediction module 42 is used to call the machine learning model generated by the training method of the machine learning model as described in any embodiment of the first aspect of the present application to predict the printing features of the three-dimensional model of the object and output the temperature history and / or stress history data of the printing material of the three-dimensional model of the object changing over time during printing.

[0237] In some embodiments, the prediction system further includes a display module for visualizing the temperature history and / or stress history of the printing material during printing of the three-dimensional model of the object.

[0238] The display module can be used to provide a user interface or call other applications with user interfaces to display the temperature history and / or stress history of the printing material of the object's three-dimensional model over time in the user interface.

[0239] In some implementations, during the machine learning model's prediction of printing, the display module can select to display one of the following based on received instructions: a stress cloud map corresponding to historical stress data, a displacement cloud map of deformation, or a temperature cloud map corresponding to historical temperature data. In some examples, the temperature field and residual stress field of the object's three-dimensional model are fully coupled in the training data of the machine learning model. As a result, the machine learning model trained in this way can simultaneously obtain the transient temperature field and stress field in the calculation at each moment during the prediction of printing. The prediction results include the time-varying, i.e., dynamic temperature field and stress field. The display module can also select to display both simultaneously.

[0240] The display module can be implemented via a graphics module and its controller in an electronic device. The graphics module includes various known software components for presenting and displaying graphics on the display screen. Note that the term "graphics" includes any object that can be displayed to the user, including but not limited to text, web pages, icons (e.g., user interface objects including soft keys), digital images, videos, animations, etc. The display screen, for example, is a touchscreen, providing both output and input interfaces between the device and the user. The touchscreen controller receives / sends electrical signals from / to the touchscreen. The touchscreen then displays visual output to the user. This visual output can include text, graphics, video, and any combination thereof.

[0241] In some embodiments, the prediction system further includes a finite element module for discretizing the G-code data of the object's three-dimensional model to obtain basic units, thereby calculating and obtaining the printing features described in any embodiment of the first aspect of this application. The implementation or function of the finite element module in discretizing the G-code data of the object's three-dimensional model to obtain printing features can refer to the embodiments provided in the first aspect of this application, and will not be repeated here. It should be noted that in embodiments where the prediction system has a finite element module, the prediction system can accept the G-code data of the object's three-dimensional model and printing parameter information, calculate and obtain the printing features of the object's three-dimensional model through the prediction system, and transmit the calculated printing features to the receiving module.

[0242] Each module in the prediction system can be called by one or more applications to perform tasks such as acquiring training data, training a machine learning model based on the training data, making predictions based on the target machine learning model, and displaying the prediction structure.

[0243] This application also provides a computer device in a fourth aspect; please refer to [link to relevant documentation]. Figure 10 The diagram shown is a simplified schematic of a computer device provided in the fourth aspect of this application in one embodiment.

[0244] As shown in the figure, the computer device includes a storage device 51 and a processing device 52. The storage device 51 is used to store at least one program and a machine learning model trained by the training method of the machine learning model as described in any embodiment of the first aspect of this application.

[0245] In embodiments, the storage device 51 may include high-speed random access memory and may also include non-volatile memory, such as one or more disk storage devices, flash memory devices, or other non-volatile solid-state storage devices. In some embodiments, the storage device 51 may also include memory located remotely from one or more processors, such as network-attached memory accessed via RF circuitry or external ports and communication networks, wherein the communication network may be the Internet, one or more intranets, local area networks, wide area networks, storage area networks, etc., or suitable combinations thereof. The storage device 51 controller may control access to the memory by other components of the device, such as the CPU and peripheral interfaces.

[0246] The processing device 52 is connected to the storage device 51 and is used to execute the at least one program to call the machine learning model in the storage device 51 to predict the printing features of the three-dimensional model of the object and output the temperature history and / or stress history data of the printing material of the three-dimensional model of the object over time during printing.

[0247] In an embodiment, the processing device 52 is operatively coupled to the storage device 51 and / or a non-volatile storage device. More specifically, the processing device 52 can execute instructions stored in the storage device 51 and / or the non-volatile storage device to perform operations in a computing device, such as generating image data and / or transmitting image data to an electronic display. Thus, the processing device 52 may include one or more general-purpose microprocessors, one or more dedicated processors, one or more field-programmable logic arrays, or any combination thereof.

[0248] In some embodiments, the processing device 52 includes an integrated circuit chip with signal processing capabilities; or it includes a general-purpose processor, which may be a microprocessor or any conventional processor, such as a central processing unit. For example, it may be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a discrete gate or transistor logic device, or a discrete hardware component, which can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of this application. For example, based on at least one program and a machine learning model stored in the storage device 51, the machine learning model is invoked to predict the printing features of the three-dimensional model of the object to output the temperature history and / or stress history data of the printing material of the three-dimensional model of the object over time during printing.

[0249] This application also provides a computer device in its fifth aspect; please refer to [link to relevant documentation]. Figure 11 The image shown is a simplified schematic diagram of a computer device provided in the fifth aspect of this application in one embodiment.

[0250] As shown in the figure, the computer device includes a communication device 61, a storage device 62, and a processing device 63.

[0251] The communication device 61 is used to obtain a machine learning model trained by the training method of the machine learning model provided in any embodiment of the first aspect of this application from the service system. The communication device 61 can communicate with the service system network, and the service system includes, but is not limited to, electronic devices, servers, server clusters, cloud server systems, etc. The service system can also be a combination of software and hardware; the network can be the Internet, mobile network, local area network (LAN), wide area network (WLAN), storage area network (SAN), or one or more intranets, or appropriate combinations thereof. In this application, the types of clients and servers, or the types or protocols of communication networks between publisher terminals and servers, or between responder terminals and servers, are not limited.

[0252] The storage device 62 is used to store at least one program. In embodiments, the storage device 62 may include high-speed random access memory and may also include non-volatile memory, such as one or more disk storage devices, flash memory devices, or other non-volatile solid-state storage devices. In some embodiments, the storage device 62 may include memory located remotely from one or more processors, such as network-attached memory accessed via RF circuitry or external ports and communication networks, wherein the communication network may be the Internet, one or more intranets, local area networks, wide area networks, storage area networks, etc., or suitable combinations thereof. The storage device controller can control access to the memory by other components of the device, such as the CPU and peripheral interfaces.

[0253] The processing device 63 is connected to the storage device 62 and is used to execute the at least one program to call the machine learning model obtained from the service system to predict the printing features of the three-dimensional model of the object, and output the temperature history and / or stress history data of the printing material of the three-dimensional model of the object during printing.

[0254] In an embodiment, the processing device 63 is operatively coupled to the storage device 62 and / or a non-volatile storage device. More specifically, the processing device 63 can execute instructions stored in the storage device 62 and / or the non-volatile storage device to perform operations in a computing device, such as generating image data and / or transmitting image data to an electronic display. Thus, the processing device 63 may include one or more general-purpose microprocessors, one or more dedicated processors, one or more field-programmable logic arrays, or any combination thereof.

[0255] In some embodiments, the processing device 63 includes an integrated circuit chip with signal processing capabilities; or it includes a general-purpose processor, which may be a microprocessor or any conventional processor, such as a central processing unit. For example, it may be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a discrete gate or transistor logic device, or a discrete hardware component, which can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of this application. For example, it can predict the printing features of the three-dimensional model of the object based on at least one program stored in the storage device 62 and a machine learning model obtained from the service system to output the temperature history and / or stress history data of the printing material of the three-dimensional model of the object over time during printing.

[0256] In a sixth aspect, this application also provides a computer-readable storage medium storing at least one program that, when executed by a processor, implements a training method for a machine learning model as described in any embodiment provided in the first aspect of this application.

[0257] The function of training and generating machine learning models based on the training method provided in the first aspect of this application, when implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0258] In the embodiments provided in this application, the computer-readable and writable storage medium may include read-only memory, random access memory, EEPROM, CD-ROM or other optical disc storage devices, disk storage devices or other magnetic storage devices, flash memory, USB flash drive, portable hard drive, or any other medium capable of storing desired program code in the form of instructions or data structures and accessible by a computer. Additionally, any connection may be appropriately referred to as a computer-readable medium. For example, if instructions are transmitted from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of the medium. However, it should be understood that computer-readable and writable storage media and data storage media do not include connections, carrier waves, signals, or other transient media, but are intended for non-transient, tangible storage media. The disks and optical discs used in the application include compact discs (CDs), laser discs, optical discs, digital multifunction discs (DVDs), floppy disks, and Blu-ray discs, where disks typically copy data magnetically, while optical discs use lasers to copy data optically.

[0259] In one or more exemplary aspects, the functions described in the computer program executing the training method of the machine learning model described in this application can be implemented in hardware, software, firmware, or any combination thereof. When implemented in software, these functions can be stored or transmitted as one or more instructions or code onto a computer-readable medium. The steps of the methods or algorithms disclosed in this application can be embodied in processor-executable software modules, wherein the processor-executable software modules can reside on a tangible, non-transitory computer-readable and writable storage medium. The tangible, non-transitory computer-readable and writable storage medium can be any available medium accessible to a computer.

[0260] The flowcharts and block diagrams in the foregoing figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the figures. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented by a dedicated hardware-based system that performs the specified function or operation, or by a combination of dedicated hardware and computer instructions.

[0261] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A method for training a machine learning model, characterized in that, The machine learning model is used to predict the temperature history and / or stress history data of the printing material over time in 3D printing. The training method includes the following steps: The printing features of multiple sets of 3D object models are acquired, as well as the temperature history and / or stress history data that change over time during the printing process of the multiple sets of 3D object models. The printing features include model geometric features and voxel features. The voxel features include at least two of the following: the type of basic unit in the 3D object model, the distance from the basic unit to the free surface, the printing speed of the basic unit, the printing time of the layer containing the basic unit, the coordinates of the basic unit's position, and the density of the basic unit's position. The machine learning model is obtained by supervising the learning process based on the input data and the output data, using the printing features of the multiple sets of 3D object models as input data and the temperature history and / or stress history data of the multiple sets of 3D object models as output data.

2. The training method for the machine learning model according to claim 1, characterized in that, The geometric features of the model include feature quantities formed by the G-Code data of the three-dimensional model of the object. The G-Code data includes one or more pieces of information such as printing path, print head movement speed, layer height, printing material output speed, and single-layer printing time.

3. The training method for the machine learning model according to claim 1, characterized in that, The printing features also include feature quantities formed by printing equipment information, wherein the printing equipment information includes one or more of the following: printing material heating temperature, extrusion temperature, component plate heating temperature of the printing equipment, initial printing temperature field information, and print head shape.

4. The training method for the machine learning model according to claim 3, characterized in that, The methods for obtaining the initial temperature field information for printing include: measuring the temperature distribution information of the printing cavity before printing based on a thermal imager or thermocouple during the printing experiment.

5. The training method for the machine learning model according to claim 3, characterized in that, The printing features also include feature quantities formed by the material property information of the three-dimensional model of the object, wherein the material property information includes one or more of the following: filament type, filament diameter, filament cross-sectional shape, maximum heating temperature of the material, material thermal parameters, and initial residual stress of the material.

6. The training method for the machine learning model according to claim 5, characterized in that, The method for determining the thermal parameters of the material includes at least one of the following: The thermal emissivity of printed materials is measured using a thermal emissivity tester. Based on the filament printing experiment, the temperature change of the printed filament after it was heated and placed in the printing chamber was recorded, and the equivalent convective heat transfer coefficient was calculated accordingly. A printing experiment was conducted on a geometrically structured model with a regular shape. The temperature field distribution of the printing material changed over time during the printing process was recorded. Furthermore, a finite element simulation was performed on the geometrically structured model with different convective heat transfer coefficients. The simulated temperature field of the printing process was output, and the convective heat transfer coefficient corresponding to the simulated temperature field that coincided with the temperature field of the printing experiment was taken as the equivalent convective heat transfer coefficient.

7. The training method for the machine learning model according to claim 5, characterized in that, The method for determining the initial residual stress of the material includes the following steps: Multiple sets of monofilament printing experiments were conducted with different printing parameter settings for the printing materials. Calculate or measure the residual stress of monofilament components obtained from multiple monofilament printing experiments; and A residual strain database is obtained, which includes the different printing parameter information and the residual stress of the monofilament component obtained from multiple printing experiments; wherein the residual stress of the monofilament component obtained from the printing experiments has a corresponding relationship with the printing parameter information of the monofilament component.

8. The training method for the machine learning model according to claim 7, characterized in that, In the step of calculating or measuring the residual stress of the monofilament components obtained from multiple monofilament printing experiments, the monofilament structure is processed to release residual strain, and the monofilament deformation is measured to calculate the residual stress. Alternatively, the residual stress of the monofilament component can be determined based on physical testing methods.

9. The training method for the machine learning model according to claim 1, characterized in that, The types of basic units include: a first type of basic unit containing a free surface and located on the surface of the model; a second type of basic unit containing a free surface and located inside the model; and a third type of basic unit not containing a free surface.

10. The training method for the machine learning model according to claim 9, characterized in that, The feature quantity formed by the type of the basic unit is a multi-dimensional vector formed by encoding the type of the basic unit using categorical variables.

11. The training method for the machine learning model according to claim 10, characterized in that, The multidimensional vector is a multidimensional vector formed by one-hot encoding.

12. The training method for the machine learning model according to claim 9, characterized in that, The feature quantity formed by the distance from the basic unit to the free surface in the voxel feature is determined based on the distance of the basic unit to the first type of basic unit or the second type of basic unit within the printed layer.

13. The training method for the machine learning model according to claim 12, characterized in that, The method for determining the distance of the basic unit within its printing layer to either the first or second type of basic unit includes the following steps: A basic unit is identified as the target unit. Using the target unit as the search center, the first type of basic unit and the second type of basic unit in the printing layer are searched step by step within a preset range. The step-by-step search method is to increase the search range outward from the search center within the preset range. The search stops after the first or second type of basic unit is found, or after the first or second type of basic unit is not found within the preset range. The distance between the first or second type of basic unit and the search center when the search stops is determined as the distance between the target unit and the first or second type of basic unit within the printing layer.

14. The training method for the machine learning model according to claim 13, characterized in that, The feature quantity formed by the distance from the basic unit to the free surface is a multi-dimensional vector formed by encoding the distance of the basic unit to the first type of basic unit or the second type of basic unit within the printing layer using category variables. The vector dimension of the multi-dimensional vector is equal to the number of levels in the preset range.

15. The training method for the machine learning model according to claim 14, characterized in that, The categorical variables are encoded using one-hot encoding.

16. The training method for the machine learning model according to claim 1, characterized in that, The basic unit is a regular hexahedron, and the position coordinates of the basic unit are its centroid coordinates.

17. The training method for the machine learning model according to claim 1, characterized in that, The method of forming a characteristic quantity from the density at the location of the basic unit includes the following steps: Define a basic unit as the target unit, and determine the number of non-empty units within N preset ranges centered on the target unit; where N is a positive integer; The number of non-empty units within each preset range and the total number of basic units within that preset range are determined to form a feature quantity with dimension N representing the density at the location of the target unit.

18. The training method for the machine learning model according to claim 1, characterized in that, It also includes the step of discretizing the three-dimensional model of the object using the finite element method to obtain multiple basic elements, including: A simulation domain is established based on the outline of the object's three-dimensional model; and the simulation domain is discretized into multiple basic elements according to a preset mesh type. In the simulation domain, the basic units to be activated are determined, wherein the basic units to be activated are determined to be non-empty units.

19. The training method for the machine learning model according to claim 18, characterized in that, Based on the G-code data of the object's 3D model, the printing speed of the basic unit and / or the printing time of the layer in which the basic unit is located are calculated to form the corresponding feature quantity.

20. The training method for the machine learning model according to claim 1, characterized in that, The printing feature also includes interlayer expansion features, which are voxel features of basic units at the same projection point in adjacent layers of basic units; for basic units in the Nth layer, the interlayer expansion features are voxel features of basic units at the same projection point in layers NK to N-1 and layers N+1 to N+K, where N is a positive integer and K is a positive integer less than N.

21. The training method for the machine learning model according to claim 20, characterized in that, The methods for obtaining the interlayer augmentation features include at least one of the following: When the basic unit inside the model in the adjacent layer is determined as the target unit, the voxel features of the basic unit at the same projection point in the adjacent layer are added to the printing features of the target unit. When the basic unit outside the model boundary of the adjacent layer is determined as the target unit, the voxel features of the basic unit at the same projection point in the adjacent layer outside the model boundary are supplemented based on the preset constant value, and the supplemented voxel features are added to the printing features of the target unit. When a basic unit with an empty basic unit at the same projection point in an adjacent layer is determined as the target unit, the voxel features of the empty basic unit at the same projection point in the adjacent layer are supplemented based on a preset constant value, and the supplemented voxel features are added to the printing features of the target unit.

22. The training method for the machine learning model according to claim 1, characterized in that, Also includes: The step of performing distribution processing on at least one feature among the printed features, and forming the input data of the machine learning model based on the feature that satisfies a Gaussian distribution after distribution processing.

23. The training method for the machine learning model according to claim 1, characterized in that, The temperature history and / or stress history data of the multiple sets of object 3D models were obtained in different finite element simulation environments, wherein the different finite element simulation environments were formed by setting different printing parameter information for the object 3D models; or, the temperature history and / or stress history data of the multiple sets of object 3D models were obtained in different printing experimental environments, wherein the different printing experimental environments were formed by setting different printing parameter information for the object 3D models.

24. The training method for the machine learning model according to claim 23, characterized in that, The steps of obtaining temperature history and / or stress history data of the multiple sets of three-dimensional object models in different finite element simulation environments include: performing coupled simulation calculations on the three-dimensional object models to obtain the temperature history and / or stress history data, wherein the coupled simulation calculation models include a linear viscoelastic model describing the mechanical deformation of the printing material, and / or a transversely isotropic thermal conduction model or an orthotropic thermal conduction model used to describe the thermal conduction behavior of the printing material.

25. The training method for the machine learning model according to claim 23, characterized in that, The steps for obtaining the temperature history and / or stress history data of the multiple sets of object 3D models include: determining the time starting point of the temperature history and / or stress history of the basic unit based on the G-code data of the object 3D model, and obtaining the temperature and / or stress data of the basic unit from the time starting point at a preset sampling time interval.

26. The training method for the machine learning model according to claim 25, characterized in that, It also includes a step of dimensionality reduction processing of the historical temperature data, so that the dimensionality-reduced historical temperature data can be used as the output data of the machine learning model.

27. The training method for the machine learning model according to claim 26, characterized in that, The methods for dimensionality reduction processing of the historical temperature data include at least one of the following: Convert historical temperature data obtained from finite element simulation or printing experiments into equivalent relaxation time data; Principal component analysis is performed on the historical temperature data obtained from finite element simulation or printing experiments to achieve dimensionality reduction by retaining the principal components of the historical temperature data.

28. The training method for the machine learning model according to claim 27, characterized in that, The equivalent relaxation time data is obtained by equating the dynamic temperature history data of the printing material over time to a preset temperature value based on the WLF equation and / or the Arrhenius equation during the printing experiment or simulated printing process.

29. The training method for the machine learning model according to claim 23, characterized in that, The temperature history and / or stress history data of the multiple sets of object 3D models are the temperature history and / or stress history data at the position coordinates of each basic unit in the object 3D model.

30. The training method for the machine learning model according to claim 23, characterized in that, The temperature history and / or stress history data of the multiple sets of object 3D models are the average stress data and / or temperature data of each sub-region in the object 3D model; wherein, the sub-region is obtained by dividing the object 3D model based on a preset size or a preset number of partitions.

31. The training method for the machine learning model according to claim 1, characterized in that, The three-dimensional model of the object includes a regular geometric structure and actual product components. The geometric structure includes one or more of the following: vertical thin wall, horizontal thin plate, inclined thin plate, cube, cylinder, cylindrical body, square cylinder, ring, square ring, cone, inclined column, solid block, and grid-filled structure. The actual product components include one or more of the following: connectors, transmission components, and shell.

32. The training method for the machine learning model according to claim 1, characterized in that, It also includes the step of evaluating the error between the predicted data and the output data to adjust the machine learning model; wherein the predicted data is historical temperature and / or historical stress data predicted by the machine learning model based on the input data.

33. The training method for the machine learning model according to claim 1, characterized in that, The machine learning model is a neural network model.

34. A training device for a machine learning model, characterized in that, The machine learning model is used to predict the temperature history and / or stress history data of the printing material over time in 3D printing, and the training device includes: The training sample acquisition module is used to acquire printing features of multiple sets of object 3D models, and to acquire temperature history and / or stress history data that change over time during the printing process of multiple sets of object 3D models; wherein, the printing features include model geometric features and voxel features; wherein, the voxel features include at least two of the following: the type of basic unit in the object 3D model, the distance from the basic unit to the free surface, the printing speed of the basic unit, the printing time of the layer where the basic unit is located, the position coordinates of the basic unit, and the density of the position of the basic unit. The training module is used to take the printing features of the multiple sets of object 3D models as input data and the temperature history and / or stress history data of the multiple sets of object 3D models as output data, and to obtain the machine learning model through supervised learning based on the input data and output data.

35. A prediction system for predicting the temperature history and / or stress history data of printing materials over time in 3D printing, characterized in that, include: A receiving module is used to receive the printing features of the three-dimensional model of the object; The prediction module is used to call the machine learning model generated by the training method of the machine learning model as described in any one of claims 1-33 to predict the printing features of the three-dimensional model of the object, and output the temperature history and / or stress history data of the printing material of the three-dimensional model of the object over time during printing.

36. The prediction system according to claim 35, characterized in that, It also includes a display module for visually displaying the temperature history and / or stress history of the printing material during the printing of the three-dimensional model of the object.

37. The prediction system according to claim 35, characterized in that, It also includes a finite element module for discretizing the G-code data of the three-dimensional model of the object to obtain basic units, so as to calculate and obtain the printing features as described in any one of claims 1-33.

38. A computer device, characterized in that, include: A storage device for storing at least one program and a machine learning model trained by the training method of the machine learning model as described in any one of claims 1-33; as well as A processing device, connected to the storage device, is configured to execute the at least one program to invoke the machine learning model in the storage device to predict the printing features of the three-dimensional model of the object, and output the temperature history and / or stress history data of the printing material of the three-dimensional model of the object over time during printing.

39. A computer device, network-connected to a service system, characterized in that, include: A communication device for obtaining from the service system a machine learning model trained by the training method of any one of claims 1-33; Storage device for storing at least one program; as well as A processing device, connected to the storage device, is configured to execute the at least one program to invoke the machine learning model obtained from the service system to predict the printing features of the three-dimensional model of the object, and output the temperature history and / or stress history data of the printing material of the three-dimensional model of the object over time during printing.

40. A computer-readable storage medium, characterized in that, The system stores at least one program, which, when executed by a processor, implements the training method for the machine learning model as described in any one of claims 1-33.

Images