Method for building a process database based on additive manufacturing and medium

Abstract

The application discloses a kind of process database construction method and medium based on additive manufacturing, belong to the field of additive manufacturing, method includes: S1, the process data of product obtained by additive manufacturing under different process parameters is acquired by experiment, to construct initial process database;S2, the preset neural network model is optimized and trained;S3, when the target process data of user is acquired, and there is no matching case in the current process database, execute S4-S5;S4, generate target process parameters using particle swarm optimization algorithm and input neural network model, to obtain the corresponding performance data estimate value, and adjust the particle position in particle swarm optimization algorithm according to performance data estimate value;S5, repeat S4 until the performance data estimate value obtained is greater than target performance data, or reach the set repetition number.It has important significance to the transformation of SLM process to standardization, stabilization, batch production.

Description

Technical Field

[0001] The present invention belongs to the field of additive manufacturing, and more specifically, relates to a method and medium for constructing a process database based on additive manufacturing. Background Art

[0002] Additive manufacturing technology is a technique that uses the discrete-accumulation principle to directly manufacture parts using three-dimensional part data. Currently, the additive manufacturing industry is in its infancy. For specific alloy systems, various industries conduct extensive and repetitive basic experiments to optimize the process, which can lead to wasteful resources.

[0003] With the rapid development of informatization, more and more companies are emphasizing information management and implementing a wide range of information-based software. The manufacturing industry is also increasingly using automated equipment to assist in production, generating a large amount of data that is stored in different databases and lacks sharing. Therefore, the entire additive manufacturing industry urgently needs process databases and expert knowledge bases to reduce the time and resource costs associated with repeated experiments. Summary of the Invention

[0004] In response to the defects of the existing technology and the need for improvement, the present invention provides a method and medium for constructing a process database based on additive manufacturing. The purpose is to construct a statistical analysis mathematical model and a neural network model through the data obtained from experiments, and obtain the relationship between process parameters and performance values, so as to predict the performance values ​​of corresponding materials through process parameters.

[0005] To achieve the above-mentioned objectives, according to one aspect of the present invention, a method for constructing a process database based on additive manufacturing is provided, comprising: S1, experimentally obtaining process data of products obtained by additive manufacturing under different process parameters, and constructing an initial process database using the process parameters and process data in the experimental process as cases; S2, optimizing and training a preset neural network model using the process parameters in the experimental process as input and the performance data in the process data as labels; S3, when the user's target process data is obtained, determining the weights of each feature in the target process data and then performing case matching, and executing S4-S5 when there is no case in the current process database whose process data matches the target process data; S4, generating target process parameters using a particle swarm algorithm and inputting them into the neural network model to obtain corresponding performance data estimates, and adjusting the particle positions in the particle swarm algorithm according to the performance data estimates; S5, repeatedly executing S4 until the obtained performance data estimates are greater than the target performance data in the target process data, or until the number of repetitions reaches a set value, so that the user performs additive manufacturing according to the target process parameters generated last time.

[0006] Furthermore, after S5, it also includes: obtaining actual process data of the product obtained by the user through additive manufacturing based on the target process parameters generated last time, and when the actual process data meets the indicators required by the user, adding the target process parameters generated last time and the actual process data as cases to the current process database.

[0007] Furthermore, the process data is composed of multiple features, and the method also includes: S3′, respectively calculating the first total similarity between the target process data and the process data of each case; if there is a first total similarity greater than the similarity threshold, the target process data matches the process data corresponding to the maximum first total similarity; otherwise, executing S3″; S3″, respectively calculating the second total similarity between some features in the target process data and some features in the process data of each case, the weights of the some features are all greater than the weight threshold; if there is a second total similarity greater than the similarity threshold, the target process data matches the process data corresponding to the maximum second total similarity; otherwise, there is no case in the current process database where the process data matches the target process data.

[0008] Furthermore, the first total similarity and the second total similarity are respectively:

[0009]

[0010]

[0011] Where s' is the first total similarity, s" is the second total similarity, σ is a set constant, α i , α j are the weights of the i-th and j-th features respectively, A i is the i-th feature in the target process data, B i is the i-th feature in the process data of the case, d(A i ,B i ) is A i 、B i The absolute distance between j is the jth feature in the target process data, B j is the jth feature in the process data of the case, and A j 、B j The weight of d(A j ,B j ) is A j 、B j The absolute distance between them, n is the number of feature types, m is the number of feature types with weights greater than the weight threshold, i max 、i min are the maximum and minimum values ​​of the i-th feature, jmax 、j min are the maximum and minimum values ​​of the j-th feature respectively.

[0012] Furthermore, it also includes: setting the same initial usage count for each case in the process database, and adding one to the usage count of the case when the process data of any case matches the target process data; when the number of cases in the process database exceeds the capacity threshold, eliminating the case with the least usage count until the number of remaining cases does not exceed the capacity threshold.

[0013] Furthermore, before S4, it also includes: taking process parameters as optimization parameters and performance data as optimization targets, and using weighted summation method to construct a multi-objective optimization model; using particle swarm algorithm to generate target process parameters in S4 includes: taking obtaining the target performance data as the target, and using the multi-objective optimization model to generate the target process parameters.

[0014] Furthermore, the performance data includes four characteristics: density, tensile strength, yield strength and elongation; the neural network model includes four sub-neural networks, which output each characteristic in the performance data in a one-to-one correspondence.

[0015] Furthermore, the method also includes: when new experimental data containing process parameters and performance data are obtained, using the experimental data to retrain the current neural network model; if the goodness of fit of the retrained neural network model is greater than that of the current neural network model, updating the current neural network model; otherwise, the current neural network model remains unchanged.

[0016] Furthermore, the experiment in S1 includes: preliminarily determining the optimal process window in the global process window through orthogonal experiments or general full-factor experiments; and determining the optimal process parameters and corresponding process data in the optimal process window through response surface method experiments.

[0017] According to another aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the program is executed by a processor, the method for constructing a process database based on additive manufacturing as described above is implemented.

[0018] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

[0019] (1) The process parameters and process data generated by a small number of experimental designs can be used to construct an initial process database, which can be expanded at any time to achieve the purpose of real-time optimization of process parameters and process data models;

[0020] (2) Specifically, the target process parameters corresponding to the target performance data are generated by the particle swarm algorithm. The process parameters generated by the particle swarm algorithm are evaluated by the neural network model trained with experimental data to adjust the particle swarm algorithm, thereby generating process parameters that meet user requirements to expand the process database, so that the model has higher performance prediction and process parameter recommendation capabilities;

[0021] (3) First, the optimal process window is preliminarily determined in the global process window, and customized experimental design is performed to determine the optimal process parameters and corresponding process data, so that the neural network model has better fitting ability, thereby improving the accuracy and practicality of the constructed database. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A flowchart of a method for constructing a process database based on additive manufacturing provided in an embodiment of the present invention;

[0023] Figure 2 A diagram illustrating the implementation process of the method for constructing a process database based on additive manufacturing provided by an embodiment of the present invention;

[0024] Figure 3 A schematic diagram of a case feature hierarchy model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0025] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0026] In the present invention, the terms "first", "second", etc. (if any) in the present invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0027] Figure 1 Flowchart of the method for constructing a process database based on additive manufacturing provided by an embodiment of the present invention. Figure 1 , combined with Figure 2-Figure 3 , a method for constructing a process database based on additive manufacturing in this embodiment is described in detail, and the method includes operations S1 to S5.

[0028] Operation S1: experimentally obtain process data of products obtained by additive manufacturing under different process parameters, and construct an initial process database using the process parameters and process data in the experimental process as cases.

[0029] According to an embodiment of the present invention, the experiment in operation S1 includes: preliminarily determining the optimal process window in the global process window through an orthogonal experiment or a general full-factor experiment; and determining the optimal process parameters and corresponding process data in the optimal process window through a response surface method experiment.

[0030] In operation S1, two different experimental schemes are designed in batches. Specifically, when the global process window is in operation, an orthogonal experiment or a general full-factor experiment is designed, samples are prepared through experiments, their performance values ​​are tested, and preliminary process parameters and process data are obtained; statistical analysis and data visualization methods are used to preliminarily determine the rough optimal process parameters (i.e., the optimal process window) and obtain the optimal two levels of each process parameter. When the optimal process window is in operation, a response surface method experiment corresponding to the factors and levels is designed, and the experimental star points are designed to be adjacent to the two horizontal points. The remaining experimental points, including the center point and the axis point, are then designed, and samples are prepared again through experimental methods to test their performance values ​​to obtain the optimal process parameters and corresponding process data.

[0031] Operation S2 optimizes and trains a preset neural network model using the process parameters in the experiment as input and the performance data in the process data as labels.

[0032] In this embodiment, the preset BP neural network model adopts a three-layer structure, consisting of an input layer, a hidden layer, and an output layer. Taking the process parameters including laser power, scanning speed, scanning spacing, and powder layer thickness, and the performance data including density, tensile strength, yield strength, and elongation as an example; the input layer of the neural network model has 4 nodes, and the output is the prediction target (i.e., feature) of the model. Since the relationship between the above four characteristic indicators is not linear, it is difficult to combine them with a unified indicator. Therefore, it is necessary to establish separate sub-neural networks for the four characteristic indicators. The output layer of each sub-neural network has only one node, that is, the neural network model in this embodiment includes four sub-neural networks, which output each feature one by one.

[0033] The structure of each sub-neural network is ultimately determined to be [4-X-1], where X is the number of hidden nodes determined based on the characteristics of the characteristic indicators. Preferably, an empirical formula is first used to determine the approximate range of the number of hidden nodes X for each of the four constructed sub-neural networks, ranging from 4 to 13. Then, through experimental comparison, the values ​​of X for different output indicators are determined. The final values ​​of X for the density sub-neural network are 6, 8 for the tensile strength sub-neural network, 8 for the yield strength sub-neural network, and 5 for the elongation sub-neural network.

[0034] Preferably, the normalization and transfer functions of the constructed BP neural network model use a hyperbolic tangent function with a variation range of [-1,1]; the random error is eliminated by taking the average of the model indicators through multiple training (such as 5 times), so that the model indicators are close to the model effect at the time of optimal convergence; the expected error is set to 0.0001-0.001, and the momentum coefficient is set to 0.001-0.01; the learning algorithm uses one of the gradient fastest descent optimization algorithm, the Levenberg-Marquardt (LM) algorithm based on the nonlinear least squares method, and the Bayesian regularization algorithm (BR).

[0035] According to an embodiment of the present invention, when new experimental data containing process parameters and performance data is acquired, the current neural network model is retrained using the experimental data. If the goodness of fit of the retrained neural network model is greater than that of the current neural network model, the current neural network model is updated; otherwise, the current neural network model remains unchanged. This results in a neural network model with more accurate performance evaluation.

[0036] Operation S3, when the user's target process data is obtained, the weights of each feature in the target process data are determined and case matching is performed. If there is no case in the current process database whose process data matches the target process data, operations S4-S5 are performed.

[0037] See Figure 3 Preferably, in this embodiment, the process data consists of formed part characteristics, processing costs, and mechanical properties (i.e., performance data). Formed part characteristics include structural characteristics and structural dimensions, processing costs include forming time, and performance data include density, tensile strength, yield strength, and elongation. Structural characteristics, structural dimensions, forming time, density, tensile strength, yield strength, and elongation are all characteristics of the process data.

[0038] The user inputs the features of the formed part, processing cost, and mechanical properties. These data constitute the target process data, and the hierarchical analysis method is used to determine the feature weight (i.e. weight) of each feature. The weights of structural features and structural dimensions are w 11 、w 12 The weight of forming time is w2; the weights of density, tensile strength, yield strength and elongation are w 31 、w 32 、w 33 、w 34 .

[0039] When the user's target process data is obtained, operation S3 includes sub-operation S3' to sub-operation S3".

[0040] In sub-operation S3′, the first total similarity between the target process data and the process data of each case is calculated respectively. If there is a first total similarity greater than the similarity threshold, the target process data is matched with the process data corresponding to the maximum first total similarity. Otherwise, sub-operation S3″ is executed.

[0041] The first total similarity represents the total similarity of all features between the target process data and the case's process data. When the first total similarity between an existing case and a new user requirement exceeds the similarity threshold, the existing case is considered a match. The solution to the case with the highest similarity is output as the solution to the requirement, and the usage count for that case is incremented by 1.

[0042] In sub-operation S3″, the second total similarities between some features in the target process data and some features in the process data of each case are calculated respectively, and the weights of the some features are all greater than the weight threshold. If there is a second total similarity greater than the similarity threshold, the target process data matches the process data corresponding to the maximum second total similarity. Otherwise, there is no case in the current process database where the process data matches the target process data, and operations S4-S5 are executed.

[0043] When the first total similarity between the non-existent case and the new user-required case exceeds the similarity threshold, all cases are considered to be inconsistent with the user's needs. At this time, the feature screening in the case modification is performed: features with weights below the weight threshold are considered to be unimportant features and are no longer included in the overall similarity calculation. The second total similarity between all cases and the new user-required case is recalculated. The second total similarity represents the total similarity between the target process data and the case's process data for features with weights greater than the weight threshold. At this time, when the second total similarity between the existing case and the new user-required case exceeds the similarity threshold, the existing case is considered to match the user's needs. At this time, the solution of the case with the highest similarity is output as the solution to the need, and the number of times this case is used is increased by 1.

[0044] According to an embodiment of the present invention, the first total similarity and the second total similarity are respectively:

[0045]

[0046]

[0047] Among them, s' is the first total similarity, s" is the second total similarity, σ is the set constant, α i , α j are the weights of the i-th and j-th features respectively, A i is the i-th feature in the target process data, B i is the i-th feature in the process data of the case, d(A i ,B i ) is Ai , B i j is the jth feature in the target process data, B j is the jth feature in the process data of the case, and A j , B j is greater than a weight threshold, d(A j , B j ) is the absolute distance between A j , B j , n is the number of feature types, m is the number of feature types whose weights are greater than the weight threshold, i max , i min are the maximum value and the minimum value of the ith feature, respectively, j max , j min are the maximum value and the minimum value of the jth feature, respectively.

[0048] It should be noted that in the embodiment, the features are classified into two data types, one is a numerical type, for example, a structural dimension of a formed part, a processing cost, a mechanical property, etc., and the other is an enumerated type, mainly a structural feature of a formed part. For features of only the numerical type, the calculation formulas of the first total similarity and the second total similarity can be used. The local similarity of the enumerated type feature can be given by a technician according to production experience, and the local similarity can be obtained by consulting, and the first similarity and the second similarity can be expressed as:

[0049]

[0050]

[0051] sim(A i , B i ) is the local similarity between A i , B i , for the numerical type A i , B i , for the enumerated type A i , B i , the local similarity table can be consulted to obtain sim(A i , B i ).

[0052] According to the embodiment of the application, the cases in the process database are classified according to product materials to form corresponding sub-databases; when the target process data of a user is obtained, the process data matching the target process data is searched from the same sub-database as the material required by the user in operation S3. Thus, the matching time can be reduced.

[0053] ​In operation S4 , target process parameters are generated by using a particle swarm algorithm and input into the neural network model to obtain corresponding performance data estimates, and particle positions in the particle swarm algorithm are adjusted according to the performance data estimates.

[0054] According to an embodiment of the present invention, before executing operation S4, the method further includes: constructing a multi-objective optimization model using a weighted summation method with the process parameters as optimization parameters and the performance data as optimization targets. Generating target process parameters using a particle swarm algorithm in operation S4 includes: using the multi-objective optimization model to search for and generate target process parameters with the goal of obtaining target performance data.

[0055] The optimization target of the particle swarm algorithm is all the features in the performance data. Based on the previously established neural network model, the value of each feature is obtained, and then a multi-objective optimization model is established using the weighted summation method. The weight factor follows the weight obtained by the hierarchical analysis method. At this time, the optimal process parameter combination is the position of the particle, and the output value of the multi-objective optimization model is the fitness of the particle. At the same time, the horizontal range of the orthogonal experiment is referenced to constrain the particle search space. Taking the process parameters including laser power, scanning speed, scanning spacing, and powder layer thickness as an example, in order to eliminate the dimensional influence of different process parameters and performance indicators, the process parameters and performance indicators involved in the model are normalized. The simplified model of weighted multi-objective optimization is as follows:

[0056]

[0057] Where x is the particle coordinate vector, including (x1, x2, x3, x4), which corresponds to the normalized values ​​of the process parameters of laser power, scanning speed, scanning spacing, and powder layer thickness; Z(x), L(x), Q(x), and Y(x) are the normalized prediction models of density, tensile strength, yield strength, and elongation; k1, k2, k3, and k4 are the weights of the above four performance indicators determined by the hierarchical analysis method; X is the constraint of the particle search space, that is, the value range of the four particle coordinates (normalized process parameters), which are all [-1, 1] here.

[0058] Operation S5, repeatedly performing operation S4 until the obtained performance data estimation value is greater than the target performance data in the target process data, or until the number of repetitions reaches a set value, so that the user performs additive manufacturing according to the target process parameters generated last time.

[0059] According to an embodiment of the present invention, after operation S5, it also includes: obtaining the actual process data of the product obtained by the user through additive manufacturing based on the target process parameters generated last time, and when the actual process data meets the indicators required by the user, the target process parameters and actual process data generated last time are added as cases to the current process database.

[0060] Specifically, after the user performs additive manufacturing based on the target process parameters generated last time, the user needs to evaluate the target process parameters generated last time based on the actual process data of the product obtained by additive manufacturing. If the actual process data meets the user's requirements, it means that the case obtained based on the target process parameters generated last time is qualified. The target process parameters generated last time and the actual process data are added as cases to the current process database. Otherwise, the target process parameters generated last time are discarded, such as Figure 2 shown.

[0061] According to an embodiment of the present invention, the same initial usage count is set for each case in the process database, and when the process data of any case matches the target process data, the usage count of the case is increased by one; when the number of cases in the process database exceeds the capacity threshold, the case with the least usage count is eliminated until the number of remaining cases does not exceed the capacity threshold.

[0062] The embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, which, when executed by a processor, implements the following Figure 1-Figure 3 The method for constructing a process database based on additive manufacturing is shown.

[0063] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for constructing a process database based on additive manufacturing, characterized in that: include: S1, experimentally obtain the process data of products obtained by additive manufacturing under different process parameters, and build an initial process database using the process parameters and process data in the experimental process as cases; S2, using the process parameters in the experiment as input and the performance data in the process data as labels, optimizes and trains the preset neural network model; S3, when the user's target process data is obtained, the weights of each feature in the target process data are determined and case matching is performed. If there is no case in the current process database whose process data matches the target process data, S4-S5 are executed; S4, generating target process parameters using a particle swarm algorithm and inputting the target process parameters into the neural network model to obtain corresponding performance data estimates, and adjusting particle positions in the particle swarm algorithm according to the performance data estimates; S5, repeatedly executing S4 until the estimated value of the obtained performance data is greater than the target performance data in the target process data, or until the number of repetitions reaches a set value, so that the user performs additive manufacturing according to the target process parameters generated last time; The experiment in S1 includes: preliminarily determining the optimal process window in the global process window through orthogonal experiments or general full-factor experiments; Through response surface method experiments, the optimal process parameters and corresponding process data are determined within the optimal process window.

2. The method for constructing a process database based on additive manufacturing according to claim 1, wherein: The S5 further includes: The actual process data of the product obtained by the user through additive manufacturing based on the target process parameters generated last time is obtained, and when the actual process data meets the indicators required by the user, the target process parameters generated last time and the actual process data are added as cases to the current process database.

3. The method for constructing a process database based on additive manufacturing according to claim 1, wherein: Process data consists of multiple features, the method Also includes: S3′, respectively calculating the first total similarity between the target process data and the process data of each case. If there is a first total similarity greater than a similarity threshold, the target process data matches the process data corresponding to the maximum first total similarity. Otherwise, executing S3″; S3″, respectively calculating the second total similarities between some features in the target process data and some features in the process data of each case, wherein the weights of the some features are all greater than a weight threshold. If there is a second total similarity greater than the similarity threshold, the target process data matches the process data corresponding to the maximum second total similarity. Otherwise, there is no case in the current process database whose process data matches the target process data.

4. The method for constructing a process database based on additive manufacturing according to claim 3, wherein: The first total similarity and the second total similarity are respectively: Where s' is the first total similarity, s" is the second total similarity, σ is a set constant, α i , α j are the weights of the i-th and j-th features respectively, A i is the i-th feature in the target process data, B i is the i-th feature in the process data of the case, d(A i ,B i ) is A i 、B i The absolute distance between j is the jth feature in the target process data, B j is the jth feature in the process data of the case, and A j 、B j The weight of d(A j ,B j ) is A j 、B j The absolute distance between them, n is the number of feature types, m is the number of feature types with weights greater than the weight threshold, i max 、i min are the maximum and minimum values ​​of the i-th feature, j max 、j min are the maximum and minimum values ​​of the j-th feature respectively.

5. The method for constructing a process database based on additive manufacturing according to any one of claims 1 to 4, characterized in that: Also includes: Setting the same initial usage count for each case in the process database, and increasing the usage count for any case by one when the process data of the case matches the target process data; When the number of cases in the process database exceeds the capacity threshold, the cases with the least number of uses are eliminated until the number of remaining cases does not exceed the capacity threshold.

6. The method for constructing a process database based on additive manufacturing according to claim 1, wherein: The step before S4 also includes: using process parameters as optimization parameters and performance data as optimization targets, and constructing a multi-objective optimization model using a weighted summation method; The generating target process parameters by using the particle swarm algorithm in S4 includes: generating the target process parameters by using the multi-objective optimization model with the goal of obtaining the target performance data.

7. The method for constructing a process database based on additive manufacturing according to claim 1, wherein: The performance data includes four characteristics: density, tensile strength, yield strength and elongation; the neural network model includes four sub-neural networks, which output each characteristic in the performance data in a one-to-one correspondence.

8. The method for constructing a process database based on additive manufacturing according to claim 1, wherein: The method further comprises: When new experimental data including process parameters and performance data is obtained, the current neural network model is retrained using the experimental data; If the goodness of fit of the retrained neural network model is greater than that of the current neural network model, the current neural network model is updated; otherwise, the current neural network model remains unchanged.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method for constructing a process database based on additive manufacturing according to any one of claims 1 to 8 is implemented.

Images