Method for predicting electron beam selective melting forming defects based on machine learning

Abstract

The invention discloses a machine learning-based electron beam selective melting defect prediction method, which is characterized in that electron beam selective melting forming is a technology of utilizing a high-energy electron beam as a heat source to melt metal powder layer by layer in a vacuum environment so as to construct a three-dimensional metal part. Due to the fact that the absorption rate of the material to electron beam energy is high, the electron beam energy density is high, in the forming process, defects are prone to being generated, and the defects mainly comprise powder collapsing, spheroidizing, holes, cracks and the like. Through a mode of developing machine learning, global search and optimization capability are realized, previous defect data can be classified and normalized, and possible defects can be predicted and analyzed in advance for next manufacturing, so that technological parameters of electron beam selective melting are optimized, and technological adjustment and optimization are assisted.

Description

Technical Field

[0001] The present invention relates to a prediction method for defects in electron beam selective melting forming based on machine learning, belonging to the field of additive manufacturing. Background Art

[0002] The heat source of the electron beam selective forming technology is an electron beam, which is a processing heat source generated by a large number of electrons obtaining acceleration through an electric field to increase the energy density. The environment of electron beam selective melting forming is a vacuum environment, which can effectively avoid the influence of elements such as O and H on the manufacturing process and achieve high-quality forming. Electron beam selective melting forming has been successfully applied in the fields of aviation, aerospace, medical treatment, etc. However, during the process of electron beam selective melting forming, due to the instability of the forming process, there are still many defects. The main defect types include balling, porosity, cracks, etc. Many studies have been done on the formation reasons, classification and solutions of these defects, but the relevant studies are all carried out after or during the process of the process, and it is difficult to predict defects before the process, which will obviously lead to material loss and time waste. There are many categories of reasons and a large parameter space for the formation of defects in electron beam selective melting forming, including irregular powder layer, spatter, balling, porosity, surface quality, cracks, geometric deformation, etc.

[0003] The existing electron beam selective melting defect analysis technologies mostly focus on the analysis of the formation mechanism of typical defects and their intelligent monitoring and process control. Based on the collected signals, features can be extracted, the relationship between the signals and defects or processes can be established, and the defects or processing states can be classified or predicted. However, at present, the monitoring signal processing is mainly carried out after the electron beam selective melting forming process. At this time, the defects have already occurred, and only after-the-fact analysis can be solved, and the advance prediction and avoidance guidance of defects cannot be achieved. Summary of the Invention

[0004] The technical problem to be solved by the present invention is: overcoming the deficiencies of the prior art, providing a prediction method for defects in electron beam selective melting forming based on machine learning, solving the problem that defects cannot be avoided in advance in the prior art, being able to realize the auxiliary analysis and decision-making of electron beam selective melting parameters, achieving high-quality preparation of materials, reducing the manufacturing time and saving the labor cost.

[0005] The technical solution of the present invention is: a prediction method for defects in electron beam selective melting forming based on machine learning, including:

[0006] S1: Count the defects of electron beam selective melting technology;

[0007] S2: According to the defects of electron beam selective melting technology, analyze the process parameters affecting the defects, and give an initial discretized influence factor to the process parameters of each defect;

[0008] S3: Obtain the historical data of electron beam selective melting, and form a standardized defect data set;

[0009] S4: Construct a defect prediction model based on a BP neural network optimized by a genetic algorithm, and train and optimize it based on the defect data set to obtain the defect prediction model;

[0010] S5: Process the parameters to be processed by the electron beam selective melting process to form standardized input parameters in the same form as the defect data set; input the input parameters into the constructed defect prediction model, adjust the process parameters according to the prediction results, and input them into the defect prediction model after adjustment. Repeat multiple times to obtain optimized process parameters;

[0011] S6: Implement the electron beam selective melting process according to the optimized process parameters obtained in S5.

[0012] Preferably, collect the defect data after each implementation of the electron beam selective melting process, and expand it to the defect data set according to the standardized defect data set format;

[0013] Use the expanded defect set data to further iteratively optimize the defect prediction model each time manufacturing is carried out. Repeat steps S5 - S6 to complete defect prediction using the iteratively optimized defect prediction model, and adjust the process parameters used when implementing the electron beam selective melting process according to the predicted defects.

[0014] Preferably, the defects of the electron beam selective melting technology include:

[0015] Material limitations, pores, cracks, balling, powder collapse.

[0016] Preferably, the process parameters affecting defects include:

[0017] Defect type, material type, electron beam current, scanning speed, scanning spacing, scanning path, electron beam spot diameter, substrate preheating temperature, acceleration voltage, powder layer thickness, overhang angle.

[0018] Preferably, the discretized influence factors include:

[0019] B = {extremely large, relatively large, large, relatively small, extremely small, small, none}.

[0020] Preferably, when forming the standardized defect data set: use the defects and the process parameters affecting the defects to form an S-rough set defect discernibility matrix, and assign values to the elements of the S-rough set defect discernibility matrix according to the objects using the discretized influence factors, that is, give specific discretized influence factors to each process parameter of each defect.

[0021] Preferably, denote the BP neural network optimized by a genetic algorithm as GA-BP. When constructing the defect prediction model:

[0022] Divide the defect data set into a training set and a test set;

[0023] Determine the network structure, including the number of input, output and hidden layer nodes of GA-BP;

[0024] Perform parameter configuration, including the number of iterations, learning rate, and activation function of GA-BP. Among them, the activation function uses the Sigmoid function;

[0025] Input the training set data for model training to determine the trained prediction model;

[0026] Input the test set into the prediction model, detect the accuracy of the model, and further adjust the network structure and parameters to optimize the prediction model.

[0027] Preferably, 85% of the defect data set is used as the training set, and the remaining 15% is used as the test set.

[0028] The present invention has the following advantages compared with the prior art:

[0029] (1) Different from the traditional processing method, the method for predicting defects in electron beam selective melting by using the genetic algorithm to optimize the BP neural network algorithm proposed by the present invention innovatively attempts the technical path from post-analysis to advance prediction. The research and application of defect prediction technology can promote the innovation and development of electron beam selective melting technology. For example, by establishing a new pore defect prediction model for electron beam selective melting formed parts based on the geometry of the molten pool, the pore defects under different materials and conditions can be accurately predicted, providing a theoretical basis for the rapid preliminary screening of processing parameters;

[0030] (2) The present invention innovatively applies the genetic algorithm to optimize the BP neural network algorithm in the electron beam selective melting process flow, and can make full use of the advantages of the genetic algorithm to optimize the BP neural network algorithm in the fields of function optimization and combinatorial optimization. By reasonably designing the fitness function according to the objective function of the specific problem, the accuracy of the prediction result can be adaptively adjusted and optimized, so as to guide the prediction process towards the optimal solution;

[0031] (3) The defect prediction model constructed by the present invention can predict in advance the possible defects in the electron beam selective melting formed parts, such as pores, cracks, etc., enabling technicians to take corresponding measures during the manufacturing process to reduce or eliminate these defects, thereby ensuring the quality and performance of the product; on the other hand, through defect prediction, potential problems can be discovered and solved in advance during the production process, avoiding rework or scrapping caused by defects in the later stage, thereby improving production efficiency;

[0032] (4) The present invention can reduce waste and rework caused by defects, thus reducing production costs. At the same time, by optimizing process parameters and selecting appropriate materials, production efficiency can be further improved and costs can be reduced;

[0033] (5) The method of the present invention can make full use of historical data and process data, can automatically make pre-judgments using computer computing power resources, realizes decision-making in advance from the perspective of simulation, the entire system has strong adaptability, the prediction time is relatively controllable, the experiment has strong repeatability, and the economic value is high. Description of the Drawings

[0034] Figure 1 is the flow chart of the genetic algorithm optimized BP neural network algorithm of the present invention;

[0035] Figure 2 is the metallographic picture of the TA15 titanium alloy specimen formed by electron beam selective melting with hole prediction parameters of the present invention. Detailed Embodiments

[0036] This method classifies and processes the influencing factors and improvement factors of electron beam selective melting defects by introducing machine learning algorithms, specifically the genetic algorithm optimized BP neural network algorithm (hereinafter referred to as the GA-BP algorithm), and makes full use of historical defect data, so as to predict the defects formed under different process parameter conditions.

[0037] The present invention belongs to the defect prediction technology in the process of electron beam selective melting forming, and particularly relates to the intelligent preprocessing and prediction method of data, which can classify and normalize historical defect data, and at the same time can predict and analyze possible defects in the future manufacturing process in advance, so as to realize process adjustment and optimization

[0038] The principle of the present invention is as follows:

[0039] The defect characteristics of electron beam selective melting technology are mainly reflected in aspects such as complex process and many influencing factors. The typical defect characteristics mainly include balling, pores and cracks, etc. The formation of these defects is related to various factors such as process parameters, powder properties, and vacuum environment

[0040] Step 1: Correlate and statistically analyze the defects and influencing factors of electron beam selective melting technology. The influencing factors of the defects are as follows:

[0041] 1) Material limitations:

[0042] At present, the main printing raw materials used in electron beam selective melting forming technology include structural steel, stainless steel, nickel-based alloy, titanium alloy, cobalt-chromium alloy, composite materials, etc.

[0043] 2) Typical internal defects:

[0044] Holes: They are internal defects that are very likely to occur in electron beam selective melting. They are mainly formed due to the aggregation and growth of gas elements in the liquid melt pool or the evaporation of light elements, resulting in hole defects.

[0045] Cracks: They are defects caused by the combined action of stress and other embrittlement factors. They are gaps formed by the destruction of the metal atom bonding force in a local area of the component, resulting in the formation of a new interface.

[0046] Spheroidization: During the process of electron beam selective melting forming, when metal powder is scanned by the electron beam and absorbs the energy of the electron beam, it quickly melts to form a metal solution. Under the combined action of surface tension, gravity, and the surrounding medium, the metal solution solidifies and shrinks to form intermittent spherical particles.

[0047] 3) Other defects:

[0048] Powder dispersion: It refers to the behavior that when the powder receives the energy of the electron beam, due to the sudden increase in kinetic energy, it is converted into kinetic energy. Coupled with the very small volume of the powder itself, it will rapidly escape from its original position.

[0049] Step 2: According to the statistical results of the defects and influencing factors of electron beam selective melting technology, form a complete defect set and establish an S-rough set defect discrimination matrix for the defect set.

[0050] 1) Attribute characteristics of S-rough set data:

[0051] Let the data universe be W, where W has m sub-data w 1 , w 2 , w 3 , w 4 Lw m , and the sub-data w i is composed of several elements, that is: w 1 ={x 1.1 , x 1.2 , Lx 1.l}, w 2 ={x 2.1 , x 2.2 , Lx 2.r}, L, w m ={x m.1 , x m.2 , Lx m.l}; a 1 ={a 1.1 , a 1.2 , La 1.l}, a 2 ={a 2.1 , a 2.2 , La 2.q}, L, a m ={am.1 , a m.2 , La m.h} are the sub - data w 1 , w 2 , w 3 , w 4 Lw m 's set of attributes.

[0052] 2) Establish the S - rough set defect discrimination matrix:

[0053] According to the requirements of the present invention, it can be known that first, a complete set of defects is established. Therefore, the complete set of defects can be used as the S - rough set data complete set W, and each defect element is used as its sub - data w 1 , w 2 , w 3 , w 4 Lw m . Given that the basic attributes of each defect type are relatively similar, in order to establish a complete set of attributes, the number of elements of each sub - data is set to be the same. The degree of association of attributes is characterized by the influence factor. According to the magnitude of the association degree, the influence factor is discretized as: B = {extremely large, relatively large, large, relatively small, extremely small, small, none}. Based on the defect - influencing process parameters mentioned above, we set the attribute values {x 1 , x 2 , Lx m} as:

[0054]

[0055] It can be seen from the above formula that the influencing factors of each defect include 10 attributes, and the attribute values of different defect types are different, and the influence factors are also different. Therefore, a state table of the defect set elements can be established as shown in Table 1.

[0056] Table 1 State table of defect set elements

[0057]

[0058] According to each conditional attribute in Table 1, each piece of data in the defect set is given an initial attribute value (that is, a specific discretized influence factor is given to each process parameter of each defect) and stored in the database. In this way, the initial defect set attributes are initially established.

[0059] Step 3: Construct the defect set attribute data set

[0060] Construction of the defect set attribute historical data set: Obtain the relevant parameters of the electron beam selective melting historical data, and label the attribute values in the S - rough set defect discrimination matrix according to the object, that is, give a specific discretized influence factor to each process parameter of each defect.

[0061] Step 4: Construction of a Defect Prediction Model Based on a BP Neural Network Optimized by a Genetic Algorithm

[0062] During the electron beam selective melting process, the generation of defects is often not caused by a single factor, and the change of process parameters does not only affect one type of defect. There is a complex many-to-many correlation between the two. Therefore, this part introduces a genetic algorithm to optimize the BP neural network algorithm, and its process is as follows Figure 1 . 85% of the data set is used as the training set, and the remaining 15% is used as the test set. The BP neural network optimized by the genetic algorithm is denoted as GA-BP. The steps for model construction are as follows

[0063] 1) Parameter configuration. It mainly includes the number of iterations, learning rate, and activation function of GA-BP. Among them, the activation function is the Sigmoid function

[0064] 2) Determine the network structure, including the number of input, output, and hidden layer nodes of the GA-BP algorithm

[0065] 3) Model training. Input the data for model training to determine the prediction model

[0066] 4) Predict data. Input the sample prediction index into the prediction model, and use the output of the trained model. The output value is the predicted value

[0067] 5) Optimize the model. Input the test set into the prediction model, detect the accuracy of the model, adjust the parameters in 1)-2), and repeat the process in 3) to optimize the prediction model

[0068] Step 5: Application of the Defect Prediction Model Based on the Genetic Optimization Algorithm

[0069] In step 4, a relatively feasible prediction model has been obtained. In actual application, the parameters to be processed by the electron beam selective melting process are processed according to the data in steps 1-3 to form standardized input parameters. Input the parameters into the defect prediction model, and the prediction results can be obtained. Adjust the process parameters according to the prediction results, input them into the defect prediction model after adjustment, and repeat multiple times to obtain process parameters with fewer types and quantities of defects

[0070] Step 6: Implementation of the Electron Beam Selective Melting Process and Data Collection

[0071] Implement the electron beam selective melting process according to the process parameters optimized in step 5, collect the defect data after implementation, and repeat steps 1-3 to expand the defect set data

[0072] Step 7: Iterative Optimization of the Prediction Algorithm

[0073] Return to step 4, and use the expanded defect set data to iteratively optimize the prediction model. Each time, use the iteratively optimized defect prediction model to predict defects before manufacturing, and adjust the process parameters used when implementing the electron beam selective melting process accordingly. In this way, continuously enrich the prediction model data and iteratively optimize the defect prediction model, so as to obtain more reasonable process parameters using the defect prediction model.

[0074] Example:

[0075] This example provides a prediction method for crack defects during the electron beam selective melting forming process of TA15 titanium alloy based on machine learning, including the following steps:

[0076] Step 1: Select spherical TA15 titanium alloy powder with a particle size of 53 - 105 μm, and dry it using a vacuum oven to eliminate the interference of moisture in the powder on defect formation. The oven settings are: temperature 60 - 120 °C, holding time 0.5 - 3 h.

[0077] Step 2: Load the dried TA15 titanium alloy powder into the powder bin. The powder bin has a weight detection and feedback control function, which can realize real-time control of the powder supply amount. The powder quality control accuracy should be ≤ ±1 g.

[0078] Step 3: Select a heat-resistant stainless steel substrate as the forming substrate, and the surface of the substrate should be kept flat and smooth; select high-purity helium as the backfill gas to ensure the stability of the additive manufacturing process.

[0079] Step 4: Use 3D model design software such as SolidWorks and UG to complete the design of the specimen model, and use model processing software such as Magics to add support structures;

[0080] Step 5: Establish a process parameter status table according to the types and influencing factors of defects during the previous electron beam selective melting forming process of TA15 titanium alloy

[0081] Table 2 Defect set element status table

[0082]

[0083]

[0084] Step 6: Establish a corresponding defect prediction model based on the genetic algorithm optimized BP neural network. The selection interval for the number of iterations of GA-BP is between 50 and 1000. In this implementation step, the number of iterations of GA-BP is selected as 100; the selection interval for the learning rate is 0.01 - 0.1. In this implementation step, the number of iterations of GA-BP is selected as 0.05; the activation function is selected as the Sigmoid function.

[0085] Step 7: Determine the network structure, including the number of input, output, and hidden layer nodes of the GA-BP algorithm. The number of input layer nodes is selected as 10, the number of output layer nodes is selected as 3, and the number of hidden layer nodes is selected as 65.

[0086] Step 8: Substitute the relevant data in Table 2 into the model for training to obtain the output values, perform data processing, select the process parameters for predicting hole defects, and obtain the results shown in Table 3. Print the specimens according to the prediction table and conduct metallographic observation on them. The results are as Figure 2 shown. It can be seen that there are hole defects at the near-surface position of the specimens, verifying the correctness of the model.

[0087] Table 3 Process Parameters for Predicting Hole Defects

[0088]

[0089]

[0090] It should be specifically noted that in addition to the hole defects listed in this embodiment, other related defects can also be verified using the present invention, which should be known to those skilled in the art.

[0091] The present invention has been described in detail above in combination with specific embodiments and exemplary examples, but these descriptions should not be construed as limitations on the present invention. Those skilled in the art understand that without departing from the spirit and scope of the present invention, various equivalent substitutions, modifications, or improvements can be made to the technical solutions of the present invention and their implementation manners, and all of these fall within the scope of the present invention. The protection scope of the present invention is subject to the appended claims.

[0092] The content not described in detail in the specification of the present invention belongs to the prior art well-known to those skilled in the art.

Claims

1. A method for predicting defects in electron beam selective melting based on machine learning, characterized in that: include: S1: Statistics on defects of electron beam selective melting technology; S2: According to the defects of electron beam selective melting technology, the process parameters affecting the defects are analyzed, and the initial discretized influencing factors are given for the process parameters of each defect; S3: Obtain the electron beam selective melting history data to form a standardized defect data set; S4: Construct a defect prediction model of BP neural network based on genetic algorithm, and train and optimize it based on defect data set to obtain the defect prediction model; S5: Processing the parameters to be processed by the electron beam selective melting process to form standardized input parameters consistent with the defect data set format; inputting the input parameters into the constructed defect prediction model, adjusting the process parameters according to the prediction results, and inputting the adjusted process parameters into the defect prediction model, and obtaining optimized process parameters after multiple repetitions; S6: Implementing electron beam selective melting process according to the optimized process parameters obtained in S5.

2. The method for predicting forming defects of electron beam selective melting based on machine learning according to claim 1, characterized in that: Collect defect data after each electron beam selective melting process, expand it into a defect data set according to a standardized defect data set format, and use the expanded defect data set to further iteratively optimize the defect prediction model; Steps S5 to S6 are repeated each time during manufacturing to complete defect prediction using the defect prediction model after further iterative optimization, and the process parameters used in implementing the electron beam selective melting process are adjusted according to the predicted defects.

3. The method for predicting forming defects of electron beam selective melting based on machine learning according to claim 1, characterized in that: The defects of electron beam selective melting technology include: Material limitations, holes, cracks, spheroidization, powder collapse.

4. The method for predicting defects in electron beam selective melting based on machine learning according to claim 1, characterized in that: Process parameters that affect defects include: Defect type, material type, electron beam current, scanning speed, scanning pitch, scanning path, electron beam spot diameter, substrate preheating temperature, acceleration voltage, powder layer thickness, overhang angle.

5. The method for predicting forming defects of electron beam selective melting based on machine learning according to claim 1, characterized in that: Discretization influencing factors include: B = {extremely large, relatively large, relatively small, extremely small, small, none}.

6. The method for predicting forming defects of electron beam selective melting based on machine learning according to claim 5, characterized in that: When forming a standardized defect data set: use defects and process parameters that affect defects to form an S-rough set defect resolution matrix, and use discretized influencing factors to assign values ​​to the elements of the S-rough set defect resolution matrix according to the object, that is, give a specific discretized influencing factor to each process parameter of each defect.

7. The method for predicting forming defects of electron beam selective melting based on machine learning according to claim 1, characterized in that: The BP neural network optimized by genetic algorithm is denoted as GA-BP. When constructing the defect prediction model: Divide the defect dataset into training set and test set; The network structure is determined, including the input, output and number of hidden layer nodes of GA-BP; Configure parameters, including the number of iterations, learning rate, and activation function of GA-BP. The activation function uses the Sigmoid function. Input training set data for model training and determine the trained prediction model; The test set is input into the prediction model to detect the accuracy of the model, and the network structure and parameters are further adjusted to optimize the prediction model.

8. The method for predicting defects in electron beam selective melting based on machine learning according to claim 7, characterized in that: 85% of the defect data set is used as the training set, and the remaining 15% is used as the test set.

Images