Machine learning based laser additive manufacturing shape and property active control method

Abstract

This invention provides a machine learning-based method for active shape control in laser additive manufacturing. The method includes: simulating temperature field distribution data and microstructure images perpendicular to the printing direction during the additive manufacturing process using a physical model, and preprocessing these to obtain the mechanical properties of the formed part corresponding to the microstructure images; predicting the forming quality at the next moment using a trained neural network; and when the predicted forming quality does not meet requirements, actively adjusting printing process parameters to achieve real-time active control of the macroscopic temperature field, thereby regulating the microstructure and mechanical properties of the formed part online. Compared to calculation methods that directly construct physical constitutive models, this method can significantly shorten the solution time. Furthermore, this invention establishes a temporal and spatial relationship between printing process parameters, temperature field distribution, microstructure images, and mechanical properties, which is beneficial for promoting comprehensive technological innovation in the field of additive manufacturing.

Description

Technical Field

[0001] This invention relates to the application of machine learning technology in the field of materials science, and in particular to a machine learning-based method for active control of shape properties in laser additive manufacturing. Background Technology

[0002] When 3D printing is used to manufacture aerospace components, defects such as coarse columnar grain structures, porosity, spheroidization, incomplete melting, and cracks are prone to occur, leading to fatigue failure, fracture failure, reduced service life, and ultimately unstable performance of the printed parts. To improve the safety of aerospace vehicles, in-depth research is needed on shape control strategies in the additive manufacturing process of critical aerospace metal materials. However, the additive manufacturing process involves numerous processing parameters, and these parameters, along with material parameters, often exhibit a certain probability distribution, causing significant fluctuations in the microstructure and mechanical properties of the printed parts.

[0003] In recent years, with the rapid development of machine learning technology, machine learning methods can quickly construct approximate models to approximate the true solution based on a sufficient number of representative samples (data), avoiding direct construction of the physical relationships involved. Therefore, it has significant advantages in dealing with such uncertain problems. However, traditional machine learning models cannot predict the forming performance at each moment during the printing process, and thus cannot actively adjust parameters to control product quality. The LSTM model, proposed by Jürgen Schmidhuber, uses a CNN-LSTM model to predict the forming performance at the next moment based on the current temperature field distribution and printing process parameters, thereby achieving real-time active control of printed product quality. However, there are currently no reports on using CNN-LSTM models to predict the active control of form characteristics in laser additive manufacturing.

[0004] The existing technology, titled "A Prediction and Control System and Method for Additive Manufacturing Part Forming Quality," discloses a method that uses an LSTM neural network to predict the quality of formed parts during additive manufacturing by acquiring images of three regions: the molten pool, the plume, and the spatter. However, this method only uses macroscopic factors such as the molten pool, plume, and spatter for prediction, without considering the influence of factors such as temperature and microstructure. This method requires experimental data to train the neural network model, necessitating a large amount of experimentation and material consumption. Furthermore, the macroscopic images of the molten pool, plume, and spatter provide limited information and cannot comprehensively reflect the influencing factors on the quality of formed parts during additive manufacturing. Summary of the Invention

[0005] The technical problem to be solved by this invention is to construct a machine learning-based active control method for the shape properties of laser additive manufacturing. A CNN-LSTM model is used to construct a data-driven machine learning model. By combining a multi-scale physical model, a data-driven machine learning model and experimental measurements, a data-driven machine learning simulation and analysis system based on a multi-scale physical model is built to predict the microstructure and mechanical properties at different times during the printing process, and to quickly and accurately predict the real-time quality of the product.

[0006] To address the above problems, this invention provides a machine learning-based active shape control method for laser additive manufacturing, comprising the following steps:

[0007] S1. Construct a database and use finite element method to generate temperature field distribution data during additive manufacturing under different printing process parameters; use phase field method to simulate and generate microstructure images at different locations along the vertical direction of laser printing under different printing process parameters; use crystal plasticity theory to calculate the mechanical properties corresponding to the microstructure obtained by phase field method.

[0008] S2. Process the microstructure images obtained in step S1, and calculate the average grain size and number of grains using statistical methods; group and match the temperature distribution, average grain size, number of grains and mechanical properties at different locations under different printing parameters to construct a dataset that meets the format requirements and matches the network structure.

[0009] S3. Use CNN neural network to capture spatial features in the additive manufacturing process, construct a two-dimensional feature matrix containing temporal and spatial information, establish a CNN-LSTM neural network model, use the data of the spatiotemporal feature matrix as input to train the neural network model and save the trained model, and update the state and output information of the LSTM memory module.

[0010] S4. Actively control the shape of the formed parts;

[0011] S41. Install a high-temperature colorimeter on the laser cladding head to record the temperature distribution during the printing process in real time.

[0012] S42. Introduce a quality qualification standard for molded parts, and use the predicted stress-strain curve to determine whether the quality of the molded parts is qualified.

[0013] S43. The temperature distribution and printing process parameters monitored in real time are input into the trained neural network model to predict the mechanical properties of the formed part. If the predicted stress-strain range of the formed part is within the 95% confidence interval of the stress-strain curve of the formed part quality qualification standard, the formed part is qualified and printing continues. If the predicted stress-strain curve of the formed part is not within the 95% confidence interval, it is considered unqualified. The printing process parameters are adjusted to control the quality of the formed part.

[0014] S44. To achieve active control of the shape and properties of the formed parts during the laser additive manufacturing process.

[0015] Furthermore, the specific process of generating the temperature field using finite element simulation in step S1 is as follows: set the material parameters, analysis steps, interactions, and mesh conditions using ABAQUS finite element simulation software; simulate the laser heat source using ABAQUS subroutines and calculate the temperature field in the laser additive manufacturing process under different printing process parameters.

[0016] Furthermore, the specific process of generating microstructure images along the vertical direction of laser printing using the phase-field method in step S1 is as follows: based on the existing phase-field model constructed according to the Ginzburg-Landau equation, the grain growth process in laser additive manufacturing is simulated.

[0017] Furthermore, the specific process of constructing the dataset used to train the model in step S2 includes the following steps:

[0018] S21. Divide the formed part into n positions along the printing direction and mark each position in sequence;

[0019] S22. Group the printed parts under different printing process parameters, take 10 equally spaced positions along the laser printing direction for each printed part, and then obtain the temperature distribution data, the number and average size of grains in the microstructure image, and mechanical properties of each position. Group and match to build a database for training CNN-LSTM neural network models.

[0020] Furthermore, the two-dimensional feature matrix containing time and spatial information mentioned in step S3 has the following structure: the horizontal axis represents the data of printing process parameters and temperature distribution ordered by time, and the vertical axis represents the order of the positions of the formed parts in the parallel additive manufacturing direction.

[0021] Preferably, the steps for updating the LSTM memory module state and outputting information in step S3 are as follows:

[0022] The forget gate forgets useless historical information; the output state of the forget gate is f at time t. t for:

[0023] f t =σ(Wf ·x t +U f ·h t-1 +b f )

[0024] Where σ is the logistic sigmoid function, W f U f The output state f of the forget gate at time t is represented. t The weight matrix, x t Let h represent the input vector at time t. t-1 b represents the output at time t-1. f This represents the bias vector of the forget gate;

[0025] The input gate updates its state based on the input data and historical information. At time t, the output state of the input gate is i. t for:

[0026] i t =σ(W i ·x t +U i ·h t-1 +b i )

[0027] Among them, W i U i Indicates the output state i of the input gate at time t. t The weight matrix, b i This represents the bias vector of the input gate;

[0028] Candidate memory units Represented as:

[0029]

[0030] Among them, W c U c Indicates the output state of the candidate memory cell at time t. The weight matrix, b c The bias vector representing the candidate memory cell;

[0031] Memory cell state C at time t t Represented as:

[0032]

[0033] Among them, C t-1 This represents the state of the memory cell at time t-1;

[0034] The output gate outputs the information at the current moment, and the output state at time t is o. t Represented as:

[0035] o t =σ(W o ·x t +U o ·h t-1 +b o )

[0036] Among them, W o U o Indicates the output state of the output gate at time t. t The weight matrix, b o This represents the bias vector of the output gate;

[0037] The output at time t is h. t =o t ·tanh(C t ).

[0038] Preferably, in step S3, the construction of the CNN-LSTM neural network prediction model includes setting the number of convolutional layers, pooling layers, and LSTM network layers in the CNN model, as well as adjusting the model parameters; in the CNN-LSTM model, the number of CNN convolutional layers is 1, the number of convolutional kernels is 20, the kernel size is 3*3, and the activation function is the SReLU function; the number of pooling layers is 1; the number of neurons in the LSTM layer is 50, the loss function is MAE, the backpropagation process uses the Adam optimization algorithm, the learning rate is 0.001, and the number of hidden neurons in the LSTM layer is 100.

[0039] Preferably, the basic unit of the LSTM model in the CNN-LSTM neural network prediction model is a memory module, which includes a memory unit and three gate structures that control the state of the memory unit, namely a forget gate, an input gate, and an output gate. Each input unit in the input gate adopts the tanh activation function, and the corresponding input is the spatiotemporal feature matrix constructed in step S3. Each output unit in the output gate adopts the tanh activation function, and the corresponding output is the number of grains, the average size, and the mechanical properties corresponding to the microstructure image in the microstructure image perpendicular to the laser printing direction.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] 1. This invention takes experimental costs into consideration and uses multiple physical models to simulate the data needed to train neural network models, saving both the cost and time of conducting experiments. In addition, it can simulate data that cannot be obtained under conventional experiments, providing richer data for training models and resulting in better models.

[0042] 2. This invention links printing process parameters, temperature field distribution, microstructure, and mechanical properties together, taking into account the microscopic influencing factor of microstructure, and constructing data with spatiotemporal characteristics. By combining the temperature field, microstructure, and mechanical properties at different times and locations, it considers the influence of printing process parameters, temperature field distribution, and microstructure on the quality of formed parts in the laser additive manufacturing process from both temporal and spatial perspectives. This is an advanced method for predicting the quality of formed parts in laser additive manufacturing. Attached Figure Description

[0043] Figure 1 This is a structural block diagram of the laser additive manufacturing shape active control method based on machine learning according to the present invention;

[0044] Figure 2 This is a schematic diagram of the spatiotemporal feature matrix constructed using CNN;

[0045] Figure 3 This is a diagram of the memory module structure of an LSTM neural network;

[0046] Figure 4 It is the stress-strain curve that meets the quality requirements of the formed part in additive manufacturing. Detailed Implementation

[0047] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in the present application can be combined with each other.

[0048] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings.

[0049] Figure 1 This invention discloses a machine learning-based active shape control method for laser additive manufacturing, which includes the following steps:

[0050] S1. Constructing a database: First, use finite element method (FEM) simulation to generate temperature field data during additive manufacturing under different printing process parameters. Specifically, use ABAQUS finite element simulation software to set the material parameters, analysis steps, interactions, and mesh conditions. Use ABAQUS subroutines to simulate the laser heat source and calculate the temperature field during laser additive manufacturing under different printing process parameters. Second, use the phase-field method to simulate and generate microstructure images along the vertical direction of laser printing at different locations under different printing process parameters. Based on the existing phase-field model constructed according to the Ginzburg-Landau equation, simulate the grain growth process during laser additive manufacturing. Third, use crystal plasticity theory to calculate the mechanical properties corresponding to the microstructure obtained by the phase-field method.

[0051] S2. Process the microstructure images obtained in step S1, and calculate the average grain size and number of grains using statistical methods; group and match the temperature distribution, average grain size, number of grains, and mechanical properties at different locations under different printing parameters to construct a dataset that meets the format requirements and matches the network structure.

[0052] S3. Utilize a CNN neural network to capture spatial features during the additive manufacturing process, constructing a two-dimensional feature matrix that includes both temporal and spatial information, such as... Figure 2 As shown, the horizontal axis of the matrix represents the printing process parameters and temperature distribution data ordered chronologically, while the vertical axis represents the chronological order of the formed parts in the parallel additive manufacturing direction. A CNN-LSTM neural network model is established, and the data of the spatiotemporal feature matrix is ​​used as input to train the neural network model and save the trained model. The LSTM memory module is then updated with state and information output.

[0053] Construct a CNN-LSTM neural network prediction model and train it. The training process includes the following steps:

[0054] In the CNN-LSTM model, the number of CNN convolutional layers is set to 1, the number of convolutional kernels is 20, the kernel size is 3*3, the activation function is SReLU, the number of pooling layers is 1, the number of neurons in the LSTM layer is 50, the loss function is MAE, the backpropagation process uses the Adam optimization algorithm, the learning rate is 0.001, and the number of hidden neurons in the LSTM layer is 100.

[0055] like Figure 3 As shown, the basic unit of the LSTM model includes a memory unit and three gate structures that control the state of the memory unit, namely the forget gate, the input gate, and the output gate. The input gate includes multiple input units, each of which uses the tanh activation function, and the corresponding input is the spatiotemporal feature matrix constructed in S3.

[0056] The output gate includes multiple output units, each of which uses the tanh activation function. The corresponding outputs are the number of grains, average size, and mechanical properties of the microstructure image along the direction perpendicular to the laser printing direction.

[0057] The steps for updating the state of the LSTM memory module and outputting information are as follows:

[0058] The forget gate forgets useless historical information; the output state of the forget gate is f at time t. t for:

[0059] f t =σ(W f ·x t +U f ·h t-1 +b f )

[0060] Where σ is the logistic sigmoid function, W f U f The output state f of the forget gate at time t is represented. t The weight matrix, x t Let h represent the input vector at time t. t-1 b represents the output at time t-1. f This represents the bias vector of the forget gate.

[0061] The input gate updates its state based on the input data and historical information. At time t, the output state of the input gate is i. t for:

[0062] i t =σ(W i ·x t +U i ·h t-1 +b i )

[0063] Among them, W i U i Indicates the output state i of the input gate at time t. t The weight matrix, b i This represents the bias vector of the input gate.

[0064] Candidate memory units Represented as:

[0065]

[0066] Among them, W c U c Indicates the output state of the candidate memory cell at time t. The weight matrix, b cThis represents the bias vector of the candidate memory cell.

[0067] Memory cell state C at time t t Represented as:

[0068]

[0069] Among them, C t-1 This represents the state of the memory cell at time t-1.

[0070] The output gate outputs the information at the current moment, and the output state at time t is o. t Represented as:

[0071] o t =σ(W o ·x t +U o ·h t-1 +b o )

[0072] Among them, W o U o Indicates the output state of the output gate at time t. t The weight matrix, b o This represents the bias vector of the output gate.

[0073] The output at time t is h. t =o t ·tanh(C t ).

[0074] S4. Record the temperature distribution of the formed part during the printing process, and introduce standards for whether the quality of the formed part is qualified, such as... Figure 4 As shown, the mechanical properties of the molded part are predicted using the neural network model trained in step S3, and then the shape of the molded part is actively controlled.

[0075] S41. Install a high-temperature colorimeter on the laser cladding head to record the temperature distribution during the printing process in real time.

[0076] S42. Introduce a quality standard for molded parts, and use the predicted stress-strain curve to determine whether the quality of the molded parts is qualified.

[0077] S43. The temperature distribution and printing process parameters monitored in real time are input into the trained neural network model to predict the mechanical properties of the formed part. If the predicted stress-strain range of the formed part is within the 95% confidence interval of the stress-strain curve of the formed part quality qualification standard, the formed part is qualified and printing continues. If the predicted stress-strain curve of the formed part is not within the 95% confidence interval, it is considered unqualified. The printing process parameters are adjusted to control the quality of the formed part.

[0078] S44. To achieve active control of the shape and properties of the formed parts during the laser additive manufacturing process.

[0079] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A machine learning-based active control method for the shape properties of laser additive manufacturing, characterized in that, It includes the following steps: S1. Construct a database and use finite element method to generate temperature field data during additive manufacturing under different printing process parameters; use phase field method to simulate and generate microstructure images of different locations along the vertical direction of laser printing under different printing process parameters. Mechanical properties of microstructures obtained by the phase-field method were calculated using crystal plasticity theory. S2. Process the microstructure images obtained in step S1, and calculate the average grain size and number of grains using statistical methods; group and match the temperature distribution, average grain size, number of grains and mechanical properties at different locations under different printing parameters to construct a dataset that meets the format requirements and matches the network structure. S3. Use CNN neural network to capture spatial features in the additive manufacturing process, construct a two-dimensional feature matrix containing temporal and spatial information, establish a CNN-LSTM neural network model, use the data of the spatiotemporal feature matrix as input to train the neural network model and save the trained model, and update the state and output information of the LSTM memory module. S4. Actively control the shape of the formed parts; S41. Install a high-temperature colorimeter on the laser cladding head to record the temperature distribution during the printing process in real time. S42. Introduce a quality qualification standard for molded parts, and use the predicted stress-strain curve to determine whether the quality of the molded parts is qualified. S43. The temperature distribution and printing process parameters monitored in real time are input into the trained neural network model to predict the mechanical properties of the formed part. If the predicted stress-strain range of the formed part is within the 95% confidence interval of the stress-strain curve of the formed part quality qualification standard, the formed part is qualified and printing continues. If the predicted stress-strain curve of the formed part is not within the 95% confidence interval, it is considered unqualified. The printing process parameters are adjusted to control the quality of the formed part. S44. To achieve active control of the shape and properties of the formed parts during the laser additive manufacturing process.

2. The machine learning-based active shape control method for laser additive manufacturing according to claim 1, characterized in that, The specific process of generating the temperature field using finite element simulation in step S1 is as follows: set the material parameters, analysis steps, interactions, and mesh conditions using ABAQUS finite element simulation software; use ABAQUS subroutines to simulate the laser heat source and calculate the temperature field in the laser additive manufacturing process under different printing process parameters.

3. The machine learning-based active shape control method for laser additive manufacturing according to claim 1, characterized in that, The specific process of generating microstructure images along the vertical direction of laser printing using the phase-field method in step S1 is as follows: based on the existing phase-field model constructed according to the Ginzburg-Landau equation, the grain structure growth process in laser additive manufacturing is simulated.

4. The machine learning-based active shape control method for laser additive manufacturing according to claim 1, characterized in that, The specific process of constructing the dataset used to train the model in step S2 includes the following steps: S21. Divide the formed part into n positions along the printing direction and mark each position in sequence; S22. Group the printed parts under different printing process parameters, take 10 equally spaced positions along the laser printing direction for each printed part, and then obtain the temperature distribution data, the number and average size of grains in the microstructure image, and mechanical properties of each position. Group and match to build a database for training CNN-LSTM neural network models.

5. The machine learning-based active shape control method for laser additive manufacturing according to claim 1, characterized in that, The two-dimensional feature matrix containing time and space information mentioned in step S3 has the following dimensions: the horizontal axis represents the printing process parameters and temperature distribution data ordered chronologically, and the vertical axis represents the order of the formed parts positions in the parallel additive manufacturing direction.

6. The machine learning-based active shape control method for laser additive manufacturing according to claim 1, characterized in that, The steps for updating the LSTM memory module state and outputting information in step S3 are as follows: The forget gate forgets useless historical information; the output state of the forget gate is f at time t. t for: f t =σ(W f ·x t +U f ·h t-1 +b f ) Where σ is the logistic sigmoid function, W f U f The output state f of the forget gate at time t is represented. t The weight matrix, x t Let h represent the input vector at time t. t-1 b represents the output at time t-1. f This represents the bias vector of the forget gate; The input gate updates its state based on the input data and historical information. At time t, the output state of the input gate is i. t for: i t =σ(W i ·x t +U i ·h t-1 +b i ) Among them, W i U i Indicates the output state i of the input gate at time t. t The weight matrix, b i This represents the bias vector of the input gate; Candidate memory units Represented as: Among them, W c U c Indicates the output state of the candidate memory cell at time t. The weight matrix, b c The bias vector representing the candidate memory cell; Memory cell state C at time t t Represented as: Among them, C t-1 This represents the state of the memory cell at time t-1; The output gate outputs the information at the current moment, and the output state at time t is o. t Represented as: the t =σ(W o ·x t +U o ·h t-1 +b o ) Among them, W o U o Indicates the output state of the output gate at time t. t The weight matrix, b o This represents the bias vector of the output gate; The output at time t is h. t =o t ·tanh(C t ).

7. The machine learning-based active shape control method for laser additive manufacturing according to claim 1, characterized in that, In step S3, the construction of the CNN-LSTM neural network model includes setting the number of convolutional layers, pooling layers, and LSTM network layers in the CNN model, as well as adjusting the model parameters. In the CNN-LSTM model, the number of CNN convolutional layers is 1, the number of convolutional kernels is 20, the kernel size is 3*3, and the activation function is the SReLU function. The number of pooling layers is 1. The number of neurons in the LSTM layer is 50, the loss function is MAE, the backpropagation process uses the Adam optimization algorithm, the learning rate is 0.001, and the number of hidden neurons in the LSTM layer is 100.

8. The machine learning-based active shape control method for laser additive manufacturing according to claim 1, characterized in that, The basic unit of the LSTM model in the CNN-LSTM neural network model is the memory module, which includes a memory unit and three gate structures that control the state of the memory unit, namely the forget gate, the input gate, and the output gate. Each input unit in the input gate adopts the tanh activation function, and the corresponding input is the spatiotemporal feature matrix constructed in step S3. Each output unit in the output gate adopts the tanh activation function, and the corresponding output is the number of grains, the average size, and the mechanical properties of the microstructure image along the direction perpendicular to the laser printing direction.

Images