A method for monitoring the porosity of different materials in laser powder bed fusion additive manufacturing

Abstract

This invention proposes a method for monitoring the porosity of different materials in laser powder bed fusion additive manufacturing, which relates to the field of additive manufacturing. The method comprises: obtaining source domain data and target domain data; constructing a pre-trained model, including a feature extraction module and a first classification and recognition module; iteratively training the pre-trained model to obtain a first parameter for the feature extraction module; locking the first parameter, adjusting the first classification and recognition module to obtain a second classification and recognition module, and constructing a fine-tuning model based on the feature extraction module and the second classification and recognition module of the pre-trained model; iteratively training the fine-tuning model; obtaining target domain data to be monitored, and using the trained fine-tuning model to identify and classify the data to be monitored to obtain a classification result. The method proposed in this invention improves the recognition accuracy of porosity monitoring of parts made of different materials, fully utilizes historical data, and avoids resource waste.

Description

Technical Field

[0001] The present invention relates to the field of additive manufacturing, and in particular to a method for monitoring the porosity of different materials in laser powder bed melting additive manufacturing. Background Art

[0002] Laser powder bed fusion (LPBF) additive manufacturing technology is capable of printing a variety of different types of metal powder materials, but existing porosity monitoring methods mostly establish monitoring models for a single printing material, such as titanium alloys, nickel-based high-temperature alloys, stainless steel, etc. However, different powder materials have different material properties. Even if the same process parameters are used, different types of porosity defects may be generated when printing with different materials, and the collected sensor signals will also change due to changes in materials, process conditions, and defect formation. This means that the monitoring model established under the printing conditions of one material cannot be directly used to monitor another material, and the generalization ability of the monitoring model is poor.

[0003] To address the challenge of monitoring porosity in different printing materials, existing methods often rely on conducting separate process experiments to obtain sufficient labeled sensor data for each material and then re-modeling the model. For example, Drissi-Daoudi et al. combined acoustic signals collected by a microphone with a machine learning algorithm to monitor porosity defects in three materials during the LPBF process. They achieved a classification accuracy exceeding 86% when identifying the three different materials, demonstrating that the acoustic emission signal characteristics are related to the interaction between the laser and the material, and proposed a method capable of classifying different materials. This technology was published in the journal "VIRTUAL AND PHYSICAL PROTOTYPING," titled "Differentiation of materials and laser powder bed fusion processing regimes from airborne acoustic emission combined with machine learning." However, this existing technology relies on experimentally re-acquiring a large amount of labeled data to build the monitoring model. This is time-consuming and requires a large amount of data. Furthermore, historical material resources cannot be fully utilized, resulting in resource waste. Summary of the Invention

[0004] In view of this, the present invention proposes a method for monitoring the porosity of different materials in laser powder bed fusion additive manufacturing. By organically combining the methods and ideas of transfer learning with the characteristics of metal materials, a simple and efficient pre-training model is designed. The historical data of the source domain is fully utilized to train the model, and then the pre-trained model is transferred to improve the generalization performance of the model. The obtained fine-tuning model has very excellent performance and high accuracy in monitoring the porosity in the target domain, so as to solve the technical problems in the existing technology that the porosity monitoring model of different materials has poor generalization ability and historical data resources are not fully utilized.

[0005] The technical solution of the present invention is achieved as follows: The present invention provides a method for monitoring the porosity of different materials in laser powder bed fusion additive manufacturing, comprising:

[0006] S1 obtains source domain data and target domain data, and divides the source domain data and target domain data into a source domain training dataset, a source domain test dataset, a target domain training dataset, and a target domain test dataset. Both the source domain data and the target domain data contain true labels.

[0007] S2 builds a pre-training model, including a feature extraction module and a first classification recognition module;

[0008] S3 randomly initializes all parameters of the pre-trained model, inputs the source domain training dataset into the pre-trained model for iterative training, and tests the trained pre-trained model using the source domain test dataset until the pre-trained model reaches a first expected accuracy, thereby obtaining first parameters of the feature extraction module;

[0009] S4 locks the first parameter and adjusts the first classification and recognition module to obtain a second classification and recognition module. The feature extraction module based on the pre-trained model and the second classification and recognition module constitute a fine-tuning model;

[0010] S5 randomly initializes the parameters of the second classification recognition module, inputs the target domain training data set into the fine-tuning model for iterative training, and uses the target domain test data set to test the trained fine-tuning model until the fine-tuning model reaches the second expected accuracy, thereby obtaining a trained fine-tuning model;

[0011] S6 obtains the data to be monitored in the target domain, uses the trained fine-tuning model to monitor the data to obtain the monitoring results.

[0012] Based on the above technical solution, preferably, step S1 includes:

[0013] Design multiple sets of process experiments with different materials, including source domains and target domains. Conduct process experiments based on a laser powder bed melting monitoring platform to obtain source domain data and target domain data. The source domain is 316L stainless steel, and the target domain is TC4 titanium alloy.

[0014] The true labels of the source and target domain data are obtained based on the porosity and density measurement results of the source and target domain materials. The porosity results are obtained by metallographic microscope observation, and the density results are obtained by Archimedean method measurement.

[0015] The source domain data is divided into a source domain training dataset and a source domain test dataset in a ratio of 4:1, and the target domain data is divided into a target domain candidate dataset and a target domain test dataset in a ratio of 4:1. A preset proportion of data is selected from the target domain candidate dataset as the target domain training dataset.

[0016] Based on the above technical solution, preferably, the true label of the source domain data is calculated as follows:

[0017]

[0018] Where y s is the true label of the source domain data, Indicates that the true label of the source domain data is of low quality, Indicates that the true label of the source domain data is of medium quality, Indicates that the true label of the source domain data is of high quality, n s is the porosity of the source material, is the first porosity threshold of the source domain material, is the second threshold of the porosity of the source material, ρ s is the density of the source material, is the first density threshold of the source domain material, is the second density threshold of the source domain material;

[0019] The true label of the target domain data is calculated as follows:

[0020]

[0021] Where y t is the true label of the target domain data, Indicates that the true label of the target domain data is of low quality, Indicates that the true label of the target domain data is of medium quality, Indicates that the true label of the target domain data is high quality, n t is the porosity of the target domain material, is the first porosity threshold of the target domain material, is the second threshold of the porosity of the target domain material, ρ t is the density of the target domain material, is the first density threshold of the target domain material, is the second density threshold of the target domain material.

[0022] On the basis of the above technical solution, preferably, the laser powder bed melting monitoring platform includes three in-situ sensors and laser powder bed melting technology processing equipment. The three in-situ sensors are an industrial camera, a microphone and a photodiode. Among them, the industrial camera is located outside the laser powder bed melting technology processing equipment, and the microphone and the photodiode are located inside the laser powder bed melting technology processing equipment.

[0023] Based on the above technical solution, preferably, the source domain data and the target domain data are both fused image signals, acoustic signals or photodiode signals, wherein:

[0024] The fusion image signal is acquired based on industrial cameras;

[0025] The acoustic signal is obtained based on microphone sensing;

[0026] The photodiode signal is collected based on the photodiode.

[0027] On the basis of the above technical solutions, preferably, the laser powder bed melting monitoring platform is provided with a high-level trigger. During the process of different material process experiments, the high-level trigger automatically triggers the industrial camera, microphone and photodiode to automatically acquire source domain data and target domain data.

[0028] On the basis of the above technical solution, preferably, in step S2, the feature extraction module includes a 5*5 convolution layer and a maximum pooling layer, and the first classification recognition module includes a fully connected layer, an activation function layer and a random dropout layer, wherein the random dropout layer is provided after the activation function layer;

[0029] Correspondingly, in step S4, the second classification recognition module includes a fully connected layer, a batch normalization layer and an activation function layer, wherein the batch normalization layer is arranged before the activation function layer.

[0030] More preferably, step S3 includes:

[0031] S31 randomly initializes all parameters of the pre-trained model;

[0032] S32 optimizes the learning rate using the Adam optimizer, iteratively trains the pre-trained model according to the source domain training dataset, extracts a first feature using the feature extraction module, then classifies and identifies the first feature according to the first classification and recognition module to obtain a predicted label of the source domain training dataset, and adjusts the parameters of the pre-trained model using the source domain test dataset until the pre-trained model reaches a first expected accuracy, thereby obtaining first parameters of the feature extraction module;

[0033] Among them, the first feature is the domain invariant feature. During training, the first loss function is introduced to calculate the gap between the predicted label of the source domain training dataset and the true label of the source domain training dataset.

[0034] More preferably, step S5 includes:

[0035] S51 randomly initializes the parameters of the second classification recognition module;

[0036] S52 optimizes the learning rate using the Adam optimizer, iteratively trains the fine-tuning model based on the target domain training dataset, extracts a second feature using the feature extraction module, then classifies and identifies the second feature using the second classification and recognition module to obtain a predicted label for the target domain training dataset, and adjusts the parameters of the fine-tuning model using the target domain test dataset until the fine-tuning model reaches the second expected accuracy, thereby obtaining a trained fine-tuning model;

[0037] During training, the first parameter of the feature extraction module is kept unchanged, and a second loss function is introduced to calculate the gap between the predicted labels of the target domain training dataset and the true labels of the target domain training dataset.

[0038] More preferably, the amount of data in the target domain training dataset is smaller than the amount of data in the target domain test dataset.

[0039] The present invention has the following beneficial effects compared to the prior art:

[0040] (1) By organically combining the methods and ideas of transfer learning with the characteristics of metal materials, a simple and efficient pre-training model is designed. The model is trained by making full use of historical data in the source domain. The pre-trained model is then transferred to improve the generalization performance of the model. The resulting fine-tuning model has excellent performance and high accuracy in monitoring porosity in the target domain.

[0041] (2) By adjusting the network structure of the second classification recognition module, the random inactivation layer is replaced by a batch normalization layer, and the layer is moved before the activation function layer. The batch normalization layer can normalize each batch of data to the same distribution, and a larger learning rate can be selected during training, which accelerates the convergence speed of the fine-tuning model, increases the robustness of the fine-tuning model, and reduces overfitting;

[0042] (3) By setting a high-level trigger, the three types of sensor data at each stage can be collected more timely and accurately, and the manual input is reduced;

[0043] (4) Both the pre-trained model and the fine-tuned model can be used for monitoring and classification of corresponding materials, avoiding waste of resources and making the most of historical material experimental data. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0046] Figure 2 A technical flow chart of an embodiment of the present invention;

[0047] Figure 3 This is a schematic structural diagram of a laser powder bed melting monitoring platform according to an embodiment of the present invention;

[0048] Figure 4 This is a network structure diagram of the pre-training model and fine-tuning model according to an embodiment of the present invention;

[0049] Figure 5 Schematic diagram of migration results based on three types of sensor data in an embodiment of the present invention;

[0050] Figure 6 Schematic diagram of the comparison of classification accuracy of different methods in the embodiments of the present invention. DETAILED DESCRIPTION

[0051] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0052] like Figure 1 As shown, the present invention provides a method for monitoring the porosity of different materials in laser powder bed fusion additive manufacturing, comprising:

[0053] S1 obtains source domain data and target domain data, and divides the source domain data and target domain data into a source domain training dataset, a source domain test dataset, a target domain training dataset, and a target domain test dataset. Both the source domain data and the target domain data contain true labels.

[0054] S2 builds a pre-training model, including a feature extraction module and a first classification recognition module;

[0055] S3 randomly initializes all parameters of the pre-trained model, inputs the source domain training dataset into the pre-trained model for iterative training, and tests the trained pre-trained model using the source domain test dataset until the pre-trained model reaches a first expected accuracy, thereby obtaining first parameters of the feature extraction module;

[0056] S4 locks the first parameter and adjusts the first classification and recognition module to obtain a second classification and recognition module. The feature extraction module based on the pre-trained model and the second classification and recognition module constitute a fine-tuning model;

[0057] S5 randomly initializes the parameters of the second classification recognition module, inputs the target domain training data set into the fine-tuning model for iterative training, and uses the target domain test data set to test the trained fine-tuning model until the fine-tuning model reaches the second expected accuracy, thereby obtaining a trained fine-tuning model;

[0058] S6 obtains the data to be monitored in the target domain, uses the trained fine-tuning model to monitor the data to obtain the monitoring results.

[0059] like Figure 2As shown, the specific technical implementation process of an embodiment of the present invention is as follows: 1) designing an experimental method to obtain source domain data and target domain data, wherein the source domain is material one, which can be specifically 316L stainless steel material, and the target domain is material two, which can be specifically TC4 titanium alloy material; 2) dividing the source domain data and the target domain data into training data sets and test data sets respectively, wherein when the source domain data is divided and selected, the number of its training data sets should be greater than the number of its test data sets, so as to fully train the model and obtain a pre-trained model with higher accuracy, and when the target domain data is divided and selected, since the target domain data is overall less, In order to maximize the performance of the model under the most limited resource environment, the number of training data sets in the target domain is set to be smaller than the number of test data sets; 3) The pre-trained model is trained using the source domain training data set, and the first loss function is introduced to calculate the loss between the predicted label and the true label. Then, the pre-trained model is adjusted using the source domain test data set to make the model reach the first expected accuracy. After the model is trained, the first parameter of the feature extraction module is locked; 4) The pre-trained model is migrated, that is, the structure of the feature extraction module and its first parameter remain unchanged, and the structure of the first classification recognition module is adjusted to obtain a fine-tuning model. At this time, the fine-tuning model and the pre-training model share the parameters of the feature extraction module, that is, there are domain-invariant features that can be transferred between the source domain and the target domain, and model migration is achieved on this basis; 5) The second classification recognition module in the fine-tuning model adjusts the network structure, specifically replacing the random inactivation layer with a batch normalization layer, and moving the layer before the activation function layer. The batch normalization layer can normalize each batch of data to the same distribution, and a larger learning rate can be selected during training, which speeds up the convergence of the model, increases the robustness of the model, and reduces overfitting; 6) Use a small amount of target domain training The fine-tuning model is trained using the dataset, and a second loss function is introduced to calculate the loss between the predicted label and the true label. The fine-tuning model is then adjusted using the target domain test dataset. It should be noted that during the entire training process, only the second classification and recognition module is adjusted, while the parameters of the feature extraction module remain unchanged until the fine-tuning model reaches the second expected accuracy, thereby obtaining the trained fine-tuning model; 7) the trained fine-tuning model is used to monitor the porosity of the target domain data to obtain the monitoring results. At the same time, the pre-trained model trained before can also monitor the porosity of the source domain data to ensure that materials, data, and models are fully utilized.

[0060] Specifically, in one embodiment of the present invention, step S1 includes:

[0061] Design multiple sets of material processing experiments, including source domains and target domains. Conduct material processing experiments based on a laser powder bed melting monitoring platform to obtain source domain data and target domain data. The source domain is 316L stainless steel, and the target domain is TC4 titanium alloy.

[0062] The true labels of the source and target domain data are obtained based on the porosity and density measurement results of the source and target domain materials. The porosity results are obtained by metallographic microscope observation, and the density results are obtained by Archimedean method measurement.

[0063] The source domain data is divided into a source domain training dataset and a source domain test dataset in a ratio of 4:1, and the target domain data is divided into a target domain candidate dataset and a target domain test dataset in a ratio of 4:1. A preset proportion of data is selected from the target domain candidate dataset as the target domain training dataset.

[0064] First, it is necessary to build a laser powder bed melting monitoring platform. Figure 3 Schematic diagram of the overall structure of the laser powder bed fusion monitoring platform, where LPBF refers to laser powder bed fusion technology. The LPBF equipment is used for additive manufacturing process experiments. In this embodiment, the LPBF equipment is a metal 3D printer. The complete laser powder bed fusion monitoring platform includes three in-situ sensors, laser powder bed fusion technology processing equipment, a data acquisition module and a monitoring system. Specifically, the data acquisition module includes two data acquisition cards, the monitoring system is a computer, and the three in-situ sensors are used to collect data and transmit the data to the monitoring system, which monitors the results.

[0065] The LPBF equipment specifically includes a fiber laser, a powder bin, a powder feeding system, a molding platform, a scraper, a powder recovery bin, and a scanning galvanometer. The three in-situ sensors are an industrial camera, a photodiode, and a microphone. There is a molding cabin door at the front of the LPBF equipment, and windows at the rear and top. The industrial camera is located outside the LPBF equipment, specifically placed near the upper window. The upper window is made of glass. The industrial camera collects melting image signals layer by layer through the upper glass window. The pixel size is 3.69μm, and the image resolution is 3384×2710 pixels. It can collect melting image signals of the entire processing area. The photodiode and microphone are located inside the LPBF equipment, specifically placed on a magnetic base bracket. The microphone has a sensing frequency range of 3.15Hz–20kHz and a sensitivity of 50mV / Pa. It is used to collect acoustic signals. The photodiode can collect photodiode signals in the 500–1700nm band released by a circular area with a diameter of 3mm.

[0066] Specifically, after building a laser powder bed melting monitoring platform, multiple sets of different material process experiments, including active and target domains, were designed. The materials were fed into the laser powder bed melting monitoring platform for 3D printing. During each layer of printing, acoustic signals and photodiode signals were collected. After each layer of part processing was completed, a melting image signal was collected. After printing, the printed parts of the material were measured for porosity and density. The porosity results were obtained through metallographic microscope observation, and the density results were obtained through Archimedes method measurement. The results for porosity and density were divided into low quality, medium quality, and high quality. Combining the porosity and density results, the three categories of low quality, medium quality, and high quality were used as the true labels for the corresponding printed parts and sensor data.

[0067] Specifically, in one embodiment of the present invention, the true label of the source domain data is calculated as follows:

[0068]

[0069] Where y s is the true label of the source domain data, Indicates that the true label of the source domain data is of low quality, Indicates that the true label of the source domain data is of medium quality, Indicates that the true label of the source domain data is of high quality, n s is the porosity of the source material, is the first porosity threshold of the source domain material, is the second threshold of the porosity of the source material, ρ s is the density of the source material, is the first density threshold of the source domain material, is the second density threshold of the source domain material.

[0070] In a specific example of the present invention, 0.6%, 0.1%, 98% It is 99%.

[0071] Specifically, in one embodiment of the present invention, the true label of the target domain data is calculated as follows:

[0072]

[0073] Where y t is the true label of the target domain data, Indicates that the true label of the target domain data is of low quality, Indicates that the true label of the target domain data is of medium quality, Indicates that the true label of the target domain data is high quality, n t is the porosity of the target domain material, is the first porosity threshold of the target domain material, is the second threshold of the porosity of the target domain material, ρ t is the density of the target domain material, is the first density threshold of the target domain material, is the second density threshold of the target domain material.

[0074] In a specific example of the present invention, 0.2%, 0.1%, 98% It is 99%.

[0075] It should be noted that due to the inconsistency of the material properties of the source domain material and the target domain material, when determining the first and second thresholds of porosity and density, they will be determined based on the actual material properties. Therefore, the first porosity threshold, the second porosity threshold, the first density threshold, and the second density threshold corresponding to the source domain and the target domain may be the same or different.

[0076] In one embodiment of the present invention, the laser powder bed fusion monitoring platform is also equipped with a high-level trigger. During and after each layer of part processing, the high-level trigger automatically triggers the industrial camera, microphone, and photodiode to collect fusion image signals, acoustic signals, and photodiode signals. The microphone collects data via the NI-9218 data acquisition card, while the photodiode collects data via the NI-9221 data acquisition card. The sampling frequency is 10 kHz, and data is collected every ten layers. By providing this high-level trigger, this embodiment allows for more timely and accurate collection of the three sensor data at each stage, while reducing manual effort.

[0077] Specifically, both the source and target domain data are melt image signals, acoustic signals, or photodiode signals. When determining the source and target domain data, the data collected by one in-situ sensor is selected. For example, if the source domain data is the melt image signal of 316L stainless steel, the target domain data is the melt image signal of TC4 titanium alloy. The true labels of the three sensor data for the same source or target material are shared. For example, in a set of material processing experiments, if the true label of the 316L stainless steel printing material is "high quality," then regardless of which source domain's melt image signal, acoustic signal, or photodiode signal is selected as the source domain data, its true label is always "high quality."

[0078] After selecting a type of sensor data as source domain data and target domain data, the source domain data is divided into a source domain training dataset and a source domain test dataset in a ratio of 4:1, and the target domain data is divided into a target domain candidate dataset and a target domain test dataset in a ratio of 4:1. A preset ratio of data is selected from the target domain candidate dataset as the target domain training dataset. The preset ratio can be 0.5%, 1%, 2.5%, or 5%. The number of source domain training datasets is greater than the number of source domain test datasets, and the number of target domain training datasets is less than the number of target domain test datasets. Since one of the purposes of the present invention is to achieve model migration and generalization of the model, so that it can be migrated from the source domain to the target domain at a limited cost and can achieve better recognition results in the target domain, when preparing data, sufficient source domain data is obtained to ensure the accuracy of the pre-trained model. In the target domain data, in order to achieve generalization of small sample training, historical data information is fully mined and only a small amount of data is selected for model training, so that the model can converge quickly and ensure accuracy.

[0079] Specifically, in one embodiment of the present invention, step S2 includes:

[0080] The feature extraction module includes a 5*5 convolution layer and a maximum pooling layer. The first classification recognition module includes a fully connected layer, an activation function layer and a random dropout layer, wherein the random dropout layer is arranged after the activation function layer.

[0081] like Figure 4 As shown, in one embodiment of the present invention, the feature extraction module has 4 5*5 convolutional layers and 4 maximum pooling layers. The convolutional layers and the maximum pooling layers are alternately connected and serve to extract shallow and general features, i.e., the first features, from the source domain data. Each convolution operation has a convolution kernel with a step size of 1 and a size of 5×5, and the maximum pooling layer has a step size of 2 and a size of 2×2. The number of convolution kernels of the four convolutional layers is 16, 32, 64, and 64 respectively according to the data flow direction. The first classification recognition module includes a first fully connected layer, an activation function layer (ReLU), a random dropout layer (Dropout), and a second fully connected layer according to the data flow direction, which is used to output the classification results. The number of neurons in the first fully connected layer is 256, the number of neurons in the second fully connected layer is 3, and the dropout probability of the dropout layer is set to 0.5.

[0082] Specifically, in one embodiment of the present invention, step S3 includes:

[0083] S31 randomly initializes all parameters of the pre-trained model;

[0084] S32 optimizes the learning rate using the Adam optimizer, iteratively trains the pre-trained model according to the source domain training dataset, extracts a first feature using the feature extraction module, then classifies and identifies the first feature according to the first classification and recognition module to obtain a predicted label of the source domain training dataset, and adjusts the parameters of the pre-trained model using the source domain test dataset until the pre-trained model reaches a first expected accuracy, thereby obtaining first parameters of the feature extraction module;

[0085] Among them, the first feature is the domain invariant feature. During training, the first loss function is introduced to calculate the gap between the predicted label of the source domain training dataset and the true label of the source domain training dataset.

[0086] Specifically, the Adam optimizer formula is as follows:

[0087] m t =β1×m t-1 +(1-β1)×g t

[0088]

[0089]

[0090]

[0091]

[0092] Where, the two hyperparameters β1 and β2 are the exponential decay rates of the first-order moment estimate and the second-order moment estimate of the gradient, respectively, which play the role of weight distribution and the influence of the gradient square; g t represents the gradient at time t; m t and v t They are the first-order moment estimation and second-order moment estimation of the gradient respectively; t represents the time step; and They are respectively for m t and v t The paranoid correction result of θ t Represents the value of the optimizer output parameter at time t; η t is the learning rate; ε represents a very small number to prevent the denominator from being 0.

[0093] The parameters at the beginning of training are set as follows: β1 is 0.9, β2 is 0.999, bach size is 32, the number of epochs during training is 30, and the learning rate is set to 0.0001.

[0094] Specifically, the formula of the first loss function L1 is as follows:

[0095]

[0096] Where x represents the value after LogSoftMax operation, class1 represents the true label of the source domain training dataset, and j1 represents the predicted label of the source domain training dataset.

[0097] In this embodiment, the first feature is a domain-invariant feature. After the training of the pre-trained model is completed, the feature extraction module has greatly improved the extraction performance of the domain-invariant feature. By freezing its first parameter and using it to extract the features of the target domain data, a more representative domain-invariant feature, i.e., the second feature, can still be obtained, so as to realize model migration and improve the generalization performance of the model.

[0098] In this embodiment, when the first loss function tends to a stable value, the iterative training of the pre-trained model using the source domain training dataset is stopped, and the accuracy of the model is evaluated using the source domain test dataset. The first expected accuracy can be that the accuracy of the model classification result reaches 90%, or it can be set to other values ​​or other evaluation criteria, depending on the actual situation.

[0099] It should be noted that if the accuracy of the pre-trained model does not reach the first expected accuracy during testing, the parameter settings during initial training are adjusted, and iterative training and testing are performed again until the pre-trained model reaches the first expected accuracy.

[0100] Specifically, in step S4, the fine-tuning model is composed of a feature extraction module and a second classification recognition module in the pre-trained model. The feature extraction module and its first parameter remain unchanged. The second classification recognition module includes a fully connected layer, a batch normalization layer and an activation function layer, wherein the batch normalization layer is arranged before the activation function layer.

[0101] like Figure 4 As shown in FIG, in one embodiment of the present invention, the feature extraction module of the fine-tuning model is the same as the feature extraction module of the pre-training model. The second classification recognition module includes a first fully connected layer, a batch normalization layer (BN layer), an activation function layer (ReLU), and a second fully connected layer according to the data flow direction, and is used to output the classification result. The number of neurons in the first fully connected layer is 256, and the number of neurons in the second fully connected layer is 3. That is, the second classification recognition module is a fine-tuning based on the first classification recognition module, specifically replacing the random dropout layer (Dropout) with a batch normalization layer (BN layer), and placing it before the activation function layer.

[0102] Specifically, step S5 includes:

[0103] S51 randomly initializes the parameters of the second classification recognition module;

[0104] S52 optimizes the learning rate using the Adam optimizer, iteratively trains the fine-tuning model based on the target domain training dataset, extracts a second feature using the feature extraction module, then classifies and identifies the second feature using the second classification and recognition module to obtain a predicted label for the target domain training dataset, and adjusts the parameters of the fine-tuning model using the target domain test dataset until the fine-tuning model reaches the second expected accuracy, thereby obtaining a trained fine-tuning model;

[0105] During training, the first parameter of the feature extraction module is kept unchanged, and a second loss function is introduced to calculate the gap between the predicted labels of the target domain training dataset and the true labels of the target domain training dataset.

[0106] In this embodiment, the fine-tuning model also uses the Adam optimizer to optimize the learning rate during training, and the optimizer formula is the same as that in the pre-training model.

[0107] In this embodiment, the fine-tuning model maintains the first parameter of the feature extraction module unchanged throughout the training and testing process, and only adjusts the parameters of the second classification recognition module. At the beginning of training, the parameters of the second classification recognition module are set to: β1 is 0.9, β2 is 0.999, the bach size is 32, the number of epochs during training is 30, and the learning rate is set to 0.001. Because the batch normalization layer can normalize each batch of data to the same distribution, a higher learning rate can be set, which will greatly reduce the training time of the fine-tuning model and improve the robustness of the fine-tuning model.

[0108] Specifically, the formula of the second loss function L2 is as follows:

[0109]

[0110] Where x represents the value after LogSoftMax operation, class2 represents the true label of the target domain training dataset, and j2 represents the predicted label of the target domain training dataset.

[0111] It should be noted that in the embodiments of the present invention, both the pre-trained model and the fine-tuned model use the Adam optimizer to optimize the learning rate. However, in other embodiments, other methods can also be used to optimize the learning rate, for example, using the AdamW optimizer combined with the cosine annealing algorithm to optimize the learning rate. The present invention is not limited to this.

[0112] In this embodiment, when the second loss function approaches a stable value, iterative training of the fine-tuning model using the target domain training dataset is stopped, and the model accuracy is evaluated using the target domain test dataset. The second expected accuracy can be 85% accuracy of the model classification results, or other values ​​or evaluation criteria can be set, depending on the specific situation. The first expected accuracy and the second expected accuracy can be the same or different.

[0113] It should be noted that if the accuracy of the fine-tuning model during testing does not reach the second preset accuracy, the parameter settings of the second classification and recognition module during the initial training are adjusted, and the iterative training and testing are re-performed. During the re-iterative training and testing, the feature extraction module and its first parameters are still kept unchanged until the fine-tuning model reaches the second expected accuracy.

[0114] In this embodiment, the trained fine-tuning model can be used to monitor the porosity of the target domain data. At the same time, the pre-trained model trained in the above step S3 can also be directly applied to the porosity monitoring of the source domain data to ensure full utilization of resources.

[0115] See also Figure 5 , Figure 5 The comparison of training time and classification accuracy before and after migration when the source domain data and target domain data are respectively molten image signal (a), acoustic signal (b), and photodiode signal (c). It should be noted that Figure 5 The x-axis scale range of each bar graph in refers to the preset ratio of the target domain training dataset to the target domain candidate dataset, that is, the performance comparison of the two models when the preset ratio is 0.5%, 1%, 2%, and 5%. Figure 5 As shown in (a), when the source domain data is a molten image signal, i.e., the molten image signal of 316L stainless steel material, and the target domain data is also a molten image signal, i.e., the molten image signal of TC4 titanium alloy, and the preset ratio is 1%, before migration, the classification accuracy of the pre-trained model for the molten image signal of the target domain TC4 titanium alloy is 78.43%, while after migration, the classification accuracy of the fine-tuned model for the molten image signal of the target domain TC4 titanium alloy is 90.3%. Figure 5 As shown in (b), when the source domain data is an acoustic signal, i.e., the acoustic signal of 316L stainless steel material, and the target domain data is also an acoustic signal, i.e., the acoustic signal of TC4 titanium alloy, and the preset ratio is 1%, before migration, the classification accuracy of the pre-trained model for the acoustic signal of the target domain TC4 titanium alloy is 72.15%, while after migration, the classification accuracy of the fine-tuned model for the acoustic signal of the target domain TC4 titanium alloy is 87.93%. Figure 5As shown in (c), when the source domain data is a photodiode signal, i.e., the photodiode signal of 316L stainless steel material, and the target domain data is also a photodiode signal, i.e., the photodiode signal of TC4 titanium alloy, and the preset ratio is 1%, before migration, the classification accuracy of the pre-trained model for the photodiode signal of the target domain TC4 titanium alloy is 71.54%, while after migration, the classification accuracy of the fine-tuned model for the photodiode signal of the target domain TC4 titanium alloy is 87.76%. As for the training time, as Figure 5 As shown in Figure 3, the training time for each preset ratio is shorter after migration than before migration.

[0116] The migration method and model structure proposed in this invention are complementary to each other. In order to demonstrate the advantages of their organic combination, this invention is compared with the existing LeNet-5, VGG16, ResNet18 and GoogleNet networks. The comparison results are shown in the figure. Figure 6 shown.

[0117] from Figure 6 It can be seen that the combination of the migration method and model structure proposed in the present invention shows the best performance. Before using the migration method for model migration, the pre-trained model of the present invention is not the best performance. For example, when the sensor data is a molten image signal, GoogleNet has a higher classification accuracy, and when the sensor data is an acoustic signal and a photodiode signal, GoogleNet and VGG16 have higher classification accuracy, respectively. After using the migration method of the present invention, GoogleNet, VGG16 and ResNet18 all showed a negative migration phenomenon. This is because when the deep network model extracts features, it extracts more domain change features, and these feature information is not conducive to the porosity monitoring of the target domain material.

[0118] Through the above comparative experiments, it can be seen that the present invention proposes a method for monitoring the porosity of different materials in laser powder bed melting additive manufacturing. By organically combining the methods and ideas of transfer learning with the characteristics of metal materials, a simple and efficient pre-training model is designed. At the same time, by fine-tuning the classification and recognition module, a fine-tuning model after migration is obtained, which reduces the model training parameters and improves the model training efficiency. The fine-tuning model has very superior performance, and the monitoring method makes full use of historical data, realizes the reuse of historical data, and improves the generalization ability of the model. The method of the present invention can lay the foundation for online monitoring and control of porosity in laser powder bed melting additive manufacturing, which is of great significance to improving the quality of parts.

[0119] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. 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 monitoring the porosity of different materials in laser powder bed fusion additive manufacturing, characterized in that: include: S1: Obtain source domain data and target domain data, and divide the source domain data and target domain data into a source domain training dataset, a source domain test dataset, a target domain training dataset, and a target domain test dataset. Both the source domain data and the target domain data contain true labels. Step S1 includes: designing multiple sets of material process experiments including active domains and target domains, performing the material process experiments based on a laser powder bed melting monitoring platform, and obtaining source domain data and target domain data, wherein the source domain is 316L stainless steel material and the target domain is TC4 titanium alloy material; The true labels of the source and target domain data are obtained based on the porosity and density measurement results of the material process experiment. The porosity is obtained by metallographic microscope observation, and the density measurement results are obtained by Archimedean method. S2 builds a pre-training model, including a feature extraction module and a first classification recognition module; S3 randomly initializes all parameters of the pre-trained model, inputs the source domain training dataset into the pre-trained model for iterative training, and tests the trained pre-trained model using the source domain test dataset until the pre-trained model reaches a first expected accuracy, thereby obtaining first parameters for the feature extraction module; S4 locks the first parameter and adjusts the first classification and recognition module to obtain a second classification and recognition module. The feature extraction module based on the pre-trained model and the second classification and recognition module constitute a fine-tuning model; S5 randomly initializes the parameters of the second classification recognition module, inputs the target domain training dataset into the fine-tuning model for iterative training, and tests the trained fine-tuning model using the target domain test dataset until the fine-tuning model reaches the second expected accuracy, thereby obtaining a trained fine-tuning model; S6 obtains the data to be monitored in the target domain, uses the trained fine-tuning model to monitor the data to obtain the monitoring results.

2. The method according to claim 1, wherein Step S1 includes: The source domain data is divided into a source domain training dataset and a source domain test dataset in a ratio of 4:1, and the target domain data is divided into a target domain candidate dataset and a target domain test dataset in a ratio of 4:

1. A preset proportion of data is selected from the target domain candidate dataset as the target domain training dataset.

3. The method according to claim 2, wherein The true label of the source domain data is calculated as follows: ； Where, is the true label of the source domain data, Indicates that the true label of the source domain data is of low quality, Indicates that the true label of the source domain data is of medium quality, Indicates that the true label of the source domain data is of high quality, is the porosity of the source material, is the first porosity threshold of the source domain material, is the second threshold of the porosity of the source domain material, is the density of the source material, is the first density threshold of the source domain material, is the second density threshold of the source domain material; The true label of the target domain data is calculated as follows: ； Where, is the true label of the target domain data, Indicates that the true label of the target domain data is of low quality, Indicates that the true label of the target domain data is of medium quality, Indicates that the true label of the target domain data is of high quality, is the porosity of the target domain material, is the first porosity threshold of the target domain material, is the second threshold of the porosity of the target domain material, is the density of the target domain material, is the first density threshold of the target domain material, is the second density threshold of the target domain material.

4. The method according to claim 2, wherein The laser powder bed fusion monitoring platform includes three in-situ sensors and laser powder bed fusion technology processing equipment. The three in-situ sensors are an industrial camera, a microphone and a photodiode. Among them, the industrial camera is located outside the laser powder bed fusion technology processing equipment, and the microphone and photodiode are located inside the laser powder bed fusion technology processing equipment.

5. The method according to claim 4, wherein The source domain data and the target domain data are both fused image signals or photodiode signals, where: The fusion image signal is acquired based on industrial cameras; The acoustic signal is obtained based on microphone sensing; The photodiode signal is collected based on the photodiode.

6. The method according to claim 4, wherein The laser powder bed fusion monitoring platform is equipped with a high-level trigger. During experiments on different material processes, the high-level trigger automatically triggers the industrial camera, microphone, and photodiode to automatically acquire source domain data and target domain data.

7. The method according to claim 1, wherein In step S2, the feature extraction module includes a 5*5 convolutional layer and a maximum pooling layer, and the first classification recognition module includes a fully connected layer, an activation function layer, and a random dropout layer, wherein the random dropout layer is provided after the activation function layer; Correspondingly, in step S4, the second classification recognition module includes a fully connected layer, a batch normalization layer and an activation function layer, wherein the batch normalization layer is arranged before the activation function layer.

8. The method according to claim 1, wherein Step S3 includes: S31 randomly initializes all parameters of the pre-trained model; S32 optimizes the learning rate using the Adam optimizer, iteratively trains the pre-trained model according to the source domain training dataset, extracts a first feature using the feature extraction module, then classifies and identifies the first feature according to the first classification and recognition module to obtain a predicted label of the source domain training dataset, and adjusts the parameters of the pre-trained model using the source domain test dataset until the pre-trained model reaches a first expected accuracy, thereby obtaining first parameters of the feature extraction module; Among them, the first feature is the domain invariant feature. During training, the first loss function is introduced to calculate the gap between the predicted label of the source domain training dataset and the true label of the source domain training dataset.

9. The method according to claim 1, wherein Step S5 includes: S51 randomly initializes the parameters of the second classification recognition module; S52 optimizes the learning rate using the Adam optimizer, iteratively trains the fine-tuning model based on the target domain training dataset, extracts a second feature using the feature extraction module, then classifies and identifies the second feature using the second classification and recognition module to obtain a predicted label for the target domain training dataset, and adjusts the parameters of the fine-tuning model using the target domain test dataset until the fine-tuning model reaches the second expected accuracy, thereby obtaining a trained fine-tuning model; During training, the first parameter of the feature extraction module is kept unchanged, and a second loss function is introduced to calculate the gap between the predicted labels of the target domain training dataset and the true labels of the target domain training dataset.

10. The method according to claim 1, wherein The amount of data in the target domain training dataset is smaller than the amount of data in the target domain test dataset.

Images