Laser powder bed fusion manufacturing powder spreading defect detection system based on federated learning

Abstract

The application discloses a laser powder bed fusion manufacturing powder laying defect detection system based on federal learning, and the system comprises the following units: a first unit is used for collecting and preprocessing the forming area platform image information when each organization participating in federal learning independently carries out laser powder bed fusion additive manufacturing, and obtaining the pretreated image information; a second unit is used for training the defect identification model of the pretreated image information by using the federal learning method and obtaining the gradient change information, the gradient change information is used for updating the defect identification model of the coordination server, and the global defect detection model is obtained based on the updated defect identification model; a third unit is used for downloading the global defect detection model from the server end by the user, and detecting the defects in the laser powder bed fusion additive manufacturing powder laying process. The global defect detection model of the system has high accuracy and can guarantee the privacy of the client data.

Description

Technical Field

[0001] This disclosure relates to the field of additive manufacturing process monitoring technology, and in particular to a laser powder bed melting manufacturing powder spreading defect detection system based on federated learning. Background Technology

[0002] Laser powder bed fusion additive manufacturing (LBD) is an advanced manufacturing method that uses high-energy lasers to continuously melt layers of metal powder, stacking them layer by layer to fabricate complex-shaped metal parts. Due to its advantages such as high process flexibility, low material waste, and good production predictability, LBD has been applied in various fields including aerospace and healthcare. During processing, various factors can interact, leading to defects in the powder spreading process, such as warping and insufficient powder spreading, which can severely affect the quality of the formed parts. Powder spreading defect detection is a crucial aspect of quality assessment and monitoring in LBD processing. Timely detection and evaluation of defects during powder spreading help adjust process parameters, improve the manufacturing process, and ultimately enhance part quality.

[0003] In recent years, research on high-precision, highly automated powder-laying defect detection technologies has largely focused on computer vision technology based on machine learning, especially semantic segmentation, which has shown excellent performance in defect localization and detection. However, several serious challenges remain in this field. The biggest challenge is the lack of defect data. Most semantic segmentation models are built around convolutional neural networks, which inherently require large amounts of high-quality data for model training. However, due to the low probability and limited shape range of defects, obtaining large amounts of defect data from a single source is impractical. Furthermore, as a cutting-edge technology in the manufacturing industry, data sharing among manufacturers is difficult due to conflicts of interest and privacy concerns. Summary of the Invention

[0004] To address the aforementioned technical problems, the purpose of this disclosure is to provide a defect detection system for powder bed fusion manufacturing in laser powder bed fusion based on federated learning. This system applies the concept of federated learning to the training of a defect identification model in the powder bed fusion process of laser powder bed fusion additive manufacturing. It updates the defect identification model by centrally aggregating gradient data from various local computations rather than the original data, thus ensuring the privacy and confidentiality of private local data while solving the problem of insufficient defect data for model training. To achieve the above objective, this disclosure provides the following technical solutions.

[0005] A laser powder bed melting manufacturing powder deposition defect detection system based on federated learning, characterized in that the system comprises the following units:

[0006] The first unit is used by participating organizations in federated learning to collect image information of the forming area platform during the powder spreading process when independently carrying out laser powder bed fusion additive manufacturing, and to preprocess the collected image information to obtain preprocessed image information.

[0007] The second unit is used to train a defect recognition model on the preprocessed image information using a federated learning method and obtain gradient change information. The gradient change information is used to update the defect recognition model of the coordinating server and obtain a global defect detection model based on the updated defect recognition model.

[0008] The third unit is used by users to download a global defect detection model from the server to perform defect detection on the powder spreading process of laser powder bed molten additive manufacturing.

[0009] Optionally, in the first unit, the laser powder bed fusion additive manufacturing process specifically involves using a laser as an energy source to sinter or melt metal powder, and layer-by-layer powder spreading and sintering, or layer-by-layer powder spreading and fusion forming of a component with an established geometric model.

[0010] Optionally, the second unit specifically performs the following steps:

[0011] Step S201: System Initialization: The coordinating server, which coordinates the federated learning, sends a modeling task, seeks participating clients, and forms a federated learning consortium; the participating clients submit joint modeling applications according to their own needs. After all participating clients reach an agreement, the basic framework of the federated learning is established, which includes a defect identification model and a federated learning optimized classification algorithm; each participating client enters the joint modeling process, the coordinating server selects participating clients, and the selected participating clients download the defect identification model from the coordinating server;

[0012] Step S202: Local computation: Each participating client uses its own preprocessed image information to train the defect recognition model, update the model parameters and perform local performance evaluation to obtain gradient change information; each participating client encrypts and uploads the gradient change information to the coordinator server;

[0013] Step S203: Central Aggregation: After receiving gradient information from each participating client, the coordinating server updates the defect identification model parameters. After completion, the coordinating server returns the updated defect identification model to each participating client.

[0014] Step S204: Repeat steps S202 and S203 until the performance of the defect identification model meets the requirements, at which point the training terminates and a global defect detection model is obtained. The global defect detection model is stored on the coordinator server for subsequent defect detection work.

[0015] Optionally, in step S201, the semantic segmentation model U-Net is used as a defect identification model for training.

[0016] Optionally, after the federated learning framework is determined, training parameters need to be set, including: local and server learning rates, local and server iteration counts, local data batch size, and client score.

[0017] Optionally, in step S201, the established federated learning optimization classification algorithm is the FedAvg method.

[0018] Optionally, the objective function of the FedAvg method is:

[0019]

[0020] Where f(·) represents the abstract expression of the deep learning model, M represents the number of participating clients in the federated modeling, and ω represents the current parameters of the model. * This represents the parameters after the model update, E represents the desired function, and x represents the data held by the participating client.

[0021] Optionally, in step S202, the participating client continuously updates the model's accuracy using a local stochastic gradient descent method to obtain gradient change information; when the predetermined number of local training iterations is reached, each participating client encrypts and uploads the gradient change information obtained from training the defect identification model to the coordinator server.

[0022] Optionally, in step S203, the coordinating server samples and aggregates the gradient change information uploaded by the participating client through random sampling, calculates the increment of the gradient information and performs a weighted average, and updates the defect identification model parameters.

[0023] Optionally, the coordinating server randomly selects a subset S of participating clients in each iteration. t Perform uniform sampling and update the parameters of the m clients participating in the training. The defect identification model parameters are updated using a weighted average, while client parameters not belonging to the sample set use the current global parameters. Substitution, that is:

[0024]

[0025] Among them, w t+1 n represents the global parameters after the coordinator server model is updated. k S represents the amount of data from the k-th client, n represents the total amount of data, and S represents the total amount of data from the client. tRepresents a subset of clients. This represents the current global parameter, and k represents the client sequence.

[0026] Compared with the prior art, this disclosure has at least the following technical effects:

[0027] A deep learning model for defect detection in the laser powder bed fusion additive manufacturing (LPB) process was constructed by learning from collected powder spreading defect data. The federated learning method addresses the issues of insufficient defect data during individual training and data privacy and leakage during centralized multi-user training. By exchanging the parameters of the deep learning model instead of the original data, a high-performance global defect detection model can be jointly trained while protecting the privacy of different users' data. Attached Figure Description

[0028] Figure 1 This is a flowchart of a laser powder bed melting manufacturing powder spreading defect detection system based on federated learning, provided in one embodiment of this disclosure;

[0029] Figure 2 This is a schematic diagram illustrating the process of acquiring morphological image data and creating a dataset for the powder bed coating process in laser powder bed fusion additive manufacturing, provided in one embodiment of this disclosure.

[0030] Figure 3 This is a schematic diagram of a federated learning framework for training a defect identification model provided in one embodiment of this disclosure;

[0031] Figure 4 This is a schematic diagram of the network structure of U-Net, a deep learning semantic segmentation model for defect classification, provided in one embodiment of this disclosure;

[0032] Figure 5 This is a comparison chart showing the results of defect detection in the powder spreading process of laser powder bed fused additive manufacturing using a model obtained through federated learning on a test set, according to an embodiment of this disclosure. Detailed Implementation

[0033] The following is in conjunction with the appendix Figures 1 to 5 The present disclosure will be further described in detail below with reference to the embodiments. It is to be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the disclosure. It should also be noted that, for ease of description, only the parts relevant to the present disclosure are shown in the accompanying drawings.

[0034] It should be noted that, where there is no conflict, the embodiments and features described in this disclosure can be combined with each other. The technical solutions of this disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0035] Unless otherwise stated, the exemplary implementations / embodiments shown are to be understood as providing exemplary features of various details that provide ways in which the technical concepts of this disclosure can be implemented in practice. Therefore, unless otherwise stated, the features of various implementations / embodiments may be additionally combined, separated, interchanged and / or rearranged without departing from the technical concepts of this disclosure.

[0036] The use of crosshairs and / or shading in the accompanying drawings is generally used to clarify the boundaries between adjacent components. Thus, unless otherwise stated, the presence or absence of crosshairs or shading does not convey or indicate any preference or requirement for the specific material, material properties, dimensions, proportions, commonalities between the illustrated components, or any other characteristics, properties, etc., of the components. Furthermore, in the accompanying drawings, the dimensions and relative dimensions of components may be exaggerated for clarity and / or descriptive purposes. When exemplary embodiments can be implemented differently, a specific process sequence may be performed in a different order than that described. For example, two consecutively described processes may be performed substantially simultaneously or in the reverse order of their description. Furthermore, the same reference numerals denote the same components.

[0037] When a component is referred to as being "on" or "above" another component, "connected to," or "joined to" another component, the component may be directly on, directly connected to, or directly joined to the other component, or there may be intermediate components. However, when a component is referred to as being "directly on" another component, "directly connected to," or "directly joined to" another component, there are no intermediate components. Therefore, the term "connection" can refer to a physical connection, an electrical connection, etc., and may or may not have intermediate components.

[0038] For descriptive purposes, this disclosure may use spatial relative terms such as “below,” “under,” “below,” “down,” “above,” “above,” “higher,” and “side (e.g., in a “sidewall”)” to describe the relationship between one component and another component as shown in the accompanying drawings. In addition to the orientations depicted in the drawings, the spatial relative terms are also intended to encompass different orientations of the device during use, operation, and / or manufacture. For example, if the device in the drawings is flipped, a component described as “below” or “under” another component or feature would subsequently be positioned “above” said other component or feature. Thus, the exemplary term “below” can encompass both “above” and “below” orientations. Furthermore, the device may be otherwise positioned (e.g., rotated 90 degrees or in other orientations), thus interpreting the spatial relative descriptive terms used herein accordingly.

[0039] The terminology used herein is for the purpose of describing particular embodiments and is not intended to be limiting. As used herein, unless the context clearly indicates otherwise, the singular forms “a” and “the” are intended to include the plural forms as well. Furthermore, when the terms “comprising” and / or “including” and variations thereof are used in this specification, it indicates the presence of the stated features, integrals, steps, operations, parts, components, and / or groups thereof, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, parts, components, and / or groups thereof. It should also be noted that, as used herein, the terms “substantially,” “about,” and other similar terms are used as approximate terms rather than as terms of degree, thus explaining the inherent biases in measurements, calculated values, and / or provided values ​​that would be recognized by one of ordinary skill in the art.

[0040] In one embodiment, a laser powder bed melting manufacturing powder spreading defect detection system based on federated learning is provided, characterized in that the system comprises the following units:

[0041] The first unit is used by participating organizations in federated learning to collect image information of the forming area platform during the powder spreading process when independently carrying out laser powder bed fusion additive manufacturing, and to preprocess the collected image information to obtain preprocessed image information.

[0042] The second unit is used to train a defect recognition model on the preprocessed image information using a federated learning method and obtain gradient change information. The gradient change information is used to update the defect recognition model of the coordinating server and obtain a global defect detection model based on the updated defect recognition model.

[0043] The third unit is used by users to download a global defect detection model from the server to perform defect detection on the powder spreading process of laser powder bed molten additive manufacturing.

[0044] This embodiment employs federated learning to learn from collected powder bed fusion additive manufacturing (PLD) defect data, constructing a deep learning model for defect detection in the PLD process. The federated learning method addresses the issues of insufficient defect data during individual training and data privacy and leakage during centralized multi-user training. By exchanging the parameters of the deep learning model instead of the original data, a high-performance global defect detection model can be jointly trained while protecting the privacy of different users' data.

[0045] In one embodiment, the laser powder bed fusion additive manufacturing process in the first unit specifically involves using a laser as an energy source to sinter or melt metal powder, and layer-by-layer powder spreading and sintering, or layer-by-layer powder spreading and fusion forming of a component with an established geometric model.

[0046] In one embodiment, the second unit specifically performs the following steps:

[0047] Step S201: System Initialization: The coordinating server, which coordinates the federated learning, sends a modeling task, seeks participating clients, and forms a federated learning consortium; the participating clients submit joint modeling applications according to their own needs. After all participating clients reach an agreement, the basic framework of the federated learning is established, which includes a defect identification model and a federated learning optimized classification algorithm; each participating client enters the joint modeling process, the coordinating server selects participating clients, and the selected participating clients download the defect identification model from the coordinating server;

[0048] Step S202: Local computation: Each participating client uses its own preprocessed image information to train the defect recognition model, updates the model parameters, and performs local performance evaluation to obtain gradient change information; each participating client encrypts and uploads the gradient change information to the coordinator server;

[0049] Step S203: Central Aggregation: After receiving gradient information from each participating client, the coordinating server updates the defect identification model parameters. After completion, the coordinating server returns the updated defect identification model to each participating client.

[0050] Step S204: Repeat steps S202 and S203 until the performance of the defect identification model meets the requirements, at which point the training terminates and a global defect detection model is obtained. The global defect detection model is stored on the coordinator server for subsequent defect detection work.

[0051] In one embodiment, in step S201, the semantic segmentation model U-Net is used as a defect identification model for training.

[0052] In one embodiment, after the federated learning framework is determined, training parameters need to be set, including: local and server learning rates, local and server iteration counts, local data batch size, and client score.

[0053] In one embodiment, the federated learning optimization classification algorithm established in step S201 is the FedAvg method.

[0054] In one embodiment, the objective function of the FedAvg method is:

[0055]

[0056] Where f(·) represents the abstract expression of the deep learning model, M represents the number of participating clients in the federated modeling, and ω represents the current parameters of the model. * This represents the parameters after the model update, E represents the desired function, and x represents the data held by the participating client.

[0057] In one embodiment, in step S202, the participating client continuously updates the model's accuracy using a local stochastic gradient descent method to obtain gradient change information; when a predetermined number of local training iterations are reached, each participating client encrypts and uploads the gradient change information obtained from training the defect identification model to the coordinating server.

[0058] In one embodiment, in step S203, the coordinating server randomly samples and aggregates the gradient change information uploaded by the participating client, calculates the increment of the gradient information and performs a weighted average, and updates the parameters of the defect identification model.

[0059] In one embodiment, the coordinating server randomly selects a subset S of participating clients in each iteration. t Perform uniform sampling and update the parameters of the m clients participating in the training. The defect identification model parameters are updated using a weighted average, while client parameters not belonging to the sample set use the current global parameters. Substitution, that is:

[0060]

[0061] Among them, w t+1 n represents the global parameters after the coordinator server model is updated. k S represents the amount of data from the k-th client, n represents the total amount of data, and S represents the total amount of data from the client. t Represents a subset of clients. This represents the current global parameter, and k represents the client sequence.

[0062] In one embodiment, such as Figure 1 As shown, this disclosure provides a laser powder bed melting manufacturing powder spreading defect detection system based on federated learning, comprising the following units:

[0063] Unit 1: When participating organizations independently conduct laser powder bed fusion additive manufacturing, they collect and process image information of the forming area platform during the powder spreading process. That is, they use image acquisition equipment to collect image information of the forming area platform and perform certain preprocessing on the images.

[0064] In this unit, the acquisition of images of the forming area platform includes topographic images of the forming platform area before and after the powder spreading process. Image preprocessing includes image selection, cropping, rotation, mirroring, and annotation.

[0065] The laser powder bed fusion additive manufacturing process in the first unit of the system is a process of layer-by-layer powder spreading, sintering, or fusion forming of a component with an established geometric model. The layer-by-layer powder spreading process involves using a scraper to push metal powder from the powder supply system and spread it in the part processing system to cover the forming plane. The sintering or fusion forming process involves using a laser as an energy source to sinter or melt the metal powder, solidifying it into a specific geometric shape.

[0066] Unit 2: The collected data is used to train the model using federated learning. When the various indicators of the model in the federated learning framework tend to stabilize during the training process and have a high accuracy in identifying defects in the powder spreading process, the model is considered to meet the requirements for defect detection, and federated learning stops.

[0067] Unit 3: Using the trained model to detect defects in actual manufacturing processes.

[0068] In this unit, after obtaining a high-performance federated learning model, this model can be deployed to the detection equipment at the required laser powder bed fusion additive manufacturing site to make real-time online predictions on the input manufacturing process images.

[0069] This embodiment employs federated learning to learn from collected powder bed fusion additive manufacturing (PLD) defect data, constructing a deep learning model for defect detection in the PLD process. The federated learning method addresses the issues of insufficient defect data during individual training and data privacy and leakage during centralized multi-user training. By exchanging the parameters of the deep learning model instead of the original data, a high-performance global defect detection model can be jointly trained while protecting the privacy of different users' data.

[0070] In one embodiment, such as Figure 2 As shown, the image information acquisition unit includes the following parts:

[0071] Part 1: Before the start of a certain layer of powder spreading process, use an image acquisition device to collect an image of the shape of the forming area platform before powder spreading and save it in a computer storage device;

[0072] Part Two: After the same layer of powder spreading process is completed, use an image acquisition device to capture the morphological image of the forming area platform after powder spreading and save it in a computer storage device;

[0073] Part 3: Organize the morphological images of the forming area platform before and after powder spreading, store them in different directories, and name two images of the same layer with the same name;

[0074] Part Four: Use computer image annotation software to annotate the obtained images, using different labels to mark the powder background, parts, and various defects in each image, and save the labels with the same names as the images.

[0075] This embodiment specifies the image acquisition and processing methods to obtain a defect image dataset that can be used for subsequent deep learning network loading and training.

[0076] In another embodiment, such as Figure 3 As shown, the second unit performs the following steps:

[0077] Step S201: System initialization: The coordinator of the federated learning sends a modeling task, seeks participating clients, and forms a federated learning consortium; the participating clients submit joint modeling applications according to their own needs. After all participating clients reach an agreement, the basic framework of the federated learning is established, which includes a defect identification model and a federated learning optimized classification algorithm; each participating client enters the joint modeling process, the coordinator server selects participating clients, and the selected participating clients download the defect identification model from the coordinator server.

[0078] Among them, the organizations participating in federated learning are called coordinators, those that formulate modeling content and perform central operations, and those that train defect identification models and submit local parameters.

[0079] The coordinator is responsible for coordinating the participating organizations in federated learning, overseeing the federated learning workflow, formulating federated learning tasks and methods, and providing servers for data aggregation and computation. The participating organizations are those that participate in federated learning training and are capable of collecting data locally, independently training defect identification models, and having a secure and isolated training environment.

[0080] Specifically, in the method for detecting powder spread defects in laser powder bed fusion manufacturing, the coordinating party needs to provide the task and plan for powder spread defect detection, find qualified federated learning participants, provide a server for central aggregation and computation, and establish communication methods with clients. Participants need to already possess or be able to obtain a dataset of laser powder bed fusion powder spread defects, have equipment capable of running deep learning training programs, and be able to independently protect data privacy. For example, after clarifying the defect detection requirements and developing a plan and performance indicators for using the semantic segmentation model U-Net for powder spread defect detection and segmentation, the coordinating party seeks organizations that possess or can obtain powder spread defect image data and are capable and willing to participate in the defect detection task. After forming a federated learning consortium, the coordinating party sends the server address and key used for central aggregation to the participating client. The participating client attempts to connect to the server and evaluates the stability of the connection through several data exchanges.

[0081] In this step, to detect and segment defects in the laser powder bed fusion additive manufacturing powder spreading process, the deep learning semantic segmentation model U-Net is used as the defect recognition model. For example... Figure 4 As shown, U-Net is a classic fully convolutional network containing 19 convolutional operations, 4 pooling operations, 4 upsampling operations, and 4 cropping and copying operations. Its unique symmetrical structure enables the acquisition of contextual information and precise localization. Its structure is based on an encoder-decoder architecture. The encoder consists of multiple convolutional and pooling layers, used to generate contextual information and features of the image. For a batch of input images, these layers in the encoder generate feature maps with progressively smaller sizes and more channels. The decoder gradually restores the image resolution through upsampling and convolutional operations, and uses skip links to fuse low-level features with corresponding high-level features, better preserving detail information while having greater context awareness. The last layer of U-Net is a fully convolutional layer that transforms the feature maps into segmentation results of the same size as the input image.

[0082] In this step, the defect segmentation model uses the average intersection-union ratio (mIoU) and the average Dice coefficient (mDice) as indicators to evaluate the defect detection performance.

[0083] The average Intersection over Union (mIoU) is the average of the Intersection over Union (IoU) values. It measures the number of pixels shared by the ground truth mask and the predicted mask divided by the total number of pixels occupied by the two masks. The formula is:

[0084]

[0085] The average Dice coefficient is a set similarity measure function, calculated as follows:

[0086]

[0087] Wherein, TP represents the true class, meaning the sample's true class is positive, and the model's identification result is also positive; FP represents the false negative class, meaning the sample's true class is positive, but the model identifies it as negative; FN represents the false positive class, meaning the sample's true class is negative, but the model identifies it as positive; K represents the total number of classification categories.

[0088] In this step, before training the federated learning model, the training parameters need to be determined, including the local and server learning rates η, the number of local iterations E, the number of server iterations N, the local data batch size B, and the client score C. Specifically, the training parameters are initially determined using a gridded optimization method, and then continuously adjusted using a controlled variable method to achieve faster convergence or higher training accuracy.

[0089] In this step, due to the slow communication speed and lack of stability in federated learning, the heterogeneous nature of participating client devices with varying computing capabilities, and the potential for incomplete trust in participating clients, it is necessary to reduce the number of clients and lower communication bandwidth. To address these practical issues, the FedAvg algorithm was chosen as the framework for federated learning. The core idea of ​​FedAvg is to optimize the defect identification model locally on the data holder side using stochastic gradient descent, while performing random sampling and aggregation operations on the central server. Its objective function is:

[0090]

[0091] Where f(·) represents the abstract expression of the deep learning model, M represents the number of participating clients in the federated modeling, and ω represents the current parameters of the model. * This represents the parameters after the model update, E represents the desired function, and x represents the data held by the participating client.

[0092] In this step, after the system initialization is complete, the coordinating server calculates the number m of randomly sampled participating clients according to the formula, forming the participating client set S. t :

[0093] m = max(P × M, 1)

[0094] Where P represents the selection rate, and the selected participating client downloads the defect identification model from the coordinator.

[0095] Step S202: Perform local calculations: Each participating client uses its own collected image data to train a defect recognition model locally, calculates the loss between the output result and the true label, uses programming methods to calculate the partial derivative of the loss with respect to the model parameters to obtain the gradient, updates the model parameters according to the gradient descent method, saves the gradient change information of each update, and performs local performance evaluation; after completion, each participating client encrypts and uploads the saved gradient change information to the coordinator server.

[0096] Specifically, each participating client adjusts its collected image data to fit the model's reading requirements based on the defect recognition model and parameters downloaded from the central server. Images that do not conform to the model's input size are cropped or downsampled, RGB color space images are converted to single-channel grayscale images, and the read image data is normalized. The image data is then input into the local training defect recognition model for training, updating the model parameters, and monitoring changes in mIoU and average Dice coefficients during training. When the changes in mIoU and average Dice coefficients stabilize, indicating that model training has stabilized, training is stopped, and the current model parameters are saved. Data not used in the training process is used as the test set, and the updated defect recognition model is used to predict the test set data. The prediction results are compared with the original labels, and mIoU and average Dice coefficients are calculated as indicators to evaluate the defect recognition model's performance. After completion, each participating client encrypts the gradient change information obtained during model training and uploads it to the coordinating server.

[0097] Step S203, Central Aggregation: After receiving gradient change information from multiple participating clients, the coordinating server updates the defect identification model parameters. After completion, the coordinating server returns the updated defect identification model to the participating clients.

[0098] In this step, the coordinating server randomly selects a subset S of participating clients in each iteration. t Perform uniform sampling and update the parameters of the m clients participating in the training. The global parameters are updated by performing a weighted average (client parameters not belonging to the sample set use the current global parameters). (replace), that is:

[0099]

[0100] Among them, w t+1 n represents the global parameters after the coordinator server model is updated. k S represents the amount of data from the k-th client, n represents the total amount of data, and S represents the total amount of data from the client. t Represents a subset of clients. This represents the current global parameter, and k represents the client sequence.

[0101] Step S204: Repeat steps S202 and S203 until the performance of the defect identification model meets the requirements, then training terminates and a global defect detection model is obtained. The global defect detection model is stored on the coordinator server for subsequent defect detection work.

[0102] In this step, an image acquisition system can be built based on the actual conditions of the industrial site. The model with the best defect detection performance obtained by federated learning can be deployed to the laser powder bed fusion additive manufacturing site to perform real-time online detection of defects generated during the manufacturing process.

[0103] Specifically, the effectiveness of the method was verified through simulation experiments using an open-source dataset of powder bed fusion additive manufacturing defects. Multiple computers were used to simulate the central server and participating clients, establishing a connection via a local area network and conducting several data transmissions to verify the stability of the connection. The central server grouped the dataset and distributed it to different clients, with each client using it as a local training data sample. Clients then updated the defect identification model parameters based on the distributed data.

[0104] Before training, each client should use gridded optimization and controlled variable methods to set training parameters. Based on the problem difficulty, the required number of local iterations E and server iterations are estimated. Based on device performance and network complexity, candidate values ​​for the local and server learning rates η and the local data batch size B are given, as shown in Table 1.

[0105] Table 1

[0106]

[0107] The program provides a list of candidate values ​​for the training parameters and a grid search space for the training parameters. The GridSearch CV function is used to perform a grid search, which iterates through all possible parameter combinations, evaluates the training, and finally outputs the optimal combination of training parameters. The final determined training parameters are shown in Table 2.

[0108] Table 2

[0109]

[0110] Each client selectively performs cropping, downsampling, and normalization operations on the image data it possesses to ensure that the data can be correctly input into the model before starting local training. Training stops when the changes in mIoU and average Dice coefficient stabilize, and the current model parameters and gradient change information are saved. The saved model is then used to perform defect detection on the test set images, and the mIoU and average Dice coefficient on the test set are calculated. The stabilized local evaluation results are shown in Table 3.

[0111] Table 3

[0112]

[0113] After the model stabilizes or reaches the required number of local iterations, the local client encrypts and uploads the gradient change information to the central server. The central server updates the defect identification model using a weighted average of the gradient information and distributes the updated defect identification model to the clients to begin the next iteration. After the server reaches the aforementioned number of iterations, the average intersection-over-union ratio and average Dice coefficient during the training process are shown in Table 4.

[0114] Table 4

[0115]

[0116] like Figure 5 As shown, the trained global defect detection model is used to test the test set images. It can be seen that, due to the increase in the amount of data, the model obtained by federated learning has higher accuracy than centralized learning. Moreover, since only encrypted gradient data is uploaded locally during federated learning, the privacy of client data is also guaranteed.

[0117] In real-world industrial scenarios, each participating client first downloads the final global defect detection model from the server and deploys it on local detection equipment, cloud servers, or edge computing devices. Next, a data transmission system is built to ensure the smooth acquisition and transmission of powder-spreading images from the industrial site. After transmission, the images are preprocessed according to the specific model. Then, the image data is input into the defect detection model for defect detection, the output results are saved, and corresponding visualization or control systems are designed to utilize the results. Simultaneously, the model can be updated locally, introducing a feedback mechanism to adapt to constantly changing industrial scenarios.

[0118] In the description of this specification, the references to terms such as "one embodiment / mode," "some embodiments / modes," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment / mode or example is included in at least one embodiment / mode or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment / mode or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments / modes or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments / modes or examples described in this specification, as well as the features of different embodiments / modes or examples.

[0119] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0120] Those skilled in the art should understand that the above embodiments are merely for illustrating the present disclosure and are not intended to limit the scope of the disclosure. Those skilled in the art can make other changes or modifications based on the above disclosure, and these changes or modifications still fall within the scope of the present disclosure.

Claims

1. A laser powder bed melting manufacturing powder spreading defect detection system based on federated learning, characterized in that, The system includes the following units: The first unit is used by participating organizations in federated learning to collect image information of the forming area platform during the powder spreading process when independently carrying out laser powder bed fusion additive manufacturing, and to preprocess the collected image information to obtain preprocessed image information. The second unit is used to train a defect recognition model on the preprocessed image information using a federated learning method and obtain gradient change information. The gradient change information is used to update the defect recognition model of the coordinating server and obtain a global defect detection model based on the updated defect recognition model. The third unit is used by users to download a global defect detection model from the server to perform defect detection on the powder spreading process of laser powder bed molten additive manufacturing. The second unit specifically performs the following steps: Step S201: System Initialization: The coordinating server, which coordinates the federated learning, sends a modeling task, seeks participating clients, and forms a federated learning consortium; the participating clients submit joint modeling applications according to their own needs. After all participating clients reach an agreement, the basic framework of the federated learning is established, which includes a defect identification model and a federated learning optimized classification algorithm; each participating client enters the joint modeling process, the coordinating server selects participating clients, and the selected participating clients download the defect identification model from the coordinating server; Step S202: Local computation: Each participating client uses its own preprocessed image information to train the defect recognition model, update the model parameters and perform local performance evaluation to obtain gradient change information; each participating client encrypts and uploads the gradient change information to the coordinator server; Step S203: Central Aggregation: After receiving gradient information from each participating client, the coordinating server updates the defect identification model parameters. After completion, the coordinating server returns the updated defect identification model to each participating client. Step S204: Repeat steps S202 and S203 until the performance of the defect identification model meets the requirements, at which point the training terminates and a global defect detection model is obtained. The global defect detection model is stored on the coordinator server for subsequent defect detection work. In step S203, the coordinating server samples and aggregates the gradient change information uploaded by the participating client through random sampling, calculates the increment of the gradient information and performs a weighted average, and updates the parameters of the defect identification model. In each iteration, the coordinating server randomly selects a subset of participating clients. Perform uniform sampling on the received Updated parameters for each client participating in training The defect identification model parameters are updated by performing a weighted average, and the client parameters that are not part of the sample set use the current global parameters. Substitution, that is: ； in, Represents the global parameters after the coordinator server model is updated. Representing the Data volume per client Represents the total amount of data. Represents a subset of clients. Represents the current global parameters. This indicates the client sequence.

2. The system according to claim 1, characterized in that, In the first unit, the laser powder bed fusion additive manufacturing process specifically involves using a laser as an energy source to sinter or melt metal powder, and layer-by-layer powder spreading and sintering, or layer-by-layer powder spreading and melting to form a component with an established geometric model.

3. The system according to claim 1, characterized in that, In step S201, the semantic segmentation model U-Net is used as a defect identification model for training.

4. The system according to claim 1, characterized in that, Once the federated learning framework is determined, training parameters need to be set, including: local and server learning rates, local and server iteration counts, local data batch size, and client score.

5. The system according to claim 1, characterized in that, In step S201, the federated learning optimization classification algorithm is established as the FedAvg method.

6. The system according to any one of claims 1-5, characterized in that, In step S202, the participating client continuously updates the model's accuracy using a local stochastic gradient descent method to obtain gradient change information; when the predetermined number of local training iterations is reached, each participating client encrypts and uploads the gradient change information obtained from training the defect identification model to the coordinator server.

Images