Method and system for detecting defects of laser surface alloying molten pool based on physical model

Abstract

The invention belongs to the technical field of defect detection, and discloses a laser surface alloying molten pool defect detection method based on a physical model, and the process from molten pool forming and flowing to final solidification can be well reacted in the numerical simulation and computer simulation process. According to the simulation result, the coordinates of the solidified material can be obtained, and three-dimensional coordinates can be obtained by using the coordinates to analyze the morphology of the strengthened part and the powder contact part, so that the relevant physical characteristics of the edge defects of the molten pool are obtained. In order to improve the recognition accuracy of an existing molten pool infrared image, coordinates occupied by all components obtained through molten pool simulation can be combined with the features of the molten pool infrared image. The most common problems of less experimental data and insufficient defect feature diversity of a neural network method are solved, a learning set of the neural network is further expanded by combining a result obtained by computer simulation after a certain amount of experimental data exists, and the accuracy of molten pool defect identification is fundamentally improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of defect detection, and particularly relates to a method and system for detecting laser surface alloying molten pool defects based on a physical model. Background Technique

[0002] Laser alloying is a typical additive manufacturing technology, and the research on laser alloying process parameters has been quite mature at present. However, due to factors such as the instability of the molten liquid flow in the molten pool and the uncertainty of powder arrangement, defects such as pores and unfused holes are likely to occur. In traditional infrared image recognition of the molten pool, due to the limited angle of photographing the molten pool, the shape of the molten pool edge in contact with the unfused metal powder cannot be recognized, which affects the success rate of defect recognition in traditional infrared images of the molten pool. At the same time, since the analysis basis of infrared image recognition of the molten pool is only the image, there is a lack of relevant theoretical explanations.

[0003] With the development of computer simulation technology, some simulations with reasonable theoretical analysis and control equation designs can well reflect the actual processing process. However, a reasonable CFD simulation setup and calculation process are very time-consuming and laborious. However, the control equations used in the simulation process have a certain theoretical background and interpretability. At the same time, the simulation results can be observed from various results, and the interaction between the molten pool edge and the unfused metal powder can be clearly analyzed, and the simulation results can be used to enhance the ability to identify molten pool defects.

[0004] Through the above analysis, the problems and defects existing in the prior art are as follows:

[0005] In traditional infrared image recognition of the molten pool, due to the limited angle of photographing the molten pool, the shape of the molten pool edge in contact with the unfused metal powder cannot be recognized, which affects the success rate of defect recognition in traditional infrared images of the molten pool. At the same time, since the analysis basis of infrared image recognition of the molten pool is only the image, there is a lack of relevant theoretical explanations. Summary of the Invention

[0006] In view of the problems existing in the prior art, the present invention provides a method for detecting laser surface alloying molten pool defects based on a physical model.

[0007] The present invention is implemented as follows. A method for detecting laser surface alloying molten pool defects based on a physical model includes:

[0008] Step 1, establish a defect prediction model based on the infrared image of the molten pool. Since the convolutional neural network will degenerate as the network depth increases;

[0009] Step 2, establish a laser surface alloying molten pool defect recognition model, and build a laser surface alloying molten pool defect prediction model on the basis of the ResNet50 network;

[0010] Step 3, CFD simulation modeling and simulation result analysis of the laser cladding molten pool. Use Flow3d software to simulate the laser surface alloying process;

[0011] Step 4, a prediction model for defects in the laser surface alloying molten pool based on a physical model. Based on the ResNet50 network system for infrared image defect recognition that has been established, establish a prediction model Net-1 for laser surface alloying based on a physical model.

[0012] Furthermore, for the establishment of the defect prediction model based on the infrared images of the molten pool, since the convolutional neural network will degenerate as the network depth increases:

[0013] Refer to the network structures commonly used for tool wear and defect detection, and select the ResNet network structure to establish a basic defect prediction model; due to the limitations of the imaging angle of the infrared imaging system, as well as phenomena such as molten pool splashing and unmolten powder occlusion during the laser surface alloying process, preprocess the infrared images.

[0014] Furthermore, for the establishment of the laser surface alloying molten pool defect recognition model, build a laser surface alloying molten pool defect prediction model on the basis of the ResNet50 network. The network structure is as follows:

[0015] Obtain the infrared images of the molten pool with labels from the laser cladding experiment, and divide them into a training set and a test set according to a ratio of 8:2;

[0016] Set relevant parameters during the training process.

[0017] Furthermore, the relevant parameter settings during the training process are as follows:

[0018] Select the stochastic gradient descent method as the optimizer, set the mini-batch, and the learning rate is 0.001. Train at batch sizes of 64, 32, and 16 respectively.

[0019] Furthermore, for the CFD simulation modeling and simulation result analysis of the laser cladding molten pool, use Flow3d software to simulate the laser surface alloying process:

[0020] First, use EDEM software to obtain a particle layer with random powder sizes and random distributions. Place the powder layer on the substrate in Flow3d software, and thus establish the geometric model of surface alloying;

[0021] Then set the heat source model, boundary conditions, and the vapor recoil force and Marangoni effect force generated by the molten pool under the action of the laser;

[0022] Analysis of simulation results and defect image recognition. Obtain the molten pool flow conditions and the interaction between the molten pool and the surrounding unfused particles from the simulation results. Similarly, relevant molten pool defect images need to be obtained and marked at the defect locations for later use as a network training set.

[0023] Furthermore, the laser surface alloying molten pool defect prediction model based on a physical model is established based on the ResNet50 network system for infrared image defect recognition that has been established, and a laser surface alloying prediction model Net-1 based on the physical model is established:

[0024] Since feature extraction needs to be performed on experimental images and simulation results simultaneously, the network consists of two feature extractors and a classifier;

[0025] The laser surface alloying molten pool defect prediction model based on a physical model is as follows:

[0026] Optimizer SGD, learning rate 0.001, batch size (mini-batch) 16 / 8.

[0027] Another object of the present invention is to provide a laser surface alloying molten pool defect detection system based on a physical model, including:

[0028] A defect prediction model establishment module for establishing a defect prediction model based on the infrared image of the molten pool. Since the convolutional neural network will degenerate as the network depth increases;

[0029] A defect recognition model establishment module for establishing a laser surface alloying molten pool defect recognition model and building a laser surface alloying molten pool defect prediction model on the basis of the ResNet50 network;

[0030] An analysis module for CFD simulation modeling and simulation result analysis of the laser cladding molten pool, and using Flow3d software to simulate the laser surface alloying process;

[0031] A laser surface alloying prediction model establishment module for the laser surface alloying molten pool defect prediction model based on a physical model, and establishing a laser surface alloying prediction model Net-1 based on the ResNet50 network system for infrared image defect recognition that has been established.

[0032] Another object of the present invention is to provide a computer device, which includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the laser surface alloying molten pool defect detection method based on a physical model.

[0033] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to execute the steps of the method for detecting defects in a laser surface alloying molten pool based on a physical model.

[0034] Another object of the present invention is to provide an information data processing terminal for implementing the system for detecting defects in a laser surface alloying molten pool based on a physical model.

[0035] Combined with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0036] The numerical simulation and computer simulation processes of the present invention can well reflect the process from molten pool formation, flow to final solidification. According to the simulation results, the coordinates of the material after solidification can be obtained, and using these coordinates, the three-dimensional coordinates can be used to analyze the morphology of the strengthened part and the powder contact part, so as to obtain the physical characteristics of relevant molten pool edge defects. To improve the accuracy of the existing infrared image recognition of the molten pool, the coordinates of each component obtained from the molten pool simulation can be combined with the characteristics of the infrared image of the molten pool. It solves the most common problems in using the neural network method: insufficient experimental data and lack of diversity of defect characteristics. After having a certain amount of experimental data, the learning set of the neural network is further expanded by combining the results obtained from computer simulation, fundamentally improving the correct rate of molten pool defect recognition.

[0037] The present invention has a certain degree of use value. After the transformation of the technical solution, it can improve the success rate of predicting laser surface and evolution defects to a certain extent, and improve the yield rate of laser surface alloying of parts. Compared with the existing method of using industrial CT to detect the inside of parts, it requires less cost and has no harm to the human body, which is a cost-effective choice.

[0038] Due to its detection principle, infrared detection can only detect the surface part of the part. Since it uses the method of image processing for detection, there is a lack of theoretical explanation. This solution can improve the theoretical basis of the entire detection method through physical model simulation, and can also predict and analyze to a certain extent the parts that cannot be detected by infrared images. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is a flowchart of the method for detecting defects in a laser surface alloying molten pool based on a physical model provided by an embodiment of the present invention.

[0040] Figure 2 is a structural block diagram of the system for detecting defects in a laser surface alloying molten pool based on a physical model provided by an embodiment of the present invention.

[0041] Figure 3It is a diagram of relevant parameter settings during the training process provided by an embodiment of the present invention.

[0042] Figure 4 It is a diagram of CFD simulation modeling and simulation result analysis of a laser cladding molten pool provided by an embodiment of the present invention.

[0043] Figure 5 It is a diagram of a laser surface alloying molten pool defect prediction model based on a physical model provided by an embodiment of the present invention.

[0044] Figure 6 : Comparison curve of success rates between the physics-driven model and the pure image detection model.

[0045] Figure 7 : Schematic diagram of the application of simulation results in defect detection optimization. Specific implementation manners

[0046] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0047] As Figure 1 shown, a method for detecting defects in a laser surface alloying molten pool based on a physical model provided by an embodiment of the present invention includes the following steps:

[0048] S101, establish a defect prediction model based on the infrared image of the molten pool. Since the convolutional neural network will degenerate as the network depth increases;

[0049] S102, establish a laser surface alloying molten pool defect recognition model, and build a laser surface alloying molten pool defect prediction model on the basis of the ResNet50 network.

[0050] S103, perform CFD simulation modeling and simulation result analysis of the laser cladding molten pool, and use the Flow3d software to simulate the laser surface alloying process;

[0051] S104, a laser surface alloying molten pool defect prediction model based on a physical model. Based on the ResNet50 network system for infrared image defect recognition that has been established, establish a laser surface alloying prediction model Net-1 based on the physical model.

[0052] The defect prediction model based on the infrared image of the molten pool provided by the embodiment of the present invention. Since the convolutional neural network will degenerate as the network depth increases:

[0053] Referring to the network structures commonly used for tool wear and defect detection, the ResNet network structure is selected to establish a basic defect prediction model; due to the limitations of the imaging angle of the infrared imaging system and phenomena such as molten pool splashing and unmelted powder occlusion during the laser surface alloying process, preprocessing of the infrared images is performed.

[0054] For the laser surface alloying molten pool defect recognition model provided in the embodiments of the present invention, a laser surface alloying molten pool defect prediction model is built on the basis of the ResNet50 network, and the network structure is as follows:

[0055] Obtain the molten pool infrared images containing labels from the laser cladding experiment and divide them into a training set and a test set according to a ratio of 8:2;

[0056] Setting of relevant parameters during the training process.

[0057] The settings of relevant parameters during the training process provided in the embodiments of the present invention are as follows:

[0058] The optimizer selects the stochastic gradient descent method, sets the mini-batch, and the learning rate is 0.001, and training is carried out respectively when the batch sizes are 64, 32, and 16.

[0059] For the CFD simulation modeling and simulation result analysis of the laser cladding molten pool provided in the embodiments of the present invention, the Flow3d software is used to simulate the laser surface alloying process:

[0060] First, use the EDEM software to obtain a particle layer with random powder sizes and random distributions, and place the powder layer on the substrate in the Flow3d software, thus establishing the geometric model of surface alloying;

[0061] Then set the heat source model, boundary conditions, and the vapor recoil force and Marangoni effect force generated by the molten pool under the action of the laser;

[0062] Analysis of simulation results and defect image recognition, obtain the molten pool flow condition and the interaction condition between the molten pool and the surrounding unfused particles from the simulation results; it is also necessary to obtain relevant molten pool defect images and mark them at the defect sites for later use as the network training set.

[0063] For the laser surface alloying molten pool defect prediction model based on the physical model provided in the embodiments of the present invention, based on the ResNet50 network system for infrared image defect recognition that has been established, establish a laser surface alloying prediction model Net-1 based on the physical model:

[0064] Since it is necessary to perform feature extraction on the experimental images and simulation results simultaneously, the network consists of two feature extractors and a classifier;

[0065] The physical model-based laser surface alloying molten pool defect prediction model is as follows:

[0066] Optimizer: SGD, learning rate: 0.001, batch size (mini-batch): 16 / 8.

[0067] First, through the defect prediction model establishment module, the present invention constructs a preliminary defect prediction model using the infrared image data of the molten pool. Considering the possible degradation phenomenon of the convolutional neural network (CNN) when the network depth increases, the ResNet network structure commonly used in tool wear and defect detection is selected as the basis. To address the limitations of the infrared imaging system in imaging angle and problems such as molten pool splash and unmelted powder occlusion during the laser surface alloying process, the acquired infrared images are preprocessed. This preprocessing step includes image denoising, enhancement, and alignment to ensure that the model can accurately extract key information, thereby improving the initial accuracy of defect prediction.

[0068] Based on the defect prediction model, the defect recognition model establishment module further optimizes the accuracy and robustness of defect detection. This module builds a deep learning model dedicated to laser surface alloying molten pool defect recognition based on the ResNet50 network structure. Specifically, infrared images of the molten pool with labels are obtained from laser cladding experiments, and the dataset is divided into a training set and a test set in a ratio of 8:2. During the training process, the stochastic gradient descent method (SGD) is selected as the optimizer, the learning rate is set to 0.001, and training is carried out under different batch sizes (64, 32, 16) to find the best combination of training parameters. This process ensures that the model can maintain efficient learning ability and recognition accuracy when processing complex defect images.

[0069] The analysis module is responsible for computational fluid dynamics (CFD) simulation modeling and result analysis of the molten pool during the laser cladding process. First, the EDEM software is used to generate a particle layer with random powder size and distribution, and it is placed in the Flow3D software to establish a geometric model of laser surface alloying. Then, the heat source model, boundary conditions, and the vapor recoil force and Marangoni effect force generated by the molten pool under the action of the laser are set. Through these simulation parameters, the Flow3D software simulates the temperature distribution, fluid flow, and molten pool morphology changes in the molten pool, generating high-precision simulation results. These simulation data not only provide real process data support for the establishment of the physical model but also provide important physical references for subsequent defect prediction and recognition, ensuring the stability and accuracy of the system under complex process conditions.

[0070] The laser surface alloying prediction model establishment module combines the simulation results of the physical model with the deep learning defect recognition model to form a comprehensive defect detection and prediction system. Specifically, first, the molten pool temperature and morphology data obtained from Flow3D simulation are processed to generate simulation images that supplement the infrared image information. Then, using the established ResNet50 defect recognition model, these simulation images are trained and optimized to generate a prediction model Net-1 based on the physical model. Net-1 can not only identify the defects visible in the infrared image but also predict the defect areas that cannot be directly detected by the infrared image through physical simulation data, significantly improving the coverage rate and accuracy of the overall detection system.

[0071] The entire defect detection system realizes efficient and intelligent operation through the close cooperation between the computer device and the information data processing terminal. The computer device includes a memory and a processor. A specially designed computer program is stored in the memory. When the processor executes these programs, it can complete the various steps of the laser surface alloying molten pool defect detection method based on the physical model. The information data processing terminal serves as the user interface, responsible for receiving and displaying the detection results, providing real-time defect warning information, and supporting the operator to configure and manage the detection system. Through system integration and optimization, the automation and intelligence of the detection process are ensured, and the overall detection efficiency and reliability are improved.

[0072] The laser surface alloying molten pool defect detection system based on the physical model of the present invention shows significant advantages in practical applications. First, by combining the physical model with deep learning technology, defect detection not only depends on the actually acquired infrared images but also can predict hidden defects through simulation data, greatly improving the comprehensiveness and accuracy of detection. Second, through the optimization of the ResNet50 network structure, the robustness and precision of the model in processing complex images are enhanced. In addition, the high-precision physical data provided by Flow3D simulation provides a solid theoretical basis and technical support for defect detection, improving the interpretability and credibility of the system. In summary, the present invention realizes the efficient and accurate detection of laser surface alloying molten pool defects through multimodal data fusion and deep learning optimization, providing strong technical guarantee for the quality control and optimization of related manufacturing processes.

[0073] As Figure 2 shown, a laser surface alloying molten pool defect detection system based on the physical model provided by an embodiment of the present invention includes:

[0074] A defect prediction model establishment module for establishing a defect prediction model based on the infrared image of the molten pool. Since the convolutional neural network will degenerate as the network depth increases;

[0075] The defect recognition model establishment module is used to establish a laser surface alloying molten pool defect recognition model, and build a laser surface alloying molten pool defect prediction model based on the ResNet50 network;

[0076] The analysis module is used for CFD simulation modeling and simulation result analysis of the laser cladding molten pool, and uses Flow3d software to simulate the laser surface alloying process;

[0077] The laser surface alloying prediction model establishment module is used to establish a laser surface alloying molten pool defect prediction model based on a physical model. Based on the ResNet50 network system for infrared image defect recognition that has been established, a laser surface alloying prediction model Net-1 based on the physical model is established.

[0078] The laser surface alloying molten pool defect detection system of the present invention consists of four core modules: a defect prediction model establishment module, a defect recognition model establishment module, an analysis module, and a laser surface alloying prediction model establishment module. Through coordinated operation, the system realizes real-time detection and prediction of molten pool defects during the laser surface alloying process. The defect prediction model establishment module is responsible for constructing a preliminary defect prediction model based on the infrared image data of the molten pool. On this basis, the defect recognition model establishment module uses the optimized ResNet50 network architecture to further improve the accuracy and robustness of defect recognition. The analysis module uses Flow3D software to perform computational fluid dynamics (CFD) simulation modeling of the laser cladding molten pool, providing real process data support for the physical model. Finally, the laser surface alloying prediction model establishment module combines the simulation results of the physical model with the infrared image defect recognition model to form a comprehensive defect detection and prediction system.

[0079] The main task of the defect prediction model establishment module is to construct a model that can preliminarily predict defects by collecting and processing the infrared images of the molten pool generated during the laser surface alloying process. First, the system collects infrared images at different time points during the laser cladding process as training data. Since the convolutional neural network (CNN) may exhibit degradation phenomena when the network depth increases, this module adopts an optimization strategy to ensure that the model can still maintain good performance after the depth increases by adjusting the network structure and parameters. Then, using these infrared image data, a preliminary defect prediction model is trained to identify potential defect areas at an early stage, providing basic support for subsequent defect recognition.

[0080] Based on the defect prediction model, the defect recognition model building module further optimizes and improves the accuracy and efficiency of defect detection. Based on the ResNet50 network structure, this module constructs a deep learning model specifically for laser surface alloying molten pool defect recognition. By introducing residual modules, ResNet50 can effectively alleviate the degradation problem of deep networks, improving the training efficiency and recognition accuracy of the model. During the training process, the module combines the preliminary defect regions output by the defect prediction model to perform more detailed image feature extraction and classification, thereby achieving precise recognition and positioning of molten pool defects.

[0081] The analysis module is responsible for detailed simulation and analysis of the physical state of the molten pool during the laser cladding process. Using Flow3D software, based on actual process parameters and equipment conditions, this module establishes a computational fluid dynamics (CFD) model of the laser surface alloying process. By simulating the temperature distribution, fluid flow, and changes in molten pool morphology within the molten pool, the analysis module generates high-precision simulation results. These simulation data not only provide a reliable basis for the establishment of physical models but also provide important physical references for subsequent defect prediction and recognition, ensuring the stability and accuracy of the system under complex process conditions.

[0082] The laser surface alloying prediction model building module combines the simulation results of the physical model with the deep learning defect recognition model to form a comprehensive defect detection and prediction system. Specifically, this module first processes the molten pool temperature and morphology data obtained from Flow3D simulation to generate simulation images that can supplement the infrared image information. Then, using the established ResNet50 defect recognition model, these simulation images are trained and optimized to generate a prediction model Net-1 based on the physical model. Net-1 can not only identify the defects visible in the infrared images but also predict the defect regions that cannot be directly detected by the infrared images through physical simulation data, improving the coverage rate and accuracy of the overall detection system.

[0083] The defect detection system of the present invention also includes a computer device and an information data processing terminal for implementing and supporting the entire detection process. The computer device includes a memory and a processor. The memory stores specially designed computer programs. When the processor executes these programs, it can complete the various steps of the method for laser surface alloying molten pool defect detection based on the physical model. The information data processing terminal serves as the user interface, responsible for receiving and displaying the detection results, providing real-time defect warning information, and supporting operators to configure and manage the detection system. Through the close cooperation of the computer device and the information data processing terminal, the entire defect detection system can achieve efficient and intelligent operation, ensuring that defects in the laser surface alloying process can be detected and processed in a timely manner, safeguarding product quality and production safety.

[0084] The main implementation process of this method is as follows:

[0085] 1. Establish a convolutional neural network defect prediction model based on the infrared image of the molten pool;

[0086] 2. Establish a laser surface alloying simulation model. According to the whole process of powder melting to solidification obtained from the model, the neural network analyzes the molten pool flow behavior and the edge of the molten pool and unmelted powder;

[0087] 3. Obtain and analyze the morphology of the molten pool obtained from the simulation, fuse the obtained theoretically interpretable features with the infrared image defect prediction model of the molten pool, and add the simulation model as a supervision or correction module to the convolutional network structure to improve the accuracy of defect prediction.

[0088] The implementation of this method is mainly divided into four steps. First, a defect prediction model based on the infrared image of the molten pool needs to be established. First, perform image preprocessing on the infrared image of the molten pool obtained from the experiment to exclude noise interference, and use the ResNet network structure to establish a defect prediction model for the infrared image of the molten pool. Then, establish a laser surface alloying simulation model similar to the processing environment according to the actual processing situation. It is necessary to consider the shape of the laser heat source loaded in the actual situation, the physical properties of the processed material, and the vapor recoil force and Marangoni effect force generated by the laser on the molten pool. Then, intercept and mark the morphology of the molten pool in the simulation results for model training. Finally, use the ResNet50 network model to combine the infrared image and the simulation image to establish a laser surface alloying molten pool defect detection model based on the physical model.

[0089] The first step: It is necessary to establish a defect prediction model based on the infrared image of the molten pool. Since the convolutional neural network will degenerate as the network depth increases, this method refers to the network structures commonly used for tool wear and defect detection, and selects the ResNet network structure to establish a basic defect prediction model. Due to the limitations of the imaging angle of the infrared imaging system, as well as phenomena such as molten pool splashing and unmelted powder occlusion during the laser surface alloying process, preprocess the infrared image. To reduce the noise influence in the infrared image of the molten pool, use bilateral filtering to process the infrared image of the molten pool.

[0090] The second step: Establish a laser surface alloying molten pool defect recognition model. Build a laser surface alloying molten pool defect prediction model on the basis of the ResNet50 network. The network structure is as follows. Obtain the infrared image of the molten pool containing labels from the laser cladding experiment, and divide it into a training set and a test set according to the ratio of 8:2. As Figure 3 shown, the relevant parameter settings during the training process are as follows: Select the stochastic gradient descent method as the optimizer, set the mini-batch, and the learning rate is 0.001. Train at batch sizes of 64, 32, and 16 respectively.

[0091] As Figure 4 shown, the third step: CFD simulation modeling and simulation result analysis of the laser cladding molten pool. Use Flow3d software to simulate the laser surface alloying process. First, use EDEM software to obtain a particle layer with random powder size and random distribution. Place the powder layer on the substrate in Flow3d software, thus establishing the geometric model of surface alloying. Then set the heat source model, boundary conditions, and the vapor recoil force and Magnus effect force generated by the molten pool under the action of the laser. Simulation result analysis and defect image recognition. Obtain the molten pool flow condition and the interaction between the molten pool and the surrounding unfused particles from the simulation results. It is also necessary to obtain relevant molten pool defect images and mark the defect parts, which will be used as the network training set in the later stage.

[0092] As Figure 5 shown, the fourth step: a physical model-based laser surface alloying molten pool defect prediction model. Based on the ResNet50 network system for infrared image defect recognition that has been established, establish a physical model-based laser surface alloying prediction model Net-1. Since it is necessary to extract features from experimental images and simulation results simultaneously, the network consists of two feature extractors and a classifier. The physical model-based laser surface alloying molten pool defect prediction model is as follows: optimizer SGD, learning rate 0.001, batch size (mini-batch) 16 / 8.

[0093] This physical model-based laser surface alloying molten pool defect detection method realizes the accurate detection of molten pool defects in the laser cladding process by combining the establishment of a deep learning model and CFD simulation.

[0094] The working principle of this method is as follows:

[0095] 1. Establishment of the molten pool infrared image defect prediction model

[0096] First, by collecting the infrared images of the molten pool during the laser surface alloying process, establish a model for predicting molten pool defects. Use a convolutional neural network (CNN) for image feature extraction. However, considering the possible degradation problem of the CNN network when the depth increases, this step uses an optimized structure convolutional neural network to ensure the prediction accuracy during deep feature extraction, laying a foundation for subsequent defect detection.

[0097] 2. Defect recognition model based on ResNet50

[0098] To further improve the accuracy of defect prediction, a laser surface alloying molten pool defect recognition model is built based on the ResNet50 network architecture. ResNet50 solves the problems of gradient disappearance and degradation in deep networks through residual modules, enabling the model to extract deeper image features and thus more accurately identify molten pool defects in infrared images. The model in this step can process complex image features without loss of accuracy, achieving high-precision defect recognition.

[0099] 3. CFD Simulation Modeling of Laser Cladding Molten Pool

[0100] To further understand the behavior of the molten pool during the laser alloying process, CFD simulation modeling is carried out using Flow3d software. The simulation model reproduces the fluid flow, heat transfer, and alloy material distribution during the laser cladding process. By analyzing these simulation results, the temperature field, flow velocity field, and morphological changes of the molten pool under different process parameters can be obtained, providing data support for the establishment of the physical model.

[0101] 4. Defect Prediction Model Based on Physical Model

[0102] Combining the physical simulation results with the image analysis ability of the ResNet50 model, a molten pool defect prediction model Net-1 based on the physical model is established. This model synthesizes the physical simulation data and the results of image recognition, and can more accurately predict the occurrence of molten pool defects under different laser parameter conditions. The Net-1 model uses physical information to optimize defect detection, effectively improving the generalization and prediction accuracy of the model.

[0103] 5. Model Optimization and Testing

[0104] During the model training process, the parameters of the Net-1 model are continuously optimized to adapt to the laser cladding process under different working conditions. By using experimental data to test and verify the model, its defect detection performance under different environments and conditions is observed. The optimized Net-1 model can monitor the molten pool state in real-time during practical applications and identify potential defects in a timely manner.

[0105] 6. Defect Detection in Practical Applications

[0106] Finally, this detection method can be applied to the real-time monitoring of the laser surface alloying process. The system obtains the real-time image of the molten pool through an infrared camera and inputs it into the Net-1 model for defect prediction. The model combines the results of real-time simulation and image recognition, outputs the detection results and generates an alarm, enabling the operator to adjust the laser parameters or cladding process in a timely manner, avoid the formation of defects, and improve the workpiece quality and production efficiency.

[0107] The following are two specific embodiments of the laser surface alloying molten pool defect detection method based on the physical model:

[0108] Example 1: Laser Surface Alloying Defect Detection for Automotive Parts

[0109] In automotive manufacturing, laser surface alloying technology is often used to enhance the surface hardness and wear resistance of engine blocks, pistons and other components. Since these components will be exposed to high-temperature and high-pressure environments during operation, it is very important to avoid defects during the surface alloying process. This detection method is applied to the following steps:

[0110] 1. Infrared image acquisition: During the laser surface alloying process, an infrared camera captures the temperature distribution image of the molten pool in real time.

[0111] 2. Defect recognition model: A molten pool defect recognition model based on ResNet50 analyzes the infrared image in real time to identify possible molten pool defects.

[0112] 3. CFD simulation and optimization: Use Flow3d software for CFD simulation, analyze the flow behavior of the molten pool, and optimize the Net-1 defect prediction model in combination with the actual working conditions to help identify defects caused by uneven temperature.

[0113] 4. Real-time monitoring and adjustment: In actual production, the system can adjust laser parameters such as power and scanning speed through a feedback mechanism when defects occur, thereby reducing the generation of defects and improving the surface quality and durability of automotive parts.

[0114] This example significantly improves the alloying quality of automotive parts, reduces the occurrence of surface defects, and thus extends the service life of the parts.

[0115] Example 2: Laser Surface Alloying Quality Control in Die Manufacturing

[0116] In die manufacturing, laser surface alloying is used to improve the surface hardness and wear resistance of dies, but defects in the molten pool will directly affect the surface finish and accuracy of the dies. The application steps of this detection method in die manufacturing are as follows:

[0117] 1. Molten pool image acquisition: The temperature distribution of the laser alloying molten pool is monitored in real time through an infrared imaging device to obtain the infrared image of the molten pool in each processing step.

[0118] 2. Defect detection based on ResNet50: Use a deep learning model based on ResNet50 to analyze the infrared image data and detect molten pool defects caused by temperature fluctuations.

[0119] 3. Simulation model verification and prediction: Analyze the molten pool temperature distribution and flow characteristics by combining with the Flow3d simulation model, and optimize the defect prediction model Net-1. The simulation results are used to verify the accuracy of defect prediction and adjust the laser alloying process parameters.

[0120] 4. Online feedback and quality control: The system adjusts the laser power and moving speed in real time through sensors and feedback mechanisms to ensure uniform temperature distribution in the molten pool and reduce the formation of defects such as pores and cracks during the alloying process.

[0121] Through this embodiment, the surface quality of the mold is significantly improved, the rework caused by surface defects is reduced, and the production efficiency and the durability of the mold are improved.

[0122] Figure 6 The expected success rate is shown. The physical-driven model will definitely have a slower improvement speed of the correct rate in the early stage of model iteration due to the addition of new simulation result images compared to the original pure image detection, but after the number of iterations reaches a certain level, its correct rate will exceed the traditional pure image detection due to its more comprehensive model data.

[0123] The equipment and technology related to infrared image acquisition have been quite mature, and scholars have continuously optimized the structure and model of the ResNet network structure, enabling this method to be based on the ResNer50 model. The biggest feature of this method is that it can combine the images obtained in reality with the simulation images of useful physical models and optimize the neural network, allowing the entire defect detection model to predict to a certain extent the areas that cannot be detected by infrared images. At the same time, with the support of the physical model, the entire defect detection model is more interpretable.

[0124] This behavior of combining virtual simulation and actual images is similar to digital twin technology, but actually due to the large amount of computation and the lack of rapid and accurate response in digital simulation. The simulation results as shown in Figure 7 can be used to obtain the molten pool temperature and morphology from different angles at different time periods, obtain the images that cannot be obtained by the infrared camera, and bring these morphological characteristics into the existing model, which can help us further optimize our model and further improve the success rate and theoretical basis of defect recognition.

[0125] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated designed hardware. Those of ordinary skill in the art can understand that the above-mentioned devices and methods can be implemented using computer-executable instructions and / or included in processor control code, such as provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and their modules of the present invention can be implemented by hardware circuits of programmable hardware devices such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above hardware circuits and software such as firmware.

[0126] As described above, the above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall all be covered by the protection scope of the present invention.

Claims

1. A laser surface alloying molten pool defect detection method based on a physical model, characterized in that: The following steps are involved: Step 1: Establish a defect prediction model based on the molten pool infrared image, because the convolutional neural network will degrade as the network depth increases; Step 2: Establish a laser surface alloying molten pool defect recognition model, and build a laser surface alloying molten pool defect prediction model based on the ResNet50 network; Step 3: CFD simulation modeling of laser cladding molten pool and analysis of simulation results, using Flow3d software to simulate the laser surface alloying process; Step 4: A laser surface alloying molten pool defect prediction model based on a physical model is established. Based on the ResNet50 network system for infrared image defect recognition, a laser surface alloying prediction model Net-1 based on a physical model is established.

2. The laser surface alloying molten pool defect detection method based on a physical model as claimed in claim 1, characterized in that: The defect prediction model based on the molten pool infrared image is established, because the convolutional neural network will degenerate as the network depth increases: Referring to the network structure commonly used in tool wear and defect detection, the ResNet network structure was selected to establish a basic defect prediction model. Due to the limitations of the imaging angle of the infrared imaging system and the phenomena of molten pool splashing and unmelted powder occlusion during the laser surface alloying process, the infrared image was preprocessed.

3. The laser surface alloying molten pool defect detection method based on physical model as claimed in claim 1, characterized in that: The laser surface alloying molten pool defect recognition model is established, and the laser surface alloying molten pool defect prediction model is built on the basis of the ResNet50 network. The network structure is as follows: The labeled molten pool infrared images were obtained from the laser cladding experiment and divided into training set and test set in the ratio of 8:

2. Relevant parameter settings during training.

4. The laser surface alloying molten pool defect detection method based on physical model as claimed in claim 3, characterized in that: The relevant parameters in the training process are set as follows: The optimizer selects stochastic gradient descent, sets mini-batch, and the learning rate is 0.001, and trains with batch sizes of 64, 32, and 16, respectively.

5. The laser surface alloying molten pool defect detection method based on physical model as claimed in claim 1, characterized in that: The CFD simulation modeling and simulation result analysis of the laser cladding molten pool used Flow3d software to simulate the laser surface alloying process: First, use EDEM software to obtain a particle layer with random powder size and distribution, and place the powder layer on the substrate in Flow3d software, so that the geometric model of surface alloying is established; Then, the heat source model, boundary conditions, and steam recoil force and Margolan effect force generated by the molten pool under the action of laser are set; Analysis of simulation results and defect image recognition: the flow conditions of the molten pool and the interaction between the molten pool and the surrounding un-clad particles are obtained from the simulation results. It is also necessary to obtain relevant molten pool defect images and mark the defective parts, which will be used as network training sets later.

6. The laser surface alloying molten pool defect detection method based on physical model as claimed in claim 1, characterized in that: The laser surface alloying molten pool defect prediction model based on the physical model is based on the ResNet50 network system for infrared image defect recognition, and a laser surface alloying prediction model Net-1 based on the physical model is established: Since it is necessary to extract features from both experimental images and simulation results, the network consists of two feature extractors and a classifier; The prediction model of laser surface alloying molten pool defects based on the physical model is as follows: The optimizer is SGD, the learning rate is 0.001, and the batch size (mini-batch) is 16 / 8.

7. A laser surface alloying molten pool defect detection system based on a physical model for implementing the laser surface alloying molten pool defect detection method based on a physical model as claimed in any one of claims 1 to 6, characterized in that: The laser surface alloying molten pool defect detection system based on the physical model includes: The defect prediction model building module is used to build a defect prediction model based on the molten pool infrared image, because the convolutional neural network will degenerate as the network depth increases; The defect recognition model building module is used to establish a laser surface alloying molten pool defect recognition model and build a laser surface alloying molten pool defect prediction model based on the ResNet50 network; Analysis module, used for CFD simulation modeling and simulation result analysis of laser cladding molten pool, using Flow3d software to simulate the laser surface alloying process; The laser surface alloying prediction model establishment module is used for the laser surface alloying molten pool defect prediction model based on the physical model. Based on the ResNet50 network system for infrared image defect recognition, the laser surface alloying prediction model Net-1 based on the physical model is established.

8. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the laser surface alloying molten pool defect detection method based on the physical model as described in any one of claims 1-6.

9. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the laser surface alloying molten pool defect detection method based on a physical model as claimed in any one of claims 1 to 6.

10. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the laser surface alloying molten pool defect detection system based on the physical model as described in claim 7.

Images