Method for accurately obtaining contour of laser melting deposition molten pool
By building a melt pool image acquisition platform and FFC-ResNet model, the accuracy and real-time problems of melt pool profile detection in laser melt deposition are solved, and high-precision melt pool profile extraction is achieved, which improves the deposition quality.
Patent Information
- Application Number
- CN202510393330.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-31
- Publication Date
- 2025-07-01
AI Technical Summary
In the prior art During laser melting and deposition process, the melt pool profile detection method has problems such as limited accuracy, susceptibility to environmental interference, poor real-time performance, and high equipment costs, which is difficult to meet the needs of high-precision additive manufacturing.
An industrial camera is used to build a melt pool image acquisition platform, combined with the FFC-ResNet model for image preprocessing and segmentation, and by constructing a convolutional neural network integrating attention mechanism and Fourier features, the precise extraction of the melt pool profile is achieved.
It realizes high-precision monitoring of the contour of the laser melt deposition molten pool, improves the deposition quality and real-time, and has high accuracy and wide applicability.
Smart Images
Figure CN120235897A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical fields of machine vision target recognition and semantic segmentation, and particularly relates to a method for accurately obtaining the contour of a laser melting deposition molten pool. Background Technique
[0002] As an advanced additive manufacturing (3D printing) technology, laser melting deposition technology melts metal powder through a high-energy laser beam and deposits it layer by layer to form metal parts with complex geometric shapes and excellent properties. It has a wide range of applications in many fields such as aerospace, automotive manufacturing, and medical devices. During the laser melting deposition process, the molten pool is a key area, and its morphological, dimensional and other characteristics will directly affect the organizational structure and mechanical properties of the deposited layer. For example, changes in the morphology and size of the molten pool will directly affect the geometric shape and dimensional accuracy of the deposited layer. By real-time monitoring the state of the molten pool, process parameters can be timely detected and adjusted, thereby effectively predicting and avoiding the generation of defects and improving the quality of the deposited layer. Traditional methods for detecting the molten pool contour, such as the direct observation method based on optical imaging and using thermocouples to measure the temperature gradient for calculation, have problems such as limited accuracy, susceptibility to environmental interference, poor real-time performance, and high equipment costs, and are difficult to meet the actual production requirements of high-precision additive manufacturing. Therefore, a method for accurately obtaining the contour of a laser melting deposition molten pool is needed to solve the above problems. Summary of the Invention
[0003] The purpose of the present invention is to provide a method for accurately obtaining the contour of a laser melting deposition molten pool to solve the problems existing in the prior art as mentioned in the above background technique.
[0004] To achieve the above purpose, the present invention provides the following technical solutions:
[0005] A method for accurately obtaining the contour of a laser melting deposition molten pool includes the following steps:
[0006] S1: Build a molten pool image acquisition platform;
[0007] S2: Use an industrial camera to take pictures of the molten pool during the laser melting deposition process to obtain the original image of the molten pool;
[0008] S3: Perform grayscale conversion and filter noise preprocessing on the obtained molten pool pictures;
[0009] S4: Divide the processed pictures into a training set and a validation set, input the training set data into the FFC-ResNet model, perform multiple iterative trainings, and use the validation set for verification and calibration;
[0010] S5: Apply the trained FFC-ResNet model to a new molten pool image, segment the molten pool area, and output the contour information of the molten pool area.
[0011] Preferably, the molten pool image acquisition platform includes an illumination light source, an industrial camera, and a filter. The illumination light source is 808 nm, the industrial camera is an industrial camera compliant with the GigE Vision standard, with a frame rate of 48 frames per second, a resolution of 1920*1200 pixels, and a dynamic range of 12 bits. The filter is 808 nm and 1070 nm.
[0012] Preferably, the filtering algorithm uses a median filtering algorithm, and the dynamic adjustment range of the filtering window size is 3*3 - 7*7 pixels.
[0013] Preferably, the specific steps of S4 are as follows:
[0014] S41: Construct an FFC-ResNet model incorporating an attention mechanism and optimizing the parameters of the Fast Fourier Convolution (FFC) module;
[0015] S42: Use the preprocessed image and the corresponding molten pool contour annotation as data, and train using the cross-entropy loss function and the stochastic gradient descent algorithm;
[0016] S43: Input the preprocessed image into the trained model to obtain a predicted probability map of the molten pool area, and binarize it to obtain a binarized contour image;
[0017] S44: Finally, refine it to a single-pixel width using the Zhang-Suen algorithm, and optimize it by least squares fitting to remove noise and burrs.
[0018] Preferably, the method of applying the attention mechanism is as follows: When the image is processed by the previous convolutional layer to obtain a feature map, for this feature map, the channel attention module performs global average pooling on each channel, and sends this vector into a multi-layer perceptron (MLP) to extract features. Finally, the weight values obtained after MLP processing are multiplied by the corresponding channels of the original feature map. Then, the channel features determined by the MLP to be important are amplified in subsequent model calculations, while the unimportant channel features are weakened, enabling the model to focus on key channel information when processing molten pool images and accurately capture the molten pool contour features.
[0019] Preferably, the threshold for the binarization process is 0.5.
[0020] Preferably, the number of iterations for the adaptive training of the FFC-ResNet model is 100 - 500 times.
[0021] Compared with the prior art, the beneficial effects of the present invention are:
[0022] The present invention constructs a convolutional neural network FFC-ResNet model improved on the basis of ResNet. ResNet solves the problems of gradient disappearance and gradient explosion in the training process of deep neural networks by introducing residual blocks, enabling the network to train deeper levels. FFC introduces Fourier features in the convolutional operation, which can better capture the global and local information of images, adds an attention mechanism to improve the model's attention to the molten pool features, and at the same time adjusts the parameters of the FFC module to better adapt to the characteristics of the molten pool image, providing precise monitoring for the laser melting deposition process and ensuring the deposition quality. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] Figure 1 It is a schematic diagram of the FFC-ResNet model, a deep learning model used in the present invention. DETAILED DESCRIPTION OF THE INVENTION
[0024] In order to make the technical means, creative features, achieved purposes and effects of the present invention easy to understand, the present invention will be further described below in conjunction with specific embodiments.
[0025] Please refer to Figure 1 , the present invention provides the following technical solutions:
[0026] A method for accurately obtaining the contour of a molten pool in laser melting deposition, comprising the following steps:
[0027] First, put the metal powder into a vacuum drying oven for drying to remove the moisture adsorbed on the powder surface in preparation for laser melting deposition. Grind the surface of the substrate to remove surface impurities and oxides. Finally, place the substrate in an ultrasonic cleaning machine, add acetone solution, and clean for 10 minutes to degrease the substrate.
[0028] S1: Build a molten pool image acquisition platform; the molten pool image acquisition platform consists of an 808nm illumination source, a GigE Vision vision standard industrial camera, and 808nm and 1070nm filter plates; during the laser melting deposition process, due to the influence of natural light, lamp light filtering out radiation light and plasma radiation light, etc., the shooting effect is seriously affected. The filter plates can filter out these influencing light sources, and at the same time can avoid overexposure of the camera caused by the reflected light brought by the strong laser, and finally obtain relatively clear molten pool information.
[0029] S2: Use an industrial camera to shoot the molten pool during the laser melting deposition process to obtain the original image of the molten pool; turn on the laser generating device, and the laser beam is accurately focused on the surface of the pre-treated substrate through an optical focusing system. At the same time, the powder feeding system uses argon gas to synchronously transport the metal powder to the laser action area for laser melting deposition. During this process, the image acquisition system takes real-time shots of the molten pool part.
[0030] S3: Preprocess the acquired molten pool images by grayscaling and filtering to reduce noise; the median filtering algorithm is used for the filtering algorithm, and the dynamic adjustment range of the filtering window size is 3*3 - 7*7 pixels; during the laser melting deposition process, the dispersed gas and dust particles around seriously affect the actual shooting. At the same time, due to the high brightness of the laser melting deposition area during the laser melting deposition process, it poses a great challenge to the shooting of ordinary industrial cameras. There are inevitably many noise points in the images collected by industrial cameras. Graying and median filtering the images can better optimize the molten pool extraction effect. After the images are processed, they are divided into a training set and a validation set to prepare for subsequent processing by the deep learning model.
[0031] S4: Divide the processed images into a training set and a validation set, input the training set data into the FFC-ResNet model, and perform multiple iterative trainings. The number of iterations for the adaptive training of the FFC-ResNet model is 100 - 500 times, and use the validation set for verification and calibration; FFC-ResNet is a convolutional neural network model improved on the basis of ResNet (residual network). ResNet solves the problems of gradient disappearance and gradient explosion during the training process of deep neural networks by introducing residual blocks, enabling the network to train deeper levels. FFC introduces Fourier features in the convolutional operation, which can better capture the global and local information of the image, adds an attention mechanism to improve the model's attention to the molten pool features. At the same time, the parameters of the FFC module are adjusted to better adapt to the characteristics of the molten pool images.
[0032] The specific steps of S4 are as follows:
[0033] S41: Construct an FFC-ResNet model that integrates an attention mechanism and optimizes the parameters of the fast Fourier convolution FFC module;
[0034] FFC-ResNet is based on the ResNet framework. ResNet solves the problems of gradient disappearance and explosion in deep neural networks by virtue of residual connections, allowing the construction of deeper network levels to mine the deep features of images. The FFC module incorporates the Fourier formula in the convolutional operation:
[0035]
[0036] For the input image (x, y), its expression in the frequency domain is realized by means of the fast Fourier transform (FFT), where j is the imaginary unit, u = 0, 1, …, M - 1), v = 0, 1, …, N - 1). This transformation can map the image from the spatial domain to the frequency domain to mine global information, and the subsequent inverse transformation can restore the image. By cleverly setting the parameters of the FFC module to adapt to the characteristics of the molten pool images, it can complement each other in capturing local fine contours and the overall contour situation.
[0037] Combined with the attention mechanism, the information of the molten pool can be further focused; the application method of the attention mechanism is as follows: when the image is processed by the previous convolutional layer, a feature map is obtained. For this feature map, the channel attention module performs global average pooling on each channel, and sends this vector into a multi-layer perceptron (MLP) to extract features. Finally, the weight value obtained after being processed by the MLP is multiplied by the corresponding channel of the original feature map, and the channel features determined by the MLP to be important are amplified in the subsequent model calculation, while the unimportant channel features are weakened, so that the model can focus on the key channel information when processing the molten pool image and can accurately capture the contour features of the molten pool.
[0038] S42: Using the preprocessed image and the corresponding molten pool contour annotation as data, training is performed using the cross-entropy loss function and the stochastic gradient descent algorithm;
[0039] Cross-entropy loss function:
[0040]
[0041] To measure the difference between the model prediction and the true molten pool contour annotation, where N is the number of training samples, C is the number of categories (two categories of molten pool and non-molten pool, i.e., C = 2)), y ij is the true label (0 or 1) of the j-th category of the i-th sample, and p ij is the probability that the model predicts that the sample belongs to the j-th category; the more accurate the model prediction, the higher the matching degree between p ij and y ij and the smaller the loss value; the weights and biases of the model are updated using stochastic gradient descent (SGD), and the formula for SGD is:
[0042]
[0043] where: θ represents the model parameters, t represents the current iteration number, η is the learning rate, which controls the step size. f i is the loss function. For the i-th data, this formula means that in each iteration, the gradient t under the current parameter θ is calculated, and then the parameters are updated in the opposite direction of the gradient to reduce the value of the loss function.
[0044] S43: Input the preprocessed image into the trained model to obtain a molten pool region prediction probability map, and binarize it to obtain a binarized contour image;
[0045] After the model training is completed, the performance of the model is evaluated using the test set, and then a threshold (the threshold is 0.5) is set for the predicted probability map. Pixels with a probability greater than the threshold are marked as the molten pool area and assigned a value of 1, and those less than or equal to the threshold are marked as the non-molten pool area and assigned a value of 0 to obtain a binary contour image of the molten pool, initially outlining the contour shape of the molten pool.
[0046] S44: Finally, it is thinned to a single-pixel width using the Zhang-Suen algorithm, optimized by least-squares fitting, and noise burrs are removed;
[0047] The Zhang-Suen algorithm is used to operate on the binary contour image. By iteratively judging conditions such as the connectivity of boundary pixel points, removable pixels are gradually deleted until the contour is thinned to a single-pixel width, accurately restoring the molten pool boundary. Using the least-squares method, for the molten pool contour points ((x i , y i )) ((i = 1, 2,..., n)) on the two-dimensional plane, based on the objective function:
[0048]
[0049] Find the fitting curve (y = f(x)), adjust the curve parameters to minimize the sum of the squares of the distances from the contour points to the curve, remove noise and burrs, make the molten pool contour smoother and more accurate. Finally, the model is adjusted and optimized according to the evaluation results. Through this series of fine processes, the FFC-ResNet model can efficiently and accurately obtain the laser melting deposition molten pool contour.
[0050] S5: Apply the trained FFC-ResNet model to a new molten pool image to segment the molten pool area. During subsequent shooting, the captured image information is input into the FFC-ResNet model, and the molten pool contour is output by the FFC-ResNet model.
[0051] Through the above method, the present invention has successfully realized a method for accurately extracting the laser melting deposition molten pool contour using the FFC-ResNet deep learning model. The accuracy of this method exceeds 95%, and it has the advantages of high accuracy, high robustness, and wide applicability, providing strong support for the optimization and development of the laser melting deposition technology.
[0052] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A method for accurately obtaining a molten pool profile of laser melting deposition, characterized in that: The following steps are involved: S1: Build a melt pool image acquisition platform; S2: Use an industrial camera to photograph the molten pool during the laser melting deposition process to obtain the original image of the molten pool; S3: grayscale, filter and denoise preprocessing of the obtained molten pool image; S4: Divide the processed images into training set and validation set, input the training set data into the FFC-ResNet model, iterate the training multiple times, and use the validation set for verification and calibration; S5: Apply the trained FFC-ResNet model to the new melt pool image, segment the melt pool area, and output the contour information of the melt pool area.
2. The method for accurately obtaining the contour of a laser melting deposition molten pool according to claim 1, characterized in that: The molten pool image acquisition platform includes an illumination light source, an industrial camera and a filter. The illumination light source is 808nm, the industrial camera is a GigE Vision visual standard industrial camera with a frame rate of 48 frames per second, a resolution of 1920*1200 pixels, a dynamic range of 12bit, and the filters are 808nm and 1070nm.
3. The method for accurately obtaining the contour of a laser melting deposition molten pool according to claim 1, characterized in that: The filtering algorithm uses a median filtering algorithm, and the dynamic adjustment range of the filtering window size is 3*3-7*7 pixels.
4. The method for accurately obtaining the contour of a laser melting deposition molten pool according to claim 1, characterized in that: The specific steps of S4 are: S41: Construct an FFC-ResNet model that incorporates the attention mechanism and optimizes the parameters of the fast Fourier convolution FFC module; S42: Using the preprocessed image and the corresponding melt pool contour annotation as data, the cross entropy loss function and the stochastic gradient descent algorithm are used for training; S43: inputting the preprocessed image into the trained model to obtain a predicted probability map of the molten pool area, and binarizing it to obtain a binary contour image; S44: Finally, the Zhang-Suen algorithm is used to refine the image into a single pixel width and the least squares method is used to fit and optimize the image to remove noise and glitches.
5. The method for accurately obtaining the contour of a laser melting deposition molten pool according to claim 4, characterized in that: The application method of the attention mechanism is as follows: when the image is processed by the previous convolution layer, a feature map is obtained. For this feature map, the channel attention module performs a global average pooling operation on each channel, and sends this vector to the multi-layer perceptron MLP to extract features. Finally, the weight value obtained after MLP processing is multiplied by the corresponding channel of the original feature map, and then the channel features determined by MLP as important are amplified in subsequent model calculations, and unimportant channel features are weakened, so that the model can focus on key channel information when processing the molten pool image, and can accurately capture the molten pool contour features.
6. The method for accurately obtaining the contour of a laser melting deposition molten pool according to claim 4, characterized in that: The threshold of the binarization process is 0.
5.
7. The method for accurately obtaining the contour of a laser melting deposition molten pool according to claim 1, characterized in that: The number of iterations of the FFC-ResNet model adaptive training is 100-500 times.