CN-122023943-A - Method and system for identifying laser coaxial powder feeding and material adding dual-channel defocusing amount

CN122023943ACN 122023943 ACN122023943 ACN 122023943ACN-122023943-A

Abstract

The invention belongs to the technical field of additive manufacturing and deep learning, and provides a method and a system for identifying laser coaxial powder feeding and additive dual-channel defocusing amount, wherein the method comprises the steps of collecting coaxial molten pool images in an additive manufacturing process under different defocusing amounts, and constructing an original data set; and extracting deep semantic features of the gray level image, fusing the deep semantic features with the multi-dimensional features, and inputting MLP training. The method and the device realize identification of the defocus amount in the additive manufacturing process by applying the image processing method and the feature fusion method to the industrial image (the single-pass experimental coaxial image for additive manufacturing), save experimental materials and processing time, ensure that the obtained image features and the defocus amount have high correlation, remarkably improve the identification precision and overcome the problem of insufficient characterization capability of the single morphological feature.

Inventors

GAO JIALI
QIU PENGFEI
WANG JIAYU
YANG KE
DONG QIN

Assignees

上海理工大学

Dates

Publication Date: 20260512
Application Date: 20260403

Claims (10)

1. A method for identifying laser coaxial powder feeding and material adding dual-channel defocusing amount is characterized by comprising the following steps: S1, acquiring coaxial molten pool images of an additive manufacturing process under different defocus amounts to form an original data set; S2, generating a gray image according to the coaxial molten pool image, and extracting multidimensional feature distribution; And S3, extracting deep semantic features of the gray level image, fusing the deep semantic features with the multidimensional features, inputting the fused features into an MLP classifier for training, and identifying the defocus amount.
2. The method for identifying the dual-channel defocus amount of the laser coaxial powder feeding and material adding according to claim 1, wherein in the step S1, fixed technological parameters are set, different defocus amounts are adjusted, corresponding coaxial molten pool images are collected as data, corresponding defocus amount labels are marked, and an original data set is constructed; The defocus amount is based on zero defocus amount, and positive and negative defocus amounts are symmetrically selected; In the step S2, gray level RGB channel weight based on OpenCV automatic center cutting and setting is used for generating gray level images with fewer noise points than the original coaxial molten pool images from the coaxial molten pool images, and morphological features, intensity distribution features and texture features are extracted from the gray level images; In the step S3, the feature extraction layer of the CNN model extracts deep semantic information of the gray level image, performs feature fusion with the multidimensional features, inputs the fused features into the MLP classifier for learning, and identifies the defocus amount according to the MLP classifier after learning.
3. The method for identifying the defocus amount of the laser coaxial powder feeding and material adding dual channels according to claim 2, wherein the step S2 comprises the following steps: Step S2a, carrying out automatic center cutting on the coaxial molten pool image based on OpenCV to obtain a cut molten pool image, setting the background outside a circular area to be transparent, and reserving RGB information in the circular area; Step S2b, processing the cut molten pool image according to the set gray RGB channel weight gray to obtain a gray image, and extracting morphological characteristics, intensity distribution characteristics and texture characteristics based on a bright area leading principle; In the step S3, PCA processing is performed on the gray level image, and after the main component is extracted, the deep semantic information is extracted by inputting a CNN model.
4. The method for identifying the defocus amount of the laser coaxial powder feeding and material adding dual channels according to claim 3, wherein the maximum value of the defocus amount is selected to be 10mm, the minimum value is selected to be-4 mm, and the interval is 2mm; the frame rate of equipment used for acquisition is higher than 30fps, and the spectral response range is staggered with a laser wave band; the set gray RGB channel weight reduces noise points of gray images and reflection of smooth inner walls of a cladding head and a nozzle; the size of the cut molten pool image is 384 multiplied by 384, the format is PNG, and the image color mode is RGBA; wherein the A channel is a transparent channel; R, G, B channels are respectively red, green and blue channels; The morphological characteristics comprise area, perimeter, circularity, aspect ratio of circumscribed rectangle and convexity; the intensity distribution characteristics include average intensity, standard deviation of intensity, maximum intensity, minimum intensity, median intensity, skewness, and kurtosis; The texture features include contrast, correlation, energy, and homogeneity.
5. The method for identifying the defocus amount of the laser coaxial powder feeding and material adding dual channels according to claim 2, wherein the step S3 comprises the following steps: S3a, performing principal component analysis on the gray level image, retaining principal components with accumulated variance contribution rate more than or equal to 95%, and inputting the processed gray level image into a CNN model; S3b, extracting depth semantic information of the gray level image based on a CNN algorithm, flattening the extracted depth semantic information, and carrying out secondary feature fusion with morphological features, intensity distribution features and texture features; Step S3c, constructing and training an MLP classifier based on a full-connection layer, taking the fusion characteristics after the secondary characteristic fusion as the input of the MLP, training and storing the MLP classifier after training; the MLP comprises an input layer, two hidden layers and an output layer, wherein the hidden layer activation functions are all ReLU, the optimizer is Adam, and the loss function is cross entropy loss; The training introduces macro-average and weighted average, and optimizes the MLP classifier by multi-fold crossover, hyper-parameter adjustment or access ResNet residual network.
6. A laser coaxial powder feeding and material adding dual-channel defocusing amount identification system is characterized by comprising: The method comprises the steps that a module M1 collects coaxial molten pool images of an additive manufacturing process under different defocus amounts to form an original data set; the module M2 generates a gray image according to the coaxial molten pool image and extracts multidimensional feature distribution; And a module M3, extracting deep semantic features of the gray level image, fusing the deep semantic features with the multidimensional features, inputting the fused features into an MLP classifier for training, and identifying the defocus amount.
7. The system for identifying the defocus amount of the laser coaxial powder feeding and material adding dual channels according to claim 6, wherein the module M1 collects coaxial molten pool images corresponding to different defocus amounts as data, marks corresponding defocus amount labels and constructs an original data set; The defocus amount is based on zero defocus amount, and positive and negative defocus amounts are symmetrically selected; The module M2 generates a gray image with fewer noise points than the original coaxial molten pool image from the coaxial molten pool image based on the automatic center clipping and the set gray RGB channel weight of OpenCV, and extracts morphological characteristics, intensity distribution characteristics and texture characteristics from the gray image; the module M3 comprises a CNN model and an MLP classifier; And the feature extraction layer of the CNN model extracts deep semantic information of the gray level image, performs feature fusion with the multidimensional features, inputs the fused features into an MLP classifier for learning, and identifies the defocus amount according to the learned MLP classifier.
8. The laser coaxial powder feed and additive dual-channel defocus amount identification system of claim 7, wherein the module M2 comprises: The module M2a performs automatic center clipping on the coaxial molten pool image based on OpenCV to obtain a clipped molten pool image, the background outside the circular area is set to be transparent, and RGB information is reserved in the circular area; the module M2b is used for carrying out gray processing on the cut molten pool image according to the set gray RGB channel weight to obtain a gray image, and extracting morphological characteristics, intensity distribution characteristics and texture characteristics based on a bright area leading principle; and the module M3 performs PCA processing on the gray level image, extracts main components, inputs the main components into a CNN model and extracts deep semantic information.
9. The laser coaxial powder feeding and material adding dual-channel defocusing amount identification system according to claim 8, wherein the maximum value of the defocusing amount is selected to be 10mm, the minimum value is selected to be-4 mm, and the interval is 2mm; the set gray RGB channel weight reduces noise points of gray images and reflection of smooth inner walls of a cladding head and a nozzle; the size of the cut molten pool image is 384 multiplied by 384, the format is PNG, and the image color mode is RGBA; wherein the A channel is a transparent channel; R, G, B channels are respectively red, green and blue channels; The morphological characteristics comprise area, perimeter, circularity, aspect ratio of circumscribed rectangle and convexity; the intensity distribution characteristics include average intensity, standard deviation of intensity, maximum intensity, minimum intensity, median intensity, skewness, and kurtosis; The texture features include contrast, correlation, energy, and homogeneity.
10. The system for identifying the dual-channel defocusing amount of the laser coaxial powder feeding and material adding according to claim 7, wherein the module M3 analyzes the main component of the gray level image, reserves the main component with the accumulated variance contribution rate of more than or equal to 95 percent, and inputs the processed gray level image into a CNN model; The CNN model extracts depth semantic information of the gray level image based on a CNN algorithm, and flattens the extracted depth semantic information to perform secondary feature fusion with morphological features, intensity distribution features and texture features; Training an MLP classifier based on a full connection layer, taking the fusion characteristics after the secondary characteristic fusion as the input of the MLP, training and storing the MLP classifier after training; the MLP comprises an input layer, two hidden layers and an output layer, wherein the hidden layer activation functions are all ReLU, the optimizer is Adam, and the loss function is cross entropy loss; The training introduces macro-average and weighted average, and optimizes the MLP classifier by multi-fold crossover, hyper-parameter adjustment or access ResNet residual network.

Description

Method and system for identifying laser coaxial powder feeding and material adding dual-channel defocusing amount Technical Field The invention belongs to the technical field of additive manufacturing and deep learning, and particularly relates to a laser coaxial powder feeding and additive dual-channel defocus amount identification method and system, in particular to a laser coaxial powder feeding and additive dual-channel defocus amount identification method based on a multidimensional feature fusion-deep learning algorithm. Background In the laser directional energy deposition manufacturing process, high-energy focused laser beams are utilized to melt metal powder, and the metal powder is printed and piled layer by layer to manufacture parts with required shapes. Due to the advantages of high material utilization rate, strong complex structure forming capability and the like, the material has been widely applied to the fields of aerospace, high-end equipment, medical appliances and the like. In the laser directional energy deposition process, the variation of the defocusing amount of the laser processing optical head causes the light spot powder spots to be mismatched, and the final forming quality is directly affected. Therefore, the defocus amount during processing needs to be closed-loop controlled, and the precondition is to realize the correct identification of defocus amount. The patent literature (CN 117593255A) discloses a laser additive manufacturing defect monitoring method based on temporal and spatial information fusion, which comprises an information extraction module, a feature fusion module and a decision classification layer which are sequentially connected, but solves the defect monitoring problem of additive manufacturing parts, relies on acquisition of multi-sensor temporal and spatial data, and particularly has extremely high requirement on multi-sensor time sequence monitoring. Patent literature (CN 118371734A) discloses a deep learning-based additive manufacturing defect detection system and method (CN 118371734A) for realizing defect detection based on multi-source data fusion of infrared imaging images, ultrasonic detection defect data, CT scanning images and the like, but the requirement on data alignment is extremely high, and CT images in the system are derived from parts after additive manufacturing and forming, so that the problem of defect detection of the additive manufacturing is solved, and the system cannot be used for real-time monitoring and real-time control in the forming process. Patent literature (CN 120839099A) discloses that an industrial camera is adopted to collect molten pool and splash plume images in the laser additive manufacturing process in real time, and a three-dimensional convolution self-coding neural network is utilized to extract characteristics. The extracted characteristic forms comprise molten pool morphology, molten pool size, splash plume size and splash plume distribution characteristics. But its extracted features focus mainly on the morphology, a single dimensional feature. And the secondary fusion of the multidimensional features such as the morphology, the intensity, the texture and the like of the molten pool image and the CNN depth semantic features is beneficial to overcoming the problem of insufficient single feature characterization capability. The patent document of on-line monitoring of defocus amount of laser fused deposition powder flow and negative feedback state identification method (CN 111390168A) discloses that the degree of automation of defocus amount identification is improved through steps of image preprocessing, depth of field judgment, neural network classification and the like. But the neural network model mainly depends on visual feature classification of molten pool images, does not further integrate multidimensional features, and does not introduce a feature fusion mechanism to enhance the characterization capability and generalization performance of the model. Therefore, a method for identifying laser coaxial powder feeding and material increasing defocus amount by fusing multidimensional features of molten pool images, such as morphology, strength and texture, and depth semantic features is needed at present. Disclosure of Invention Aiming at the defects in the prior art, the invention aims to provide a method and a system for identifying the coaxial image types of different defocus amounts by using laser coaxial powder feeding and material adding dual-channel defocus amount, which effectively judge coaxial image types of different defocus amounts and provide preconditions for real-time regulation and control of defocus amounts. The invention provides a method for identifying laser coaxial powder feeding and material adding dual-channel defocusing amount, which comprises the following steps: S1, acquiring coaxial molten pool images of an additive manufacturing process under different defocus amounts to form an ori