Laser coaxial powder feeding additive dual-channel defocusing amount identification method and system

Abstract

This invention belongs to the field of additive manufacturing and deep learning technology, and provides a method and system for identifying defocus amount in dual-channel laser coaxial powder feeding additive manufacturing. The method includes: acquiring coaxial molten pool images of the additive manufacturing process under different defocus amounts to construct an original dataset; generating grayscale images based on the coaxial molten pool images and extracting multi-dimensional feature distributions; extracting deep semantic features from the grayscale images and fusing them with the multi-dimensional features for inputting into an MLP for training. This invention achieves defocus amount identification during additive manufacturing by applying image processing and feature fusion methods to industrial images (coaxial images of single-channel additive manufacturing experiments), saving experimental materials and processing time; ensuring a high correlation between the obtained image features and the defocus amount, significantly improving recognition accuracy; and overcoming the problem of insufficient representation capability of single morphological features.

Description

Technical Field

[0001] This invention belongs to the field of additive manufacturing and deep learning technology. Specifically, it relates to a method and system for identifying the defocus amount of dual-channel additive manufacturing using laser coaxial powder feeding, and in particular, it relates to a method for identifying the defocus amount of dual-channel additive manufacturing using laser coaxial powder feeding based on multi-dimensional feature fusion and deep learning algorithms. Background Technology

[0002] Laser-directed energy deposition (LDED) manufacturing utilizes a high-energy focused laser beam to melt metal powder, which is then deposited layer by layer to create parts of the desired shape. Due to its advantages such as high material utilization and strong ability to form complex structures, it has been widely used in aerospace, high-end equipment, and medical devices. During LDED, variations in the defocusing amount of the laser processing optical head cause a mismatch between the laser spot and the powder spot, directly affecting the final forming quality. Therefore, closed-loop control of the defocusing amount during processing is necessary, which requires accurate identification of the defocusing amount.

[0003] The patent document "A method for monitoring defects in laser additive manufacturing based on spatiotemporal information fusion" (CN117593255A) discloses a neural network model based on spatiotemporal information fusion, which includes an information extraction module, a feature fusion module, and a decision classification layer connected in sequence. However, it solves the problem of monitoring defects in additive manufacturing parts and relies on the acquisition of spatiotemporal data from multiple sensors, especially with extremely high requirements for multi-sensor time-series monitoring.

[0004] The patent document "A Deep Learning-Based Additive Manufacturing Defect Detection System and Method" (CN118371734A) discloses a defect detection method based on the fusion of multiple source data such as infrared imaging images, ultrasonic defect detection data, and CT scan images. However, it also has extremely high requirements for data alignment, and the CT images in the system are derived from the parts after additive manufacturing. Therefore, it also solves the problem of additive manufacturing defect detection, but cannot be used for real-time monitoring and real-time control during the forming process.

[0005] The patent document "Additive Defocus Amount Control Device and Method Based on Image Recognition and Machine Learning" (CN120839099A) discloses the use of an industrial camera to acquire real-time images of the molten pool and spatter plumes during the laser additive manufacturing process, and to extract features using a three-dimensional convolutional autoencoder neural network. The extracted features include molten pool morphology, molten pool size, spatter plume size, and spatter plume distribution characteristics. However, the extracted features mainly focus on the single-dimensional feature of morphology. A secondary fusion of multi-dimensional features such as morphology, intensity, and texture of the molten pool image with deep semantic features of a CNN helps overcome the problem of insufficient representation ability of a single feature.

[0006] The patent document "Online Monitoring and Negative Feedback State Recognition Method for Defocus Amount of Laser Melting Deposition Powder" (CN111390168A) discloses a method to improve the automation level of defocus amount recognition through steps such as image preprocessing, depth judgment, and neural network classification. However, its neural network model mainly relies on the visual feature classification of the melt pool image, without further integrating multi-dimensional features or introducing a feature fusion mechanism to enhance the model's representation ability and generalization performance.

[0007] Therefore, there is a need for a laser coaxial powder feeding additive defocusing amount recognition method that integrates multi-dimensional features of the molten pool image, such as morphology, intensity, and texture, with deep semantic features. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for identifying the defocus amount in dual-channel laser coaxial powder feeding additive manufacturing, which effectively determines the coaxial image category with different defocus amounts, providing a prerequisite for real-time control of the defocus amount.

[0009] A method for identifying the defocus amount in dual-channel laser coaxial powder feeding additive manufacturing according to the present invention includes:

[0010] Step S1: Collect coaxial molten pool images of the additive manufacturing process under different defocusing amounts to form the original dataset;

[0011] Step S2: Generate a grayscale image based on the coaxial molten pool image and extract multi-dimensional feature distribution;

[0012] Step S3: Extract deep semantic features from the grayscale image, fuse them with multi-dimensional features, input them into MLP training, and identify the defocus amount.

[0013] Preferably, in step S1, fixed process parameters are set, different defocusing amounts are adjusted, corresponding coaxial molten pool images are collected as data, corresponding defocusing amount labels are marked, and an original dataset is constructed.

[0014] The defocus amount is based on zero defocus amount, and positive and negative defocus amounts are symmetrically selected.

[0015] In step S2, based on OpenCV's automatic center cropping and the set grayscale RGB channel weights, a grayscale image with less noise is generated from the coaxial melt pool image, and morphological features, intensity distribution features, and texture features are extracted from the grayscale image.

[0016] In step S3, the feature extraction layer of the CNN model extracts the deep semantic information of the grayscale image, fuses it with multi-dimensional features, inputs the fused features into the MLP classifier for learning, and identifies the defocus amount based on the learned MLP classifier.

[0017] Preferably, step S2 includes:

[0018] Step S2a: Perform automatic center cropping on the coaxial molten pool image based on OpenCV to obtain the cropped molten pool image. Set the background outside the circular area to transparent, and retain RGB information within the circular area.

[0019] Step S2b: Obtain a grayscale image from the molten pool image after grayscale processing and cropping according to the set grayscale RGB channel weights, and extract morphological features, intensity distribution features, and texture features based on the principle of bright area dominance.

[0020] In step S3, the grayscale image is processed by PCA to extract principal components, which are then input into a CNN model to extract deep semantic information.

[0021] Preferably, the maximum defocusing amount is selected as 10mm, the minimum defocusing amount is selected as -4mm, and the interval is 2mm.

[0022] The equipment used for data acquisition has a frame rate higher than 30fps, and its spectral response range intersects with the laser band.

[0023] The grayscale RGB channel weights set reduce noise in grayscale images and reflectivity from the smooth inner wall of the cladding head and nozzle.

[0024] The cropped molten pool image has a size of 384×384, is in PNG format, and has an RGBA color mode.

[0025] Among them, channel A is the transparent channel; channels R, G, and B are the red, green, and blue channels, respectively.

[0026] The morphological features include area, perimeter, roundness, aspect ratio of the circumscribed rectangle, and convexity.

[0027] The intensity distribution characteristics include average intensity, intensity standard deviation, maximum intensity, minimum intensity, median intensity, skewness, and kurtosis.

[0028] The texture features include contrast, correlation, energy, and homogeneity.

[0029] Preferably, step S3 includes:

[0030] Step S3a: Perform principal component analysis on the grayscale image, retain the principal components with a cumulative variance contribution rate ≥ 95%, and input the processed grayscale image into the CNN model;

[0031] Step S3b: Extract the depth semantic information of the grayscale image based on the CNN algorithm, flatten the extracted depth semantic information, and perform secondary feature fusion with morphological features, intensity distribution features, and texture features;

[0032] Step S3c: Construct and train an MLP classifier based on a fully connected layer, using the fused features after the secondary feature fusion as the input to the MLP, and train and save the trained MLP classifier.

[0033] The MLP consists of an input layer, two hidden layers, and an output layer. The activation function of the hidden layers is ReLU, the optimizer is Adam, and the loss function is cross-entropy loss.

[0034] The training process incorporates macro-averaging and weighted averaging, and optimizes the MLP classifier through multi-fold crossover, hyperparameter adjustment, or integration with a ResNet residual network.

[0035] A laser coaxial powder feeding additive dual-channel defocusing amount recognition system according to the present invention includes:

[0036] Module M1: Acquires coaxial molten pool images of the additive manufacturing process under different defocusing amounts to form the original dataset;

[0037] Module M2 generates a grayscale image based on the coaxial molten pool image and extracts multi-dimensional feature distribution;

[0038] Module M3 extracts deep semantic features from grayscale images, fuses them with multi-dimensional features, and inputs them into an MLP classifier for training to identify the amount of defocus.

[0039] Preferably, module M1 acquires coaxial melt pool images corresponding to different defocus amounts as data, labels the corresponding defocus amounts, and constructs the original dataset.

[0040] The defocus amount is based on zero defocus amount, and positive and negative defocus amounts are symmetrically selected.

[0041] The module M2, based on OpenCV's automatic center cropping and set grayscale RGB channel weights, generates a grayscale image with less noise from a coaxial melt pool image, and extracts morphological features, intensity distribution features, and texture features from the grayscale image.

[0042] The module M3 includes a CNN model and an MLP classifier.

[0043] The feature extraction layer of the CNN model extracts deep semantic information from the grayscale image, fuses it with multi-dimensional features, and inputs the fused features into the MLP classifier for learning. The amount of defocus is then identified based on the learned MLP classifier.

[0044] Preferably, the module M2 includes:

[0045] Module M2a performs automatic center cropping on the coaxial molten pool image based on OpenCV to obtain the cropped molten pool image, sets the background outside the circular area to transparent, and retains RGB information within the circular area;

[0046] Module M2b obtains a grayscale image from the molten pool image after grayscale processing and cropping according to the set grayscale RGB channel weights, and extracts morphological features, intensity distribution features, and texture features based on the principle of bright area dominance.

[0047] The module M3 performs PCA processing on the grayscale image, extracts the principal components, and then inputs them into the CNN model to extract deep semantic information.

[0048] Preferably, the maximum defocusing amount is selected as 10mm, the minimum defocusing amount is selected as -4mm, and the interval is 2mm.

[0049] The grayscale RGB channel weights set reduce noise in grayscale images and reflectivity from the smooth inner wall of the cladding head and nozzle.

[0050] The cropped molten pool image has a size of 384×384, is in PNG format, and has an RGBA color mode.

[0051] Among them, channel A is the transparent channel; channels R, G, and B are the red, green, and blue channels, respectively.

[0052] The morphological features include area, perimeter, roundness, aspect ratio of the circumscribed rectangle, and convexity.

[0053] The intensity distribution characteristics include average intensity, intensity standard deviation, maximum intensity, minimum intensity, median intensity, skewness, and kurtosis.

[0054] The texture features include contrast, correlation, energy, and homogeneity.

[0055] Preferably, module M3 performs principal component analysis on the grayscale image, retains principal components with a cumulative variance contribution rate ≥ 95%, and inputs the processed grayscale image into the CNN model.

[0056] The CNN model extracts deep semantic information from grayscale images based on the CNN algorithm, flattens the extracted deep semantic information, and performs secondary feature fusion with morphological features, intensity distribution features, and texture features.

[0057] Train an MLP classifier based on a fully connected layer by using the fused features obtained from the fusion of secondary features as input to the MLP, and then train and save the trained MLP classifier.

[0058] The MLP consists of an input layer, two hidden layers, and an output layer. The activation function of the hidden layers is ReLU, the optimizer is Adam, and the loss function is cross-entropy loss.

[0059] The training process incorporates macro-averaging and weighted averaging, and optimizes the MLP classifier through multi-fold crossover, hyperparameter adjustment, or integration with a ResNet residual network.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] 1. This invention applies image processing and feature fusion methods to industrial images (coaxial images of single-pass additive manufacturing experiments) to identify the defocus amount during additive manufacturing, thereby avoiding forming defects caused by improper defocus setting and saving experimental materials and processing time.

[0062] 2. This invention only changes the defocus amount in a single-channel experiment while keeping other process parameters constant, ensuring that the obtained image features are highly correlated with the defocus amount, thus significantly improving recognition accuracy.

[0063] 3. This invention constructs a primary feature engineering processing chain for coaxial images from laser additive manufacturing, highlighting the multi-dimensional essential features of the molten pool, such as morphology, intensity, and texture, providing high-quality input for subsequent deep neural networks and overcoming the problem of insufficient representation capability of single morphological features. Attached Figure Description

[0064] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0065] Figure 1 A schematic diagram of a dual-channel defocusing amount identification method for laser coaxial powder feeding additive manufacturing;

[0066] Figure 2 This is a schematic diagram illustrating the noise reduction effect of the RGB ratio in an embodiment of the present invention. Detailed Implementation

[0067] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0068] This invention provides a method for identifying the defocus amount in a dual-channel laser-coaxial powder-feed additive manufacturing process. Through a single-pass additive manufacturing experiment, only the defocus amount is changed, without altering other process parameters such as laser power, powder feeding rate, and scanning speed. The coaxial images acquired during the process are then processed to construct a primary feature engineering processing chain for laser additive manufacturing coaxial images. Figure 1 For example, the specific steps include:

[0069] Step S1: Conduct additive manufacturing experiments. Set fixed process parameters (laser power, scanning speed, and powder feeding rate), adjust different defocusing amounts, collect coaxial molten pool images during the process as data, label each image with the corresponding defocusing amount, and construct a raw dataset for defocusing amount recognition.

[0070] In more preferred examples, since the molten pool is a typical target with strong light and strong interference characteristics, the image acquisition process not only needs to effectively suppress strong light but also needs to meet frame rate requirements. Regarding frame rate requirements, the commonly used scanning speed in actual deposition processes is 5-10 mm / s. Subsequent training images need to meet a minimum acquisition requirement of 1 mm / s, meaning the output of the molten pool image needs to be 5-10 Hz. Considering the safety margin in industrial scenarios, according to Shannon's sampling theorem, the frame rate usually needs to be significantly higher than the theoretical minimum, and the camera frame rate needs to be higher than 30 fps. Regarding strong light suppression and spectral matching requirements, there is interference from molten pool divergent light, laser beams, powder reflections, and ambient light during deposition. The camera needs to be equipped with appropriate filters to accurately select the target wavelength band and maintain efficient spectral acquisition. It is known that the wavelength band of a typical laser beam is 1030-1090 nm, requiring the camera's spectral response range to intersect with the laser wavelength band. Simultaneously, the filter needs to be selected according to the camera's spectral response curve to ensure the camera's photoelectric conversion efficiency and acquire clear images of the molten pool.

[0071] Based on the above requirements, in more preferred examples, the a2A1920-51gcBAS Basler ace 2 GigE camera is selected. The camera has a maximum frame rate of 51fps and a sensitivity range (spectral response range) between 400-700nm, which can block other unnecessary light interference.

[0072] The process parameters are: laser power of 1200W, scanning speed of 8mm / s, and powder feeding speed of 6.78g / min. The defocusing amount was selected with zero defocusing as the baseline, and positive and negative defocusing amounts were symmetrically selected as experimental design values, with an interval of 2mm. Excessive negative defocusing may damage the nozzle and its internal optical instruments; therefore, the maximum negative defocusing value was selected as -4mm. Excessive positive defocusing would prevent the laser energy received by the molten pool from completely melting the powder; therefore, the maximum positive defocusing value was selected as 10mm.

[0073] The coaxial CCD camera was used to acquire coaxial images of the molten pool during the laser additive manufacturing process, and the acquired raw images were labeled.

[0074] A dual-channel approach is used to perform grayscale processing and feature extraction on the acquired molten pool image, which is suitable for industrial images. The deep semantic information of the image is extracted by utilizing the image processing characteristics of CNN and fused with the features extracted in the first step. MLP learns the fused features and classifies these features.

[0075] The extraction of morphological features, intensity distribution features, and texture features from the grayscale image of the molten pool includes molten pool monitoring and image feature processing, which relies on manual extraction based on technical experience. CNN is used to extract deep semantic features. The extracted image features are directly related to the coaxial powder feeding additive manufacturing process. As the defocusing amount changes, the geometry of the molten pool, temperature distribution, etc., change significantly.

[0076] Step S2: Construct a primary feature engineering processing chain for a coaxial image from laser additive manufacturing.

[0077] Specifically, using OpenCV-based automatic center cropping and a set grayscale RGB channel ratio, a grayscale image with less noise than the original image is generated. Lower noise results in higher image quality. If multiple grayscale images are generated, the one with the least noise is selected for further processing. Following the principle of "brightness-dominated" (prioritizing or emphasizing areas with higher brightness, i.e., larger grayscale values, during feature extraction), morphological features, intensity distribution features, and texture features are extracted from the grayscale image. This includes the following steps:

[0078] During the conversion to grayscale image, when the R channel accounts for more than 50%, compared to when the G and B channels account for more than 50%, the noise of the grayscale image is significantly reduced and the reflectivity is lower. That is, under the set ratio of grayscale RGB channels, the noise of the grayscale image and the reflection of the smooth inner wall of the cladding head and nozzle are reduced.

[0079] Step S2a: Import the image acquired by the coaxial CCD camera into the program, and perform automatic center cropping on the acquired image based on OpenCV to obtain the cropped molten pool image. The cropped image size is 384×384, the format is PNG, and the image color mode is RGBA, where the A channel is the transparency channel, and the R, G, and B channels are the red, green, and blue channels, respectively. Since the molten pool image is similar to a circle, the inscribed circle of the cropped 384×384 image is taken as the final image before grayscale processing, that is, the background outside the circular area is set to transparent, while the RGB information of the original image is retained inside the circle.

[0080] Step S2b: Perform grayscale processing on the cropped image using the set grayscale RGB channel weights to obtain a grayscale image. Based on the principle of "bright area dominance", optimize the feature extraction logic to extract morphological features, intensity distribution features, and texture features from the molten pool grayscale image.

[0081] By setting the grayscale RGB channel ratio to reduce noise in the molten pool grayscale image and reflections from the smooth inner wall of the cladding head and nozzle, and by cropping to minimize the influence of bright areas outside the molten pool, the processed areas with higher brightness are more likely to contain the features that need to be extracted. The principle of prioritizing bright areas means that during feature extraction, attention should be given priority to or emphasized to the molten pool areas with higher brightness, i.e., larger grayscale values.

[0082] In more preferred examples, Figure 2 For example, the proportions of the RGB channels are 0.2, 0.6, and 0.2, respectively, to reduce noise around the molten pool image. This proportion is specifically optimized based on the molten pool's radiation spectrum characteristics and can effectively reduce noise interference such as spatter.

[0083] The morphological features include area, perimeter, roundness, aspect ratio of the circumscribed rectangle, and convexity; the intensity distribution features include average intensity, intensity standard deviation, maximum intensity, minimum intensity, median intensity, skewness, and kurtosis; the texture features include contrast, correlation, energy, and homogeneity. These features belong to the characteristics of the molten pool during the process. Combining these features for defocusing identification helps improve the identification accuracy.

[0084] Step S3: Perform secondary feature fusion on the features extracted in the first step and classify them using MLP (Multilayer Perceptron).

[0085] Step S3a, preprocessing, i.e. PCA data dimensionality reduction: perform principal component analysis (PCA) on the grayscale image, retain the principal components with a cumulative variance contribution rate of ≥95%, remove redundant information and reduce data dimensionality while retaining the key features of the melt pool grayscale image, and improve the efficiency of subsequent feature extraction and model training.

[0086] Performing PCA on the processed grayscale image can reduce the training time of the CNN (Convolutional Neural Network) model, improve training efficiency, and preserve the main components of the grayscale image. The grayscale image after extracting the principal components can then be used to train a CNN model.

[0087] This process involves extracting deep semantic information from grayscale images using the feature extraction layer of a CNN (Convolutional Neural Network) model. The manually extracted features (multi-dimensional features) are then fused with the CNN-extracted features (deep semantic information). The fused features are then input into an MLP (Multi-Level Processing) for learning, completing a task to recognize defocus. Specifically, this includes:

[0088] Step S3b: Extract deep semantic information from the image based on the CNN algorithm, flatten the extracted semantic information, and perform secondary feature fusion with the features extracted in the first step.

[0089] Step S3c, MLP: Construct and train an MLP classifier based on a fully connected layer, using the fused image features as input to the MLP.

[0090] The structure of the MLP is: input layer + two hidden layers + output layer. The activation function for all hidden layers is ReLU, the optimizer is Adam, and the loss function is cross-entropy loss.

[0091] In many other optimized examples, model performance and generalization ability can be improved by adjusting hyperparameters. Methods such as ResNet residual network can also be used to solve the gradient vanishing and gradient exploding problems in deep network training.

[0092] In more optimized examples, macro average and weighted average are introduced to reduce the impact of the number of samples in individual experimental categories on the training results. By using a multi-fold crossover method, any experimental category is used as the test set to exclude individual groups from affecting the training results due to the instability of actual experiments.

[0093] The MLP training results showed an accuracy of 0.9744 on the training set and 0.9850 on the test set. The training and test sets were divided into eight categories based on eight different defocus values. The training set contained 860 images, and the test set contained 267 images.

[0094] The present invention also provides a laser coaxial powder feeding additive dual-channel defocus amount identification system. The laser coaxial powder feeding additive dual-channel defocus amount identification system can be implemented by executing the process steps of the laser coaxial powder feeding additive dual-channel defocus amount identification method. That is, those skilled in the art can understand the laser coaxial powder feeding additive dual-channel defocus amount identification method as a preferred embodiment of the laser coaxial powder feeding additive dual-channel defocus amount identification system.

[0095] A laser coaxial powder feeding additive dual-channel defocusing amount recognition system according to the present invention includes:

[0096] Module M1: Acquires coaxial molten pool images of the additive manufacturing process under different defocusing amounts to form the original dataset;

[0097] Module M2 generates a grayscale image based on the coaxial molten pool image and extracts multi-dimensional feature distribution;

[0098] Module M3 extracts deep semantic features from grayscale images, fuses them with multi-dimensional features, and inputs them into an MLP classifier for training to identify the amount of defocus.

[0099] In more preferred embodiments, module M1 acquires coaxial molten pool images corresponding to different defocus amounts as data, labels the corresponding defocus amounts, and constructs the original dataset.

[0100] The defocus amount is based on zero defocus amount, and positive and negative defocus amounts are symmetrically selected.

[0101] The module M2, based on OpenCV's automatic center cropping and set grayscale RGB channel weights, generates a grayscale image with less noise from a coaxial melt pool image, and extracts morphological features, intensity distribution features, and texture features from the grayscale image.

[0102] The module M3 includes a CNN model and an MLP classifier.

[0103] The feature extraction layer of the CNN model extracts deep semantic information from the grayscale image, fuses it with multi-dimensional features, and inputs the fused features into the MLP classifier for learning. The amount of defocus is then identified based on the learned MLP classifier.

[0104] In more preferred embodiments, module M2 includes:

[0105] Module M2a performs automatic center cropping on the coaxial molten pool image based on OpenCV to obtain the cropped molten pool image, sets the background outside the circular area to transparent, and retains RGB information within the circular area;

[0106] Module M2b obtains a grayscale image from the molten pool image after grayscale processing and cropping according to the set grayscale RGB channel weights, and extracts morphological features, intensity distribution features, and texture features based on the principle of bright area dominance.

[0107] The module M3 performs PCA processing on the grayscale image, extracts the principal components, and then inputs them into the CNN model to extract deep semantic information.

[0108] In more preferred embodiments, the maximum defocusing amount is selected as 10mm, the minimum defocusing amount is selected as -4mm, and the interval is 2mm.

[0109] The grayscale RGB channel weights set reduce noise in grayscale images and reflectivity from the smooth inner wall of the cladding head and nozzle.

[0110] The cropped molten pool image has a size of 384×384, is in PNG format, and has an RGBA color mode.

[0111] Among them, channel A is the transparent channel; channels R, G, and B are the red, green, and blue channels, respectively.

[0112] The morphological features include area, perimeter, roundness, aspect ratio of the circumscribed rectangle, and convexity.

[0113] The intensity distribution characteristics include average intensity, intensity standard deviation, maximum intensity, minimum intensity, median intensity, skewness, and kurtosis.

[0114] The texture features include contrast, correlation, energy, and homogeneity.

[0115] In more preferred embodiments, module M3 performs principal component analysis on the grayscale image, retains principal components with a cumulative variance contribution rate ≥ 95%, and inputs the processed grayscale image into the CNN model.

[0116] The CNN model extracts deep semantic information from grayscale images based on the CNN algorithm, flattens the extracted deep semantic information, and performs secondary feature fusion with morphological features, intensity distribution features, and texture features.

[0117] Train an MLP classifier based on a fully connected layer by using the fused features obtained from the fusion of secondary features as input to the MLP, and then train and save the trained MLP classifier.

[0118] The MLP consists of an input layer, two hidden layers, and an output layer. The activation function of the hidden layers is ReLU, the optimizer is Adam, and the loss function is cross-entropy loss.

[0119] The training process incorporates macro-averaging and weighted averaging, and optimizes the MLP classifier through multi-fold crossover, hyperparameter adjustment, or integration with a ResNet residual network.

[0120] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function as logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0121] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for identifying the defocusing amount in dual-channel laser coaxial powder feeding additive manufacturing, characterized in that, include: Step S1: Collect coaxial molten pool images of the additive manufacturing process under different defocusing amounts to form the original dataset; Step S2: Based on OpenCV's automatic center cropping and the set grayscale RGB channel weights, generate a grayscale image from the coaxial melt pool image and extract multi-dimensional feature distribution; Step S3: The feature extraction layer of the CNN model extracts the deep semantic features of the grayscale image, fuses them with multi-dimensional features, and inputs them into the MLP classifier for training. The amount of defocus is identified based on the learned MLP classifier. Step S3 includes: Step S3a: Perform principal component analysis on the grayscale image, retain the principal components with a cumulative variance contribution rate ≥ 95%, and input the processed grayscale image into the CNN model; Step S3b: Extract the depth semantic information of the grayscale image based on the CNN algorithm, flatten the extracted depth semantic information, and perform secondary feature fusion with morphological features, intensity distribution features, and texture features; Step S3c: Construct and train an MLP classifier based on a fully connected layer. Use the fused features after the secondary feature fusion as the input of the MLP, train and save the trained MLP classifier.

2. The laser coaxial powder feeding additive dual-channel defocusing amount identification method according to claim 1, characterized in that, In step S1, fixed process parameters are set, different defocusing amounts are adjusted, corresponding coaxial molten pool images are collected as data, corresponding defocusing amount labels are labeled, and the original dataset is constructed. The defocus amount is based on zero defocus amount, with a maximum value of 10mm, a minimum value of -4mm, and an interval of 2mm. In step S2, a grayscale image with less noise than the original coaxial molten pool image is generated from the coaxial molten pool image, and morphological features, intensity distribution features, and texture features are extracted from the grayscale image. In step S3, the feature extraction layer of the CNN model extracts the deep semantic information of the grayscale image, fuses it with multi-dimensional features, inputs the fused features into the MLP classifier for learning, and identifies the defocus amount based on the learned MLP classifier.

3. The laser coaxial powder feeding additive dual-channel defocusing amount identification method according to claim 2, characterized in that, Step S2 includes: Step S2a: Perform automatic center cropping on the coaxial molten pool image based on OpenCV to obtain the cropped molten pool image. Set the background outside the circular area to transparent, and retain RGB information within the circular area. Step S2b: Obtain a grayscale image from the molten pool image after grayscale processing and cropping according to the set grayscale RGB channel weights; extract morphological features, intensity distribution features, and texture features based on the principle of bright area dominance. In step S3, the grayscale image is processed by PCA to extract principal components, which are then input into a CNN model to extract deep semantic information.

4. The laser coaxial powder feeding additive dual-channel defocusing amount identification method according to claim 3, characterized in that, The equipment used for data acquisition has a frame rate higher than 30fps, and its spectral response range intersects with the laser band. The set grayscale RGB channel weights reduce noise in grayscale images and reflectivity on the smooth inner wall of the cladding head and nozzle. The cropped molten pool image has a size of 384×384, is in PNG format, and has an RGBA color mode. Among them, channel A is the transparent channel; The R, G, and B channels are the red, green, and blue channels, respectively. The morphological features include area, perimeter, roundness, aspect ratio of the circumscribed rectangle, and convexity; The intensity distribution characteristics include average intensity, intensity standard deviation, maximum intensity, minimum intensity, median intensity, skewness, and kurtosis; The texture features include contrast, correlation, energy, and homogeneity.

5. The laser coaxial powder feeding additive dual-channel defocusing amount identification method according to claim 2, characterized in that, The MLP includes an input layer, two hidden layers, and an output layer. The activation function of the hidden layers is ReLU, the optimizer is Adam, and the loss function is cross-entropy loss. The training process incorporates macro-averaging and weighted averaging, and optimizes the MLP classifier through multi-fold crossover, hyperparameter adjustment, or integration with a ResNet residual network.

6. A dual-channel defocusing amount recognition system for laser coaxial powder feeding additive manufacturing, characterized in that, include: Module M1: Acquires coaxial molten pool images of the additive manufacturing process under different defocusing amounts to form the original dataset; Module M2, based on OpenCV, automatically centers and crops the grayscale image and sets the RGB channel weights, generates a grayscale image from the coaxial melt pool image, and extracts multi-dimensional feature distribution; Module M3 includes a CNN model and an MLP classifier; The module M3 performs principal component analysis on the grayscale image, retains the principal components with a cumulative variance contribution rate of ≥95%, and inputs the processed grayscale image into the CNN model. The feature extraction layer of the CNN model extracts deep semantic features of the grayscale image, fuses them with multi-dimensional features, and inputs them into the MLP classifier for training. The amount of defocus is identified based on the learned MLP classifier. The CNN model extracts the deep semantic information of grayscale images based on the CNN algorithm, flattens the extracted deep semantic information, and performs secondary feature fusion with morphological features, intensity distribution features, and texture features; Train an MLP classifier based on a fully connected layer by using the fused features obtained from the fusion of secondary features as input to the MLP, and then train and save the trained MLP classifier.

7. The laser coaxial powder feeding additive dual-channel defocusing amount recognition system according to claim 6, characterized in that, The module M1 acquires coaxial molten pool images corresponding to different defocus amounts as data, labels the corresponding defocus amounts, and constructs the original dataset. The defocus amount is based on zero defocus amount, with a maximum value of 10mm, a minimum value of -4mm, and an interval of 2mm. The module M2 generates a grayscale image with less noise than the original coaxial molten pool image from the coaxial molten pool image, and extracts morphological features, intensity distribution features and texture features from the grayscale image; The module M3 includes a CNN model and an MLP classifier; The feature extraction layer of the CNN model extracts deep semantic information from the grayscale image, fuses it with multi-dimensional features, and inputs the fused features into the MLP classifier for learning. The amount of defocus is then identified based on the learned MLP classifier.

8. The laser coaxial powder feeding additive dual-channel defocusing amount identification system according to claim 7, characterized in that, The module M2 includes: Module M2a performs automatic center cropping on the coaxial molten pool image based on OpenCV to obtain the cropped molten pool image, sets the background outside the circular area to transparent, and retains RGB information within the circular area; Module M2b obtains a grayscale image from the molten pool image after grayscale processing and cropping according to the set grayscale RGB channel weights, and extracts morphological features, intensity distribution features and texture features based on the principle of bright area dominance; The module M3 performs PCA processing on the grayscale image, extracts the principal components, and then inputs them into the CNN model to extract deep semantic information.

9. The laser coaxial powder feeding additive dual-channel defocusing amount identification system according to claim 8, characterized in that, The set grayscale RGB channel weights reduce noise in grayscale images and reflectivity on the smooth inner wall of the cladding head and nozzle. The cropped molten pool image has a size of 384×384, is in PNG format, and has an RGBA color mode. Among them, channel A is the transparent channel; The R, G, and B channels are the red, green, and blue channels, respectively. The morphological features include area, perimeter, roundness, aspect ratio of the circumscribed rectangle, and convexity; The intensity distribution characteristics include average intensity, intensity standard deviation, maximum intensity, minimum intensity, median intensity, skewness, and kurtosis; The texture features include contrast, correlation, energy, and homogeneity.

10. The laser coaxial powder feeding additive dual-channel defocusing amount recognition system according to claim 7, characterized in that, The MLP includes an input layer, two hidden layers, and an output layer. The activation function of the hidden layers is ReLU, the optimizer is Adam, and the loss function is cross-entropy loss. The training process incorporates macro-averaging and weighted averaging, and optimizes the MLP classifier through multi-fold crossover, hyperparameter adjustment, or integration with a ResNet residual network.

Images