An Asymmetric Self-Supervised Melt Pool Defect Identification Method and Related System Based on Feature Redundancy Loss

Abstract

This invention provides an asymmetric self-supervised molten pool defect identification method and related system based on feature redundancy loss, belonging to the field of industrial quality monitoring and defect identification. By applying a view augmentation strategy to an unlabeled molten pool dataset, the model's ability to learn invariant and discriminative features is enhanced. A self-supervised learning network architecture is introduced to extract low-dimensional representations from the augmented views of the molten pool image. Feature redundancy loss is introduced as the objective loss function to minimize redundancy among the components of these vectors, increase the independence of each component representation, and thus decouple the component representations, ensuring correct convergence of the network learning. This method can learn invariant feature knowledge from various unlabeled molten pool images, fully utilize all useful information, achieve identification based on the state of molten pool defects, and output a highly indicative and accurate health status assessment of a single molten pool image.

Description

Technical Field

[0001] This invention belongs to the field of industrial quality monitoring and defect identification, specifically relating to an asymmetric self-supervised melt pool defect identification method and related system based on feature redundancy loss. Background Technology

[0002] In Selective Laser Melting (SLM), monitoring the state of the molten pool is crucial for ensuring manufacturing quality and efficiency. As a core component of the forming process, the molten pool is directly related to all fundamental parameters of SLM, such as laser power, beam characteristics, and scanning speed. Therefore, the state of the molten pool reflects the quality information of the constructed part. However, due to differences in parameters among different processing equipment, the formation and evolution of the molten pool exhibit high sensitivity and variability. In actual industrial production, molten pool monitoring faces significant data annotation challenges. A large amount of molten pool data is rapidly generated during actual production, and the data volume is enormous. Accurately annotating this data requires substantial manpower, resources, and time. Furthermore, the formation and changes of the molten pool are influenced by multiple factors, including the performance of the processing equipment, the setting of process parameters, and the properties of the materials. These factors are intertwined, making accurate annotation of molten pool data exceptionally difficult.

[0003] In the field of industrial quality monitoring, traditional supervised learning methods rely on a large amount of accurately labeled training data to build models. However, in practical applications of molten pool monitoring, the lack of labeled information makes it difficult to provide sufficient training samples for supervised learning methods. This results in the difficulty of accurately building effective monitoring models and predicting the state of the molten pool. Therefore, how to effectively monitor and analyze molten pools even in the absence of labeled information is a pressing problem in current industrial production. Summary of the Invention

[0004] The purpose of this invention is to overcome the problem that existing molten pool monitoring methods are difficult to implement in practical applications due to the lack of labeled information, which makes it difficult to establish an effective monitoring model and thus difficult to detect and analyze the molten pool. This invention provides an asymmetric self-supervised molten pool defect identification method and related system based on feature redundancy loss.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] In a first aspect, the present invention provides an asymmetric self-supervised melt pool defect identification method based on feature redundancy loss, comprising the following steps:

[0007] Obtain a dataset of molten pool images of different quality levels, including unlabeled molten pool images and their corresponding labels;

[0008] Image enhancement of unlabeled molten pool images;

[0009] Based on the enhanced view of the unlabeled molten pool image, a self-supervised learning network framework is constructed, and a feature redundancy loss function is introduced for backpropagation training to obtain a low-dimensional feature representation of the molten pool image.

[0010] The low-dimensional feature representations and their corresponding labels are divided into training and test sets;

[0011] After the self-supervised learning network is trained, an SVM classifier is connected. The SVM classifier is trained based on the training set. After the SVM classifier is trained, it performs defect classification and identification on the test samples in the test set, and judges the quality level of the defect based on relevant reliability evaluation indicators.

[0012] The datasets of molten pool images of different quality levels were acquired through the SLM process molten pool monitoring system.

[0013] The image enhancement of the unlabeled molten pool image yields two views of the image, v. A and v B .

[0014] The enhanced view based on the unlabeled melt pool image is constructed using a self-supervised learning network framework. A feature redundancy loss function is introduced for backpropagation training to obtain a low-dimensional feature representation of the melt pool image. The specific training process is as follows:

[0015] Construct an online network and a target network, where the online network consists of a set of weights. The target network weights are defined as follows: ;

[0016] View v A As for obtaining image v in online networks A Forecast indicates View v B As for the target network, the image v is obtained. B The low-dimensional representation of z B ;

[0017] right and z B Perform L2 standardization, and calculate the loss function based on the standardization results;

[0018] By introducing a feature redundancy loss function and measuring the cross-correlation matrix between the outputs of the two networks, the final loss function is obtained.

[0019] The gradient descent method is used to minimize the loss function, and the parameters of the online network and the target network are updated during the process of minimizing the loss function to obtain a low-dimensional feature representation of the melt pool image.

[0020] The target network provides a regression target for the online network, and its parameters are... Through online parameters It is updated using an exponential moving average, given a target decay rate. After each training step, we update as follows: :

[0021] .

[0022] The process of calculating the loss function based on the standardized results is as follows:

[0023] L2 standardization The loss function is calculated, and the initial default loss function is as follows:

[0024]

[0025] After calculating the loss function, then v B Placed on an online network, v A Placed in the target network, the symmetric loss of the loss function is calculated. ;

[0026] The resulting loss function is: 。

[0027] The introduced feature redundancy loss function is:

[0028]

[0029] in It is a positive constant used to weigh the first term of the feature redundancy loss function. With the second item The importance of Outputs of the online network and the target network in the batch dimension and z B The cross-correlation matrix between them can be calculated as follows:

[0030]

[0031] Here, 'b' represents different samples within the same batch, meaning each element of the cross-correlation matrix is ​​calculated with respect to the batch dimension; 'i' and 'j' are indices of different dimensions in the network output vector. It is a square matrix whose values ​​are between -1 and 1, corresponding to completely uncorrelated and completely correlated, respectively;

[0032] The final loss function obtained after introducing feature redundancy loss into the self-supervised network architecture is:

[0033]

[0034] in, The introduced feature redundancy loss function, This is the symmetric loss corresponding to the feature redundancy loss function.

[0035] The formula for minimizing the loss function using the gradient descent method is as follows:

[0036]

[0037]

[0038] in, Describe the objective function The gradient at the current parameter value θ, This represents the learning rate used for weight updates during network training.

[0039] The specific indicators for determining the quality level of defects based on reliability evaluation indicators are as follows:

[0040] Recognition accuracy (ACC): ;

[0041] JS divergence between different categories of melt pool image features:

[0042] in This represents the number of correctly classified samples. This indicates the total number of samples in the identification and classification process. and This represents the probability distribution of the same feature dimension for two different categories of melt pool images, where i is the index of the feature dimension. express and The average distribution express Divergence, its JS divergence value range is within The smaller the value, the higher the overlap between the two features; conversely, the larger the value, the lower the overlap.

[0043] Secondly, the present invention provides an asymmetric self-supervised melt pool defect identification system based on feature redundancy loss, comprising:

[0044] The acquisition module is used to acquire molten pool image datasets of different quality levels, including unlabeled molten pool images and their corresponding labels;

[0045] The image enhancement module is used to enhance the image of the unlabeled molten pool.

[0046] The model training module is used to construct an enhanced view based on the unlabeled melt pool image. In the self-supervised learning network framework, a feature redundancy loss function is introduced for backpropagation training to obtain a low-dimensional feature representation of the melt pool image.

[0047] The dataset construction module is used to divide low-dimensional feature representations and their corresponding labels into training and test sets;

[0048] The classification module is used to connect the SVM classifier after the self-supervised learning network has been trained. The SVM classifier is trained based on the training set. After the SVM classifier is trained, it performs defect classification and identification on the test samples in the test set, and judges the quality level of the defect based on relevant reliability evaluation indicators.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] This invention provides an asymmetric self-supervised molten pool defect identification method and related system based on feature redundancy loss, comprising the following steps: acquiring molten pool image datasets of different quality levels, including unlabeled molten pool images and their corresponding labels; performing image enhancement on the unlabeled molten pool images; constructing a self-supervised learning network framework based on the enhanced view of the unlabeled molten pool images, introducing a feature redundancy loss function for backpropagation training to obtain a low-dimensional feature representation representing the molten pool image; dividing the low-dimensional feature representation and its corresponding labels into a training set and a test set; connecting an SVM classifier after the trained self-supervised learning network, training the SVM classifier based on the training set, and after the SVM classifier is trained, classifying and identifying defects in the test samples in the test set, and determining the quality level of the defects based on reliability evaluation indicators. Training the neural network using a self-supervised learning network architecture eliminates the need for labeled data, enabling application to downstream tasks and meeting the generalization requirements of practical industries; by introducing feature redundancy loss as the target loss function, the redundancy between the components of the cross-correlation function matrix vector is minimized, increasing the independence of each component representation, thereby decoupling the component representations, ensuring correct convergence of network learning, and improving the accuracy of defect identification. This invention can learn invariant feature knowledge from various unlabeled molten pool images, make full use of all useful information, identify the defect state of the molten pool, and output a health status evaluation of a single molten pool image with strong indicativeness and high accuracy.

[0051] Furthermore, based on the view augmentation strategy, the unlabeled melt pool image dataset is randomly transformed and scaled within a specific angular range to construct a preprocessed augmented image dataset, which helps to enhance the model's ability to learn invariant and distinguish features, and provides diverse representations of samples for subsequent model training. Attached Figure Description

[0052] Figure 1This invention presents an asymmetric self-supervised melt pool defect identification method based on feature redundancy loss.

[0053] Figure 2 This is a schematic diagram of the layout of the SLM process melt pool monitoring system used in the experiments of this invention;

[0054] Figure 3 These are sample images of the four quality levels of melt pool images used in this invention;

[0055] Figure 4 This is a diagram of the BYOL self-supervised learning network based on a dual-network architecture used in this invention.

[0056] Figure 5 The diagram shows the loss curve and the visualization process of t-SNE nonlinear dimensionality reduction during the experimental training of this invention.

[0057] Figure 6 This is a confusion matrix diagram of the classifier part and comparison method of the present invention on the melt pool identification task;

[0058] Table 1 shows the average JS divergence among different quality level categories under multiple dimensions calculated in the experiment of this invention. Detailed Implementation

[0059] To further understand the content of this invention, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention.

[0060] like Figure 1 As shown, an asymmetric self-supervised melt pool defect identification method based on feature redundancy loss includes the following steps:

[0061] S1: Obtain a dataset of molten pool images of different quality levels, including unlabeled molten pool images and their corresponding labels;

[0062] S2; Image enhancement of the unlabeled molten pool image;

[0063] S3: Based on the enhanced view of the unlabeled melt pool image, a self-supervised learning network framework is constructed, and a feature redundancy loss function is introduced for backpropagation training to obtain a low-dimensional feature representation of the melt pool image.

[0064] S4: Divide the low-dimensional feature representations and their corresponding labels into training and test sets;

[0065] S5: After the self-supervised learning network is trained, an SVM classifier is connected. The SVM classifier is trained based on the training set. After the SVM classifier is trained, it performs defect classification and identification on the test samples in the test set, and judges the quality level of the defect based on the relevant reliability evaluation indicators.

[0066] Specifically, in S1, since the morphology of the molten pool varies with the processing task, an enhanced view of each molten pool image is obtained from the collected and combined datasets of molten pool images from different categories through a view augmentation strategy for subsequent model training. The specific process is as follows:

[0067] The obtained melt pool image dataset consists of labeled defect images, while the model training process of this invention uses unlabeled images without labels.

[0068] First, the molten pool image dataset used in this study was acquired using the SLM process molten pool monitoring system, and its layout is as follows: Figure 2 As shown, a coaxial imaging setup was used for the machining area. The powder used in the experiment was Ti6Al4V titanium alloy powder, and a coaxial color charge-coupled device (CCD) camera from Watec Corporation, Japan, was employed. This camera operated at a frame rate of 30 frames per second to capture the morphology of the machining area. The captured image consisted of three RGB channels with a resolution of 640×480 pixels. ImageJ software was used to extract the captured images from a continuous video stream, and the resulting molten pool image samples are shown below. Figure 3 As shown, by maintaining a constant scanning speed of 2000 mm / min, variations in the molten pool and process zone morphology were captured for each quality level under four different laser power conditions. Laser powers of 175 W, 275 W, 375 W, and 475 W were used to achieve four different build qualities. The processed part qualities were labeled P1, P2, P3, and P4, and their porosity was measured using CT scanning, yielding 6.2%, 0.01%, 0.002%, and 0.001%, respectively. The molten pool image dataset used contains defect images for the four quality levels.

[0069] Specifically, in S2, two different view enhancement transformations are performed on each image, such as cropping, rotation, and grayscale conversion, to obtain two views v of the image. A and v B .

[0070] Specifically, in S3, based on the enhanced view of the unlabeled melt pool image, the encoder embedding and mapping of the online network and the target network in the self-supervised learning framework are constructed. The training process of backpropagation with feature redundancy loss is introduced to obtain a low-dimensional feature representation of the melt pool image. The specific method is as follows:

[0071] First, build Figure 4 The BYOL (Bootstrap Your Own Latent) self-supervised learning network, based on a dual-network architecture, consists of an online network and a target network, both learned through neural networks. The online network comprises a set of weights. This network architecture, as defined, integrates multiple components such as an encoder, projector, and predictor. The target network employs a similar architecture to the online network, but it does not include a predictor. The weights of the target network... It is unique and updated independently.

[0072] Secondly, model training is performed, with the target network providing the regression target and target parameters for the online network. Through online parameters It is updated using an exponential moving average. Specifically, given a target decay rate... After each training step, we update as follows: :

[0073]

[0074] use The current value, combined with the decay rate Calculate The new value is used to ensure a balance between the stability of the target network and the real-time performance of the online network.

[0075] The specific training methods are as follows:

[0076] 1) For each input raw unlabeled melt pool image x, obtain two views v of the image according to S2. A and v B The two views of the image are placed in the online network and the target network, respectively. View v A The image x is predicted by passing through an encoder, projector, and predictor in an online network. v B After passing through the encoder and projector in the target network, the image v is obtained. B The low-dimensional representation of z B ;

[0077] 2) To and z B Perform L2 standardization;

[0078] 3) The result of L2 standardization The loss function is calculated, and the initial default loss function is as follows:

[0079]

[0080] in, They are respectively The L2 normalization is obtained. The loss function is used to calculate the image v within a batch. A The prediction indicates and image v B The low-dimensional representation of zB The root mean square error between them.

[0081] 4) After calculating the loss function, then v B Placed on an online network, v A Placed in the target network, the initial default loss function from step 3) is used for calculation to obtain the symmetric loss of the loss function. .

[0082] The resulting loss function is:

[0083] 5) To avoid the constant solution problem common in dual-network structures, ensure correct convergence of network results, learn essentially useful feature vectors, and enhance model accuracy, a feature redundancy loss function considering cross-correlation matrix loss is proposed. This function measures the cross-correlation matrix between the outputs of the two networks and makes it as close as possible to the identity matrix. This results in similar feature representation vectors in the enhanced view of the samples, while minimizing redundancy between the components of these vectors, increasing the independence of each component representation, and thus decoupling the component representations. The feature redundancy loss function is as follows:

[0084]

[0085] in It is a positive constant used to weigh the first term of the feature redundancy loss function. With the second item The importance of Outputs of the online network and the target network in the batch dimension and z B The cross-correlation matrix between them can be calculated as follows:

[0086]

[0087] Here, 'b' represents different samples within the same batch, meaning each element of the cross-correlation matrix is ​​calculated with respect to the batch dimension; 'i' and 'j' are indices of different dimensions in the network output vector. It is a square matrix with values ​​between -1 and 1, corresponding to completely uncorrelated and completely correlated, respectively.

[0088] After calculating the feature redundancy loss function, the symmetric loss of the feature redundancy loss function is obtained by exchanging the data transmission paths in the target network and the online network.

[0089] In summary, the final loss function obtained after introducing feature redundancy loss into the self-supervised network architecture is:

[0090]

[0091] in, The introduced feature redundancy loss function, To introduce a symmetric loss for the feature redundancy loss function, the final loss function should be defined as the sum of the feature redundancy loss and its symmetric loss.

[0092] 6) Minimize the loss function using gradient descent, and update the parameters of the online network and the target network during the process. The main principle of this network framework can be summarized by two formulas:

[0093]

[0094]

[0095] in, Describe the objective function The gradient at the current parameter value θ, This represents the learning rate used for weight updates during network training; this formula allows for parameter updates between the online network and the target network.

[0096] Specifically, in S4, based on the obtained self-supervised network model trained on feature redundancy loss, the encoder and mapper of the online network of the trained self-supervised learning network are evaluated using labeled melt pool images. An SVM classifier is then connected after the mapper to divide the labeled melt pool images into training and test sets.

[0097] Specifically, in S5, the SVM classifier is trained based on the training set and tested based on the test set, thereby realizing the identification of defect types in each melt pool image.

[0098] Specifically, the following indicators are used:

[0099] Relevant index values ​​for the molten pool defect identification process include:

[0100] Accuracy of recognition (ACC): ;

[0101] JS divergence between different categories of melt pool image features:

[0102] in This represents the number of correctly classified samples. This indicates the total number of samples in the identification and classification process. and This represents the probability distribution of the same feature dimension for two different categories of melt pool images, where i is the index of the feature dimension. express and The average distribution express Divergence, its JS divergence value range is within The smaller the value, the higher the overlap between the two features; conversely, the larger the value, the lower the overlap.

[0103] By implementing the above four-part process, a high level of accuracy in identifying different types of defects can be guaranteed. In addition, by comparing the JS divergence calculation and probability distribution overlap region plot of molten pool image features with different part porosities, it is verified that the proposed feature redundancy loss effectively reduces the feature redundancy stability between the two categories. This can fully guide the state identification and maintenance process of molten pool defects.

[0104] Example 2:

[0105] An asymmetric self-supervised melt pool defect identification system based on feature redundancy loss includes:

[0106] The acquisition module is used to acquire molten pool image datasets of different quality levels, including unlabeled molten pool images and their corresponding labels;

[0107] The image enhancement module is used to enhance the image of the unlabeled molten pool.

[0108] The model training module is used to construct an enhanced view based on the unlabeled melt pool image. In the self-supervised learning network framework, a feature redundancy loss function is introduced for backpropagation training to obtain a low-dimensional feature representation of the melt pool image.

[0109] The dataset construction module is used to divide low-dimensional feature representations and their corresponding labels into training and test sets;

[0110] The classification module is used to connect the SVM classifier after the self-supervised learning network has been trained. The SVM classifier is trained based on the training set. After the SVM classifier is trained, it performs defect classification and identification on the test samples in the test set, and judges the quality level of the defect based on relevant reliability evaluation indicators.

[0111] Example 3:

[0112] The publicly available dataset of molten pool images from the SLM process molten pool monitoring system includes 62,194 images representing four quality levels. The dataset was randomly divided into three subsets: training, validation, and testing. The first subset contained 60% of the training images, the second subset (10%) was used for validation, and the remaining images were used to test the performance of the self-supervised algorithm. An asymmetric self-supervised molten pool defect identification evaluation method based on feature redundancy loss was implemented on this dataset. The self-supervised learning architecture was trained using unlabeled molten pool data, and the encoder and mapper of the trained online network were evaluated using labeled molten pool data. An SVM classifier was then added after the mapper for training and testing. The loss curves during the training process of the self-supervised network architecture and the visualization process of t-SNE nonlinear dimensionality reduction are shown below. Figure 5 As shown, the low-dimensional features of the molten pool image extracted using the trained encoder are used for downstream tasks to identify and classify molten pool categories. The classifier is set to a support vector machine (SVM) classification model. Figure 6 (Left) Shows the confusion matrix of the classifier on the melt pool recognition task. Based on the proposed model, ablation experiments were conducted, replacing the feature redundancy loss with the common mean squared error loss. Figure 6 (Right) is the confusion matrix of the ablation experiment. In addition, to further quantify the degree of overlap between different categories of melt pool image features, JS divergence was used for calculation. Table 1 shows the distribution overlap of different categories in each dimension of features.

[0113] from Figure 5 It can be seen that the asymmetric self-supervised melt pool defect identification and evaluation method based on feature redundancy loss proposed in this invention has good convergence in the training process, can achieve dimensionality reduction of features under different quality levels, and effectively distinguishes different types of melt pools.

[0114] from Figure 6 It can be seen that the asymmetric self-supervised melt pool defect identification and evaluation method based on feature redundancy loss proposed in this invention has a better identification effect than the self-supervised defect identification method based on mean square error loss. It learns different feature embeddings between multiple categories without relying on explicit labels, indicating that the proposed feature redundancy loss effectively reduces feature redundancy between the two categories. Higher accuracy can be obtained by using lower-level feature mapping.

[0115] As shown in Table 1, the JS divergence between P3 and P4 is small, indicating a small difference in feature distribution between the two categories and a large overlapping area. This is unfavorable for the classifier to classify based on features, leading to more common classification errors in P3 and P4. This corresponds to the recognition confusion phenomenon between P3 and P4 in actual identification. Furthermore, it also reflects that the asymmetric self-supervised melt pool defect identification and evaluation method based on feature redundancy loss proposed in this invention has good defect identification accuracy even when the features of multiple melt pool categories have a certain degree of overlap.

[0116] Table 1. Mean JS divergence of feature distributions among different categories

[0117]

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An asymmetric self-supervised melt pool defect identification method based on feature redundancy loss, characterized in that, Includes the following steps: Obtain a dataset of molten pool images of different quality levels, including unlabeled molten pool images and their corresponding labels; Image enhancement is performed on the unlabeled molten pool image to obtain two views of the image, v. A and v B ; Based on the enhanced view of the unlabeled molten pool image, a self-supervised learning network framework is constructed, and a feature redundancy loss function is introduced for backpropagation training to obtain a low-dimensional feature representation of the molten pool image. Specifically: Construct an online network and a target network, where the online network consists of a set of weights. The target network weights are defined as follows: ; View v A As for obtaining image v in online networks A Forecast indicates View v B As for the target network, the image v is obtained. B The low-dimensional representation of z B ; right and z B Perform L2 standardization, and calculate the loss function based on the standardization results; L2 standardization The first loss function is calculated, with the initial default loss function as follows: After calculating the loss function, then v B Placed on an online network, v A Placed in the target network, the symmetric loss of the loss function is calculated. ; The resulting loss function is: ; By introducing a feature redundancy loss function and measuring the cross-correlation matrix between the outputs of the two networks, the final loss function is obtained. The introduced feature redundancy loss function is: in It is a positive constant used to weigh the first term in the feature redundancy loss function. With the second item The importance of Outputs of the online network and the target network in the batch dimension and z B The cross-correlation matrix between them is calculated using the following formula: Here, 'b' represents different samples within the same batch, meaning each element of the cross-correlation matrix is ​​calculated with respect to the batch dimension; 'i' and 'j' are indices of different dimensions in the network output vector. It is a square matrix whose values ​​are between -1 and 1, corresponding to completely uncorrelated and completely correlated, respectively; The final loss function obtained after introducing feature redundancy loss into the self-supervised network architecture is: in, The introduced feature redundancy loss function, To introduce a symmetric loss for the feature redundancy loss function; The gradient descent method is used to minimize the loss function, and the parameters of the online network and the target network are updated during the process of minimizing the loss function to obtain a low-dimensional feature representation of the melt pool image. The low-dimensional feature representations and their corresponding labels are divided into training and test sets; After the self-supervised learning network is trained, an SVM classifier is connected. The SVM classifier is trained based on the training set. After the SVM classifier is trained, it performs defect classification and identification on the test samples in the test set, and judges the quality level of the defect based on relevant reliability evaluation indicators.

2. The asymmetric self-supervised melt pool defect identification method based on feature redundancy loss according to claim 1, characterized in that, The datasets of molten pool images of different quality levels were acquired through the SLM process molten pool monitoring system.

3. The asymmetric self-supervised melt pool defect identification method based on feature redundancy loss according to claim 1, characterized in that, The target network provides a regression target for the online network, and its parameters are... Through online parameters It is updated using an exponential moving average, given a target decay rate. After each training step, we update as follows: : 。 4. The asymmetric self-supervised melt pool defect identification method based on feature redundancy loss according to claim 1, characterized in that, The formula for minimizing the loss function using the gradient descent method is as follows: in, Describe the objective function The gradient at the current parameter value θ, This represents the learning rate used for weight updates during network training.

5. The asymmetric self-supervised melt pool defect identification method based on feature redundancy loss according to claim 1, characterized in that, The specific indicators for determining the quality level of defects based on reliability evaluation indicators are as follows: Recognition accuracy (ACC): ; JS divergence between different categories of melt pool image features: in This represents the number of correctly classified samples. This indicates the total number of samples in the identification and classification process. and This represents the probability distribution of the same feature dimension for two different categories of melt pool images, where i is the index of the feature dimension. express and average distribution express Divergence, its JS divergence value range is within The smaller the value, the higher the overlap between the two features; conversely, the larger the value, the lower the overlap.

6. An asymmetric self-supervised molten pool defect identification system based on feature redundancy loss, based on the asymmetric self-supervised molten pool defect identification method based on feature redundancy loss according to any one of claims 1 to 5, characterized in that, include: The acquisition module is used to acquire molten pool image datasets of different quality levels, including unlabeled molten pool images and their corresponding labels; The image enhancement module is used to enhance the image of the unlabeled molten pool. The model training module is used to construct an enhanced view based on the unlabeled melt pool image. In the self-supervised learning network framework, a feature redundancy loss function is introduced for backpropagation training to obtain a low-dimensional feature representation of the melt pool image. The dataset construction module is used to divide low-dimensional feature representations and their corresponding labels into training and test sets; The classification module is used to connect the SVM classifier after the self-supervised learning network has been trained. The SVM classifier is trained based on the training set. After the SVM classifier is trained, it performs defect classification and identification on the test samples in the test set, and judges the quality level of the defect based on relevant reliability evaluation indicators.

Images