Feature redundancy loss-based asymmetric self-supervised molten pool defect identification method and related system

Abstract

The invention provides an asymmetric self-supervised molten pool defect identification method based on feature redundancy loss and a related system, and belongs to the field of industrial quality monitoring and defect identification. By performing a view enhancement strategy on a label-free molten pool data set, the learning invariance and feature distinguishing capabilities of the model are enhanced; a self-supervised learning network architecture is introduced, and low-dimensional representation is extracted from the enhanced view of the molten pool image; feature redundancy loss is introduced to serve as a target loss function, redundancy between components of the vectors is reduced to the maximum extent, independence of component representation is improved, coupling between component representation is relieved, and correct convergence of network learning is guaranteed. According to the method, invariant feature knowledge of various label-free molten pool images can be learned, all useful information is fully utilized, identification based on the molten pool defect state is realized, and single molten pool image health state evaluation with strong indication and high precision can be output.

Description

Technical Field

[0001] The invention belongs to the field of industrial quality monitoring and defect identification, and specifically relates to an asymmetric self-supervised molten pool defect identification method based on feature redundancy loss and a related system. Background Art

[0002] In the Selective Laser Melting (SLM) process, the state monitoring of the molten pool is the key to ensure the quality and efficiency of the manufacturing process. As the core link in the forming process, the molten pool is directly related to all the basic parameters of the selective laser melting process, such as laser power, beam characteristics, scanning speed, etc. Therefore, the state of the molten pool can reflect the quality information of the constructed parts. However, due to the differences in parameters of different processing equipment, the formation and evolution of the molten pool show high sensitivity and variability. In actual industrial production, molten pool monitoring faces serious data annotation problems. A large amount of molten pool data is generated rapidly in the actual production process, and the amount of data is huge. Accurately annotating these data requires a lot of manpower, material resources and time costs. In addition, the formation and change of the molten pool are affected by many factors, including the performance of the processing equipment, the setting of process parameters, the characteristics of the material, etc. These factors are intertwined, making it extremely difficult to accurately annotate the molten pool data.

[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 the actual application of molten pool monitoring, due to the lack of labeled information, it is difficult to provide enough training samples for supervised learning methods. This makes it difficult to accurately establish an effective monitoring model and make accurate predictions about the state of the molten pool. Therefore, how to effectively monitor and analyze the molten pool in the absence of labeled information is an urgent problem to be solved in current industrial production. Summary of the invention

[0004] The purpose of the present invention is to overcome the problem that in practical applications of existing molten pool monitoring, it is difficult to establish an effective monitoring model due to the lack of annotation information, resulting in difficulty in detecting and analyzing the molten pool, and to provide an asymmetric self-supervised molten pool defect recognition method and related system based on feature redundancy loss.

[0005] In order to achieve the above object, the present invention adopts the following technical solution: In a first aspect, the present invention provides an asymmetric self-supervised molten pool defect recognition method based on feature redundancy loss, comprising the following steps: Obtain melt pool image datasets of different quality levels, including unlabeled melt pool images and their corresponding labels; Perform image enhancement on unlabeled melt pool images; 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 back-propagation training to obtain a low-dimensional feature representation representing the melt pool image; Divide the low-dimensional feature representation and its corresponding labels into training and test sets; After the training of the self-supervised learning network is completed, the SVM classifier is connected and trained based on the training set. After the SVM classifier is trained, the defects of the samples to be tested in the test set are classified and identified, and the quality level of the defects is determined according to the reliability evaluation related indicators.

[0006] The molten pool image data sets of different quality levels are acquired through an SLM process molten pool monitoring system.

[0007] The unlabeled melt pool image is enhanced to obtain two views v of the image A and v B .

[0008] The enhanced view based on the unlabeled melt pool image is constructed in a supervised learning network framework, and a feature redundancy loss function is introduced for back propagation training to obtain a low-dimensional feature representation representing the melt pool image. The specific training process is as follows: Construct an online network and a target network. The online network consists of a weight set Definition, the target network weight is defined as ; Set the view v A As for the online network, we get the image v A Prediction ; View v B As for the target network, we get the image v B The low-dimensional representation z B ; right and z B Perform L2 normalization and calculate the loss function based on the normalized results; Introduce the feature redundancy loss function, measure the cross-correlation matrix between the outputs of the two networks, and obtain the final 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 in the process of minimizing the loss function to obtain a low-dimensional feature representation representing the melt pool image.

[0009] The target network provides regression targets for the online network, and its parameters is through the online parameters The exponential moving average is updated, given a target decay rate , after each training step, we update : .

[0010] The process of calculating the loss function based on the standardized result is as follows: The L2 normalized Calculate the loss function. The initial default loss function is as follows:

[0011] After calculating the loss function, v B Put it on the online network and A Place it in the target network and calculate the symmetric loss of the loss function ; The loss function obtained is: 。

[0012] The introduced feature redundancy loss function is:

[0013] in Is a positive constant used to weigh the first term of the feature redundancy loss function With the second The importance of Output of the online network and target network in batch dimension and z B The cross-correlation matrix between can be calculated as follows:

[0014] At this time, here, b represents different samples in the same batch, that is, each element of the cross-correlation matrix is ​​calculated based on the batch dimension; i and j are the indices of different dimensions of the network output vector, is a square matrix whose values ​​are between -1 and 1, corresponding to completely irrelevant and completely relevant, respectively; The final loss function obtained after introducing feature redundancy loss in the self-supervised network architecture is:

[0015] in, is the introduced feature redundancy loss function, It is the symmetric loss corresponding to the feature redundancy loss function.

[0016] The formula for minimizing the loss function using the gradient descent method is:

[0017]

[0018] in, Represents the objective function The gradient at the current parameter value θ, Represents the learning rate used for weight updates during network training.

[0019] In the above-mentioned determination of the quality level of defects based on the reliability evaluation related indicators, the specific indicators are as follows: Recognition accuracy ACC: ; JS divergence between different categories of melt pool image features:

[0020] in Indicates the number of samples classified correctly, Represents the total number of samples in the recognition and classification process, and represents the probability distribution of the same feature dimension of two different categories of melt pool images, where i is the index of the feature dimension, express and The average distribution of express Divergence, the JS divergence value range is The smaller the value, the higher the overlap between the two features. Conversely, the larger the value, the lower the overlap between the two features.

[0021] In a second aspect, the present invention provides an asymmetric self-supervised molten pool defect recognition system based on feature redundancy loss, comprising: An acquisition module is used to acquire melt pool image datasets of different quality levels, including unlabeled melt pool images and their corresponding labels; An image enhancement module, used to enhance the image of the unlabeled melt pool; The model training module is used to build an enhanced view based on the unlabeled melt pool image in a self-supervised learning network framework, introduce a feature redundancy loss function for back-propagation training, and obtain a low-dimensional feature representation representing the melt pool image; A dataset construction module that divides low-dimensional feature representations and their corresponding labels into training and test sets; The classification module is used to access the SVM classifier after the self-supervised learning network is trained, and the SVM classifier is trained based on the training set. After the SVM classifier is trained, the defect classification and identification of the samples to be tested in the test set are performed, and the quality level of the defects is determined according to the reliability evaluation related indicators.

[0022] Compared with the prior art, the present invention has the following beneficial effects: The present invention provides an asymmetric self-supervised melt pool defect recognition method and related system based on feature redundancy loss, including the following steps: obtaining melt pool image data sets of different quality levels, including unlabeled melt pool images and their corresponding labels; performing image enhancement on the unlabeled melt pool images; constructing a self-supervised learning network framework based on the enhanced view of the unlabeled melt pool images, introducing a feature redundancy loss function for back propagation training, and obtaining a low-dimensional feature representation representing the melt pool image; dividing the low-dimensional feature representation and its corresponding label 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 for samples to be tested in the test set, and judging the quality level of defects according to reliability evaluation related indicators. The neural network is trained by the self-supervised learning network architecture, and does not need to use labeled labels, can be applied to downstream tasks, and meet the actual industrial generalization requirements; by introducing feature redundancy loss as the target loss function, the redundancy between the components of the cross-correlation function matrix vector is minimized, the independence of each component representation is increased, and then the coupling between the component representations is released, the correct convergence of network learning is ensured, and the defect recognition accuracy is improved. The present invention can learn the invariant feature knowledge of multiple unlabeled melt pool images, make full use of all useful information, realize the identification of the defect state of the melt pool, and output a single melt pool image health status evaluation with strong indicativeness and high accuracy.

[0023] Furthermore, based on the view enhancement strategy, the unlabeled melt pool image dataset is randomly transformed and scaled within a specific angle range to construct a preprocessed enhanced image dataset, which helps to enhance the model's ability to learn invariant and distinguish features, and provides diverse expressions of samples for subsequent model training. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 The asymmetric self-supervised molten pool defect recognition method based on feature redundancy loss proposed by the present invention; Figure 2 This is a schematic diagram of the arrangement of the SLM process molten pool monitoring system used in the experiment of the present invention; Figure 3 Sample images of melt pool image data sets with four quality levels used in the present invention; Figure 4 This is a BYOL self-supervised learning network diagram based on a dual network architecture adopted by the present invention; Figure 5 The loss curve in the experimental training process of the present invention and the t-SNE nonlinear dimension reduction visualization process diagram; Figure 6 It is a confusion matrix diagram of the classifier part of the present invention and the comparative method on the molten pool identification task; Table 1 shows the average JS divergence between different quality level categories under multiple dimensions calculated by the experiment of the present invention. DETAILED DESCRIPTION

[0025] In order to further understand the content of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the embodiments are only for explaining the present invention and are not intended to limit it.

[0026] like Figure 1 As shown, an asymmetric self-supervised melt pool defect recognition method based on feature redundancy loss includes the following steps: S1: Obtain melt pool image datasets of different quality levels, including unlabeled melt pool images and their corresponding labels; S2; image enhancement of unlabeled melt pool image; 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 back-propagation training to obtain a low-dimensional feature representation representing the melt pool image; S4: Divide the low-dimensional feature representation and its corresponding labels into training set and test set; S5: After the training of the self-supervised learning network is completed, the SVM classifier is connected and trained based on the training set. After the SVM classifier is trained, the defects of the samples to be tested in the test set are classified and identified, and the quality level of the defects is determined according to the reliability evaluation related indicators.

[0027] Specifically, in S1, since the morphology of the melt pool varies with the processing task, each melt pool image enhancement view is obtained from the collected and combined melt pool image datasets from different categories through the view enhancement strategy for subsequent model training. The specific process is as follows: The acquired molten pool image data set is a defect image with labeled data, and the model training process of the present invention adopts unlabeled images with labels removed.

[0028] First, the melt pool image dataset used this time was obtained using the SLM process melt pool monitoring system, which is arranged as follows: Figure 2 As shown in Figure 1, a coaxial processing area imaging setup was used. The powder used in the experiment was Ti6Al4V titanium alloy powder, and a coaxial color charge-coupled device (CCD) camera from Watec of Japan was used. The camera ran at a frame rate of 30 frames per second to capture the morphology of the processing area. The captured image consisted of three RGB channels with a resolution of 640×480 pixels. The captured image was extracted from the continuous video stream using Image J software, and the obtained melt pool image sample is shown in Figure 1. Figure 3As shown. By keeping the scanning speed constant at 2000mm / min, the changes in the morphology of the molten pool and process zone of each quality level are captured under four different laser power conditions. The laser powers used are 175W, 275W, 375W and 475W, respectively, to achieve four different build qualities. The quality of the processed parts is marked as P1, P2, P3, and P4, and the porosity of the parts is measured by CT scanning, which are 6.2%, 0.01%, 0.002%, and 0.001%, respectively. The molten pool image dataset used contains defect images of four quality levels.

[0029] Specifically, in S2, two different view enhancement transformations are performed on each image, such as cropping, rotation, grayscale, etc., to obtain two views v of the image. A and v B .

[0030] 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, and the training process of feature redundancy loss back propagation is introduced to obtain the low-dimensional feature representation representing the melt pool image. The specific method is as follows: First, build Figure 4 The BYOL (Bootstrap Your Own Latent) self-supervised learning network based on a dual network architecture includes an online network and a target network, both of which are implemented through neural network learning. The online network consists of a set of weights. Definition: This network structure integrates multiple components such as encoder, projector and predictor. The target network uses a similar architecture to the online network, but it does not include the predictor. The weights of the target network are Are unique and updated independently.

[0031] Secondly, the model is trained. The target network provides regression targets for the online network. The target parameters is through the online parameters Specifically, given a target decay rate , after each training step, we update :

[0032] use The current value of , calculate to ensure a balance between the stability of the target network and the real-time performance of the online network.

[0033] The specific training methods are as follows: 1) For each input of an original unlabeled melt pool image x, two views v of the image are obtained according to S2 A and v B , the two views of the image are placed in the online network and the target network respectively, and the view v A In the online network, the predicted representation of image x is obtained by passing through the encoder, projector and predictor respectively. , v B After passing through the encoder and projector in the target network, the image v is obtained B The low-dimensional representation z B ; 2) Yes and z B Perform L2 normalization; 3) L2 normalization Calculate the loss function. The initial default loss function is as follows:

[0034] in, They are The loss function is used to calculate the image v in the batch A The prediction indicates and image v B The low-dimensional representation z B The root mean square error between .

[0035] 4) After calculating the loss function, v B Put it on the online network and A Place it in the target network and use the initial default loss function in step 3) to calculate the symmetric loss of the loss function .

[0036] The loss function obtained is:

[0037] 5) In order to avoid the common constant solution problem of dual network structures, make the network results converge correctly, learn the essential useful feature vectors, and enhance the accuracy of the model, an objective function considering the 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 to the unit matrix as possible. This will cause the feature representation vectors of the enhanced views of the samples to be similar, while minimizing the redundancy between the components of these vectors, increasing the independence of the representations of each component, and thus decoupling the component representations from each other. The feature redundancy loss function is as follows:

[0038] in Is a positive constant used to weigh the first term of the feature redundancy loss function With the second The importance of Output of the online network and target network in batch dimension and z B The cross-correlation matrix between can be calculated as follows:

[0039] At this time, here, b represents different samples in the same batch, that is, each element of the cross-correlation matrix is ​​calculated based on the batch dimension; i and j are the indices of different dimensions of the network output vector, is a square matrix whose values ​​are between -1 and 1, corresponding to completely uncorrelated and completely correlated, respectively.

[0040] After calculating the feature redundancy loss function, the symmetric loss of the feature redundancy loss function is obtained by exchanging the transmission paths of data in the target network and the online network.

[0041] In summary, the final loss function obtained after introducing feature redundancy loss in the self-supervised network architecture is:

[0042] in, is the introduced feature redundancy loss function, To introduce the symmetric loss of the feature redundancy loss function, the final loss function should be defined as the sum of the feature redundancy loss and its symmetric loss.

[0043] 6) Use the gradient descent method to minimize the loss function, and update the parameters of the online network and the target network in the process of minimizing the loss function. The main principles of this network framework can be summarized in two formulas:

[0044]

[0045] in, Represents the objective function The gradient at the current parameter value θ, Represents the learning rate used to update weights during network training; this formula can be used to update the parameters of the online network and the target network.

[0046] Specifically, in S4, according to the obtained self-supervised network model based on feature redundancy loss after training, the encoder and mapper of the online network of the trained self-supervised learning network are evaluated using labeled melt pool images, and an SVM classifier is connected after the mapper to divide the labeled melt pool images into a training set and a test set; Specifically, in S5, the SVM classifier is trained based on the training set, and the trained SVM classifier is tested based on the test set, so as to realize the defect type recognition of each molten pool image.

[0047] Specifically through the following indicators: The relevant index values ​​of the molten pool defect identification process include: Recognition accuracy (ACC): ; JS divergence between different categories of melt pool image features:

[0048] in represents the number of samples classified correctly, Represents the total number of samples in the recognition and classification process, and represents the probability distribution of the same feature dimension of two different categories of melt pool images, where i is the index of the feature dimension, express and The average distribution of express Divergence, the JS divergence value range is The smaller the value, the higher the overlap between the two features. Conversely, the larger the value, the lower the overlap between the two features.

[0049] By implementing the above four-part process, a high level of accuracy in identifying defects of different categories can be guaranteed. In addition, by performing JS divergence calculation and probability distribution overlapping area map drawing comparison between the melt pool image features with different part porosity, it is verified that the proposed feature redundancy loss can better reduce the feature redundancy stability between the two categories, which can fully guide the state identification and maintenance process of the melt pool defects.

[0050] Embodiment 2: An asymmetric self-supervised melt pool defect recognition system based on feature redundancy loss, comprising: An acquisition module is used to acquire melt pool image datasets of different quality levels, including unlabeled melt pool images and their corresponding labels; An image enhancement module, used to enhance the image of the unlabeled melt pool; The model training module is used to build an enhanced view based on the unlabeled melt pool image in a self-supervised learning network framework, introduce a feature redundancy loss function for back-propagation training, and obtain a low-dimensional feature representation representing the melt pool image; A dataset construction module that divides low-dimensional feature representations and their corresponding labels into training and test sets; The classification module is used to access the SVM classifier after the self-supervised learning network is trained, and the SVM classifier is trained based on the training set. After the SVM classifier is trained, the defect classification and identification of the samples to be tested in the test set are performed, and the quality level of the defects is determined according to the reliability evaluation related indicators.

[0051] Embodiment 3: The public dataset of melt pool images based on the SLM process melt pool monitoring system includes a total of 62,194 melt pool images representing four quality levels. The dataset is randomly divided into three training, validation, and test subsets. The first subset contains 60% of the training images, the second subset (10%) is used for validation, and the remaining images are used to test the performance of the self-supervised algorithm. An asymmetric self-supervised melt pool defect recognition evaluation method based on feature redundancy loss is implemented on this dataset. The self-supervised learning architecture is trained using unlabeled melt pool data, and the encoder and mapper of the trained online network are evaluated using labeled melt pool data. The SVM classifier is connected after the mapper for training and testing. The loss curve during the training process of the self-supervised network architecture of the evaluation method and the t-SNE nonlinear dimensionality reduction visualization process are shown in the figure. Figure 5 As shown in the figure, the low-dimensional features of the melt pool image extracted by the trained encoder are used for downstream tasks to identify and classify the melt pool category, and 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, an ablation experiment is set up, replacing the feature redundancy loss with the common mean square error loss. Figure 6 (Right) is the confusion matrix of the ablation experiment. In addition, to further quantify the degree of overlap between the features of different categories of melt pool images, JS divergence is used to calculate the distribution overlap of different categories in each dimension feature in Table 1.

[0052] from Figure 5 It can be seen that the asymmetric self-supervised molten pool defect recognition and evaluation method based on feature redundancy loss proposed in the present invention has good convergence in the training process, can achieve dimensionality reduction division of features at different quality levels, and effectively distinguish different types of molten pools.

[0053] from Figure 6It can be seen that the asymmetric self-supervised melt pool defect recognition and evaluation method based on feature redundancy loss proposed in the present invention has better recognition effect than the self-supervised defect recognition method based on mean square error loss. It learns different feature embeddings between multiple categories without relying on explicit labels, which shows that the proposed feature redundancy loss can better reduce the feature redundancy between the two categories, and higher accuracy can be obtained by using lower-level feature mapping.

[0054] It can be seen from Table 1 that the JS divergence between P3 and P4 is small, that is, the difference in the feature distribution of the two categories is small, and the overlapping area is large, which is not conducive to the classifier to classify according to the features, resulting in more common classification errors in P3 and P4, which corresponds to the recognition confusion phenomenon between P3 and P4 in the actual recognition of the classifier. In addition, it also reflects that the asymmetric self-supervised melt pool defect recognition and evaluation method based on feature redundancy loss proposed in the present invention also has a good defect recognition accuracy when the multi-category melt pool features have a certain degree of overlap.

[0055] Table 1 Average JS divergence of feature distribution between different categories

[0056] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does 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 molten pool defect recognition method based on feature redundancy loss, characterized in that: The steps include: Obtain melt pool image datasets of different quality levels, including unlabeled melt pool images and their corresponding labels; Perform image enhancement on unlabeled melt pool images; 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 back-propagation training to obtain a low-dimensional feature representation representing the melt pool image; Divide the low-dimensional feature representation and its corresponding labels into training and test sets; After the training of the self-supervised learning network is completed, the SVM classifier is connected and trained based on the training set. After the SVM classifier is trained, the defects of the samples to be tested in the test set are classified and identified, and the quality level of the defects is determined according to the reliability evaluation related indicators.

2. The asymmetric self-supervised molten pool defect identification method based on feature redundancy loss according to claim 1 is characterized in that: The molten pool image data sets of different quality levels are acquired through an SLM process molten pool monitoring system.

3. The asymmetric self-supervised molten pool defect identification method based on feature redundancy loss according to claim 1 is characterized in that: The unlabeled melt pool image is enhanced to obtain two views v of the image A and v B .

4. The asymmetric self-supervised molten pool defect identification method based on feature redundancy loss according to claim 3 is characterized in that: The enhanced view based on the unlabeled melt pool image is constructed in a supervised learning network framework, and a feature redundancy loss function is introduced for back propagation training to obtain a low-dimensional feature representation representing the melt pool image. The specific training process is as follows: Construct an online network and a target network. The online network consists of a set of weights. Definition, the target network weight is defined as ; Set the view v A As for the online network, we get the image v A Prediction ; View v B As for the target network, we get the image v B The low-dimensional representation z B ; right and z B Perform L2 normalization and calculate the loss function based on the normalized results; Introduce the feature redundancy loss function, measure the cross-correlation matrix between the outputs of the two networks, and obtain the final 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 in the process of minimizing the loss function to obtain a low-dimensional feature representation representing the melt pool image.

5. The asymmetric self-supervised molten pool defect identification method based on feature redundancy loss according to claim 4 is characterized in that: The target network provides regression targets for the online network, and its parameters is through the online parameters The exponential moving average is updated, given a target decay rate , after each training step, we update : 。 6. The asymmetric self-supervised molten pool defect identification method based on feature redundancy loss according to claim 4 is characterized in that: The process of calculating the loss function based on the standardized result is as follows: The L2 normalization results in Calculate the first loss function. The initial default loss function is as follows: After calculating the loss function, v B Put it on the online network and A Place it in the target network and calculate the symmetric loss of the loss function ; The loss function obtained is: .

7. The asymmetric self-supervised molten pool defect identification method based on feature redundancy loss according to claim 6 is characterized in that: The introduced feature redundancy loss function is: in Is a positive constant used to weigh the first term in the feature redundancy loss function With the second The importance of Output of the online network and target network in batch dimension and z B The cross-correlation matrix between can be calculated as follows: At this time, here, b represents different samples in the same batch, that is, each element of the cross-correlation matrix is ​​calculated based on the batch dimension; i and j are the indices of different dimensions of the network output vector, is a square matrix whose values ​​are between -1 and 1, corresponding to completely irrelevant and completely relevant, respectively; The final loss function obtained after introducing feature redundancy loss in the self-supervised network architecture is: in, is the introduced feature redundancy loss function, To introduce the symmetric loss of feature redundancy loss function.

8. The asymmetric self-supervised molten pool defect identification method based on feature redundancy loss according to claim 4 is characterized in that: The formula for minimizing the loss function using the gradient descent method is: in, Represents the objective function The gradient at the current parameter value θ, Represents the learning rate used for weight updates during network training.

9. The asymmetric self-supervised molten pool defect identification method based on feature redundancy loss according to claim 1 is characterized in that: In the above-mentioned determination of the quality level of defects based on the reliability evaluation related indicators, the specific indicators are as follows: Recognition accuracy ACC: ; JS divergence between different categories of melt pool image features: in Indicates the number of samples classified correctly, Represents the total number of samples in the recognition and classification process, and represents the probability distribution of the same feature dimension of two different categories of melt pool images, where i is the index of the feature dimension, express and The average distribution of express Divergence, the JS divergence value range is The smaller the value, the higher the overlap between the two features. Conversely, the larger the value, the lower the overlap between the two features.

10. An asymmetric self-supervised molten pool defect recognition system based on feature redundancy loss, characterized in that: include: An acquisition module is used to acquire melt pool image datasets of different quality levels, including unlabeled melt pool images and their corresponding labels; An image enhancement module, used to enhance the image of the unlabeled melt pool; The model training module is used to build an enhanced view based on the unlabeled melt pool image in a self-supervised learning network framework, introduce a feature redundancy loss function for back-propagation training, and obtain a low-dimensional feature representation representing the melt pool image; A dataset construction module that divides low-dimensional feature representations and their corresponding labels into training and test sets; The classification module is used to access the SVM classifier after the self-supervised learning network is trained, and the SVM classifier is trained based on the training set. After the SVM classifier is trained, the defect classification and identification of the samples to be tested in the test set are performed, and the quality level of the defects is determined according to the reliability evaluation related indicators.

Images