A semi-supervised multi-label feature selection algorithm for laser metal deposition manufacturing quality monitoring

Abstract

The application provides a semi-supervised multi-label feature selection algorithm for laser metal deposition manufacturing quality monitoring. In order to improve the performance and efficiency of LMD quality monitoring, multiple features of the molten pool, temperature field and spatter are inputted, considering that artificial labeling is time-consuming and laborious and unlabeled data is easy to obtain, a semi-supervised multiple regression quality model is proposed to generate different quality level labels on unlabeled data, and a quality correlation evaluation function is proposed to calculate the correlation probability between the features and the quality indicators to evaluate the contribution of the features, so as to reduce the feature dimension and perform feature selection. The results show that the method can improve the accuracy and efficiency of multi-sensor monitoring in the manufacturing process, and solve the complexity and difficulty of application of multi-feature monitoring in LMD manufacturing.

Description

Technical Field

[0001] This invention belongs to the fields of additive manufacturing and machine vision, and specifically relates to a semi-supervised multi-label feature selection algorithm for quality monitoring in laser metal deposition manufacturing. Background Technology

[0002] Additive manufacturing (AM), also known as 3D printing, is a novel manufacturing technology. AM technology boasts advantages such as short production cycles, high economic efficiency, and unrestricted shape creation, solving many complex structural component forming problems and significantly shortening processing cycles, leading to its widespread application in important fields such as aerospace and machinery manufacturing. Laser metal deposition (LMD) is a laser metal AM technology. It uses metal powder as raw material and high-energy laser as energy to deposit metal powder layer by layer, thus achieving direct manufacturing of metal parts. However, LMD also suffers from various defects and forming quality issues, and the post-processing is time-consuming, making it difficult to improve the quality of manufactured parts. Therefore, quality monitoring throughout the entire manufacturing process is currently one of the key issues in LMD.

[0003] Machine vision inspection can intelligently monitor the entire forming and manufacturing process without the need for post-processing manual inspection, greatly optimizing the cost of formed parts and improving the efficiency of forming and manufacturing. Therefore, vision-based additive manufacturing quality monitoring has been extensively studied. Fang et al. proposed a molten pool image segmentation model based on fuzzy C-means to extract the length features of the molten pool for quality monitoring. Repossini et al. extracted features related to spatter from high-speed images to describe the suitability of laser powder bed fusion (LPBF) quality. Chen et al. extracted key features of the temperature field and established a mathematical model for quality monitoring based on these key features. While these methods have many advantages, when using multiple sensors and multiple features for quality monitoring, an excessively large input feature dimension can easily lead to model overfitting, thereby reducing the accuracy and efficiency of monitoring.

[0004] The purpose of this invention is to overcome these problems by proposing a novel semi-supervised multi-label feature selection algorithm for quality monitoring in laser metal deposition (LMD) manufacturing. This algorithm considers the time-consuming and labor-intensive nature of manual labeling, while unlabeled data is readily available. It also attempts to focus on the correlation between features and forming quality, thereby reducing feature dimensionality and facilitating feature selection. This method can improve the accuracy and efficiency of multi-sensor monitoring during the manufacturing process, addressing the complexity and difficulty in applying multi-feature monitoring in LMD manufacturing. Summary of the Invention

[0005] To address the aforementioned technical challenges, a semi-supervised multi-label feature selection algorithm for quality monitoring in laser metal deposition (LMD) manufacturing is proposed. This semi-supervised feature selection method exhibits stronger generalization capabilities compared to traditional methods, requires less manual annotation, and improves the correlation between features and forming quality, making it more suitable for online quality monitoring in LMD manufacturing processes.

[0006] The technical solution adopted by this invention to solve its technical problem is: a semi-supervised multi-label feature selection (SSMLFS) algorithm for quality monitoring of laser metal deposition manufacturing, including inputting multivariate features, a semi-supervised multivariate regression quality model, and calculating the correlation between multivariate features and quality indicators; specifically including the following steps:

[0007] Step S1: Input multi-dimensional features: Based on machine vision, use multiple sensors such as high-speed cameras and thermal imagers to collect images of the molten pool, temperature field and spatter during the manufacturing process, and extract key features from them: Molten pool features include area, perimeter, length, width, aspect ratio and molten pool intensity (defined as average gray value), spatter features: total area and number of spatter, temperature field features: maximum temperature, minimum temperature, average temperature and temperature gradient.

[0008] Step S2, Semi-supervised multivariate regression quality model: First, based on the microstructure (pores and cracks) of the formed parts after cutting, grinding, and chemical etching, the quality is divided into four levels according to the severity, and a portion of data with real quality labels is obtained; second, this labeled data is used to train the multivariate regression quality model; finally, the trained model is used to generate different quality level labels on unlabeled data.

[0009] Step S3: Calculate the correlation between multivariate features and quality indicators: A quality correlation evaluation function based on mutual information is proposed. This function uses a dynamic feature weight to address the high feature dimensionality problem in LMD imbalanced data. The evaluation function assesses the contribution of features by calculating the correlation probability between features and quality indicators, thereby selecting features highly correlated with forming quality and eliminating features with low correlation, achieving dimensionality reduction and optimization of multiple features.

[0010] Compared with existing technologies, the advantages of this invention are: based on the semi-supervised approach, we use a multivariate regression quality model to generate quality labels without requiring excessive real label data, greatly saving manual time and improving the model's generalization ability. Furthermore, for the microstructure of formed parts with pores and cracks, we provide quality grading standards and propose a quality correlation evaluation function based on these standards, giving a ranking of feature contribution and importance, which particularly improves the efficiency and performance of LMD multi-feature quality monitoring. Attached Figure Description

[0011] Figure 1 These are high-speed camera images and thermal imager images captured during the LMD manufacturing process;

[0012] Figure 2 This is a flowchart of a multivariate quality regression model;

[0013] Figure 3 It is a ranking of the contribution of features. Detailed Implementation

[0014] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0015] Figure 1 Images were captured using high-speed cameras and thermal imagers during the LMD manufacturing process. We extracted 12 key features from these images, as shown in Table 1: Molten pool features include area, perimeter, length, width, aspect ratio, and molten pool intensity (defined as average grayscale value); spatter features include the total area and number of spatters; and temperature field features include the maximum temperature, minimum temperature, average temperature, and temperature gradient.

[0016] Table 1. Symbols and meanings of multiple features

[0017]

[0018] like Figure 2 As shown, based on the microstructure (pores and cracks) of the formed part, the quality is divided into four levels: if any cracks are observed, it is considered "Grade D" (the worst quality, unsuitable for industrial production); if the calculated porosity is less than 0.5%, it is considered "Grade A" (the best quality); if the porosity is between 0.5% and 1%, it is considered "Grade B"; and if the porosity is greater than 1%, it is considered "Grade C". Here, porosity is defined as follows:

[0019]

[0020] We obtain data with true quality labels based on the quality levels, and use this data to train a multivariate regression quality model. The trained model is then used to generate different quality levels on unlabeled data. The model is defined as follows:

[0021]

[0022] Formula Q represents the function of the model, m is the total number of features, and θ i F represents the characteristic coefficient of the i-th feature. i Let λ represent the category of the i-th feature, and λ be the coefficient of the L2 regularization term.

[0023] Finally, in order to screen out features that are highly correlated with quality indicators, a quality correlation evaluation function based on mutual information is proposed, which is defined as follows:

[0024]

[0025] Formula J SSMLFS F is the expression for the function. i Q represents the category of the i-th feature. S For the s-th quality class, p(F) i ) and p(Q s ) represent discrete variables F i and Q s The probability distribution, p(F) i Q s ) is a variable (F) i Q s The joint probability distribution of ). The dynamic feature weights are defined as:

[0026]

[0027] formula y n For the nth training data, p(y) n Q s ) is a variable (y) n Q s The joint probability distribution of ), p(y) n Q s |F i ) is satisfied by the known variable F i The lower (y) n Q s The joint conditional probability distribution of y, p(y) n |F i ) is satisfied by the known variable F i The variable y n The conditional probability distribution.

[0028] The numerator of the dynamic feature weights represents the complementarity of features; a higher value indicates a greater impact from imbalanced data. The denominator is used to make different features and quality levels comparable, reducing the influence of high-dimensional data. For high-dimensional, multi-label data, dynamic feature weights can address the problem of data imbalance under multiple quality levels, further improving the stability and applicability of the SSMLFS algorithm.

[0029] like Figure 3 As shown, Figure 3 This illustrates the ranking of feature contributions. Figure 3The results show that: 1) The highest temperature feature has the highest contribution rate, which means that it best reflects the forming quality when monitored online in laser deposition manufacturing; 2) Compared with other features, the molten pool feature has the highest total contribution rate, and therefore reflects the forming quality better than the temperature field and spatter; 3) The feature contribution rate of the number of spatters has the lowest ranking, indicating that this feature has the highest correlation with quality.

[0030] Labeled data is time-consuming, labor-intensive, and difficult to obtain during quality monitoring in LMD manufacturing, while unlabeled data is readily available. Furthermore, when using multiple sensors and multiple features for quality monitoring, excessively large input feature dimensions can easily lead to overfitting of the monitoring model. Therefore, this invention proposes a semi-supervised multi-label feature selection algorithm for quality monitoring in laser metal deposition manufacturing. This algorithm provides quality grade standards for the microstructure of formed parts with porosity and cracks, and proposes a quality correlation evaluation function based on these standards, ultimately ranking the contribution and importance of features. This algorithm can improve the accuracy and efficiency of multi-sensor monitoring in LMD manufacturing, solving the problems of complexity and difficulty in application of multi-feature monitoring.

[0031] In summary, this invention provides a semi-supervised multi-label feature selection algorithm for quality monitoring in laser metal deposition (LMD) manufacturing. LMD forming and manufacturing suffers from various defects and forming quality problems. Quality monitoring throughout the entire manufacturing process is one of the key issues in LMD. Traditional methods rely on extensive manual annotation, which is time-consuming and labor-intensive. Furthermore, when performing multi-sensor, multi-feature quality monitoring, excessively large input feature dimensions can easily lead to model overfitting and low accuracy. To improve the performance and efficiency of LMD quality monitoring, this invention proposes a semi-supervised multi-label feature selection algorithm. We input multivariate features of the molten pool, temperature field, and spatter. Considering the time-consuming and labor-intensive nature of manual annotation and the readily available unlabeled data, a semi-supervised multivariate regression quality model is proposed to generate different quality level labels on the unlabeled data. A quality correlation evaluation function is proposed to calculate the correlation probability between features and quality indicators to evaluate the contribution of features, thereby reducing feature dimensionality and facilitating feature selection. Results show that this method can improve the accuracy and efficiency of multi-sensor monitoring during the manufacturing process, solving the complexity and application difficulties of multi-feature monitoring in LMD manufacturing.

Claims

1. A semi-supervised multi-label feature selection method for quality monitoring in laser metal deposition manufacturing, characterized in that, Including the following steps: Step 1: Input the visual images acquired by the high-speed camera and thermal imager to the molten pool, temperature field and spatter multi-dimensional characteristics; Step 2: Establish a semi-supervised multiple regression quality model for the unlabeled data generated during the forming process; We obtain data with true quality labels based on the quality levels, and use this data to train a multivariate regression quality model. The trained model is then used to generate different quality levels on unlabeled data. The model is defined as follows: （2）； Formula Q represents the function of the multiple regression quality model, m is the total number of features, and θ i F represents the characteristic coefficient of the i-th feature. i Let λ represent the category of the i-th feature, and λ be the coefficient of the L2 regularization term. Step 3: Calculate the correlation between the multivariate features and quality indicators of the input based on the quality correlation evaluation function of mutual information; The quality correlation evaluation function based on mutual information calculates the correlation between multiple input features and quality indicators, which includes the following steps: A quality correlation evaluation function based on mutual information is proposed. The evaluation function uses a dynamic feature weight to solve the high feature dimensionality problem in LMD imbalanced data. The evaluation function evaluates the contribution of features by calculating the correlation probability between features and quality indicators, thereby screening out features that are highly correlated with forming quality and eliminating features with low correlation, thus achieving dimensionality reduction and optimization of multiple features. A quality correlation evaluation function based on mutual information is proposed: （3）； Formula J SSMLFS F is the expression for the quality correlation evaluation function based on mutual information. i Q represents the category of the i-th feature. S For the s-th quality class, p(F) i ) and p(Q s ) represent discrete variables F i and Q s The probability distribution, p(F) i Q s ) is a variable (F) i Q s The joint probability distribution of ), where the dynamic feature weights are defined as: （4）； Formula y n For the nth training data, p(y) n Q s ) is a variable (y) n Q s The joint probability distribution of ), p(y) n Q s |F i ) is satisfied by the known variable F i The lower (y) n Q s The joint conditional probability distribution of y, p(y) n |F i ) is satisfied by the known variable F i The variable y n The conditional probability distribution.

2. The semi-supervised multi-label feature selection method for quality monitoring in laser metal deposition manufacturing according to claim 1, characterized in that, The input of visual images from multiple sensors, including high-speed cameras and thermal imagers, to the molten pool, temperature field, and spatter features includes the following steps: using multiple sensors, high-speed cameras and thermal imagers, to acquire images of the molten pool, temperature field, and spatter during the manufacturing process, and extracting key features from them: molten pool features include area, perimeter, length, width, aspect ratio, and molten pool intensity; spatter features include the total area and number of spatters; and temperature field features include the maximum temperature, minimum temperature, average temperature, and temperature gradient.

3. The semi-supervised multi-label feature selection method for quality monitoring in laser metal deposition manufacturing according to claim 1, characterized in that, The steps for establishing a semi-supervised multivariate regression quality model for unlabeled data generated during the forming process are as follows: First, based on the microstructure of the formed part after cutting, grinding, and chemical etching, the quality is divided into four levels according to the severity, and a portion of data with real quality labels is obtained; second, the labeled data is used to train the multivariate regression quality model; finally, the trained model is used to generate different quality level labels on the unlabeled data.

Images