CN-116305975-B - Semi-supervised multi-label feature selection algorithm for laser metal deposition manufacturing quality monitoring

Abstract

The invention provides a semi-supervised multi-label feature selection algorithm for monitoring laser metal deposition manufacturing quality. In order to improve the performance and efficiency of LMD quality monitoring, the multi-element characteristics of a molten pool, a temperature field and splashing are input, the fact that unlabeled data are time-consuming and labor-consuming due to manual labeling is considered to be easily obtained, a semi-supervised multi-element regression quality model is provided to generate different quality grade labels on unlabeled data, a quality correlation evaluation function is provided to calculate the correlation probability between the characteristics and quality indexes so as to evaluate the contribution degree of the characteristics, and therefore the characteristic dimension is reduced and characteristic selection is carried out. The result shows that the method can improve the accuracy and efficiency of multi-sensor monitoring in the manufacturing process, and solves the problems of complexity and difficult application of multi-feature monitoring in LMD manufacturing.

Inventors

• WU ZIQIAN
• XU ZHENYING
• HAN LILING
• JIANG PAN
• Li linhang
• ZHANG YUXUAN

Assignees

• 江苏大学

Dates

Publication Date

20260512

Application Date

20230324

Claims (3)

1. 1. A semi-supervised multi-label feature selection method for laser metal deposition fabrication quality monitoring, comprising the steps of: step 1, inputting a molten pool, a temperature field and splash multielement characteristics according to visual images acquired by a high-speed camera and a multi-sensor of a thermal imager; step 2, establishing a semi-supervised multiple regression quality model for unlabeled data generated in the forming process; The data with the real quality label is obtained according to the quality grade, so that a multiple regression quality model is trained, the trained model is used for generating different quality grades on unlabeled data, and the definition of the model is as follows: (2); Equation Q represents a function of the multiple regression quality model, m is the total number of features, θ i is a feature coefficient representing the ith feature, F i represents a class of the ith feature, and λ is an L2 regularization term coefficient; Step 3, calculating the correlation between the input multi-element characteristics and the quality index based on the quality correlation evaluation function of the mutual information; the method comprises providing a quality correlation evaluation function based on mutual information, wherein a dynamic feature weight is used in the evaluation function to solve the problem of high feature dimension in LMD unbalanced data; the evaluation function evaluates the contribution degree of the features by calculating the correlation probability between the features and the quality indexes, so that the features highly correlated with the forming quality are screened out, the features with low correlation are removed, and the dimension reduction and optimization of multiple features are realized; a quality correlation evaluation function based on mutual information is provided: (3); Equation J SSMLFS is an expression of a mutual information-based quality-relevance evaluation function, F i represents a class of the ith feature, Q S is a class s quality class, p (F i ) and p (Q s ) represent probability distributions of discrete variables F i and Q s , respectively, p (F i , Q s ) is a joint probability distribution of the variable (F i ,Q s ), wherein the definition of dynamic feature weights is: (4); Equation y n is the nth training data, p (y n , Q s ) is the joint probability distribution of the variable (y n , Q s ), p (y n , Q s |F i ) is the joint conditional probability distribution of the lower (y n , Q s ) that satisfies the known variable F i , and p (y n |F i ) is the conditional probability distribution of the variable y n that satisfies the known variable F i .
2. 2. The method of semi-supervised multi label feature selection for laser metal deposition manufacturing quality monitoring of claim 1, wherein inputting molten pool, temperature field, and splash multi-element features from visual images acquired by high speed cameras, multiple sensors of a thermal imager includes acquiring molten pool, temperature field, and splash images during manufacturing using high speed cameras, multiple sensors of a thermal imager, and extracting key features from the molten pool features including area, perimeter, length, width, aspect ratio, and molten pool intensity, splash features including total area and quantity of splashes, temperature field features including maximum temperature, minimum temperature, average temperature, and temperature gradient.
3. 3. The method for semi-supervised multi-label feature selection for quality monitoring for laser metal deposition manufacturing of claim 1, wherein the creation of a semi-supervised multiple regression quality model for unlabeled data generated during the formation process includes the steps of first classifying the quality into four classes according to severity and obtaining a portion of data with a true quality label based on the microstructure of the formed part after cutting, sample grinding and chemical etching, second using the labeled data to train the multiple regression quality model, and finally using the trained model to generate different quality class labels on the unlabeled data.

Description

Semi-supervised multi-label feature selection algorithm for laser metal deposition manufacturing quality monitoring Technical Field The invention belongs to the field of additive manufacturing and machine vision, and particularly relates to a semi-supervised multi-label feature selection algorithm for monitoring laser metal deposition manufacturing quality. Background Additive Manufacturing (AM), also known as 3D printing, is a novel manufacturing technique. The AM technology has the advantages of short production period, good economy, unlimited molding shape and the like, solves the molding problem of a plurality of complex structural members, greatly shortens the processing period, and enables the AM technology to be widely applied in important fields such as aviation, aerospace, mechanical manufacturing and the like. Laser Metal Deposition (LMD) fabrication is a laser metal AM technique. The method takes metal powder as a raw material, and high-energy laser as an energy source, and deposits the metal powder layer by layer, thereby realizing the direct manufacture of metal parts. However, LMD also suffers from various drawbacks and formation quality problems, and the post-treatment process is time consuming and still difficult to use to improve the quality of the manufactured parts. Quality monitoring of the entire manufacturing process is therefore one of the key issues of current LMDs. The machine vision detection can intelligently monitor the whole forming manufacturing process without post manual detection, so that the cost of a formed part is greatly optimized, and the forming manufacturing efficiency is improved. Accordingly, extensive research has been conducted on vision-based additive manufacturing quality monitoring. Fang et al propose a pool image segmentation model based on fuzzy C-means to extract pool length features to monitor quality. Repossini et al extract features related to splatter from the acquired high-speed images to describe the suitability of the Laser Powder Bed Fusion (LPBF) quality. Chen et al extracted key features of the temperature field and established a quality monitoring mathematical model based on the key features. Although the method has a plurality of advantages, when multi-sensor and multi-feature quality monitoring is used, the oversized dimension of the input features is easy to cause model over fitting, so that the accuracy and efficiency of monitoring are reduced. The invention aims to overcome the problems and provides a novel semi-supervised multi-label feature selection algorithm for monitoring the quality of laser metal deposition manufacturing. The algorithm takes time and effort in manual labeling into account, while unlabeled data is readily available and attempts to focus on the correlation between features and forming quality, thereby reducing feature dimensions and making feature selections. The method can improve the accuracy and efficiency of multi-sensor monitoring in the manufacturing process, and solves the problems of complexity and difficult application of multi-feature monitoring in LMD manufacturing. Disclosure of Invention Aiming at the technical problems, a semi-supervised multi-label feature selection algorithm for monitoring the laser metal deposition manufacturing quality is provided. Compared with the traditional feature selection method, the algorithm semi-supervised feature selection method is stronger in generalization, does not need too many manual labels, improves the correlation between features and forming quality, and is more suitable for quality on-line monitoring in the LMD manufacturing process. The invention solves the technical problems by adopting a technical scheme that a semi-supervised multi-label feature selection (SSMLFS) algorithm for monitoring the quality of laser metal deposition manufacturing comprises the steps of inputting a multi-element feature, a semi-supervised multi-regression quality model and calculating the correlation between the multi-element feature and a quality index, and specifically comprises the following steps: S1, inputting multiple characteristics, namely acquiring molten pool, temperature field and splash images in the manufacturing process by using a high-speed camera, a thermal imager and other multiple sensors based on machine vision, and extracting key characteristics from the molten pool, wherein the molten pool characteristics comprise area, perimeter, length, width, length-width ratio and molten pool strength (defined as average gray values), the splash characteristics comprise total area and quantity of splashes, and the temperature field characteristics comprise maximum temperature, minimum temperature, average temperature and temperature gradient. Step S2, semi-supervised multiple regression quality model, namely firstly dividing the quality into four grades according to the severity according to microstructure (pores and cracks) of a formed part after cutting, sample grinding and ch