Research on a bimodal fusion detection method for surface defects of metal AM components based on LIBS

Metal Additive Manufacturing (AM) holds significant importance in advancing intelligent manufacturing and sustainable development. However, due to the unique manufacturing process of AM, defect detection in AM components has always been a challenging issue. This study employed Laser-Induced Breakdown Spectroscopy (LIBS) technology to capture spectral information and utilized a high-speed camera to record plasma images, comprehensively extracting pertinent details from each laser event. LIBS spectral scores were obtained via principal component analysis (PCA) and plasma image features were extracted to generate a bimodal fusion descriptor. This descriptor was employed to enhance the detection capability of three common surface defects in metal AM, specifically holes, cracks and bulges. Building on this foundation, a mid-level data fusion technique was employed to integrate the scores of LIBS spectra derived from PCA with seven features extracted from plasma images, resulting in the development of a bimodal fusion approach. Subsequently, the distribution of spectral data, plasma image features and bimodal fusion descriptors was discussed. Finally, three models, namely Random Forest (RF), Support Vector Machine (SVM) and Linear Discriminant Analysis (LDA), were used to evaluate the recognition accuracy of component defects. Additionally, the application scenarios of these three different models in spectral data, plasma image features and bimodal fusion descriptors were compared. The results indicate that the LDA model, utilizing bimodal fusion descriptors, yields the most effective classification. For samples #1–#100, the accuracy increased from 99.08% and 97.94% for spectral data and plasma image features to 99.92% for fusion data. Similarly, for samples #101–#120, the accuracy increases from 97.19% and 96.21% for spectral data and plasma image features to 97.34% for fusion data. This method improves the recognition of different defects of metal AM components. This study represents a first attempt to enhance the capability of LIBS in distinguishing various surface defects of metal AM components by inputting laser plasma image data and spectral information to generate statistical descriptors. The bimodal fusion approach offers an efficient method for detecting surface defects of metal AM components, characterized by low data complexity.