Hybrid fuzzy-transformer-CNN network combined with fs-LA-SIBS for accurate classification of magnesium alloy components

Abstract

Accurate classification of magnesium alloys is frequently compromised by the intrinsic complexity of plasma emission spectra, characterized by subtle peak overlap between alloying elements and minor Mg emission lines as well as high background noise. In this work, a femtosecond laser-ablation spark-induced breakdown spectroscopy (fs-LA-SIBS) technique was developed and coupled with a novel Hybrid Fuzzy-Transformer-CNN (HybridTFNet) to achieve high-fidelity magnesium alloy component classification. The fs-LA-SIBS configuration leverages the dual energy input of femtosecond ablation and spark discharge, enhancing the signal-to-noise ratio by over 40% compared with conventional laser induced breakdown spectroscopy (LIBS). Taking 190-dimensional preprocessed one-dimensional spectral feature vectors as input, the HybridTFNet model innovatively integrates three functional modules: fuzzy logic for adaptive plasma noise suppression, transformer for modeling long-range wavelength dependencies in spectral signals, and convolutional neural networks (CNNs) for extracting local spectral feature patterns. This multimodal architecture effectively mitigates the heterogeneity of spectral signals caused by microalloying element content differences and subtle peak overlap of alloying elements in complex multi-element spectra. Experimental validation on five certified standard magnesium alloys showed that the proposed method achieved an average classification accuracy of 99.63% with a near-perfect micro-AUC of 0.9998. These results confirm that the integration of discharge-enhanced fs-LA-SIBS with the physically interpretable HybridTFNet provides a robust and efficient approach for rapid qualitative classification of magnesium alloy components, which holds great potential for industrial alloy quality control.