Fuzzy-Enhanced Transformer–CNN Feature Fusion for Robust Steel Alloy Classification Using Femtosecond Laser-Ablation Spark-Induced Breakdown Spectroscopy

Accurate steel alloy classification is a core task in industrial quality control, yet laser-induced breakdown spectroscopy (LIBS) data suffers from spectral noise, matrix effects, and multi-element peak overlap. Femtosecond laser-ablation spark-induced breakdown spectroscopy (fs-LA-SIBS) enables high-sensitivity detection, but plasma fluctuations degrade the performance of conventional chemometric methods and standalone deep learning models. To address these issues, we propose HybridTFNet, a hybrid fuzzy-Transformer-convolutional neural network (CNN) framework for uncertainty-aware spectral learning and multi-scale feature fusion. A fuzzy logic layer with Gaussian membership functions quantifies spectral uncertainty and suppresses noise-dominated regions. A Transformer encoder captures long-range wavelength dependencies and global inter-element coupling, while a one-dimensional CNN extracts fine-grained local features such as characteristic peaks and wavelength shifts. Adaptive fusion of global and local representations constructs a highly discriminative feature space for classification. Experiments on nine certified steel alloy fs-LA-SIBS datasets show that HybridTFNet achieves 100% classification accuracy across four independent test sets, significantly outperforming traditional machine learning (PLS, KNN, RF) and deep learning models (CNN, LSTM, Transformer). Multi-test-set evaluation on the current dataset indicates stable classification performance; we note that all results are from a single representative run, and variability across repeated runs is not quantified in this work. With low complexity and low inference latency, the proposed framework is suitable for near-real-time industrial deployment, providing an effective solution for intelligent spectral information processing.