Deep recurrent neural networks with spectral-statistical fusion for industrial-grade steel alloy classification using femtosecond laser-ablation spark-induced breakdown spectroscopy

Abstract

Building on the spectral stability and sensitivity of femtosecond laser-ablation spark induced breakdown spectroscopy (fs-LA-SIBS), and the pattern recognition capacity of deep learning, a DeepRNN spectral-statistical fusion framework is introduced for high-precision classification of industrial-grade steel alloys. The framework fuses two-channel raw spectra with statistical descriptors, and employs configurable bidirectional recurrent encoders (GRU, LSTM, and vanilla RNN) to capture temporal dependencies and shape a robust decision space under end-to-end training. Under a unified evaluation protocol, the DeepRNN framework models are benchmarked against CNN and Transformer, and against traditional machine learning methods including RF, SVM, and PLS-DA; wavelength contribution analysis is performed to identify discriminative regions and interpretable importance profiles. Under the DeepRNN framework, the three encoders consistently outperform CNN, Transformer, and traditional machine learning baselines on core metrics including accuracy, cross-split consistency, and perturbation robustness, with average accuracy improved by approximately 2.35 to 3.50 percentage points compared to CNN and Transformer, and by 6.58 to 15.40 percentage points relative to traditional baselines. They also achieve favorable trade-offs among accuracy, efficiency, and deployability, with wavelength importance aligning with physically meaningful line structures. This sensor-intelligent system enables scenario-oriented deployment: vanilla RNN is chosen when accuracy is paramount; GRU is suitable for low-latency, energy-constrained online monitoring; and LSTM is preferred for the most conservative optimization trajectory and robustness under complex conditions, providing a scalable pathway for real-time industrial alloy identification and quality control.