Stacked Target‐Related Autoencoder‐Extreme Learning Machine: A Novel Soft Measurement Modeling Approach for Near‐Infrared Spectroscopy

ABSTRACT This paper proposes a novel quantitative modeling and prediction approach for near‐infrared (NIR) spectroscopy, combining a stacked target‐related autoencoder with an extreme learning machine (STAE‐ELM). The STAE performs hierarchical pre‐training using multiple improved target‐related autoencoders (TAEs) to extract deep spectral features highly correlated with target values. Crucially, the top‐level structure of the STAE is replaced by the ELM, which serves as the final prediction model. This integration streamlines training by reducing parameters and steps while simultaneously enhancing performance through optimized initialization of the ELM's weights and biases. Compared to conventional feature selection methods and stacked autoencoders, the STAE‐ELM extracts more comprehensive and target‐relevant deep features from spectral data, mitigating overfitting risks. The method's efficacy was validated on five open NIR datasets, benchmarking against three approaches: feature selection modeling, SAE‐based feature extraction modeling, and backpropagation‐based deep network modeling. Results demonstrate that calibration models built with STAE‐ELM achieved average reductions in RMSEP of 18.48%, 5.74%, and 12.14%, respectively compared to these benchmarks. Furthermore, modeling efficiency was significantly improved over the backpropagation‐based deep network approach.