Lightweight feature selection with statistical priors for industrial quality monitoring

Quality monitoring is essential for ensuring the stable operation of industrial production systems, where feature selection plays a critical role in identifying key variables for accurate diagnosis. However, this task remains challenging in industrial scenarios characterized by high dimensional feature spaces, limited samples, and imbalanced class distributions. Existing methods have difficulty in providing reliable feature evaluation under such conditions, as statistical priors are typically used in a static manner without interaction with model learning, while limited samples and class imbalance introduce unstable and biased feature representations. To address these limitations, this paper proposes a lightweight feature selection framework with statistical priors (SP-LFS) for industrial quality monitoring. Through a soft reweighting mechanism, statistical priors are integrated into the learning process, enabling dynamic guidance rather than static initialization. Meanwhile, a lightweight supervised autoencoder is employed to learn compact and discriminative latent representations, improving stability under small sample conditions. In addition, an imbalance weighting and progressive pruning strategy is embedded into the optimization process to alleviate bias introduced by imbalanced distributions and suppress irrelevant correlations, resulting in a more robust feature subset. Experiments on four representative industrial datasets demonstrate that the proposed method achieves competitive results in most settings and the best average Friedman rank, while maintaining stable performance under small sample and imbalanced conditions, indicating its suitability for industrial feature selection.