WSE-XIDS: A Weighted Stacking Ensemble of Neural and Gradient-Boosted Learners with SHAP Explanations for Zero-Day-Aware Intrusion Detection

Main Article Content

Namrata Nebhnani
Dr. Sudhir Agrawal

Abstract

Electroluminescence (EL) imaging exposes micro-cracks, broken fingers and inactive regions in photovoltaic (PV) cells, but grading thousands of EL cell images by hand is slow and subjective. Most recent automated graders rely on deep convolutional neural networks (CNNs) or transformer hybrids that need graphics processing units (GPUs) and large labelled sets. This paper studies a deliberately lightweight alternative, MTF-RSWS, that runs on an ordinary central processing unit (CPU). Each cell is denoised with a median filter and enhanced with contrast-limited adaptive histogram equalisation (CLAHE); a 1,893-dimensional multi-texture fusion (MTF) vector is then built from local binary patterns (LBP), histograms of oriented gradients (HOG), grey-level co-occurrence matrix (GLCM) statistics, Gabor energies and intensity statistics. ReliefF ranks the features, and an ensemble of radial-basisfunction support vector machines (RBF-SVMs) trained on ReliefF-weighted random subspaces (RSWS) makes the final decision by weighted soft voting. On a stratified 75/25 split of the public ELPV benchmark (2,624 cells, four defect-probability classes), MTF-RSWS reached 75.76% accuracy, 62.52% macro precision, 55.68% macro sensitivity and 57.99% macro F1-score, the best held-out macro F1-score among eight classical baselines; in the functional-versus-defective setting it reached 83.99% accuracy. Five-fold cross-validation, however, places its macro F1-score (55.53 ± 1.52%) within one standard deviation of a single SVM on the same features (56.14 ± 0.85%), so the held-out margin is not statistically reliable. Training took about 100 s on a two-core CPU and inference 5.8 ms per cell. The results sit below those of deep models on the same data, but they give a transparent, reproducible and hardware-light reference point, and the error analysis shows that the two uncertain classes remain the main obstacle.

Article Details

Section
Articles