Developing Robust Pattern Recognition in Imbalanced Data Using Cost-Sensitive Multi-Channel Convolutional Neural Networks

Abstract

Detecting abnormal patterns in control charts is critical for ensuring quality in smart manufacturing, where sensor proliferation generates voluminous, noisy, and imbalanced data. Minority-class abnormal patterns, though rare, signal critical process deviations that can escalate costs and compromise product quality if undetected. This study addresses three critical challenges—severe class imbalance, data noise, and sensitivity to temporal perturbations—through a novel dual-channel cost-sensitive convolutional neural network framework. We propose CS-2CCNN, which processes raw one-dimensional time series data alongside two-dimensional regression graph images to extract complementary temporal and spatial features, combined with cost-sensitive learning to prioritize minority-class detection. We further introduce DeepHybridCS-2CCNN, which enhances CS-2CCNN by integrating empirical Bayesian wavelet denoising to remove noise and XGBoost classification for robust, adaptive prediction. Evaluated on simulated datasets with varying imbalance ratios (1:20 and 1:200) and the real-world Wafer dataset, DeepHybridCS-2CCNN achieves G-mean values exceeding 0.86 for severely imbalanced patterns, representing a 123% improvement over the baseline cost-sensitive CNN and outperforming traditional resampling methods (SMOTE, ADASYN) by 28–32%. The model attains F1-scores above 0.85, Matthews Correlation Coefficient values exceeding 0.80, and Area Under the ROC Curve scores surpassing 0.93 for critical patterns, demonstrating balanced performance across normal and abnormal classes. Unlike conventional oversampling approaches, this work framework minimizes sensitivity to data perturbations and enhances minority-class detection without introducing synthetic noise, offering a scalable, computationally efficient solution for industrial quality control in smart manufacturing environments.

Article history: Received (August 19, 2025); Revised (October 30, 2025); Accepted (November 28, 2025); Published online (February 12, 2026)