Efficient Multi-Path CNN Architecture for Anomaly Detection in Control Charts with Imbalanced Data

Control Chart Pattern Recognition (CCPR) is critical for early anomaly detection in industrial processes, yet data imbalance poses a significant challenge, leading to biased models that overlook rare but critical anomalies. This study introduces a novel Multi-Path SMOTE-CNN framework to address this issue, enhancing classification performance under moderate (1:20) and severe (1:200) imbalance ratios. The methodology integrates univariate and dual-channel Synthetic Minority Over-sampling Technique (SMOTE) extensions with a multi-path Convolutional Neural Network (CNN) that processes raw time-series and recurrence plot images in parallel, extracting complementary temporal and spatial features. The model was evaluated on simulated datasets and the real-world Wafer dataset using accuracy, sensitivity, specificity, and G-mean metrics. Results demonstrate significant improvements over the baseline CSCNN model, achieving G-mean values up to 0.9928 for cyclic patterns under severe imbalance and 0.9987 on the Wafer dataset. The MP-SMOTE-CNN’s ability to handle extreme imbalance, reduce manual feature engineering, and ensure computational efficiency highlights its potential for real-time industrial applications. This work advances CCPR by offering a robust, scalable solution for anomaly detection, with implications for quality control in manufacturing and beyond.