Control chart pattern recognition with variable window size for imbalanced data based on convolutional neural network with a convolutional block attention module

In modern manufacturing processes, the complexity and variability of process conditions as well as the high degree of automation, impose greater demands on the real-time monitoring of production status. Control Chart Pattern Recognition (CCPR) is a crucial method for evaluating the condition of the production process, improving the ability to detect abnormalities and ensuring product quality and process stability. However, current CCPR methods typically employ a fixed-window design, which makes adapting to dynamically changing window requirements challenging. Additionally, these methods often exhibit low accuracy when handling imbalanced control chart pattern data. To address these challenges, this paper proposes a control chart pattern recognition method, MFF-CNN-CBAM, which combines a convolutional neural network with a convolutional block attention module and multi-feature fusion to recognize control chart patterns with variable window sizes in imbalanced data. The method first unifies the window size through a resampling technique. It then applies the synthetic minority oversampling technique (SMOTE) algorithm to balance the dataset, followed by extracting six mixed features, which include five statistical features and one shape feature. These features are fused with the convolutional features extracted by CNN-CBAM, and classification is conducted using a Softmax layer to increase recognition performance. The experimental results indicate that, compared to the raw data-only approach, the machine learning classifier approach, and other CCPR methods, this method demonstrates superior recognition performance, achieving an accuracy of 99.73 % when handling variable-window and imbalanced data. In addition, this paper develops an online monitoring system based on the CNN-CBAM model, which combined with the sliding window technique, can recognize normal and abnormal patterns in dynamic data in real time, achieving high recognition accuracy.