Advanced Anomaly Detection in Manufacturing System: An Optimized Multi-CNN Ensemble Learning Approach
Paper Title: Advanced Anomaly Detection in Manufacturing System: An Optimized Multi-CNN Ensemble Learning Approach
Journal Name: IEEE TRANSACTIONS ON EMERGING TOPICS IN COMPUTATIONAL INTELLIGENCE
Abstract: In conjunction with current deep learning networks, the potential for power anomaly detection in manufacturing systems has been turned into a realized task that serves the process efficiently and reliably. This paper proposes an ensemble model developed for anomaly detection within a manufacturing environment employing six convolutional neural network architectures: DenseNet121, EfficientNetB0, InceptionV3, ResNet50, VGG16, and Xception, over two datasets representing structural and logical anomalies, MVTec LOCO AD and MVTec AD. The proposed ensemble model can handle various issues associated with anomaly detection, which include class imbalance, poor variability of images, and model overfitting. A large dose of intensive data augmentation techniques and the use of Gaussian noise in a way that gives an overwhelming boost to the generalization of the models will be passed over. The models underwent multiple training iterations to optimize testing accuracy. The ensemble technique was employed to aggregate the results, culminating in the highest accuracy, precision, recall, and F1 scores achieved in the experiment. This approach demonstrated promising performance and increased model robustness and reliability. Additionally, it achieved approximately 98.15 and 97.56 AUROC on the MVTec LOCO AD and MVTec AD datasets, respectively.