Enhancing rail track defect detection using an ensemble approach with self-attention mechanism

Abstract

Rail transit is regarded as the most dependable means of transport. Train track maintenance is critical to the rail network’s safety and reliability. Conventional methods sometimes provide potential concerns since they are unable to identify intricate defects. For railway networks to remain secure, defects in the rail track must be detected and classified competently. Existing approaches designed for automated rail track defect detection rely on either a single deep-learning model (DL), like CNNs (convolutional neural networks), or utilize pre-trained models for transfer learning, and even ensembled models where multiple networks are combined to enhance the efficiency of the model. Nevertheless, the models can end up in unexpected efficiency as well as blockage issues if they are not appropriate for a particular problem domain. To negotiate these barriers, this research suggests a novel approach that integrates self-attention layers utilizing ensemble methods of DL to enhance the accuracy of classification for eight kinds of rail defects. Self-attention layers focus on the most relevant features of the input data by assigning higher importance to critical regions, thus enabling the model to capture intricate patterns and subtle anomalies better. Two ensemble models are built, one with self-attention and the other without, using four transfer learning techniques, namely VGG16, DenseNet121, MobileNetV2, and Inception V3 as base models. This study employs a comparative analysis to evaluate the ensemble models constructed with and without self-attention layers employing performance metrics that include precision, recall, accuracy, along F1 score. These metrics are substantiated by visual representations, such as confusion matrices, ROC curves, as well as attention maps. The results show that attention mechanisms have a considerable influence on model performance, providing a solid solution for real-world rail defect detection and classification.