Performance evaluation of pretrained deep learning architectures for railway passenger ride quality classification

Abstract

This study investigates the performance of pretrained Convolutional Neural Networks (CNNs) in predicting passenger ride comfort classes based on encoded acceleration signal images. Two approaches were explored: fine-tuning a pretrained CNN for classification and using transfer learning with deep-feature extraction, dimensionality reduction, and Bidirectional Long Short-Term Memory (BiLSTM) architecture. Four pretrained CNN architectures (DenseNet-201, ResNet-101, AlexNet, and GoogleNet) were evaluated. To demonstrate the superiority of pretrained networks, five conventional machine learning algorithms (SVM, DT, BTE, ANN, and KNN) and were trained from scratch using handcrafted features and compared to the two architectures. Results show that the second architecture consistently outperformed the first, achieving better accuracy (up to 14–18.5 %) on unseen data, with no apparent bias sensitivity to the unbalanced dataset. The overall accuracy of the two configurations was consistently above 80 % and 98 %, respectively. This demonstrates that pretrained CNN models, leveraging deep feature extraction, dimensionality reduction, and BiLSTM, can consistently achieve excellent performance in classifying passenger ride comfort. The study highlights the effectiveness of transfer learning and deep learning techniques in this application.

Keywords

• Ride comfort prediction
• Transfer learning
• Machine learning
• Railway vehicle
• Raw acceleration