Damage Identification of Railway Bridge KW51 Conditions Using Deep-Learning-Based 1D CNN Model

This paper compares the damage identification outcomes of the Machine Learning (ML) and Deep Learning (DL) algorithms. The algorithms in both approaches have employed vibration data from the benchmark railway bridge KW51. The One-Dimensional Convolutional Neural Network (1D CNN) model is exploited in the DL algorithm to classify the response measurements. The 1D CNN classification algorithm is compared with a statistical-based ML model and another CNN model. The alternative classification model uses human-derived damage-sensitive features extracted from Principal Components Analysis (PCA) in the supervised Linear Discriminant Analysis (LDA) method. The metrics of the confusion matrix are applied as a reference for arbitrating the healthier classification between the two models. The classification accuracy of the 1D CNN algorithm varies from 90% (when acceleration measurements in the lateral direction, y-axis are used) to 100% (when measurements in the vertical direction z-axis are used). In contrast, the accuracy of the comparative ML-based classification approach varies from 66% (when acceleration measurements in the lateral direction y-axis are used) to 94.4% (when data from two sensors are jointly merged). In 1D CNN model, the maximum value of Type I error (False Positive FP) reaches 4.8%, while it reaches 14.4% in ML model. The maximum value of Type II error (False Negative FN) was 8.9% for data from accelerometer aBD23Ay, whereas it was 4.8% in similar case studies analyzed by the 1D CNN method. Once again, in both ML and DL models, the error is more relatively palpable when acceleration measurements in the lateral direction, y-axis are used.