Vibration-Based Machine Learning Model Training for Railway Bridge Health Monitoring.

Area of Science:

• Civil Engineering
• Artificial Intelligence
• Machine Learning
• Structural Health Monitoring

Background:

• Bridges require continuous monitoring to prevent severe outcomes from undetected damage.
• Sensors like accelerometers collect data for analyzing bridge structural integrity.
• Machine learning algorithms are crucial for processing sensor data for anomaly detection and damage prediction.

Purpose of the Study:

• To develop and evaluate a machine learning method for detecting damage location and intensity in bridges.
• To investigate the impact of data preprocessing techniques, specifically data slicing, on model performance.
• To compare the effectiveness of different optimization algorithms for bridge health monitoring.

Main Methods:

• Feature extraction from time-series sensor data (accelerometer signals).
• Application of a deep neural network (DNN) for classification and prediction tasks.
• Experimentation with data augmentation, subdivision (slicing), and parameter tuning.
• Comparison of Adam and Stochastic Gradient Descent (SGD) optimizers.

Main Results:

• Optimal performance was achieved using one-fourth data slicing (40x40 feature matrix).
• High accuracies were reported: 93.54% for bridge scenario classification and 98.21% for damage intensity classification.
• The Adam optimizer demonstrated superior performance over SGD for both damage localization and intensity estimation, with test accuracies up to 93.7%.

Conclusions:

• Data slicing, particularly a 40x40 feature matrix, significantly enhances DNN performance in bridge health monitoring.
• The Adam optimizer is more effective and stable than SGD for this application.
• The proposed machine learning approach provides a robust method for accurate bridge damage assessment.