Multi Agent System for Railway Steel Bridges Health Monitoring by Vibration Based Sensory Data and Fuzzy Logic

The Structural Health Monitoring (SHM) of railway steel bridges is critical to ensure transportation safety, reliability, and timely maintenance. This thesis proposes a comprehensive vibration-based SHM framework that integrates signal processing, deep learning, and fuzzy reasoning to detect, classify, and interpret structural damage. This thesis proposes a comprehensive Multi-Agent System (MAS) for SHM of railway steel bridges using vibration-based sensory data, machine learning algorithms, and fuzzy logic reasoning. The system integrates diverse artificial intelligence agents; responsible for sensing, perception, classification, reasoning, and alert generation; into a cohesive architecture designed to detect, localize, and assess structural damage in real-time. Vibration signals collected from accelerometer sensors are first processed through the Hankel matrix construction and Singular Value Decomposition (SVD). These transformed signals are converted into statistical features and spectrograms using the Short-Time Fourier Transform (STFT), capturing the temporal and spectral evolution of bridge vibrations to be used in time domain SHM detection and Frequency Domain classifications respectively. These statistical features and spectrograms form the shared environment observed by multiple classifier agents. We implement and compare classical machine learning models (SVM, KNN, RF, MLP, Ensemble), a custom Radial Basis Function Network (RBF-Net), and deep learning architectures (ResNet50, DenseNet121, EfficientNetB0). Additionally, hybrid DNN-Transformer fusion models (TRNet, TDNet, TENet) are proposed to combine local feature extraction with global temporal attention. These models collectively support bridge scenario classification and damage intensity estimation, achieving accuracies exceeding 90%. An Explainable Boosting Machine (EBM) agent further enhances transparency by highlighting the influence of input features such as skewness, kurtosis, and signal energy on classification outcomes. Finally, a Fuzzy Logic Reasoning Agent is introduced, which aggregates outputs from perception agents and applies expert-defined fuzzy rules to generate human-readable alerts (Attention, Pre-Alarm, Alarm). The proposed MAS architecture ensures robustness, interpretability, and modularity, key requirements for intelligent infrastructure monitoring. The work advances SHM by demonstrating that distributed intelligence, deep learning, and fuzzy reasoning can be jointly leveraged to build reliable, interpretable, and scalable monitoring systems for safety-critical railway bridges.