Introduction to 'Artificial intelligence in failure analysis of transportation infrastructure and materials'

Transportation infrastructures such as roads, bridges, tunnels, stations, airports and subways are critical to modern society, and failures can cause substantial public harm. Artificial intelligence has recently emerged as a powerful tool to understand, detect and analyse failures in transportation infrastructure and materials. This introductory article outlines the motivation and scope of a theme issue that showcases state-of-the-art AI applications for failure analysis across pavements, tunnels and rail systems, highlighting methods that enhance detection, diagnosis, prediction and decision-making for infrastructure performance and safety.

The article surveys and contextualizes contributions within the theme issue: (1) Road crack detection via deep learning, including an optimized Mask R-CNN approach achieving 95–99% accuracy for automatic identification of pavement distress (Lv et al.). (2) A transformer-based method (TransCrack) that decomposes images into patch sequences and uses a pure transformer encoder with reduced self-attention to model long-range dependencies, delivering state-of-the-art crack detection and fine-grained profile extraction (Lin et al.). (3) A novel Falling Weight Deflectometer (FWD) dispersion curve methodology leveraging Rayleigh wave dispersion from deflection-time histories to estimate modulus profiles, with case studies on concrete pavements and recommendations for device modifications (Wang et al.). (4) AI-assisted backcalculation and optimization frameworks for asphalt pavement condition assessment using FWD tests and finite element model updating, comparing alternative schemes and settings (Deng et al.). (5) A hybrid data-driven/model-driven approach to predict fatigue life of in-service asphalt pavements by coupling AI-based FE model updating with mechanistic fatigue equations; validation against field-monitored responses (Luo et al.). (6) Recurrent neural networks (LSTM/GRU) incorporating maintenance action data to predict urban road performance in Beijing, trained on five-year maintenance and performance indicator datasets, showing significant accuracy gains (Deng et al.). (7) Dielectric property modeling of asphalt mixtures for future electrified roads across temperature (−30 to 60°C) and frequency (200–2,000,000 Hz) ranges, proposing prediction models for dielectric constant and loss factor with temperature/frequency effects (Dawei Wang et al.). (8) Lightweight few-shot Siamese learning for automatic pavement texture recognition across four texture classes, achieving high accuracy with limited data (Pan et al.). (9) A machine learning-based assessment of autonomous (CAV) versus human-driven (HDV) vehicle impacts on pavement roughness (IRI) using gradient boosting with driving pattern features, finding faster deterioration under CAV trajectories on average (Chen et al.). (10) Multi-objective data mining and neural networks (BP) for stability assessment of shield tunnel excavation faces, including analytical limit support pressure, feature reduction via rough sets, and BP model validated against TOPSIS and cloud models with low prediction error (Li et al.). (11) BPNN-based identification of segment joint failures in underground tunnels from settlement curves using equivalent axial stiffness-derived labels to predict dislocation and opening (Jin et al.). (12) Rail corrugation recognition using a particle probabilistic neural network (PPNN) with PSO optimization, leveraging in-vehicle noise and bogie acceleration features to classify corrugation wavelengths and amplitudes with high accuracy (Cai et al.). (13) Passenger flow anomaly detection in urban rail transit networks via a GCN-Informer model for spatiotemporal representation and Gaussian Naive Bayes for binary classification, outperforming baselines on a real Beijing URTN dataset (Liu et al.).

• Deep learning for pavement crack detection achieves high accuracy: Mask R-CNN reached 95–99% accuracy across datasets; transformer-based TransCrack achieved state-of-the-art performance, improving contiguous crack recognition and fine-grained profiles. - FWD dispersion curve method uses deflection-time histories and Rayleigh wave dispersion to compute modulus profiles; effective on concrete pavement segments, with device modifications recommended for shallow layer characterization. - AI-driven FE model updating supports backcalculation for pavement condition assessment, aiding automated parameter identification in high-dimensional spaces and comparing backcalculation schemes. - Hybrid data- and model-driven framework accurately predicts asphalt pavement fatigue life by updating mechanistic parameters with AI-inferred damage states; simulated responses aligned with field measurements. - RNNs (LSTM/GRU) trained with 5-year maintenance and performance data significantly improved urban road performance prediction accuracy in Beijing. - Dielectric properties of aggregates, binders and mixtures vary linearly with temperature (−30 to 60°C), with frequency-dependent growth rates (200–2,000,000 Hz); new models predict dielectric constant and loss factor under temperature/frequency effects. - Few-shot Siamese model attained 89.8% accuracy in a four-way five-shot pavement texture classification task (dense asphalt concrete, micro surface, OGFC, SMA). - Gradient boosting model predicted pavement IRI from driving patterns and context variables; connected and automated vehicles (CAVs) increased pavement deterioration rate by 8.1% on average compared to human-driven vehicles. - Shield tunnel face stability assessment achieved a BP neural network prediction error of 5.7675×10^-4, outperforming TOPSIS and cloud models. - Tunnel segment joint failure identification from settlement curves via BPNN predicted dislocation and opening, enabling automated condition assessment. - Rail corrugation classification using PPNN achieved average accuracies of 96.43% (noise-based wavelengths 30/50 mm) and 95.40% (acceleration-based amplitudes 0.1/0.2 mm). - GCN-Informer plus Gaussian Naive Bayes framework showed superior accuracy for detecting network-level passenger flow anomalies in a real URTN dataset.

Collectively, the featured studies demonstrate how AI enhances the failure analysis pipeline for transportation infrastructure: sensing and data acquisition (e.g., FWD waveforms, vehicle noise/acceleration, passenger flows), feature learning and representation (CNNs, transformers, GNNs, sequence models), inverse modeling and backcalculation (optimization, FE model updating), and decision support (stability assessment, anomaly detection). These advances address the need for faster, more accurate, and scalable detection, diagnosis, and prediction of failures and deterioration in pavements, tunnels, and rail systems, enabling proactive maintenance and improved safety. Integration of AI with physics-based models (hybrid approaches) and with operational data (maintenance histories, driving patterns) shows particular promise for interpretable, actionable insights and lifecycle management.

This introductory article synthesizes the latest AI developments in transportation infrastructure failure analysis showcased in the theme issue. Key contributions span vision and transformer models for crack detection, physics-informed and data-driven hybrid methods for modulus and fatigue life estimation, sequence models for performance prediction, dielectric modeling for electrified roads, few-shot texture recognition, machine learning for tunnel stability and joint failure identification, rail corrugation recognition from operational responses, and graph-based methods for network anomaly detection. Future research directions include: integrating multimodal sensing (images, waveforms, telemetry) with hybrid physics-informed AI; standardizing datasets and benchmarks for infrastructure AI; adapting models to evolving conditions (domain adaptation, continual learning); quantifying uncertainty and ensuring robustness; assessing emergent mobility technologies (e.g., CAVs) on infrastructure performance; and translating models into deployable, real-time decision support tools for maintenance and safety.

As an editorial introduction, the article does not present original experimental methods or a unified dataset; instead, it summarizes selected contributions within a theme issue. The scope is limited to the included studies and may not cover all relevant literature. Reported metrics and findings are drawn from individual papers and may be context- and dataset-specific, which can limit generalizability until further validated across broader conditions.