Bridging Expertise and Automation: A Hybrid Framework for Intelligent Pavement Maintenance Decision-Making

Ensuring the structural reliability of asphalt pavements has always been a cornerstone of modern transportation infrastructure, but the complexity of that task has evolved sharply in recent years. It’s not just heavier traffic or the wear of time—though both matter. It’s the pace at which demands on road systems are growing, often outpacing the capacity of conventional maintenance practices to respond efficiently or consistently. Led by Associate Professor Leilei Chen at Southeast University, a research team in intelligent transportation is tackling this challenge with a novel hybrid AI framework Traditionally, maintenance strategies have relied on the trained eyes and judgment of engineers who interpret performance data, inspect damage, and recommend repairs. That system has worked—more or less—but scaling it to the size and urgency of today’s networks is another matter entirely. The bottleneck, to put it simply, is human bandwidth. No matter how skilled, an engineer making decisions for hundreds or thousands of kilometers of roadway will inevitably face limitations. And so the appeal of data-driven solutions—especially machine learning—has grown. These tools can process vast datasets, identify patterns invisible to the human eye, and predict future deterioration based on numerous variables. The math checks out. But predictions alone aren’t enough. Anyone who’s worked in infrastructure knows that a good forecast can still lead to a poor decision if it doesn’t account for context—site history, materials quirks, climate shifts, even budget constraints.

This gap between computational power and engineering intuition is where the real challenge lies. Many predictive models, while technically impressive, remain black boxes: they output results without offering reasons that practitioners can understand or trust. On the other hand, expert systems rooted in human logic tend to lack flexibility—they don’t adapt well to noisy or shifting data. To this account, new research paper published in International Journal of Pavement Engineering and conducted by Associate Professor Leilei Chen—an expert in intelligent infrastructure systems and AI-assisted pavement maintenance—and doctoral candidate Wei Li and Xiaohu Chen, alongside Ruipeng Chen, Zhendong Qian, Daoxie Chen, and Yitao Song from the Intelligent Transportation System Research Center at Southeast University, designed a hybrid framework that combines neural networks with knowledge graphs. This framework was designed not just to predict pavement deterioration but to recommend interventions in a way that reflects how real engineers think. It’s not about removing people from the process. Quite the opposite. It’s about encoding their expertise, capturing the rationale behind their choices, and letting that guide automated systems. The goal isn’t pure efficiency. It’s better decisions—faster, yes, but also smarter and more defensible. That, in the end, is what modern infrastructure demands: tools that think with us, not just for us.

 To put their hybrid framework to the test, the research team selected a live freeway in eastern China as their case study. This wasn’t an arbitrary choice—its proximity to the coastline and exposure to shifting weather patterns offered a challenging, variable environment to evaluate how well the model could perform outside of theory. The pavement structures along the route were diverse, combining different layers of asphalt and composite bases. That heterogeneity made it a strong proving ground, particularly for a system designed to accommodate complex, real-world conditions.

Over four years, the team compiled a comprehensive dataset. It included pavement performance indicators, traffic loads, structural information, and climate records—roughly 19,000 samples in total, each representing a 100-meter segment. With this data, they trained a Convolutional Neural Network (CNN) to forecast rutting depth, a widely recognized metric for assessing structural degradation. The model’s predictive output proved encouraging: errors remained within a tight 2% margin. In pavement management, that kind of precision isn’t trivial—small deviations in rutting depth can significantly influence both safety outcomes and cost-effective timing of interventions.

Afterward, the researchers wanted to take it a step further: translating those forecasts into treatment plans. They applied a statistical approach to analyze the distribution of predicted rutting values and identified segments most at risk of deterioration. In total, 87 sections were flagged for preventive action. What set this work apart, however, was how those treatment decisions were derived—not manually, and not solely by the model’s data output, but by a domain-specific knowledge graph constructed using a BERT-BiLSTM-CRF pipeline. This tool parsed thousands of technical documents and encoded relationships between performance indices, environmental factors, and treatment types. The framework didn’t just identify where maintenance was needed; it also proposed how to do it. For all 87 flagged sections, micro-surfacing emerged as the recommended strategy. That recommendation was subsequently validated in practice. After applying micro-surfacing, follow-up measurements showed the average rutting index had improved—from 90.10 to 93.12. That shift might sound modest, but within the context of pavement degradation, it’s significant. More importantly, the decision aligned with engineering standards, even though it originated from an AI-informed process.

At its core, the authors’ work isn’t just about refining forecasts—it’s about rethinking how infrastructure decisions get made. Traditionally, engineering judgment has stood apart from computational models. Here, the two are fused. The knowledge graph doesn’t just enhance the interpretability of the CNN’s output; it restores a layer of contextual reasoning that data alone can’t offer. That’s a meaningful development, particularly in a field where decisions are rarely black and white.

We believe the real-world applications are substantial. Agencies managing road networks are constantly caught between shrinking budgets and aging infrastructure. Having a system that can identify risks early and recommend treatments grounded in both data and expert knowledge could streamline resource allocation and reduce long-term maintenance costs. The fact that this approach performed successfully in the field—rather than just in simulation—lends credibility to its scalability. On a broader level, the work led by Associate Professor Leilei Chen at Southeast University challenges the false binary often drawn between automation and professional expertise. By bridging AI with real-world engineering knowledge, his team demonstrates that AI can be used not to replace human judgment, but to amplify it—embedding years of engineering insight into tools that can operate consistently, rapidly, and transparently. In doing so, it marks a shift toward more accountable, explainable automation in civil infrastructure—something the field desperately needs as it transitions into the next era of intelligent asset management.