Deep learning models for asphalt material behavior analysis under smart city IoT infrastructures

Abstract

Asphalt is used for smart city infrastructure, which is subjected to constant travel, stress, and ultimately decay. Cannot detect real-time material activity like time, destructive testing cannot guarantee predictive maintenance. The algorithms in use today cannot handle high spatial and temporal data produced or recorded by IoT-enabled monitoring systems for proactive infrastructure management. Asphalt analysis predictive maintenance approaches depend on either limited data on simple rules based on composite models or on deep learning (DL) models—like LSTMs, to scale. This generates inefficiency in smart city network maintenance, resource investment/reallocation, and scaling. Multi-modal IoT predictive maintenance for asphalt pavement degradation requires a more flexible, accurate, and interpretable preventative maintenance strategy. The proposed IoT-based Asphalt Analysis Framework (IAAF) incorporates Vision Transformers (VITs) to collate spatiotemporal global information from high-resolution images of the pavement surface, and Temporal Convolutional Networks (TCNs) to graph long-term temporal patterns from sensor data provided by the IoT. The comprehensive hybrid attention module (HAM) connects and collates data to identify asphalt degradation and failure by utilizing both temporal and spatiotemporal data in the same network. The experimental assessment shows that the IAAF model significantly reduces prediction errors compared to baseline LSTM models. MAE lowers from 38.23 to 14.04, RMSE from 40.59 to 20.79, and MAPE from 13.03 to 4.21%. Attention mechanisms and visual information transducers minimize false alarm rates by over 25% and preserve accuracy with less than 5% variance when sensor inputs are noisy or insufficient. With inference latency of 500 ms, the framework is ideal for edge device real-time IoT deployments.