پيش‌بيني وقوع و شدت تصادفات ترافيكي در زمان واقعي: چارچوب چندوظيفه‌اي نوين مبتني بر يادگيري عميق GRAPH–LSTM با مكانيزم توجه و يادگيري برخط

Real-time prediction of road traffic accident occurrence an‎d severity remains a paramount challenge in the domain of traffic safety. Driven by the inherent rarity an‎d severe imbalance of such events, alongside intricate spatiotemporal dependencies an‎d the high operational sensitivity required for timely al‎e‎rts, addressing this challenge necessitates the deployment of advanced machine learning techniques an‎d structural data analysis. This research introduces an innovative hybrid framework that seamlessly integrates graph-based modeling of road network topology with temporal sequence analysis an‎d an Attention mechanism, all unified within an online learning paradigm. Within this framework, the online learning module is engineered to dynamically adjust model parameters following each influx of new traffic data, circumventing the computationally expensive need for complete retraining. Consequently, the system acquires the capability to continuously adapt to evolving traffic conditions, thereby maintaining its predictive accuracy at an optimal an‎d up-to-date level. In the proposed architecture, the road network was modeled as a spatial graph comprising 244 nodes an‎d 1,376 edges. Topological features were extracted utilizing a Graph Convolutional Network (GCN), while temporal dependencies were captured via Long Short-Term Memory (LSTM) networks. Furthermore, an Attention mechanism was employed to assign adaptive weights to critical temporal frames immediately preceding an accident. To counteract the challenge of severe class imbalance, a Weighted Focal Loss function was implemented. Simultaneously, the continuous online learning mechanism was realized utilizing a replay buffer of 10,000 samples coupled with a forgetting factor of 0.95. Comprehensive eva‎luations an‎d comparative analyses against a baseline model (an offline architecture lacking the attention mechanism) demonstrated substantial performance gains across both primary predictive tasks. In the Accident Occurrence Prediction task, the integration of continuous learning facilitated the refinement of decision boundaries. Notably, while maintaining an‎d enhancing the system’s event detection capabilities, false positives were significantly reduced by approximately 25 percent (decreasing from 46,809 to 35,250 instances). Concurrently, the specificity metric improved from 82 percent to 87 percent, recall increased from 30 percent to 34 percent, an‎d the ROC-AUC score improved from 0.61 to 0.65. Regarding the Accident Severity Prediction task, the baseline model proved virtually ineffective at identifying severe collisions due to its inherent bias towards the majority class (yielding a recall of merely 0.15). In contrast, the proposed hybrid model successfully mitigated this bias, elevating the critical recall metric to 0.66.

احمدرضا طالبيان

محمدمهدي بشارتي , سينا صاحبي