A Federated Learning Approach for Scalable Road Condition Detection in Connected Vehicles

Traditional centralized road monitoring systems are hindered by significant limitations, with assessment speeds of only 3 to 4 km/h, while 39% of major U.S. roads remain in poor condition, resulting in a $684 billion funding gap. This study introduces an enhanced federated learning framework, building upon the FedProx algorithm, for privacy-preserving road condition detection in connected vehicles. Our approach enables collaborative model training across distributed vehicle networks using accelerometer, gyroscope, and GPS sensor data, while reducing data exposure by keeping raw sensor data on-device. We developed the Federated Learning Model and trained it on a comprehensive dataset comprising 28,523 samples across three road conditions: bumps, potholes, and normal, which were collected from smartphone sensors. The proposed architecture outperforms existing methods, achieving 98.18% accuracy for pothole detection and 97.53% for bump detection, surpassing FedAvg and FedSGD by up to 5.6%. Experimental validation with small-scale federations (5 to 15 clients) demonstrates the framework's effectiveness and validates the core algorithmic contributions. The proposed architecture is designed to scale to larger deployments and addresses fundamental challenges of centralized systems. This approach enables real-time road condition monitoring among several vehicles, overcomes the scalability challenges of centralized systems, and offers potential annual cost savings of $2.1 billion by leveraging 90% connected vehicle penetration by 2030.