Evaluation of Flexible Pavement Performance Based on Genetic Neural Network

Abstract

The distress condition, skid resistance, and structural performance of flexible pavements exhibit simultaneous and interrelated deterioration. This study aimed to evaluate the three aspects of performance quantified by indices as the pavement surface condition index (PCI), skid resistance index (SRI), and structure strength index (PSSI), respectively, using cracking indices that can be rapidly obtained due to advanced cracking detection technologies. Considering the effect of structural design and traffic and climatic conditions, 29 structural, traffic, and climatic indices were selected as well and subsequently reduced to 12 principle components using principal component analysis, which retained more than 80% of the original information. To evaluate the PCI, SRI, and PSSI of flexible pavements, genetic neural networks (GNNs) with inputs of cracking indices and structural, traffic, and climatic principal components were developed by employing genetic algorithms to optimize the hyperparameters of artificial neural networks (ANNs). The GNNs were trained using the data of 287 flexible pavement sections from the Long-Term Pavement Performance program. Results indicate that ANNs using only cracking indices as inputs are of low accuracy, and introducing structural, traffic, and climatic indices into inputs can increase the accuracy of ANNs. Furthermore, replacing these condition indices with their corresponding principal components as inputs reduces the iterations of training ANNs by 63.9%, 45.5%, and 32.8% for flexible pavements with semirigid, granular, and asphalt-bound base layers, respectively, while the decrease in accuracy of ANNs is less than 0.5%. The training efficiency of GNNs is up to 773.8% higher than ANNs. The evaluation accuracy of GNNs for PCI, SRI, and PSSI ranges from 0.944 to 0.954, 0.899 to 0.918, and 0.895 to 0.906, due to various flexible pavement types, which is 0.2% to 4.9% higher than ANNs. This approach can assist in simultaneously evaluating the distress condition, skid resistance, and structural performance of flexible pavements using conveniently detected cracking data, thereby reducing detection cost. The distress condition, skid resistance, and structural performance of flexible pavements exhibit simultaneous deterioration, while the corresponding detections are conducted independently, increasing both detection cost and interference in traffic. Thus, this study developed a machine learning model to simultaneously evaluate the three aspects of performance using cracking detection data for time-saving and cost-effective reasons, by considering the convenience of cracking detection technologies. Cracking data, namely alligator and block cracking areas, and longitudinal and transverse cracking lengths, should be collected as the input data when using this model. Furthermore, some condition data should also be collected for ensuring the accuracy of the results: (1) structural design parameters, such as thickness and moduli of pavement layers; (2) traffic survey data, such as annual average daily traffic and equivalent single axle load number; and (3) climate data, including temperature, precipitation, air humidity, and solar radiation. The accuracy of this model has been verified using measured data from the Long-Term Pavement Performance program. This model is expected to assist highway owners in simultaneously evaluating the aforementioned three performances of flexible pavements, using easily obtained cracking and condition data, thereby reducing detection cost.