Predicting international roughness index by deep neural networks with Levenberg-Marquardt backpropagation learning algorithm
| Autor: | Massimo Losa, Nicholas Fiorentini, Pietro Leandri |
|---|---|
| Jazyk: | angličtina |
| Rok vydání: | 2021 |
| Předmět: |
International Roughness Index
Artificial neural network Mean squared error Computer science business.industry Deep learning Falling Weight Deflectometer Non-Destructive High-Performance Techniques Overfitting Levenberg Marquardt Backpropagation Algorithm Ground Penetrating Radar Backpropagation Data Fusion Non-Destructive High-Performance Techniques Data Fusion Deep Learning Artificial Neural Networks Levenberg Marquardt Backpropagation Algorithm International Roughness Index Falling Weight Deflectometer Ground Penetrating Radar Levenberg–Marquardt algorithm Falling weight deflectometer Deep Learning Artificial intelligence business Algorithm Artificial Neural Networks |
| Popis: | This paper proposes a methodology based on Artificial Neural Networks (ANNs) for integrating products derived by three on-ground Non-Destructive high-performance Techniques (NDTs) to estimate the International Roughness Index (IRI) of flexible road pavements. About that, IRI of 93 two-lane road sections has been detected by a Laser Profiler (LaP) and considered as output target to be predicted. Structural and geometrical road pavement parameters recognized by Falling Weight Deflectometer (FWD) and Ground Penetrating Radar (GPR) have been considered as input features, along with climate and rainfall information. Accordingly, different ANNs architectures have been trained (by using 70% of samples), validated (15% of samples), and tested (15% of samples) by the Levenberg-Marquardt (LM) backpropagation learning algorithm. We chose the LM since it allows handling limited data and avoiding overfitting issues. Outcomes reveal that a Deep Neural Network (DNN) recognizes hidden patterns between different road surveys and made the integration of NDTs possible and reliable. Specifically, a DNN architecture composed of two hidden layers containing 23 and 12 artificial neurons, respectively, shows a Determination Coefficient (R2) of 0.813 for the training phase, 0.761 for the validation phase, and 0.741 for the test phase. Also, the residual distribution is Gaussian with zero mean. Supported by these findings, we have deployed the DNN including all road sections, obtaining an R2 parameter of 0.762 and a Root Mean Square Error (RMSE) of 0.450 mm/km. Road authorities can consider ANNs and the LM backpropagation learning algorithm for appropriately predicting IRI and efficiently integrating NDT-based outcomes. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|