Applications of Data-Driven Intelligent Approaches for Non-Destructive Asphalt Pavement Density Measurement

Abstract

The main topic of this research is to solve the problem faced by Fulton Hogan (FH) Ltd., which is a large infrastructure construction and asphalt supplier company based in New Zealand and Australia. Their engineers and technicians found that it is difficult to apply an accurate and efficient method for measuring the density of thin-layer asphalt pavement, whose thickness is less than 50 mm.

The commonly used methods include the coring method (CM) and the methods based on two devices: the nuclear density gauge (NDG) and the pavement quality indicator (PQI). The CM is the most accurate method, but it requires core sample extraction and in-lab tests, which are time-consuming and cause damage to the pavement. On the other hand, the NDG-based and PQI-based methods do not cause damage to the pavement. The NDG is a nuclear-based device, which is relatively accurate and efficient, but it has the risk of nuclear radiation. The PQI is an electrical device, which is the most efficient and simplest device to operate, but its accuracy is affected by environmental factors such as temperature and humidity.

Considering the advantages of the PQI, this research is focused on improving its accuracy through some data pre-processing and processing approaches, including regression analysis (RA), principal component analysis (PCA), and artificial neural network (ANN).

RA is a data-driven approach used to establish the relationships between the input and target variables and to filter outliers. Hence, it has the potential to be used as a data pre-processing approach to support other approaches, such as ANN. PCA is a data pre-processing approach widely applied across numerous fields. Though it has not yet been applied to asphalt pavement density measurement, it shows great potential for this application. ANN is a very popular data-driven intelligent approach. However, it is seldom applied to asphalt pavement density measurement and has not been applied to improve the accuracy of PQI. Hence, ANN is selected as the main data processing approach, while RA and PCA are selected as the main data pre-processing approaches. In addition, data normalization is used to pre-process data for ANN, either alone or combined with PCA or RA, since the density, temperature, and humidity values are in different ranges..

Three proposed data-driven intelligent approaches are constructed by combining ANN with different data pre-processing approaches. The combined models are named ANN (with only data normalization), RA-ANN, and PCA-ANN models.

Data collection is an essential part for training the ANN models and evaluating the performance of the combined models. With technical support from FH’s engineers and technicians, 180 data samples were collected from thin-layer asphalt pavement sections of an FH project. Each sample contains temperature, humidity, and densities measured by CM, NDG, and PQI. The three densities are the so-called core density, NDG density, and PQI density, respectively. The asphalt pavement is divided into six parts with layer thicknesses ranging from 35 mm to 50 mm. These data samples are used as the input variables of the RA, PCA, and ANN models. In addition, data samples from the public domain are used to evaluate the performance of the combined models. These samples are collected from four USA states covering different climate zones.

Based on the data samples from the FH project, how to use the RA and PCA models as the data pre-processing approaches is discussed first. Three scenarios are defined to represent different cases: only PQI is used in Scenario 1, only NDG is used in Scenario 2, and both devices are used in Scenario 3. The multiple linear and polynomial RA models are established in each scenario. Among them, the polynomial RA model in Scenario 3 performs the best. Similarly, the PCA model in Scenario 3 can calculate four contribution rates that contain the most information. In addition, the RA and PCA models in Scenarios 1 and 2 can be used to quantify the effects of temperature and humidity on the accuracy of PQI and NDG. It is found that temperature and humidity greatly affect the accuracy of PQI, but slightly affect that of NDG.

Then, the factors affecting the performance of the ANN model are discussed based on the data samples from the FH project. These factors include the ANN model structure and the learning algorithm. As for the structure, the ANN model contains one or two hidden layers, and it is named ANN Model 1 or 2, respectively. The number of neurons in each hidden layer ranges from 2 to 20. As for the learning algorithm, three options are Levenberg–Marquardt (LM), Bayesian Regularization (BR), and Scaled Conjugate Gradient (SCG) algorithms. The ANN models trained by them are named the LN, BN, and SN models, respectively. The results indicate that the LN Model 2, BN Model 1, and SN Model 2 have excellent performance. The best-fit LN, BN, and SN models contain 10, 5, and 17 neurons in each hidden layer, respectively.

Finally, the performance of the ANN, RA-ANN, and PCA-ANN models is evaluated based on the data samples from the public domain. How temperature, humidity, the learning algorithm, and the data pre-processing approach affect their performance is also discussed. As temperature and humidity rise, the performance of the PQI declines due to an increase in its RMSE. The performance of the ANN, RA-ANN, and PCA-ANN models is also slightly affected. The best and second-best models are selected when data samples from different states are used. Three of the best models are PCA-LN models, and one of them is PCA-SN. Two of the second-best models are PCA-SN models; the rest of them are PCA-LN and RA-LN models. Overall, the best combination is the PCA-ANN model trained by the LM algorithm, followed by the one trained by the SCG algorithm.

Overall, three data-driven intelligent approaches are proposed in this research. In the first approach, NDG and PQI are employed as asphalt density measurement devices and ANN is also employed as the data-processing approach. The NDG density, PQI density, temperature and humidity data collected by NDG and PQI are input to the ANN model to predict the density of the thin-layer asphalt pavement. In the second and third approaches, RA and PCA are also employed along with ANN. The mentioned data are input to the RA-ANN and PCA-ANN models instead of the ANN model. The improvement of the PCA-ANN is higher than that of the RA-ANN and ANN models. By considering the ANN structure and learning algorithm, the PCA-LN model containing 2 hidden layers and 10 neurons in each layer performs the best.

The proposed approaches have been presented to the engineers and technicians of FH and have been applied to measure the density of thin-layer asphalt pavement across several projects. The engineers and technicians are impressed with their performance. Hence, further field applications of these approaches are scheduled for more upcoming projects. This positive feedback from FH demonstrates the practical significance of the proposed approaches in this research and verifies their wide applicability for thin-layer asphalt pavement density measurement.