Leveraging Statistical and Machine Learning Techniques for Pavement Instrumentation Evaluation

Abstract: Federal Aviation Administration (FAA) uses a variety of in-pavement sensors to monitor pavement responses to actual and simulated aircraft loads. Analysis of sensor data can be challenging given the dynamic nature of the response and inherent noise in the signal. Maintaining data integrity traditionally requires visual inspection of the signal to determine if the sensor behavior is logical and to identify the part of the signal that is relevant to the study. Past attempts to automate this effort at the FAA National Airport Pavement Test Facility (NAPTF) led to oversimplification of the sensor response. As an illustration of the problem, the NAPTF database contains a note regarding stored peak values, which states: “[The database] does not capture every peak in a typical strain response, so in the case of multiple peaks due to tandem wheel loads, it is recommended to examine the sensor record.” An improved method for identifying and classifying key sensor response values that does not rely on extensive human intervention is desirable.

A novel method was developed to automate the signal processing procedure for pavement instrumentation and identify values of significant interest during dynamic moving load events. The approach uses a combination of B-spline median regression with K-Nearest Neighbor clustering to filter signal noise and identify the signal of interest. Then, ordinary least squares regression combined with an Artificial Neural Network is used to predict the probability of a collected data point being a point of interest. The method was developed, tested, and validated using data collected from NAPTF Construction Cycle 9 (the most recent dataset available). Trafficking plan included various aircraft wheel loads (12, 24, and 36 kips), loading configurations (Dual, D; Dual-Tandem, 2D; and Dual-Tridem, 3D), and vehicle speeds (0.5; 1.6; and 2.5 miles per hour). The developed method was able to accurately predict the points of interest in 93% of sensor responses (out of 2,458 critical responses). Data pairs consisting of a response value with the probability of it being a point of interest were stored for future data analysis and evaluation.