Principal Investigator:	Ronald S. Harichandran, Ph.D., P.E., F.ASCE
Research Assistant:	Gang Zhang
Period:	September 2006 - December 2009

Research Objective

A major form of bridge deck deterioration is the surficial scaling damage and corrosion-induced delamination of the concrete cover above the top layer of steel reinforcement. Several techniques are currently available to detect delaminations such as sounding, impact-echo, ultrasonic pulse velocity, infrared thermography and ground-penetrating radar. The sounding test has the advantage of being fast, simple and inexpensive and therefore has been widely used by field engineers. The traditional methods for delamination detection involve: (1) bar/hammer tapping of the deck and listening to the acoustic response and (2) dragging a chain over the deck and listening to the change in the sound. The delamination is characterized by a dull, hollow sound. However there are two problems associated with the current approaches. First, the traffic noise from adjacent lanes may contaminate the sound signals and make it hard to detect the delamination. Second, the detection is performed by listening to the sound, which is highly subjective and requires extensive training. The goal of this research is to develop a refined delamination detection system with improved performance.


Research Approach

Two major issues need to be addressed to accomplish the goal. First, an efficient algorithm is needed to eliminate the traffic noise from the recordings. Second, an objective delamination detection algorithm is needed to differentiate between the solid concrete and the delaminated concrete. To solve the first problem, a modified independent component analysis (ICA) is used to separate the traffic noise and the acoustic signals. For the second problem, different features of the acoustic signals were compared and mel-frequency cepstral coffeicients (MFCC) were found to be effective features. Detection based on MFCC has good performance and can approximate human hearing. A mutual information based method was used to select the optimal MFCC for delamination detection. The selected features were used to train a classifier which was then used to classify new recordings.


Research Results

The performance of the modified ICA is shown in Figure 1. The original signal in Fig. 1(a) was mixed with the noise signal in Fig. 1(b) to create the two simulated recordings shown in Fig. 1 (c) and (d). The modified ICA was then used to obtain the recovered signal shown in Fig. 1(e). As can be seen, the original signal was successfully recovered from the noisy recordings. Table 1 compares the performance under different conditions. The error rate is the proportion of times signals were misclassified. Using the raw recordings, the detection using MFCC is very good under quiet conditions, but becomes poor with increasing noise levels. However, if the recordings are first filtered using the modified ICA, then the detection algorithm performs very well even at high noise levels.

FIGURE 1 Performance of the modified ICA

 

Table 1. Performance of the Detection Algorithm

Measurements	Error Rate (%)
	Raw Signal	Filtered Signal
Quiet	N/A	1.44
Low Noise Level	3.49	1.79
Moderate Noise Level	34.37	1.61
High Noise Level	48.49	1.71

References