Classification of impulse radar waveforms using neural networks

Area of Science:

• Civil Engineering
• Geophysics
• Computer Science

Background:

• Bridge decks with asphalt overlays present complex structures for non-destructive evaluation.
• Variations in asphalt thickness, concrete cover, and climate-induced damage create diverse reflected waveforms.
• Accurate classification of these waveforms is crucial for assessing bridge structural integrity.

Purpose of the Study:

• To evaluate the effectiveness of multilayer neural networks (MLNNs) for classifying impulse radar waveforms from asphalt-covered bridge decks.
• To assess the impact of data dimensionality reduction using principal components analysis (PCA) on classification accuracy.
• To demonstrate a robust method for non-destructive evaluation of bridge deck conditions.

Main Methods:

• Utilized multilayer neural networks trained with the backpropagation algorithm and radial basis functions.
• Applied impulse radar to collect waveforms from three distinct asphalt-covered bridge deck structures.
• Implemented principal components analysis (PCA) to reduce the dimensionality of the input data.

Main Results:

• MLNNs achieved classification accuracies ranging from 89.9% to 100% for impulse radar waveforms.
• After PCA, classification accuracies improved, ranging from 95.6% to 100%.
• The study successfully classified waveforms despite significant variations in asphalt thickness and concrete cover.

Conclusions:

• Multilayer neural networks provide a highly accurate method for classifying impulse radar waveforms from asphalt-covered bridge decks.
• PCA is an effective technique for enhancing classification performance by reducing data dimensionality.
• This approach offers a promising solution for reliable structural health monitoring of bridges.