Guided Waves Damage Identification in Beams with Test Pattern Dependent Series Neural Network Systems

In regression neural networks for pattern recognition of preprocessed guided waves signals in beams, a trained network produced large errors when identifying a test pattern not found in the training set. To improve the accuracy of results, a new neural network procedure was introduced where progressive training was performed in a series combined network with the integration of a weight-range selection (WRS) technique that was dependent on the test pattern. The WRS method was applied for a supervised multi-layer perceptron operating with one hidden layer of neurons and trained using a backpropagation algorithm. The system was able to achieve average predictions accurate to 2.5% and 7.8% of the original training range sizes for the damage location and depth respectively while the WRS provided up to 13.9% improvement compared to equivalent conventional neural networks. Key-Words: pattern recognition, combined neural networks, generalization, ultrasonic guided waves, quantitative nondestructive evaluation, structural health monitoring