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

In regression neural networks for pattern recognition, a trained network may often produce large errors when identifying a test pattern not found in the training set. This is especially true when test patterns and training patterns are obtained from two different sources, as in the case from measured and simulated data. Therefore, this paper investigates a new neural network procedure where progressive training is performed in a series network with the implementation of a weight-range selection (WRS) technique that depends on the test pattern. An integer states rejection (ISR) criterion is also introduced to monitor and select the final network outputs. The WRS and ISR methods are applied for a supervised multi-layer perceptron operating with one hidden layer of neurons and trained using a backpropagation algorithm. An example of this system has been esigned for damage identification in beams investigated with guided waves for structural health monitoring.