Experimental Evaluation of Generalization Performance of a Neural Network used by a Multiple-defect-location Discrimination System for House-model-based Health Monitoring

The purpose of this study was to evaluate the generalization performance of a health-monitoring system based on a wooden-house model. Previous studies have only validated the discrimination of a single deficiency. To improve the performance of the system, experiments were conducted with the expectation of improving the results by increasing the number of defects and the number of samples. The model, a two-story structure, is composed of braces at 28 different locations. It was divided into 28 to 378 classes according to the combination of braces that were removed as pseudo-damage. A piezoelectric sensor was attached to the model to acquire vibration waveforms. This waveform data set was split into training and evaluation data, and a neural network (NN) was trained on the former data to classify the latter data. As a result, the NN achieved a high accuracy of up to 87.24% in classifying missing-brace positions, even for vibration waveforms that were not included in the training set. These results demonstrate the high generalization performance of the trained NN used in the house-model-based health-monitoring system.