Damage-Position Identification of Wooden-House Models for Structural Health Monitoring Using Machine Learning

We used our previously proposed structural-health-monitoring system that uses machine learning and requires only one sensor to identify damage locations of braces and walls and applied it to two wooden-house model to identify damage locations. In our previous studies, we succeeded in identifying damage locations with 90% accuracy in a wooden-house model with two crossed braces by using our system. We also conducted an experiment on an actual wooden house in Oita prefecture, Japan and identified the damage locations with an accuracy of 86.0%. For this study, we used this system to identify the damage locations of a wooden-house model with only 28 diagonal braces and another wooden-house model with 26 walls. We removed only one brace and one wall from each model and assumed that they were the damage locations. Shaking was generated by attaching a motor as a vibration source to models. The vibration of models was detected using a piezoelectric sensor, and the output voltage waveform of the piezoelectric sensor was recorded using a digital oscilloscope. This output voltage waveform was analyzed using a neural network. Using a three-layer neural network, four sides of both models were identified separately and more than 95% of the braces and walls were recorded. Damage locations throughout the entire braced and walled models were then identified using a neural network with three to ten layers. As a result, the identification rate was 94.5% for the braced model with the neural network with four layers and 97.8% for the walled model with the neural network with five layers.