An Evaluation of Wooden House Health Monitoring System using PVDF Piezoelectric Sensor with 3-layer Neural Network and Inverted Binary-Data Augmentation

We propose a wooden house health monitoring system using an AI chip and polyvinylidene fluoride (PVDF) piezoelectric sensors. In our experiments, we vibrated a test bed simulating a Japanese tea room, and obtained waveform data were binarized to be trained with a 3-layer neural network as a classifier. Using this 3-layer neural network, we determined that only one of the test bed's four seismic shear walls was damaged. A comparison was made between cases where "inverted data for each bit of binarized waveform data" were added as data augmentation at the time of training and where they were not added. As a result, the accuracy rate improved by 10% at most when augmenting the data. In addition, the identification rate was a maximum of 70.3% for the data obtained by the piezoelectric sensor attached to the south side secondary member upper part located at the center of the test bed. We intend to further increase the identification rate and implement the classifier in a field-programmable gate array (FPGA).