Bayesian estimation of output combining method for bridge damage identification using multiple CNNs

This research concerns a bridge condition identification method using a convolutional neural network (CNN). Currently, bridges are mainly inspected visually by humans, and it is difficult to detect damage that does not appear on the surface. Therefore, a condition evaluation method using sensors is required. In this study, a damage identification method is proposed by classifying the images, visualized by vibration analysis such as spectrogram or FFT of acceleration response of a bridge, using CNN. The effect of analysis methods, the presence or absence of a time component, the processing of the image itself, and frequency resolution on diagnostic accuracy are clarified. The overall Identification rate is higher for spectrograms containing more information, and for damage with less effect on vibration, the FFT has a higher Identification rate. Furthermore, a method to improve accuracy by combining these multiple CNNs using Bayesian estimation is proposed. Accurately identifying damage, the degree of which varies incrementally, was a complex problem for a single CNN. Combining multiple CNNs with various characteristics using attribution probabilities has reduced misclassification and improved identification rates over a single CNN.