Vibration signal denoising for structural health monitoring by residual convolutional neural networks

In vibration based structural health monitoring (SHM), measurement noise inevitably exists in the vibration data, which significantly influences the usability and quality of measured vibration signals for structural identification and condition monitoring. As a result, there is a high demand for developing effective methods to reduce noise effect, especially in harsh and extreme environment. This paper proposes a vibration signal denoising approach for SHM based on a specialized Residual Convolutional Neural Networks (ResNet). Dropout, skip connection and sub-pixel shuffling techniques are used to improve the performance. The effectiveness and robustness of this developed approach are validated with acceleration data measured from Guangzhou New TV Tower. The results show that the proposed approach is effective in improving the quality of the acceleration data with varying levels of noises and different types of noises. Modal identifications based on signals contaminated with intensive noise and de-noised signals are conducted. Modal information of weakly excited modes masked by noise and closely spaced modes can be clearly and accurately identified from the de-noised signals, which could not be reliably identified with the original signal, indicating the effectiveness of using this developed approach for SHM. Besides white noise, a group of data contaminated with pink noise, which is not included in the training data, is also tested. Good results are obtained. The developed ResNet extracts high-level features from the vibration signal and learns the modal information of structures automatically, therefore it can well preserve the most important vibration characteristics in vibration signals, and can assist in distinguishing the physical modes from the spurious modes in structural modal identification.