VibroNet: Recurrent Neural Networks with Multi-target Learning for Image-based Vibration Frequency Measurement

Abstract Information on the vibration characteristics of machinery or civil structures is crucial in identifying or diagnosing potential malfunctions or faults. The use of cameras in vibration frequency measurement enables the acquisition of spatially high-density information of a distant measurement target. Cameras can be used for remote monitoring or as contact-less inspection sensors for machinery or civil structures owing to their low cost, high availability, and ease of installation. However, this requires complex image processing and signal processing algorithms to interpret the vibration frequency from image data. This paper proposes an image- and machine-learning-based method to measure the vibration frequency, in which long short-term memory-recurrent neural networks (LSTM-RNNs) and multi-target learning are used to predict the vibration frequency components directly. The results of the proposed method are compared against the results obtained using an accelerometer and eddy current sensor in the frequency domain. The methodology was applied in forced vibration and free vibration laboratory experiments and real-world structure experiments including a membrane structure and a cable bridge. The experimental results show that the proposed method is robust and accurate. Proposals for future applications and research directions of this methodology and limitations and challenges in real-world implementations are presented.