Deep learning-based impact locating using the power spectrum of an acceleration signal on a cantilever beam.

This study proposes a deep neural network-based impact locating method on a cantilever beam. The power spectrum of a measured acceleration signal, when an impact is exerted at an arbitrary location on the beam, contains the inherent frequency response of the beam, including resonances and anti-resonances. Especially, the anti-resonances can be a useful feature for estimating the impact location because they are dominated by both the sensor and impact locations. However, in the power spectrum expressed using a linear scale, these anti-resonances may be less noticeable due to their small values relative to the resonances. The proposed impact locating method adopts the power spectrum expressed using a dB scale to highlight the importance of the anti-resonances as the input of a deep neural network. The deep neural network was trained, validated, and tested using a simulated dataset derived from an ideal cantilever beam model with a length of 800 mm, including a single accelerometer. From the test result, the proposed method achieved a root mean square error of about 1 mm in impact locating for a total of 800 impact locations with an interval of 1 mm, a significantly improved accuracy from that using the linear scaled power spectrum. [ABSTRACT FROM AUTHOR]