Measuring flow speeds in natural waters by training an artificial neural network to analyze high-frequency flow-induced vibrations of tethered floats

摘要

Measuring water currents in natural waters is limited by the cost of sensors. Standard sonarbased acoustic current Doppler profilers (ADCPs) are high cost, about $10-20 K per unit. Tilt current meters (TCMs) are much cheaper. They consist of a bottom-mounted subsurface float equipped with an inertial measurement unit (IMU) and data center that records the float's motion and attitude as a time series. The flow speed is measured by calculating the tilt angle of the float in response to the current. However, tilt-based measurements require the float system to be carefully engineered and its physical response optimized for good results. Even so, high-frequency flow-induced vibrations often dominate the motion and must be averaged and filtered out of the data and discarded. This represents the loss of potentially valuable information, but decoding the high-frequency components for such useful data is difficult. These experiments explored using an artificial neural network (ANN) approach to extract the ambient water current speed from that high-frequency data alone, after the displacement information was filtered out. The methods were informed by the ANN designs and data augmentation techniques used by neurologists to observe the tremors and other motions exhibited by patients experiencing symptoms of Parkinson's disease. Once the model was trained using carefully selected training and validation sets to prevent overfitting, the results of evaluating previously unseen data by the model are clear and promising. Water current speed was accurately calculated from the high-frequency components of the motion sensor data and agreed with corresponding current speeds measured by established methods. This novel approach could facilitate new sensor system designs that can be empirically or self-calibrated more efficiently and have a lower barrier to application than those currently available.