Recurrent neural networks for hydrodynamic imaging using a 2D-sensitive artificial lateral line

Abstract

The lateral line is a mechanosensory organ found in fish and amphibians that allows them to sense and act on their near-field hydrodynamic environment. We present a 2D-sensitive artificial lateral line (ALL) comprising eight all-optical flow sensors, which we use to measure hydrodynamic velocity profiles along the sensor array in response to a moving object in its vicinity. We then use the measured velocity profiles to reconstruct the object’s location, via two types of neural networks: feed-forward and recurrent. Several implementations of feed-forward neural networks for ALL source localisation exist, while recurrent neural networks may be more appropriate for this task. The performance of a recurrent neural network (the long short-term memory, LSTM) is compared to that of a feed-forward neural network (the online-sequential extreme learning machine, OS-ELM) via localizing a 6 cm sphere moving at 13 cm s−1. Results show that, in a 62 cm 9.5 cm area of interest, the LSTM outperforms the OS-ELM with an average localisation error of 0.72 cm compared to 4.27 cm, respectively. Furthermore, the recurrent network is relatively less affected by noise, indicating that recurrent connections can be beneficial for hydrodynamic object localisation.

Hydrodynamic Imaging with Artificial Intelligence: detecting submerged objects at a distance using a 2D-sensitive flow sensor array and neural networks

“…This type of network was also used in [121] to localize both moving and stationary vibrating sources in a 2D plane. Recently [123], the ELM architecture was compared to a recurrent network architecture (LSTM) for objects moving in a straight line in a 2D plane. To the authors' knowledge, the present work is the first effective demonstration of a CNN architecture for localizing a source with an ALL.…”

“…This could be explained by the fact that it is easier to generalize from single time steps than it is to generalizing from multiple time steps. Perhaps in a more complex setting it might still be valuable to incorporate previous time frames, as has been shown for localizing a single source using regression [123]. One such situation that may benefit from taking history into account is when objects turn more slowly than the current maximal 1 rad (57 degrees) per second or when the objects can vary their speed.…”

“…The application of hydrodynamic detection of objects can therefore not scale up indefinitely, but the range can be extended beyond the current chosen domain. It has been shown that the source can be positioned in an area next to the array and still be detected reliably using different types of artificial neural networks [6,121,123]. It is, therefore, not necessary for the whole velocity pattern to be sampled within the area directly in front of the array; the domain can extend beyond the length of the array.…”

“…To demonstrate object shape classification, we installed eight 2D-sensitive all-optical flow sensors [123,124] each 0.5 m apart at the short side of a 18×25 m swimming pool, see also figure 8.2. The 3.5 m array was placed 0.5 m from the wall and centered with respect to the 18 m wide pool area.…”

“…In both cases, the ALL measures the hydrodynamic effect of relative fluid motion. A stationary ALL can be used for object detection and tracking vibrating dipole sources [4,12,31,125,131] or moving objects [43,121,123], with potential use cases including tracking nearby moving vessels. An ALL attached to a vessel can be used for aligning with the freestream flow [35,73], or for obstacle avoidance [76,117], potentially enabling safe navigation for ships or autonomous underwater vehicles (AUVs) [73].…”