As autonomous vehicles are developed to take on more responsibilities, program algorithms will be developed to accommodate these tasks. Currently, as mentioned in the WaveGlider section, new distributed on-board collaborative autonomous vehicle control programs are being developed that will enable an individual vehicle to coordinate and control multiple vehicles. This technique enables "swarm" capabilities among multiple vehicles. With onboard collaborative control, the vehicles operate as a group, functioning together as a "swarm." The swarm processes and communicates relevant information allowing individual vehicles and the entire swarm (i.e., group) to change direction, autonomously, in response to sensor inputs. This control is one of the primary research initiatives by the military for unmanned vehicle control in the air, on the ground and underwater. The concept of swarming is also useful to science for the sampling of entire regions for a specific organism, substance or phenomenon. The control of an autonomous underwater vehicle whether powered or a glider will also in the future utilize some combination of traditional (Figure 16) and neural network (Figure 17) navigation system that uses Kalman filters4 to control the AUV.
One of the requirements for a long duration, autonomous underwater vehicle, is the need for a robust, fault tolerant, navigation system. In addition to the robustness issue, there are core issues of nonlinear control as they pertain to maneuverability and sea keeping. In both issues, neural networks offer very promising solutions. For example, the calculation of the
4 A Kalman filter is a recursive filter estimating the state of a dynamic system. It is especially useful for handling incomplete or noisy measurements.
distances and the relative velocities will be by the use of the positioning data as well as by measuring inertial sensor data. In order to increase the reliability of the data, a reconciliation of both processes must be accomplished accurately and efficiently. The coordination of the target trajectories of the AUVs can give further important information for the positioning prognosis.
Current thruster powered commercial AUV systems use a combination of internal inertial, compass, and accelerometer sensors, in conjunction with external active acoustic triangulation methods (LBL, SBL, USBL)5. These have met with some success for applications of cable following, standard grid surveying, search and rescue, or signal
5 LBL - Long Baseline, SBL - Short Baseline, USBL - Ultra Short Baseline.
following. But in each of these cases, the system is unable to respond to a) abrupt changes in external environment, b) system damage, c) uncertain or indeterminate data input. In these areas, some scattered research on the use of neural networks has been performed with success, addressing specifically the fault tolerance, docking, and ranging issues. For example, Wilson (Wilson, 1995) successfully evaluated the use of a neural network for a spaceship application providing robust navigation despite thruster failure. Most of the work in this area has been in spacecraft, but the work is directly applicable to underwater and surface vehicles. In most of the cases, a back propagating network is applied using position, rotation, or acceleration error as the training tool. In each case, changes to the vessel control system itself or in the external environment (displacement forces) causes the system to update its training, which in turn prompts it to compensate for the change in forces. Ship navigation has been evaluated using neural network based adaptive critic designs. For autonomous underwater vehicle (AUV) control, a neural network has been modeled at the University of Hawaii for the problem of depth gradient descent only. In each case, the results were very positive, indicating that if generalized, a full neural network system could provide robust navigation for an AUV.
In addition to the constituent issues above, there are many problems these vehicles are only beginning to address. Examples of these might include: search for environmental pollutants; search and analyze biological systems; locate and identify artificial acoustic sources; long term scanning for physical, biological, or chemical subjects of interest; non-inertial navigation.
The final step in the process is to use the processed multi-sensory data from the pattern recognition and data classification modules to provide control inputs for the navigation system. Thus, the system would then be able to track and monitor targets as listed above. In this phase, a simple feedback of neural network outputs will be sent to the control processor algorithm. The power of the neural network paradigm is the ability of the system to integrate the sensor input from a variety of sources into multi-sensory patterns, that is, acoustic with salinity, temperature and pressure, spectrographic with temperature, etc. But instead of traditional analytical methods where the individual datasets are correlated one by one, the neural network will be able to search for patterns in all sets together.
Was this article helpful?