This paper compares between the results obtained in manoeuvrability trials and simulated trajectories using Artificial Neural Networks. Manoeuvrability trials have been conducted with a fast patrol vessel of the Portuguese navy. The trial plan followed the IMO interim standards for ship manoeuvrability (1993 and 1994): Turning circles at full and half ruder, zigzag manoeuvres 10 deg - 10 deg and 20 deg - 20 deg, spiral manoeuvre, stopping manoeuvre. A DGPS system generated instantaneous ship co-ordinates in terms of latitude and longitude as well as the global time. For the simulations, a recursive neural net (RNN)with 16 biased computational nodes (4 input nodes, 2 output nodes, and one hidden layer consisting of 10 nodes) was used. Output parameters were the sway and yaw velocities. Input parameters were rudder angle, ship speed, heading, and recursively the output parameters of the previous time step. A standard backpropagation algorithm was used to train the network. The binary sigmoid function was used as nonlinear transfer function, operating on the inputs to the node and producing a smoothly varying output. Data collected from 15 manoeuvrability tests were used to train the RNN, each test lasting about 10 minutes. The sampling period was 1 second. All input data had to be filtered applying a low pass Butterworth digital filter. The training data vector used 70% of the total data. The remaining data were used for cross validation and tests. The learning rate and the momentum were set to 0.1 and 0.7, respectively. For considerably higher values, the training failed to converge to an acceptable error. 65500 iterations were used. The predictions of the turning circles are good; the averaged relative error for the two co-ordinates is 5% and 9%, respectively. The zigzag manoeuvre is more complex, and so the network predictions show a relative error of 24,6% based on an average peak-to-peak heading variation of 44 deg. The path length prediction for the crash-stop path and a stopping manoeuvre are good (relative error 0.8% and 3.8% respectively), but there is a considerable difference in lateral deviation (relative error 32.7% and 55.3%, respectively) due to the high sensitivity of the manoeuvre to unknown environmental perturbations. The use of 20 nodes in the hidden layer did not change accuracy significantly, while increasing computational time by 60%.
Comparison between manoeuvering trials and simulations with Recursive Neural Networks
Vergleich zwischen Manövrierversuchen und Simulationen mit rekursiven neuronalen Netzen
Ship Technology Research / Schiffstechnik ; 50 , 2 ; 77-84
2003
8 Seiten, 12 Bilder, 3 Tabellen, 16 Quellen
Aufsatz (Zeitschrift)
Englisch
Comparison between Manoeuvring Trials and Simulations with Recursive Neural Networks
Taylor & Francis Verlag | 2003
|Comparison between Manoeuvring Trials and Simulations with Recursive Neural Networks
Online Contents | 2003
|Manoeuvering trials for a modern containership
Tema Archiv | 1989
|Navigating Between Anchored Ships and Manoeuvering Difficulty
Online Contents | 2002
|Manoeuvering and reversing rudders
Engineering Index Backfile | 1920