Sideslip angle estimation through neural networks is an attractive research perspective because of its potential to overcome the limitations of filter-based approaches. While a close eye is generally kept on the training loss function value to prevent overfitting, there are limited attempts at tailoring the input vector to be qualitatively significant rather than quantitatively significant. This paper investigates this issue by factoring out the kinematic contribution of sideslip angle, and by only selecting meaningful input signals - leaving out those who are not beneficial to the network performance. The obtained RMSEs for different input combinations are compared to the standard input set, targeting the whole sideslip angle. Results show the most insightful signals can reach better validation performance than the benchmark approach, using only two instead of five signals.
Physics-Infused Neural Network-Driven Investigation of Vehicle Sideslip Angle
Lenzo B.
;
2024
Abstract
Sideslip angle estimation through neural networks is an attractive research perspective because of its potential to overcome the limitations of filter-based approaches. While a close eye is generally kept on the training loss function value to prevent overfitting, there are limited attempts at tailoring the input vector to be qualitatively significant rather than quantitatively significant. This paper investigates this issue by factoring out the kinematic contribution of sideslip angle, and by only selecting meaningful input signals - leaving out those who are not beneficial to the network performance. The obtained RMSEs for different input combinations are compared to the standard input set, targeting the whole sideslip angle. Results show the most insightful signals can reach better validation performance than the benchmark approach, using only two instead of five signals.Pubblicazioni consigliate
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.