Abstract
The research presented in this thesis focuses on the evaluation of existing ridecomfort assessment methods and investigates the use of novel approaches for the
data correlation and measurement of the subjective and objective responses of
the human body. The study has two main objectives. The first is the investigation
of the applicability of artificial neural networks for solving the objective and
subjective data correlation. The second in the evaluation of influence of road
induced vibration on human physiology with possible correlation of the
physiological response with subjectively perceived comfort. In addition to that
influence of anthropometric size on the subjectively perceived ride comfort is
investigated. The created neural network is studied parametrically, and a
sensitivity study of the network input parameters complements the investigation.
To achieve this set of objectives, two field studies were designed. Data collection
commenced on a mix of road sections located on a public road and a proving
ground at HORIBA MIRA in the UK. In the first study data from 10 participants
were collected. Objective and subjective measurements were taken. In the
second study data from 12 participants were collected. In addition to the
objective and subjective measurements the physiological measurements were
also collected. The collected data set was analysed according to the guidelines
found in the international standard for whole-body vibration. The physiological
measurements were analysed, and the essential metrics were calculated. Data
obtained from the field studies were used to train a classifier utilising artificial
neural networks.
The obtained results show that it is possible to train a classifier utilising artificial
neural networks to predict subjective comfort scores based on the objective and
anthropometric data. The trained neural network has achieved high accuracy of
R=0.92415. Parametric studies have revealed that the best performing neural
network for solving the objective-subjective data fitting problem uses the
Levenberg-Marquardt backpropagation algorithm and uses the mean squared
error method of calculating error in the learning process. The validation of the
iv
trained neural network has shown that predicted data is of high accuracy and the
evaluations are within 10% error of the measured dataset.
Additionally, a study investigating the applicability of physiological
measurements for ride comfort estimation has been conducted. It has shown that
although these measurements can be used for studies causing motion sickness,
for ride comfort studies there is no direct correlation between the subject’s
physiological response and the vibration induced by the irregularities in the road
surface. Although the electrodermal signal has shown a good correlation with
the jerk vibrations, the area under curve metric has shown no correlation with
different ride comfort states. The study showed that the physiological
measurement yielded no correlation concerning vehicle ride comfort and that
further work in this field is needed.
In conclusion, the novel findings obtained from the artificial neural network
study prove that through utilisation of already existing metrics it is possible to
create a classifier which will be characterised by a high level of accuracy in
subjective comfort prediction. Implementation of such an approach would allow
automotive manufacturers to drastically limit the development time required for
ride comfort assessment of vehicles.
| Date of Award | Jun 2020 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Sponsors | HORIBA MIRA Ltd. |
| Supervisor | Stratis Kanarachos (Supervisor), Mike Blundell (Supervisor) & Cyriel Diels (Supervisor) |