Driver Start-Up Comfort Rating of Automated Clutch Systems~An Objectification Using Artificial Neural Networks
Published May 23, 2004 by Society of Automotive Engineers of Korea in South Korea
During the product development process, dynamic features of vehicles are determined in drive tests. However, the evaluation of comfort behavior - either by engineers or by customers - can only be done at a later stage, because this requires a prototype close to the final product. Today's product development in automotive industries has moved more and more towards a virtual product development and tests have been shifted form "road to rig." Therefore, the prediction of comfort features from test-bench data and numerical simulation data is a great challenge and finding correlations between driver rating and data measured in drive tests is an important objective.
The subject of the presented study is the start-up process, i.e., the process of starting to drive from standstill and the following acceleration process. Drive tests were performed both by experts in vehicle evaluation and also amateurs representing average consumers. As the test vehicle, an intermediate-class car with an automated clutch system is used. The start-up characteristics can thereby be set to different variants by applying different types of controller software. Characteristic values are derived from data obtained in each test and the corresponding comfort ratings are captured. Similarly to the way a person would make his evaluation, an Artificial Neural Network (ANN) is used to interconnect input data (sensation) with output data (comfort rating) by "learned" connections. ANNs are information-processing systems consisting of a large number of simple units which dispatch information in the form of activation of these units via directed connections. A fundamental feature is their learning aptitude, the ability to independently learn a task from training examples without being explicitly programmed. During the training stage, the input and output teacher signals of some of the startup processes determined for every person are presented to the ANN. This way it learns how to connect the objective data with the corresponding subjective rating. During the application stage, the trained network is then able to interpret input data which has the same structure as the teacher signals but was not used for training.
The evaluation of the results is carried out by comparing the predicted comfort rating output value of the ANN for the remaining sets of data - not used for training - and the subjective rating actually given by the person in the corresponding drive test. The comparison shows that the approximation of comfort ratings of both experts and laymen is possible and thus the practicality of the described method is proven. This method may serve as a powerful tool especially for the optimization and individualization of not only automated clutch systems, but automated systems in general.