Prediction of Performance Measures for Buses: A System-Based Approach
Open Access
- Author:
- Muthiah, Saravanan
- Graduate Program:
- Mechanical Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- October 19, 2006
- Committee Members:
- Thomas Litzinger, Committee Chair/Co-Chair
Dr Mirna Urquidi Macdonald, Committee Chair/Co-Chair
Timothy William Simpson, Committee Member
Sean N Brennan, Committee Member
Shelley Marie Stoffels, Committee Member - Keywords:
- acceleration and gradeability
fuel economy
Artificial Neural Networks (ANN)
Transit buses
Performance Based Standards (PBS)
pass-by noise
reliability
multi-criteria optimization - Abstract:
- Heavy vehicles transport people and freight in an efficient manner and form the backbone of any developed economy. Historically, heavy vehicle standards (called Design Based Standards-DBS) have relied primarily on placing limits on vehicle weights and dimensions. Ease of implementation and lack of complete understanding of the complex relationships between vehicle design parameters and vehicle performance are the primary reasons such standards have been in place for decades. However, recently there has been a realization that such indirect control of vehicle performance can lead to a wide gap between intended performance and actual performance. A concept called Performance Based Standards (PBS) that assigns numerical limits to performance measures but leaves open ways of achieving the performance has now gained attention. However, for an effective implementation of PBS and for an optimal design of vehicles capable of achieving such performance, it is imperative that vehicle designers, test engineers and regulatory authorities alike have a sound understanding of the complex relationships that exist between vehicle design parameters and vehicle performance. The primary goal of this thesis was to establish and validate methods of analysis that can be used to investigate the relationships between vehicle design and performance. The other main goals include development of a reliable means of predicting the useful range of values for vehicle design parameters that would result in a vehicle with desired performance objectives (one or more) and evaluation of the effectiveness of PTI (Pennsylvania Transportation Institute) testing with an aim to suggest ways for further improvement. The thesis discusses four performance measures, namely fuel economy, acceleration and gradeability, pass-by noise and vehicle reliability primarily because PTI data were readily available for these measures. However, as the methods developed in this thesis are very general in nature, they can be used for analyzing other vehicle performance measures as well, provided reliable and accurate data are available. In this thesis, a two-stage system based model was implemented. In the first stage, functional relationships between vehicle design parameters and vehicle performance under laboratory/test track conditions were modeled using a vehicle transformation, F. In the second stage, the interdependence between test-track performance and real-life traffic performance was modeled using a traffic transformation, G. PTI data on fuel economy, acceleration and gradeability, pass-by noise and reliability for 124 diesel two-axle transit buses were used to model the vehicle transformation, F. On the other hand, the traffic transformation, G, was modeled using in-use data (at transit agencies) on fuel economy and reliability obtained from the National Transit Database (NTD). Modeling was done using two artificial neural network (ANN) based methods – N2PFA/REFANN proposed by Setiono et al. and RF5 proposed by Saito and Nakano. Both approaches have built-in rule generation capabilities and an automatic means for selection of the number of hidden neurons. Hence the problem of ANN being a black box is alleviated with either approach. A two-step method of input selection was used in this study. Firstly, correlation coefficients were evaluated between all pairs of input variables. Based on the assumption that a low correlation coefficient (< 0.7) can be used as a measure of linear independence, inputs that are least correlated to other inputs were selected. In the second step, an information theory based approach was used to narrow this input set down to a still smaller subset that contains most of the original information and is a good predictor of the output (vehicle performance measure) under study. Often it is also of interest to know vehicle configurations that can achieve desired vehicle performance objectives. To this effect, a means for generating an inverse model based on non-linear programming is also presented. In this thesis, inversions were obtained starting from each known bus (the ANN training set) as a reference point. The final solution (inversion) selected was the one closest in terms of Euclidean distance to an existing bus. The main idea behind using such an objective function was ease of manufacture. However, one can easily modify the objective function as per one’s requirement and achieve corresponding optimal solutions. Engineers often face the issue of determining a vehicle configuration that can simultaneously achieve more than one competing objectives. This study shows that Gembicki’s Goal Attainment method for multi-objective optimization can be very effective for this purpose. The results from the fuel economy and acceleration models validate the accuracy of the two ANN-based methods used in this study. This opens the way for use of these two methods for modeling other performance measures wherein knowledge may be inadequate. The two methods not only corroborate well with each other with respect to most relevant and least relevant inputs but also seem to perform better than two of the commonly used modeling methods, namely regression and decision tree. Further, unlike regression, the two methods make no assumption on the nature of the functional relationship being modeled. The nonlinear effect of vehicle weight on fuel economy of a bus is clearly brought out in this study. This is in contrast with existing literature that suggests a linear dependence. With regards to whole vehicle pass-by noise models, the methods used in this study were able to generate models that are very close in accuracies with the best such model (+/-1.1 dB(A)) described in literature while using only 1.2% of the number of data samples. This study confirms the importance of vehicle weight in determining reliability and also brings forth the possible case for vehicle height as an important determining factor.