Multidisciplinary Design Method for Off-Road Vehicles Using Bayesian Active Learning

Abstract

<div class="section abstract"><div class="htmlview paragraph">When developing an off-road vehicle, it is essential to create excellent drivability that enables the vehicle to be driven on all surfaces while ensuring passenger comfort. Since durability is another indispensable performance aspect for these vehicles, the development method must be capable of considering a high-level combination of a wide range of performance targets. This paper proposes a method to identify the region in which each performance aspect is realized through a complex domain combination problem. The proposed method is helpful in the initial design stage when the detailed specifications of the target vehicle are not determined because it is capable of considering both the specifications and usage method of the target vehicle, such as the selection of road profiles and driving speeds as design variables. The proposed method has the advantage of enabling efficient concurrent studies to search for feasible regions. By introducing a probabilistic representation of multidisciplinary constraint functions, the feasible regions of each discipline subproblem can be decoupled by the rule of product. This nature makes adding the constraint functions in the later design stage easy. In contrast, it is unclear which constraint function is active at the early design stage. Calculating inactive constraint functions yields a complex prediction model and requires a higher calculation cost. In the early stage of off-road vehicle design, many design variables and unclear cost functions exist. This paper proposes an improved modeling method represented as the combination of constraint functions. The proposed method can deal with more complex constraint functions without degradation of model accuracy. To show the effectiveness of the proposed method, a practical numerical example of a multidisciplinary vehicle design problem is presented.</div></div>