Lepine, Julien (2016) An Optimised Machine Learning Algorithm for Detecting Shocks in Road Vehicle Vibration. PhD thesis, Victoria University.

Abstract

In most countries, domestic transport of products is predominantly made on roads. Protective packaging is used to protect the freight from the shocks and vibration encountered during this type of transportation. Inefficient packaging constitutes a significant problem that costs hundreds of billions of dollars and has an important environmental impact. An insufficient level of packaging increases the occurrence of product damage during transport, whereas excessive packaging increases the packages weight and volume which is costly throughout the supply chain. In order to reduce these costs, packaging protection is currently optimised by simulating the vibration produced by road vehicles. Despite these simulations being prescribed by many standards, it is broadly acknowledged that these methods oversimplify the Road Vehicle Vibration (RVV) which imposes a significant limit on packaging optimisation. More complex RVV models have been recently developed to enhance simulations. However, the fundamental problem still remains; none of these alternatives consider the different excitation modes contained in RVV: i.e. the nonstationary random vibration induced by pavement’s profile and vehicle speed random variations; the shocks caused by pavement’s aberrations and discontinuities; and the sinusoidal (harmonic) vibration caused by unbalanced wheels and the engine-borne vibration. These excitation modes should be included in the RVV simulation to ensure that simulations are realistic and accurate and that the packaging will protect against vibration during transport without being too excessive. Each of these modes is represented by a different mathematical model and cannot be analysed with the same statistical tools. This means that they have to be characterised separately in order to create an accurate RVV simulation model. This task is challenging because all the excitation modes are simultaneously present in the acceleration signal recorded on a vehicle. This PhD thesis proposes to use machine learning to separate these modes from a signal. Being a first attempt to apply this approach to index RVV, the most common classification algorithms are used to identify the two predominant modes; i.e. the nonstationary vibration and the shocks. The important novelty of this approach is that algorithms integrate many RVV analysis methods such as moving statistics, the Discrete Wavelet Transform and the Hilbert-Huang Transform. A comprehensive evaluation and optimisation of the classification algorithms was performed using synthetically generated RVV signal. The best performing algorithm was applied on a real measurement dataset. The RVV mode decomposition will greatly increase the ability to correctly optimise the level of packaging required. An accurate model comprising all the characteristics inherent to RVV will be a considerable step forward to reduce unnecessary level of protective packaging without risking damage to products.

Search Google Scholar

Repository staff login