Application of SOM Neural Network in the Construction of Urban Ramp Driving Cycle

Abstract

In order to construct vehicle driving cycle with ramp characteristics, this paper applies SOM neural network to the construction of urban ramp driving cycle. Firstly, the data collected by the actual vehicle test is divided into micro-trips, after that the principal component analysis method is used to reduce the dimension of the selected 20 characteristic parameters, afterward the first five principal components of all the micro-trips are clustered by the SOM neural network, and then the micro-trips with the appropriate length of time are selected from each category to build a representative driving cycle with the smallest average relative error and stable slope angle-time curve. The research results show that the SOM neural network has high clustering accuracy in the construction of the ramp driving cycle. Introduction The driving cycle of the automobile is mainly used to test the performance indexes such as fuel consumption and emissions of automobiles in the development and evaluation of new technologies of automobiles, which is a common core technology of the automobile industry. The three typical driving cycles that are widely used in automotive fuel consumption testing are the US driving cycle, Japanese driving cycle and New European driving cycle(NEDC) [1]. Most of the researches on the driving cycle of automobile at home and abroad are based on the speed-time driving cycle curve. The urban ramp factors are not taken into account and cannot be used to test the performance of vehicles driving on urban ramps. Therefore, in order to construct the driving cycle of urban ramp, this paper proposes a construction method based on SOM neural network. SOM Neural Network Clustering Analysis Theory and Method The structure of SOM neural network is shown in Fig 1. It is a two-layer neural network composed of an input layer-competition layer (output layer). And it has the characteristics of unsupervised autonomous learning. Then the neurons in the competition layer of the network compete for the opportunity to respond to the input samples. Finally there is only one winning neuron, it represents the classification of the input samples[2,3]. The main process of cluster analysis consists of the following steps: 1) Initialize the network weight ∈ [0,1] ; determine the initial value of the learning rate η 0 ∈(0,1]; determine the initial field strength 0 ; set the maximum number of learning ; 2) Calculate the distance between the input vector = , , ..., and the output layer neuron as follows; d = ∑ − = 1, 2, . . . , ! (1) 3) Selecting the output layer neuron having the smallest distance from the input vector as the winning neuron; 4) Take the field strength = 0 ∗ $1 − % & ; adjust the nerves contained in the winning neurons and their fields (the field strength is ) as follows weight coefficient