Vehicle driving cycle construction method of urban roads

FullText

0. 引言

At the beginning of the 21st century, China directly adopted the European NEDC driving conditions to certify the energy consumption and emissions of automotive products^[6]Effectively promoting the development of energy-saving and emission reduction technologies for automobiles. However, in recent years, with the rapid growth of car ownership, the road traffic situation in China has undergone significant changes. China has significant differences from European countries in terms of car structure, urban layout, traffic rules, and driving habits. The results certified based on NEDC driving conditions are increasingly deviating from the actual situation. In order to construct vehicle driving conditions suitable for the local area, scholars at home and abroad have conducted a large amount of research.

The method proposed in this article breaks through the traditional research's single use of PCA processing mode in dimensionality reduction methods, innovatively introduces AE for comparison, and avoids the shortcomings of manual operation based on different index partitioning criteria in clustering operations and kinematic segment screening. It enriches the selectivity of the construction method for automobile driving conditions and forms a complete method system.

1. 数据预处理

1.1 数据采集

The data collection was conducted on the roads in Fuzhou City, consisting of three sets of data files. The latitude and longitude data were used to obtain the path distribution of the car's driving route and to locate it using a mapFigure 1~3As shown.

Figure  1.  Vehicle driving routes of data file 1

DownLoad: Full-Size Img PowerPoint

Figure  2.  Vehicle driving routes of data file 2

DownLoad: Full-Size Img PowerPoint

Figure  3.  Vehicle driving routes of data file 3

DownLoad: Full-Size Img PowerPoint

causeFigure 1~3It can be seen that the test vehicle of data file 1 departs from Cangshan District and mainly passes through Changle District. The driving area of data file 2 is mainly Putian, and the driving area of data file 3 is concentrated in Mawei District; The three sets of data files cover multiple major areas in Fuzhou City, with rich sampling samples and comprehensive sampling time, ensuring the diversity and integrity of the data in terms of time and space, reflecting the typical road traffic conditions in Fuzhou City, and meeting the requirements for constructing urban road vehicle driving conditions.

1.2 Abnormal data processing

The raw data directly recorded by the car driving data collection equipment contains abnormal data values, mainly including the following types.

(3) Abnormal data collected during long-term parking, such as parking without turning off the engine, parking with the engine turned off but the collection equipment is still in operation.

(4) Long term traffic congestion and intermittent low-speed driving (maximum speed less than 10.0 km/h)^-1）Usually, it can be handled at idle speed.

(5) It is generally believed that an idle time exceeding 180 seconds is an abnormal situation, and the longest idle time should be treated as 180 seconds.

1.3 基于多尺度小波变换的车速扰动降噪

1.3.1   多尺度小波变换原理

Wavelet transform inherits and develops the localization idea of short-time Fourier transform, while overcoming the disadvantage of window not changing with frequency. It can provide a "time-frequency" window that changes with frequency and is an ideal tool for time-frequency analysis. Its main feature is to fully highlight certain features through transformation, enable local analysis of time (space) and frequency, gradually refine the signal at multiple scales through scaling and translation operations, and ultimately achieve the requirements of time subdivision at high frequencies, frequency subdivision at low frequencies, and adaptive time-frequency signal analysis. Therefore, it can focus on any detail of the signal, becoming a major breakthrough since Fourier transform and widely applied^[33-35].

Wavelet transform decomposes the original signal into two types of information: approximate information and detail information. The approximate coefficient represents the low-frequency part of the signal, while the detail coefficient represents the high-frequency part of the signal. The single scale wavelet transform only decomposes the original signal into two parts: approximate information and detail information, while the multi-scale wavelet transform can expand the approximate information of the previous layer into new approximate information and detail information of the next layer. Considering the timeliness of the calculation, this paper adopts a 3-layer wavelet decomposition, such asFigure 4As shown.

Figure  4.  Three-level wavelet decomposition

DownLoad: Full-Size Img PowerPoint

of whichSRepresenting raw data,B₁、B₂、B₃The approximate coefficients obtained from the decomposition of layers 1-3 are respectively,D₁、D₂、D₃The detail coefficients obtained from the decomposition of layers 1-3 are respectively.

1.3.2   应用3层小波变换的车速降噪

In wavelet decomposition, the approximation coefficients obtained from the lower layer decomposition are smoother than those from the upper layer, and denoising is achieved by reconstructing the approximation coefficients of each layer. To ensure the relative integrity of information, the detail coefficients of each layer can be selectively suppressed.

Figure  5.  Wavelet decomposition coefficients of vehicle speed data file 1

DownLoad: Full-Size Img PowerPoint

The threshold is determined based on the maximum absolute value of each layer's detail coefficients. Each layer's detail coefficients only retain the parts with absolute values greater than the threshold. The processed new coefficients are reconstructed to obtain the denoised vehicle speed data information. Noise often contains detailed information. By setting a threshold to limit the intensity of details, noise can be reduced while preserving necessary details. Therefore, multi-scale wavelet transform achieves vehicle speed denoising while retaining key information about vehicle speed.

The noise reduction effect of the vehicle speed data in the three data files is similar. The comparison of noise reduction effects is also demonstrated using the noise reduction effect of the vehicle speed data in data file 1 as an example. The effects of the other two data files will not be repeated. Wavelet decomposition denoising of data file 1, for exampleFigure 6As shown, the denoised vehicle speed signal retains the main components of the original vehicle speed signal, while the disturbance information is reduced.

Figure  6.  Noise reduction comparison of vehicle speed data file 1

DownLoad: Full-Size Img PowerPoint

1.4 Kinematic fragment extraction

Table  1.  Number of kinematic clips

文件编号	文件1	文件2	文件3	总数
运动学片段数量	1 268	1 788	1 298	4 354

 | Show Table

DownLoad: CSV

2. data analysis

2.1 运动学片段特征体系构建

Building a comprehensive and reasonable system of automobile motion characteristics is a prerequisite for constructing the driving conditions of automobiles, and is the subsequent process of dimensionality reduction based on PCA and AE featuresK-The basis of means++clustering analysis is to extract representative feature parameters closely related to driving characteristics in order to better construct the driving conditions of automobiles. Each kinematic segment obtained from the solution needs to extract feature parameters to characterize the driving characteristics of the car.

The feature system constructed in this article consists of 9 typical feature parameters, namely average speed, average driving speed, idle time ratio, acceleration time ratio, deceleration time ratio, average speed deviation, average acceleration, acceleration standard deviation, and average deceleration.

2.2 Dimension reduction of kinematic segment features

2.2.1   PCA based kinematic fragment feature dimensionality reduction

The kinematic segment feature system is a typical multivariate large dataset, and there will inevitably be information overlap between the various feature parameters in the feature system. In order for the selected feature parameters to comprehensively and redundantly reflect the driving characteristics of the vehicle, it is necessary to reduce the dimensionality of the feature parameters. PCA is the most widely used feature dimensionality reduction method, which is implemented by calculating the covariance matrix of the data matrix, obtaining the eigenvectors of the eigenvalues of the covariance matrix, and selecting the one with the largest eigenvaluekThe matrix composed of feature vectors corresponding to each feature is transformed into a new space to achieve dimensionality reduction of data features. The specific steps are as follows.

(1) Central standardization

$\mathit{A}{}_{m\times n}=\left[\begin{array}{ccc}a{}_{11}& \cdots & a{}_{1n}\\ ⋮& & ⋮\\ a{}_{m1}& \cdots & a{}_{mn}\end{array}\right](1)$

In the formula:A_m×nA feature parameter matrix composed of various kinematic segments and their kinematic feature parameters;a_ij(i=1, …, m; j=1, …, n）For the thiA kinematic fragment of thejA parameter.

In this article,Table 1If the number of obtained kinematic segments is 4354, thenmFor 4 354, there are 9 types of feature parameters in this articlenFor 9.

answerA_m×nThe elements in are standardized as follows:

$z{}_{ij}=\frac{a{}_{ij}-\stackrel{-}{{\displaystyle a}}{}_j}{s{}_j}$(2)

$\begin{array}{l}\stackrel{-}{{\displaystyle a}}{}_j=\frac{{\displaystyle \sum _{i=1}^{m}a}{}_{ij}}{m}(3)\\ s{}_j^2=\frac{{\displaystyle \sum _{i=1}^m(a{}_{ij}-\stackrel{-}{{\displaystyle a}}{}_j)}{}^2}{m-1}(4)\end{array}$

Thus, a standardized matrix is obtainedZ_m×n, for

$\mathit{{\rm Z}}{}_{m\times n}=\left[\begin{array}{ccc}z{}_{11}& \cdots & z{}_{1n}\\ ⋮& & ⋮\\ z{}_{m1}& \cdots & z{}_{mn}\end{array}\right](5)$

(2) Calculate covariance matrix

Table  2.  Contribution rates and cumulative contribution rates of each principal component

主成分序号	特征值	贡献率/%	累计贡献率/%
1	3.54	39.32	39.32
2	2.44	27.06	66.38
3	1.74	19.31	85.69
4	0.61	6.74	92.43
5	0.33	3.72	96.14
6	0.15	1.67	97.81
7	0.11	1.20	99.01
8	0.07	0.79	99.81
9	0.02	0.19	100.00

 | Show Table

DownLoad: CSV

Table  3.  Load factors of each principal component

原始特征	M1	M2	M3
平均速度/(km·h-1)	0.46	-0.23	-0.16
平均行驶速度/(km·h-1)	0.43	-0.18	-0.38
怠速时间比/%	-0.39	0.17	-0.34
加速时间比/%	0.35	0.12	0.42
减速时间比/%	0.24	0.08	0.57
速度平均偏差/(km·h-1)	0.42	-0.17	-0.35
平均加速度/(m·s-2)	0.08	0.54	-0.17
加速度标准偏差/(m·s-2)	0.26	0.53	0.04
平均减速度/(m·s-2)	-0.14	-0.51	0.26

 | Show Table

DownLoad: CSV

Table  4.  Component values of each principal component

序号	M1	M2	M3
1	41.07	-5.05	-7.38
2	39.76	-0.21	8.43
3	41.05	-6.39	-17.09
︙	︙	︙	︙
2 176	-25.36	15.03	-21.59
2 177	-27.81	16.74	-21.68
2 178	-23.69	13.82	-27.01
︙	︙	︙	︙
4 352	-10.32	12.57	-45.23
4 353	19.51	-0.46	-30.84
4 354	76.81	-21.45	-21.01

 | Show Table

DownLoad: CSV

2.2.2   基于AE的运动学片段特征降维

Figure  7.  Structure of AE network

DownLoad: Full-Size Img PowerPoint

Table  5.  Eigenvalues obtained after dimension reduction of AE

运动学片段序号	隐含层节点1数值	隐含层节点2数值	隐含层节点3数值
1	95.96	263.52	18.57
2	76.54	250.75	34.53
3	117.99	281.37	-2.03
︙	︙	︙	︙
2 176	149.62	203.93	-109.34
2 177	159.89	216.69	-116.97
2 178	156.32	210.12	-113.59
︙	︙	︙	︙
4 352	201.36	274.89	-111.06
4 353	132.52	274.89	-46.84
4 354	149.03	351.07	56.87

 | Show Table

DownLoad: CSV

2.3 基于K-means++算法的聚类分析及结果选择

Figure  8.  Flow of K-means algorithm

DownLoad: Full-Size Img PowerPoint

Figure  9.  Clustering results of two categories

DownLoad: Full-Size Img PowerPoint

Figure  10.  Clustering results of three categories

DownLoad: Full-Size Img PowerPoint

Figure  11.  Clustering results of four categories

DownLoad: Full-Size Img PowerPoint

3. 汽车行驶工况的确定与评估

3.1 基于最小性能值的汽车行驶工况构建

Figure  12.  Curves of vehicle driving cycles

DownLoad: Full-Size Img PowerPoint

Table  6.  Time of vehicle driving cycle

特征降维方法	汽车行驶工况时间/s
PCA	1 212
AE	1 286

 | Show Table

DownLoad: CSV

3.2 汽车行驶工况评估

Table  7.  Comparison of characteristic parameters data based on PCA

特征参数	所采集数据源	基于PCA的数据	数据相对误差绝对值/%
平均速度/(km·h-1)	23.36	23.18	0.75
平均行驶速度/(km·h-1)	36.22	34.22	5.50
怠速时间比/%	35.52	32.26	9.14
加速时间比/%	59.51	65.35	9.80
减速时间比/%	21.16	23.27	9.98
速度标准差/(km·h-1)	25.45	19.26	24.31
平均加速度/(m·s-2)	0.54	0.49	8.45
加速度标准差/(m·s-2)	2.51	2.35	6.17
平均减速度/(m·s-2)	-0.66	-0.71	7.73

 | Show Table

DownLoad: CSV

Table  8.  Comparison of characteristic parameters data based on AE

特征参数	所采集数据源	基于AE的数据	数据相对误差绝对值/%
平均速度/(km·h-1)	23.36	25.55	9.39
平均行驶速度/(km·h-1)	36.22	42.40	17.07
怠速时间比/%	35.51	39.73	11.91
加速时间比/%	59.51	55.91	6.06
减速时间比/%	21.16	21.54	1.82
速度标准差/(km·h-1)	25.45	23.35	8.23
平均加速度/(m·s-2)	0.54	0.65	21.38
加速度标准差/(m·s-2)	2.51	2.47	1.42
平均减速度/(m·s-2)	-0.66	-0.55	16.64

 | Show Table

DownLoad: CSV

4. 结语