Analysis of commercial truck drivers’ potentially dangerous driving behaviors based on 11-month digital tachograph data and multilevel modeling approach

Highlights

• An innovative way for extracting the information hiding in daily driving data.

• 40% of the truckers tend to drive in a substantially dangerous way.

• The explained variance proportion of different truckers vary distinctly.

• Dangerous truckers are less likely to overspeed on holidays or rainy days.

Abstract

This study analyzed the potentially dangerous driving behaviors of commercial truck drivers from both macro and micro perspectives. The analysis was based on digital tachograph data collected over an 11-month period and comprising 4373 trips made by 70 truck drivers. First, different types of truck drivers were identified using principal component analysis (PCA) and a density-based spatial clustering of applications with noise (DBSCAN) at the macro level. Then, a multilevel model was built to extract the variation properties of speeding behavior at the micro level. Results showed that 40% of the truck drivers tended to drive in a substantially dangerous way and the explained variance proportion of potentially extremely dangerous truck drivers (79.76%) was distinctly higher than that of other types of truck drivers (14.70%˜34.17%). This paper presents a systematic approach to extracting and examining information from a big data source of digital tachograph data. The derived findings make valuable contributions to the development of safety education programs, regulations, and proactive road safety countermeasures and management.

Introduction

Traffic safety is a critical issue in the transportation field. Traffic safety conditions are determined by drivers, vehicles, and driving environment. Previous research revealed that over 90% of traffic accidents were associated with unsafe driving behaviors (e.g., Petridou and Moustaki, 2001; Ellison et al., 2015; Atombo et al., 2016). Thus, an enhanced understanding of unsafe driving behaviors could provide meaningful contributions to road safety research.

Driving behavior plays an important role in driving risk analysis. However, it is difficult to measure risk in real-life situations (Eboli et al., 2017). Driving simulators are often used to investigate driving behaviors in various experimental environments (Pankok and Kaber, 2018). Some vehicle instrument technologies such as the Naturalistic Driving Study (NDS) (Guo and Hankey, 2009) and the DriveCam system (Hickman et al., 2010) have been applied to monitor driving behaviors and kinematic signatures on a large scale. The NDS programs such as the Second Strategic Highway Research Program (SHRP2), the 100-Car NDS, and Europe’s UDRIVE have provided valuable insights into accident causation and driving behaviors (Guo, 2019). Most existing analyses of dangerous driving behavior have relied on crash data or self-reported questionnaire surveys (Lord and Mannering, 2010; Ellison et al., 2015; Mannering and Bhat, 2014; Huang et al., 2017). To fully explore driving behavior in traffic accidents, it is important to understand driving styles. Driving styles, or habits, are defined as the way that individuals choose to drive and are analyzed over periods of years (Constantinescu et al., 2010). Traditional attributes of driving style include choices of driving speed, acceleration, deceleration or braking, threshold for overtaking, headway, and propensity to commit traffic violations (Murphey et al., 2009; Wu et al., 2016).

Among all types of vehicles, trucks are the largest contributor to traffic accidents, injuries, and fatalities owing to their high proportion among the roadway population, as well as their size, weight, and other unique characteristics (Zhu and Srinivasan, 2011; He et al., 2019). Compared to other vehicle types, the larger size and higher center of gravity of trucks result in longer braking distances and more severe consequences when involved in accidents. Moreover, truck crashes have higher economic impacts because of the damage to high-value cargo and travel delays caused by traffic accidents. Thus, research identifying influential factors in truck accidents would facilitate the development of countermeasures that could reduce the number and severity of accidents involving trucks.

Prior research has demonstrated that driving environment has a substantial influence on driver safety awareness (e.g., Kaber et al., 2012; Faure et al., 2016; Yan et al., 2017). Driving is a complex task that involves maintaining appropriate steering and speeds while accurately perceiving, identifying, and anticipating road elements such as road type or transit route, and other dynamic conditions including traffic flow, car following situation, and weather. Given these parameters, it is crucial to advance our understanding of how driving environment affects driving behavior.

The rapid rise and prevalence of mobile technologies have enabled the collection of a massive amount of passive data, e.g., big data. Effective analysis of big data provides new opportunities to advance our understanding of critical transportation problems such as road safety (Chen et al., 2016; Bao et al., 2019). Unlike small-scale data obtained via questionnaires or surveys, most big data are initially generated for other purposes, but have high potential value for research applications. Currently, this massive amount of second-by-second data have yet to be fully utilized in road safety research.

Approached with a microcosmic perspective, most road safety data are hierarchically organized to facilitate road safety research (Dupont et al., 2013). This makes multilevel models, or hierarchical linear models, technologically effective and efficient for figuring out the heterogeneity among the complicated data structures. Billot et al. (2009) investigated the influence of rain on driving behavior using a multilevel model. Huang and Abdel-Aty (2010) proposed a 5×ST-level (S: spatial; T: temporal) hierarchy to represent the general framework of multilevel data structures in traffic safety. This is a conceptual framework for assessing structured safety data. Several case studies using Bayesian hierarchical models were summarized to improve model fitting and predictive performance over traditional models. However, all the data utilized were sourced from crash data, either frequency or severity. The authors also noted concerns about the applicability and transferability of these multilevel, or hierarchical models.

In summary, the following research gaps were identified in existing research. First, because crash data is traditionally the primary or only data source, proactive approaches to effective traffic safety measures have been ignored. Second, the influence of unobserved heterogeneities omitted by multilevel or hierarchical models on traffic safety remain unknown. Third, in Japan, installing tachograph recorders has been mandatory for trucks since 1967. Prior to then, only analogue-type recorders were available¹ . Since the 2010s, the use of digital tachograph recorders has increased and effectively replaced analogue-type recorders. However, the accumulated big data have seldom been applied to road safety research in Japan.

To fill the abovementioned research gaps, the objectives of this study are twofold. The first objective is to identify different types of truck drivers in terms of driving performance at the macro level. The second objective is to excavate the variation properties of speeding behavior across different types of truck drivers by building a multilevel model at the micro level, where the effects of multi-dimensional unobserved heterogeneities are reflected. This study used digital tachograph data collected from a major Japanese freight transport company over an 11-month period in the year 2014. Driving performance indicators included speeding, driving duration, and jerky driving indicators, which were measured using the tachograph data and further exploited to indirectly capture potentially dangerous driving behaviors. This study differs from existing studies by simultaneously considering the following aspects:

• Adopting a real-world commercial truck-driving data and capturing potentially dangerous driving behaviors based on multiple indicators;

• Using a large-scale dataset that recorded truck drivers’ driving speed-related data at an interval of 0.5 s in tandem with GPS data collected at an interval of 60 s;

• Building a multilevel model that reflects the influences of multi-dimensional unobserved heterogeneities;

• Focusing on using the above large-scale driving data with detailed temporal and spatial information in the specific context of Japan.