Home Explore Help
Sign In
ardo
/
Obsidian
1
0
Fork 0
Files Issues 0 Pull Requests 0 Wiki
Tree: a2f8402268
Branches Tags
Obsidian
/
Zotero
/
001_artiklid
/
zhangsafetyassessmenttrucksbased2022.md

zhangsafetyassessmenttrucksbased2022.md 70 KB
History Raw

category: literaturenote citekey: zhangsafetyassessmenttrucksbased2022 title: Safety assessment of trucks based on GPS and in-vehicle monitoring data authors: "Zhang, Xuxin; Wang, Xuesong; Bao, Yanli; Zhu, Xiaohui" year: 2022 date: 2022-00-00 2022 doi: 10.1016/j.aap.2022.106619 publication: "Accident Analysis & Prevention" url: "https://doi.org/10.1016/j.aap.2022.106619" zotero_key: IXMRVKN4 zotero_storage: T428FS87 collections: transpordi_planeerimine folder: 001_artiklid firstAuthor: "Zhang, Xuxin"

status: converted

Contents lists available at ScienceDirect

Accident Analysis and Prevention

journal homepage: www.elsevier.com/locate/aap

Safety assessment of trucks based on GPS and in-vehicle monitoring data

Xuxin Zhang a,b , Xuesong Wang a,b,c,* , Yanli Bao a,b , Xiaohui Zhu d

  • a College of Transportation Engineering, Tongji University, China
  • b The Key Laboratory of Road and Traffic Engineering, Ministry of Education, Shanghai, China
  • c National Engineering Laboratory for Integrated Optimization of Road Traffic and Safety Analysis Technologies, China
  • d China Pacific Property Insurance Co., Ltd, China

ARTICLE INFO

Risk assessment

Keywords: Usage-based Insurance (UBI) Manage-how-you-drive (MHYD) Telematics Driver monitoring system Advanced driver assistance systems (ADAS)

ABSTRACT

Increasingly, drivers are choosing to buy usage-based automobile insurance (UBI). Manage-how-you-drive (MHYD) insurance, a new type of UBI, incorporates active safety management to monitor driver behavior and issue warnings as needed. While researchers have introduced telematics data into automobile insurance pricing, the specific effect of in-vehicle active safety management on driver risk assessment has been neglected, especially for truck drivers, whose crashes have more serious consequences. This study uses telematics and in-vehicle monitoring features to examine the key factors underlying large commercial truck crashes, and quantifies the effect of these factors on crash risk. Data from 2,185 trucks in Shanghai, China, were collected for a total of 105,786 trips and 465,555 in-vehicle warnings to investigate three types of factors affecting risk: travel characteristics, driving behavior, and in-vehicle warnings. A zero-inflated Poisson (ZIP) regression model was built, and a ZIP model without the warning variables as well as a basic Poisson model with warnings were considered for comparison. It was found that the ZIP model considering in-vehicle warning information performed significantly better than the other models. The standardized regression coefficient method was used to identify the most important variables. In-vehicle yawn and smoking warnings had significantly more association with the number of crashes than did the travel characteristics and driving behavior variables, though freeway distance traveled, average freeway speed, percentage of trips on sunny days, and percentage of trips at night also correlated significantly with crash risk. These results can provide a reference for UBI insurance professionals considering in-vehicle active safety management, as well as support freight companies in drafting appropriate working regulations.

1. Introduction

According to the data released by the American Highway Safety Insurance Association, a total of 4,119 people died in large commercial truck crashes in the United States in 2019 (IIHS-HLDI, 2021). Among them, 16% were truck occupants, 67% were occupants of cars and other passenger vehicles, and 15% were pedestrians, cyclists, or motorcyclists. These fatalities from large trucks crashes represent 1.7 deaths per 100 million vehicle miles traveled, which, as shown in Fig. 1, is significantly higher than that of passenger vehicles' (1.1), and indicates the consequences of large truck crashes are more serious and cause more deaths. Large trucks generally carry a substantial quantity of goods, and so have great inertia. Crashes involving these vehicles endanger life and property and bring huge economic losses to society. Truck safety can not be overemphasized, and the research supporting it is vital.

Traffic crashes can be attributed to human and vehicle factors as much as to road and traffic factors (AASHTO, 2010). Data collection methods for all these factors mainly include questionnaires, driving simulator experiments, and naturalistic driving studies (NDS) (Wu and Xu, 2018). Among these methods, an NDS can best reflect the driver's behavior and the surrounding environment during daily driving, as the data is collected by installing a data acquisition system into each driver's vehicle. Because the NDS typically needs to install cameras, sensors and other equipment in vehicles, however, the cost of NDS is relatively high. Another type of database similar to an NDS database has become popular in China. In China, heavy-duty trucks, tractor semitrailers, and touring coaches, or buses, are required by the government to install and use satellite positioning devices with travel recording functions such as the Global Positioning System (GPS) (State Council of the People's Republic of China, 2014). Because substantial travel and driving behavior

* Corresponding authors at: College of Transportation Engineering, Tongji University, China. E-mail address: wangxs@tongji.edu.cn (X. Wang).

information can be obtained through GPS, some insurance companies have achieved real-time monitoring of vehicles and obtained high-value driving data on these large vehicles. Through Internet of vehicles (IoV) technology, data can be collected and analysed by insurance providers to improve their products and services. However, those data have not yet been widely nor deeply mined, so a new opportunity is available to study the data to prevent crashes and improve traffic safety.

This same type of in-vehicle data has become the basis for a new type of insurance policy: usage-based insurance (UBI). UBI is based on the customer's driving behavior. Using data from IoV technology, smartphones, and on-board diagnostics (OBD), UBI integrates driving habits, driving skills, vehicle information, and surrounding environment data, thereby establishing a multi-dimensional model of driver, vehicle, road, and environment information for insurance pricing (Tselentis et al., 2017). In general, there are two key research questions in this area: 1) how driver's behavioral data affects driving risk, and 2) consequently, how driver behavior should affect UBI pricing (Bian et al., 2018). The first question is the concern of this study.

Driver risk assessment is the foundation of any driver risk management strategy. A common assumption is that all drivers can drive satisfactorily, but driving is a skill that is learned through experience. Unfortunately, drivers may also develop bad habits. Many researchers have carried out driving risk assessments for drivers of private, or passenger, cars (Segovia-Gonzalez et al., 2009; Hultkrantz and Lindberg, 2011; Lahrmann et al., 2012; Zhuo and Huang, 2019). These drivers tend to use the same car regularly, and their behavior patterns are relatively fixed and easy to capture. When freight companies carry out transport tasks, however, the vehicles and drivers are usually not fixed, and there may also be a shift work rule, that is, a work schedule that may involve driver rotation to complete long-distance transport tasks. In addition, freight companies often have a high turnover of commercial drivers, all of which may lead to a deviation from traditional risk assessment model results that use individual attributes of drivers (such as gender, age, and driving experience) as risk assessment indicators.

Most of the existing UBI-related studies have focused on pay-as-youdrive (PAYD) and pay-how-you-drive (PHYD) insurance types. PAYD, the first development of UBI, considers a driver's travel characteristics, such as travel distance and time, and road network type, in order to build a relationship with crash risk, while PHYD (UBI's second development) mainly uses the driver's driving behavior, such as speed, braking, acceleration, and turning behavior (Tselentis et al., 2017; Ayuso et al., 2019; P´erez-Marín and Guillen, 2019; Boucher and Turcotte, 2020). Much of the UBI research has focused on the abovementioned variables, but there has been little consideration in risk assessment modeling of driver-state conditions, such as phone use, smoking, and drowsiness, and of habitual vehicle operation status, such as insufficient headway. Manage-how-you-drive (MHYD) is the third stage of UBI development, and it does take into account driver-state conditions. It not only considers drivers' travel characteristics, such as distance traveled, and driving behavior features, such as speeding, but also uses an in-vehicle active safety system to monitor driver behavior (Arumugam and Bhargavi, 2019) and issue warnings as needed. An active safety system is a system of real-time alerts and suggestions to drivers for ensuring safety by monitoring their potentially hazardous state (e.g., drowsiness and distraction) and their vehicles' running status (e.g., headway). China Pacific Property Insurance Company's "Taihaobao" project, for example, aims to reduce the loss from commercial vehicle crashes by installing active safety systems in the trucks it insures. GPS, driver monitoring systems (DMS), and advanced driver assistance systems (ADAS) collect and analyze real-time data of the driver, vehicle, road, environment, and other factors that may affect driving safety, and provide timely warnings to drivers when it is necessary.

Little driver risk assessment research has been conducted on the influence of in-vehicle warning systems, and even less has addressed the systems with regard to truck drivers. In order to determine whether active safety monitoring information can improve the accuracy of risk assessment models, this study investigated the effect on crash risk of commercial truck drivers' travel characteristics, driving behavior, and in-vehicle warning data such as yawning. Naturalistic data from 2,185 trucks was acquired by a Shanghai insurance company's GPS and invehicle monitoring devices, and a zero-inflated Poisson (ZIP) regression risk assessment model was proposed.

2. Literature review

2.1. Data collection methods

Traditional approaches used by insurance companies to estimate customers' driving risk are based on their personal information, such as age, gender, driving experience, income, and personality type (Litman, 2005; Guo and Fang, 2013; Weidner et al., 2017). Some studies of insurance company practice also indicate that annual driving distance is a

Fig. 1. Fatalities in crashes involving large trucks and passenger vehicles per 100 million vehicle miles traveled, 2010–2019 (FMCSA, 2021).

factor (White, 1976), but it is often provided by the driver, which may decrease its reliability (Winlaw et al., 2019).

The development of technology such as IoV has introduced many types of equipment to detect the state of a traveling vehicle and its surrounding environment. GPS is a widely used device to collect travel behavior, such as distance traveled, road type, trip duration, and time and location data (Ayuso et al., 2014; Ayuso et al., 2016). OBD and the telematics box, also known as the "black box," can be installed in vehicles to measure various aspects of how, when, and where it is driven; this information can then be sent to the insurance company (Troncoso et al., 2011; Bian et al., 2018; Derbel and Landry, 2018). The smartphone is the latest addition to the existing methods. By the smartphone's combination of in-built sensors (such as accelerometers, gyroscopes, and magnetometers) and GPS system, it works either as an independent device or can be linked to the in-vehicle information system to transmit a variety of information from the vehicle (Arumugam and Bhargavi, 2019).

2.2. Driver risk assessment related variables

Most driver risk assessment studies consider crash records or insurance claims as the dependent variable, and investigate the relationships among driver demographics, travel behavior, driving behavior, and crash risk (Ayuso et al., 2014; Ma et al., 2018; P´erez-Marín and Guillen, 2019). Guillen et al. (2019) confirmed the commonplace assumption that novice drivers were more likely to be involved in crashes than experienced drivers. Ayuso et al. (2014) compared novice and experienced drivers in their risk of crash involvement, and found that female novice drivers with longer nighttime driving experience were involved in their first traffic crashes earlier than those with less nighttime driving experience, and male novice drivers were involved in their first crashes sooner than female novice drivers. However, a driver crash risk study considering driving route familiarity by Jin et al. (2018) showed that, based on crash reports, gender and age had no significant correlation with crashes.

In their research on travel characteristics, Paefgen et al. (2014) found a nonlinear relationship between vehicle distance traveled and crash risk through studying data from in to vehicle data recorders; the nonlinear relationship shows that if an insurance company uses only traveled distance to predict risk and assumes they are linearly correlated, errors in drivers' risk classifications may occur. Research has also demonstrated that the greater the proportion of urban driving, the sooner the driver's first crash is likely to occur and the larger the expected number of crash claims will be (Ayuso et al., 2014; Guillen et al., 2019). Similarly, Guillen et al. (2019) found that the greater the percentage of kilometers driven per year, the more at-fault crashes in which the driver will be involved.

Winlaw et al. (2019) concluded that speeding is the most important driver behavior link to crash risk, confirming the finding of Jun et al. (2011) that drivers with histories of crash involvement tended to drive at higher speeds. Kovaceva et al. (2020) demonstrated that harsh brake frequency was correlated to an aggressive driving style, which has been associated with crashes. Bian et al. (2018) found that the execution of sudden turns, emergency braking, and rapid acceleration are useful variables for driver risk classification and can significantly improve a model's classification accuracy. Bunn et al. (2005) demonstrated that fatigue and distraction increase the risk of fatal commercial motor vehicle collision, and Sacks and Nelson (1994) found smokers were 1.5 times more likely to be involved in a traffic crash than non-smokers. Most studies have found that insufficient headway is correlated with crash occurrence (Taieb-Maimon and Shinar, 2001; Hyun et al., 2019). Nodine et al. (2017) showed that a short headway (less than 1.0 s) significantly increases the risk of rear-end crash, and Starnes (2006) found that following too closely was one of truck driver errors that caused heavy truck at-fault crashes.

2.3. Driver risk assessment models

Driver risk assessment models evaluate risk level by using the characteristics of driving behavior and crash frequency. Statistical modeling and data mining are two typical methods that have been used in previous driver risk assessment studies. Data mining refers to searching by an algorithm for information hidden in a large quantity of data. Using a bagging-based Naive Bayes classifier, Bian et al. (2018) classified drivers into five risk groups based on their number of traffic accident involvements. Guo and Fang (2013) adopted the K-means cluster method to classify drivers into three risk groups based on their crash and nearcrash rates (often used as a crash is a low probability event). Yuksel and Atmaca (2021) used fuzzy logic to calculate and output a driver risk score between 0 and 100 by considering the number of detected risky driving behaviors.

Several statistical models have been utilized in previous driver risk assessment studies, such as the generalized linear model, logistic regression, and zero-inflated Poisson model (Ma et al., 2018; Guillen et al., 2019; Winlaw et al., 2019; Sagar et al., 2020). Ma et al. (2018) examined the relationship between a variety of risk factors and the probability of vehicle crashes through generalized linear models, and Sagar et al. (2020) identified high-risk commercial vehicle drivers by binary logistic regression based on their sociodemographic characteristics. Two concerns have led to the development of zero-inflated count data models for insurance companies' assessment of driver risk. As mentioned, crashes are low probability events; additionally, crash records in insurance databases may be inaccurate because claims are often under-reported due to deductible agreements and no-claim discount policies. Both issues, crashes' low probability and their underreporting, can lead to excessive zero-crash driver data in insurance databases. Since traditional methods of Poisson and negative binomial distributions for model fitting may not be adequate due to the presence of excess zeros, zero-inflated count data models have been proposed (Lambert, 1992). Guillen et al. (2019) used a zero-inflated Poisson model to assess driver risk by analyzing telematics data on risk exposure and driving habits. They demonstrated that drivers may experience a learning effect, and suggested that insurance premiums should not be strictly proportional to driving distance.

3. Data preparation

3.1. Data acquisition

The data used in this study were collected by China Pacific Property Insurance Company. GPS and in-vehicle monitoring devices, that is, DMS and ADAS with digital video recorders (DVRs), were installed in the trucks of the insured freight companies, who were mainly involved in local cargo and express transportation. The collected vehicle trajectory and warning information data were transmitted to the insurance company cloud platform for storage and management through 4G wireless communications technology.

The data collected by GPS includes the vehicle's travel time, speed, and location (longitude and latitude). Based on video recognition technology, DMS detects a driver's behavior and status in real time, such as smoking, yawning, and using a mobile phone (see Fig. 2(a)–(c)). Also using video recognition, ADAS monitors the vehicle's forward running state, such as traveling too close to the forward vehicle (see Fig. 2(d)), crossing lane divider lines, and near crashes. Relying on these installed in-vehicle devices, the cloud platform receives satellite telemetry in real time and obtains important information such as driver status, environmental status, vehicle position, driving direction, and speed.

3.2. Database description

A total of 105,786 trips, made by 2,185 trucks, occurring on urban roads and freeways in the Shanghai area from January 1, 2019, to

Fig. 2. Examples of in-vehicle monitoring information.

December 31, 2019, were extracted for analysis from the Taihaobao database. Urban roads are defined in this study as roads with a speed limit of less than 80 km/h, and freeways are defined as roads with a maximum speed limit of 120 km/h and a minimum speed limit of 60 km/h. It should be noted, however, that the maximum speed limit for trucks is 100 km/h. A total of 465,555 early warning messages during 2019 were collected, including 213,689 smoking warnings, 109,352 yawn warnings, 8,288 phone use warnings, and 134,226 insufficient headway warnings.

Crashes were obtained from the police crash reporting system as the dependent variable. Five types of independent variables were derived from the GPS data, in-vehicle monitoring devices, and insurance company and police crash reporting systems, as shown in Table 1. The first type is travel characteristics, including exposure variables (total distance traveled and distance traveled on freeways), travel time variables (weekend and nighttime), and weather variables (sunny and rainy). The mean values for total and freeway distance traveled were 2,444.960 km and 1459.620 km, respectively.

The second type of variable is driving behavior, including average speed and standard deviation of speed on both urban roads and freeways. The average speeds on urban roads and freeways are 31.437 km/h and 64.420 km/h, respectively, and the standard deviations of speed are 21.940 km/h and 17.050 km/h, respectively.

The third type is the in-vehicle monitoring information of driver state, including smoking, yawning, and phone use. Calculated as the warning frequency per 100 km for each vehicle, the average values for smoking, yawning and phone use warnings were 4.454, 1.639, and 0.184 times per 100 km, respectively. It should be noted that the number of warnings are not necessarily equivalent to number of incidences of a

behavior. The monitoring system issues warnings every 30 s, so in a single incidence of phone use, for example, several warnings may be issued if the phone call is lengthy. The fourth type of variable is also acquired from in to vehicle monitoring: the forward collision warning system monitors the vehicle's forward running status, and warns the driver when the vehicle is too close to the forward vehicle as defined by the monitoring system. The mean value of the insufficient headway warning variable, for which the trigger threshold was *<*1.0 s, is 3.087 times per 100 km.

The last type of variable consists of the vehicle characteristics. License plate numbers and crash information were provided for all trucks in the database by the monitoring insurance company and police crash reporting system, respectively. The mean crash count was 0.029 and the standard deviation of crash count was 0.189.

4. Methodology

This study developed three statistical models for assessing key factors affecting commercial truck crashes: a Poisson regression model with exposure, and two zero-inflated Poisson models with exposure, one that included the warning variables and another that did not. The number of crashes for commercial vehicle is assumed to be Poisson distributed with mean λi.

4.1. Poisson regression with exposure

The probability of a vehicle i having yi crashes is given by Eq. (1) (Washington et al., 2003):

Table 1 Summary of the variables used in the models.

Variable Description Mean SD
Travel characteristics
Distance
traveled		 Total distance traveled (km) 2444.950 5726.392
Freeway
distance
traveled		 Total distance traveled on freeways
(km)		 1459.620 3309.037
Weekends Percentage of trips on weekends
(defined as the time between 00:00 h
on Saturday to 24:00 h on Sunday)
(%)		 28.523 26.763
Nighttime Percentage of trips driven at night
(22:00–06:00) (%)		 48.279 33.841
Sunny Percentage of trips driven on sunny
days (%)		 10.668 19.293
Overcast Percentage of trips driven on cloudy
or overcast days (%)		 58.278 29.588
Rainy Percentage of trips driven on rainy
days (%)		 28.059 27.495
Driving behavior
Speed_avg on
urban roads		 Average travel speed on urban roads
(km/h)		 31.437 15.711
Speed_avg on
freeways		 Average travel speed on freeways
(km/h)		 64.420 6.629
Speed_sd on
urban roads		 Standard deviation of speed on urban
roads (km/h)		 21.940 6.760
Speed_sd on
freeways		 Standard deviation of speed on
freeways (km/h)		 17.050 4.700
In-vehicle monitoring: Driver state
Smoking Smoking warning frequency per 100
km		 4.454 9.091
Yawn Yawn warning frequency per 100 km 1.639 3.910
Phone use Phone use warning frequency per 100
km		 0.184 1.125
In-vehicle monitoring: Forward collision warning
Too close Insufficient headway warning
frequency per 100 km		 3.087 8.004
Vehicle characteristics
Vehicle ID License plate number – –
Crashes Crash frequency for each vehicle in
2019		 0.029 0.189

$$P(Y_i = y_i) = \frac{e^{-\lambda_i} \lambda_i^{y_i}}{y_i!}, i = 0, 1, 2, 3, \dots, n$$ (1)

where P Yi = yi ) is the probability of the vehicle i having yi crashes during travel within the Shanghai area in 2019, and *λi*is the Poisson parameter for vehicle i. The relationship between explanatory variables and the expected number of crashes can be written as Eq. (2).

$$E(Y_i) = \lambda_i = exp(\beta X_i)$$ (2)

where Xi is a vector of explanatory variables and β is a vector of estimable parameters.

The role of exposure to risk in terms of distance driven must be considered in the model, so Di indicates total distance traveled among the trips. Then the model can be written as Eq. (3):

$$\log(\lambda_i) = \ln(D_i) + \beta X_i \tag{3}$$

4.2. Zero-inflated Poisson model with exposure

As mentioned in Section 2.3, crashes' low probability and their underreporting can lead to excessive zero-crash driver data in insurance databases. It is assumed that there are two processes producing the two types of zero crash counts: one type of zero is produced from drivers with fewer vehicle kilometers traveled (VKT) within the Shanghai area, their lower risk exposure leads to a crash count of zero; the other type of zero is produced by drivers who may have greater VKT but excellent driving behavior, also leading to a crash count of zero.

Zero-inflated Poisson (ZIP) regression, originally proposed by Lambert, is widely used to model crash count data that has an excess of zeros (Lambert, 1992). The ZIP model assumes there are two underlying processes producing the above zeros: one process is governed by logit distribution that predicts excess zeros; the other process is governed by a Poisson distribution that generates crash count, including zeros. Together, the two processes can be written as (Washington et al., 2003; NCSS, 2021)

$$Pr(Y_i = j) = \begin{cases} \pi_i + (1 - \pi_i)exp(-\lambda_i), ifj = 0\ \frac{(1 - \pi_i)exp(-\lambda_i)\lambda_i'}{j!}, ifj > 0 \end{cases}$$

$$(4)$$

where *πi*is the probability of an excess of zeros for the i th observation. The logistic component *ηi*includes an exposure variable Di and a set of K regressor variables (the Z's); the *γk*are coefficients of the regressors. It can be written as

$$\pi{i} = \frac{\eta{i}}{1 + \eta{i}} = \frac{exp\left(ln(D{i}) + \gamma{0} + \sum{k=1}^{K} \gamma{k} Z{ik}\right)}{1 + exp\left(ln(D{i}) + \gamma{0} + \sum{k=1}^{K} \gamma{k} Z_{ik}\right)}$$ (5)

The Poisson component *λi*can include an exposure Di and a set of M regressor variables (the X's). It can be written as

$$\lambda_i = exp\left(ln(D_i) + \beta0 + \sum{m=1}^{M} \betam X{im}\right)$$ (6)

5. Results and discussion

5.1. Model estimation results

The model estimation results for the basic Poisson regression model and the zero-inflated Poisson (ZIP) models, with and without warning variables, are presented in Table 2; highly correlated independent variables were excluded before modeling, and the ZIP model retained only significant variables. Some of the variables did not follow a normal distribution, so it was necessary to transform the data. As power transformations have been shown to be more effective in reducing the Akaike information criterion (AIC) than logarithmic transformations (Guillen et al., 2021b), several power transformations (pt) were investigated for nighttime a , yawn b , and smoking b variables, where a = 0.1 to 2.0 with increases in 0.05 increments, and b = 0.05 to 0.5 with 0.05 increases. Finally, the combination of power transformations that reduced the AICs the most (thereby increasing goodness of fit) were nighttime 1.5, smoking 0.45, and yawn 0.45.

Because there was significant interaction between yawn_pt and log of freeway distance traveled, yawn_pt × log (freeway distance traveled) was introduced as a fixed effect covariate and was used as an explanatory variable to increase model accuracy.

AIC, log-likelihood and McFadden's pseudo R-squared were used to measure fitness. As can be seen in the results shown in Table 2, both ZIP models had much smaller AIC values and larger log-likelihood values than the basic Poisson model, thus demonstrating that the ZIP models were more accurate. Table 2 also compares the ZIP model results with and without the warning variables of smoking and yawn (mobile phone use and insufficient headway do not appear in the table as their results were not significant). The pseudo R-squared of the ZIP model with warnings is 0.226. Because pseudo R-squared values between 0.2 and 0.4 indicate excellent fit (Hensher and Stopher, 1979), the 0.226 (*>*0.2) value represents the only good performance of the three models. Since a lower AIC value also indicates a better fit, a clear improvement in the ZIP model's accuracy is evident when all the significant regressors are considered, that is, the ZIP with warnings has the lowest AIC value.

Table 2 Estimation results of the three models.

Variable ZIP with warnings ZIP without
warnings		 Basic Poisson
Estimate p
value		 Estimate p
value		 Estimate p
value
Poisson part
Intercept − 0.833 0.543 − 1.151 0.446 − 7.205 0.000
Log (freeway
distance
traveled)		 − 0.324 0.006 − 0.841 0.000 − 0.197 0.025
Sunny 0.031 0.000 0.057 0.000 0.030 0.000
Nighttime_pt − 0.003 0.000 − 0.002 0.012 − 0.001 0.003
Average speed
on freeway		 − 0.097 0.000 − 0.055 0.059 − 0.028 0.035
Smoking_pt − 0.717 0.001 0.079 0.470
Yawn_pt 2.864 0.004 0.775 0.190
Yawn_pt × Log
(freeway
distance
traveled)		 − 0.287 0.021 − 0.198 0.023
Offset: Log
(distance
traveled)		 1.000 1.000 1.000
Zero-inflated part
Intercept 0.211 0.954 0.275 0.931
Log (freeway
distance
traveled)		 − 0.592 0.048 − 2.518 0.000
Sunny 0.011 0.546 0.059 0.033
Nighttime_pt − 0.025 0.003 − 0.004 0.191
Average speed
on freeway		 − 0.010 0.865 0.122 0.056
Smoking_pt − 9.713 0.003
Yawn_pt 21.755 0.003
Yawn_pt × Log
(freeway
distance
traveled)		 − 1.184 0.011
Offset: Log
(distance
traveled)		 1.000 1.000
AIC 471.861 487.074 534.140
Log-likelihood − 219.931 –233.500 − 259.069
Pseudo R
squared		 0.226 0.178 0.123

Therefore, the rest of the analysis will refer to the ZIP model with warning variables unless otherwise stated.

In general, a positive value for the estimated coefficient of a variable in the zero-inflated part of the model (bottom part of Table 2) indicates the probability will increase that the crash count is in a zero state. Meanwhile, if the estimated coefficient of a parameter in the Poisson part (top part of Table 2) is positive, then the predicted mean value of the crash count will increase (Chen et al., 2016; Guillen et al., 2019).

The coefficients for the log (freeway distance traveled), nighttime_pt, and smoking_pt in the logit model (the bottom part of Table 2, which predicts zero inflation) are negative, that is, the probability of observing excess zeros increases with decreasing log of freeway distance traveled, nighttime_pt, and smoking_pt. In other words, the decrease of the abovementioned variables would increase the probability of truck's crash is a certain zero.

Total distance traveled on freeways, percentage of trips driven on sunny days, and percentage of trips driven at nighttime comprise the three travel characteristics influencing crash count according to the Poisson part of the model. Trips driven on sunny days has a positive coefficient, which shows that more crashes may happen when more of the driving is on sunny days. The negative coefficient of nighttime_pt indicates that nighttime trips may have the benefit of fewer crashes. Log of freeway distance traveled is significant with a negative coefficient; however, there is an interaction with yawn_pt, so it is not accurate to say that the likelihood of a crash increases as the percentage of trips driven on freeways decreases. The effect of this interaction will be discussed in Section 5.3.1.

Yawning is one of the in-vehicle monitoring variables. The negative value (-0.287) for the effect of the interaction term implies that yawn_pt has a negative correlation with expected crash count, but for yawning as well as freeway distance traveled, because there is an interaction between yawn_pt and log of freeway distance traveled, it is not so clear how yawn_pt affects crash count; this relationship will be discussed in further detail in Section 5.3.3. The other monitoring variable, smoking_pt, also has a negative sign, which indicates that the likelihood of a crash decreases as smoking warning frequency per 100 km increases.

Of the driving behavior variables, average speed on freeways is the only significant factor related to crash count. The negative sign of the parameter shows that the lower the speed while driving on freeways, the more the predicted mean value of the crash count will increase. Since most studies show speed being positively correlated with crash count (Dickerson et al., 2000; Wang et al., 2018), this result is unexpected. Reduced congestion, however, is likely to improve safety (Golob et al., 2004) and suggests one possible explanation: higher average speed on a freeway indicates less traffic congestion, which could reduce the probability of a crash.

5.2. Variables' importance ranking

As discussed in the previous section and presented in Table 2, the ZIP model demonstrated that the three travel characteristics, one driver behavior, and two in-vehicle monitoring variables significantly affect crash count. Because the model cannot reveal which variables are the most important, however, their relative importance was calculated by the standardized regression coefficient (SRC) method (Siegel, 2016). The SRC compares the strength of the effect of each individual independent variable to the dependent variable, that is, the degree of influence of each of the variables on crash risk. The SRC can be computed as

$$SRC = \widehat{\beta}j \frac{S{X_j}}{S_Y} \tag{7}$$

where β ̂ *j*is the regression coefficient of the independent variable Xj in the ZIP model, and SXj and SY are the standard deviations of the independent variable Xj and the dependent variable Y, respectively. The absolute values of the independent variables' SRCs can thus be determined, and can then be compared to show the relative importance of the variables, with a higher absolute value of SRC indicating higher importance. Fig. 3, which shows the SRCs of the ZIP model, illustrates which variables are contributing the most to the regression equation.

As shown in Fig. 3, yawn warning frequency per 100 km was found to be the most influential factor, followed by the interaction term of yawn and freeway distance traveled, and smoking warning frequency per 100 km. Although the percentage of nighttime driving, average speed on freeway, freeway distance traveled, and percentage of trips driven on sunny days had significant effect on the number of crashes, their influence was smaller than these in-vehicle monitoring variables.

5.3. Safety effects of variables

5.3.1. Travel characteristics

Holding other variables in the ZIP model constant, when yawn_pt = 0 (that is, when the power transformation of yawn frequency per 100 km is 0), the estimated coefficient of − 0.324 for log of freeway distance traveled (Table 2, Poisson part) means that the crash rate per kilometer decreases by a factor of 0.723 (=exp (-0.324)) with each unit increase in log (freeway distance traveled). Similarly, when the mean value of yawn_pt is 0.700, an average 1.77 (=1/exp(-0.324–0.287*yawn_pt))

Fig. 3. Ranking results of variables' importance.

increase in crash rate per kilometer is produced by a one-unit decrease in freeway distance traveled. This result suggests that less freeway distance traveled may indicate an increasing proportion of driving on urban roads, which has been found in many studies to be significantly associated with higher expected crash risk (Ayuso et al., 2014; Guillen et al., 2019). Compared to urban roads, freeways often have better driving surface and access control, wider lanes, and fewer interruptions, which may reduce the probability of crash.

If a truck were to increase its percentage of trips driven on sunny days by 1%, the difference in the crash rate per kilometer would be expected to increase by 3.1% (=exp (0.031)-1), while holding the other variables in the model constant. This result is consistent with the findings of Brijs et al. (2008) that both the number of hours of sunshine, and the sun dazzle, or glare effect, significantly increase the number of crashes.

When holding other variables constant, if a truck were to increase its percentage of nighttime trips by 1 unit, the difference in the crash rate per kilometer would be expected to decrease by 0.30% (=exp (− 0.003)- 1). Although many studies show that nighttime travel increases the risk of crashes (Ayuso et al., 2014; Jin et al., 2018), Keall et al. (2005) found that roads with illumination at night are less risky than they are during the day. The benefits of night driving include less traffic on the road, avoidance of rush hour traffic, and ability to maintain a steady speed. The difference in results may also be because most studies investigate private, or passenger cars. Compared to drivers of passenger cars, commercial truck drivers are professional drivers who usually have more driving experience and, because of their experience and their companies' regulations, are stricter about their behavior such as avoiding the consumption of alcohol before driving. These differences may explain why more nighttime trips may reduce the risk of crash for truck drivers.

5.3.2. Driving behavior

Holding all other variables constant, the ZIP model (Table 2) used in this study predicted a 10.2% increase in crash rate per kilometer for a 1 km/h decrease in average speed during freeway driving (because 1/exp (-0.097) = 1.102). Despite prior research findings that the risk of a crash increases with an increase in speed (Ayuso et al., 2014; Wang et al., 2018), larger differences in speed between vehicles, that is, relative speeds, have also been demonstrated to correlate with higher crash rates (Aarts and van Schagen, 2006). Generally, freight vehicles are large and heavy, and their speed on freeways is usually lower than the average speed of other traffic. When a truck further reduces its speed, it produces an excessive speed difference in relation to the surrounding vehicles, which could result in more crashes.

5.3.3. In-vehicle monitoring features

Because there is an interaction effect between yawn_pt and log of freeway distance traveled, if a truck driver were to increase yawn_pt by one unit, the expected crash rate per kilometer would increase by a factor of exp (2.864–0.287 * log (freeway distance traveled)), holding all other variables in the model constant. The mean and maximum values of log of freeway distance traveled were calculated at (=5.576) and (=10.494), respectively. When log of freeway distance traveled = 5.576, a one-unit increase in yawn_pt will produce a 3.539 per-unit increase in crash rate per kilometer. In contrast, when log (freeway distance traveled) = 10.494, a one-unit increase in yawn_pt will increase the crash rate per kilometer by a per-unit increase of just 0.863. As Wierwille and Ellsworth (1994) showed, repeated yawning, along with eye closures and minimal body and eye movement, are behaviors related to drowsiness. Yawning correlates significantly with driver drowsiness; that is, a greater frequency of yawning indicates a higher level of drowsiness. Drowsiness leads to slower reaction times and impaired vigilance, attention, mental processing, judgment, and decision making (Goel et al., 2009), all of which contribute to negative outcomes such as poor performance and traffic crash. Klauer et al. (2006) demonstrated that drowsy drivers have a crash and near-crash risk that is 4 to 6 times higher than that of alert drivers.

The estimated coefficient of − 0.717 for smoking_pt (Table 2 Poisson part) means that the crash rate per kilometer decreases by a factor of 0.488 with each additional smoking warning per 100 km. Smoking is known to be a common measure adopted by drivers to maintain alertness while driving (H¨ akkanen ¨ and Summala, 2000; Tzamalouka et al., 2005); Tzamalouka et al. (2005) has also demonstrated that a greater number of cigarettes smoked per day correlates with a reduced risk of falling asleep at the wheel.

The results in Table 2 met expectations that drivers who yawn more may be involved in more crashes due to drowsiness, whereas smoking may help drivers control that drowsiness, thus reducing the probability of a crash. Nevertheless, the results shown in Fig. 2 demonstrate that drowsiness has a greater impact on crash rate than smoking, and Sacks and Nelson (1994) found smokers were more likely to be involved in crashes than non-smokers. Although smoking can reduce the incidence of drowsy driving to a certain extent, it should be acknowledged that the long-term effect on health is not likely worth this advantage.

It was surprising to find that neither cell phone use nor insufficient headway warnings were found significant. One possible explanation is that the road traffic safety law in China does not permit driving a motor vehicle while dialing and answering a cell phone, thus leading to a relatively lower occurrence of phone use which could result in the variable showing no significant effect on crash rates. Commercial truck

drivers who are regulated and monitored by their employers may be even more compliant. The professional status of truck drivers may also be the reason that the insufficient headway warnings variable was not significant. Previous studies have found that insufficient headway was correlated with crash occurrence (Taieb-Maimon and Shinar, 2001; Hyun et al., 2019), including Nodine et al. (2017) who used the same 1.0 s threshold as the current study. The drivers in this study usually have more driving experience than the average driver. They may be less likely to follow too closely, and may be more able to control their vehicles to avoid crashes when they are in that situation. On the other hand, Starnes (2006) found that, even for truck drivers, crashes and short headways were correlated. Another possible explanation is that the in-vehicle headway warnings improved driver behavior, which may also be the case for this study's low incidence of phone use warnings.

6. Conclusion

The primary aim of this study was to determine if active safety warning information could improve driver risk assessment, particularly for defining risk factors related to commercial truck crashes. Daily driving characteristics were collected through in-vehicle monitoring, and a zero-inflated Poisson (ZIP) regression model was built to associate crash count with trucks' travel characteristics, driver behaviors, and invehicle warnings. A ZIP model without warnings and a basic Poisson regression model were built for comparison. Total distance traveled was used as the offset. The ZIP model with warnings fit the data better than the other models, so was used for analysis. The results showed that active safety warning information did indeed improve driver risk assessment. Specifically, 1) including in-vehicle warning information significantly improved the ZIP model; 2) while freeway distance traveled, average speed on freeways, percentage of trips driven on sunny days, and percentage of trips driven at night also correlated significantly with crash risk, yawning and smoking were significantly more associated with crashes than the other variables.

The first practical application of this study is to provide a reference for modeling driver risk assessment for the manage-how-you-drive (MHYD) insurance strategy. When assessing truck drivers' risk levels, insurance companies should take into account the warning information collected by in-vehicle monitoring devices in addition to the more commonly used travel characteristics and driving behaviors. Second, there appears to be potential to improve truck drivers' safety through wider scale implementation of strategies suggested by this study's results. Freight company safety managers should consider adjusting transportation routes and working hours according to the risk patterns found in this study, for example, by reducing the number of routes on urban road when possible, avoiding peak hours, and increasing the frequency of nighttime travel. Managers should also interact with drivers to review any warning information collected by in-vehicle monitoring devices. Drivers who yawn frequently, for example, may perform better with changes to their rest periods.

Naturally, this study has some limitations. First, the data used in this study do not include the operating characteristics of the vehicles before the installation of active safety devices, so no comparison was possible. Whether the frequency of in-vehicle warnings reflects a lack of safety awareness among the drivers or promotes the improvement of driver safety needs further study. Second, the personal characteristics of drivers, such as gender, age, and driving experience, might also influence crash risk, but these data were not available in this study, and may produce variation in results. Future research should establish models that incorporate these characteristics, especially given the high turnover common among drivers in freight companies. Third, environmental characteristics such as traffic flow and road alignment should be investigated as they may effect driver behavior, both in direct driver response and by adding psychological stress, thereby further affecting on crash risk. Last but not least, this study only considered crashes because claims are based on crashes. While ZIP model was used to account for the excess zeros in the dataset, as many researchers have noted (Guillen et al., 2020; Guillen et al., 2021a), due to crash counts are generally very low in datasets, it will be advantageous to take near-miss, or near-crash, events into consideration in future study.

CRediT authorship contribution statement

Xuxin Zhang: Conceptualization, Methodology, Data curation, Formal analysis, Writing – original draft, Writing – review & editing. Xuesong Wang: Conceptualization, Supervision, Funding acquisition, Writing – review & editing. Yanli Bao: Data curation, Formal analysis. Xiaohui Zhu: Data curation, Resources.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

This study was jointly sponsored by the National Key R&D Program of China (2018YFB0105202) and the Outstanding Clinical Discipline Project of Shanghai Pudong (Grant No.: PWYgy2018-10). The authors are grateful to Barbara Rau Kyle and Marilyn R. P. Morgan for their helpful editing and thoughtful feedback.

References

Traffic Safety Administration.