Copyright © 2009 KSAE 1229−9138/2009/046−01 International Journal of Automotive Technology, Vol 10, No 3, pp 265−276 (2009) DOI 10.1007/s12239−009−0031−3 EVALUATION OF LOW-TEMPERATURE DIESEL COMBUSTION REGIMES WITH n-HEPTANE FUEL IN A CONSTANT-VOLUME CHAMBER U B AZIMOV , K S KIM , D S JEONG and Y G LEE 1) 1)* 2) 2) Department of Mechanical Design Engineering, Chonnam National University, Cheonnam 550-749, Korea Korea Institute of Machinery and Materials, Eco-Machinery Research Division, 171 Jang-dong, Yuseong-gu, Daejeon 305-343, Korea 1) 2) (Received 31 March 2008; Revised 24 December 2008) ABSTRACT–The concept of Low Temperature Combustion (LTC) has been advancing rapidly because it may reduce emissions of NOx and soot simultaneously Various LTC regimes that yield specific emissions have been investigated by a great number of experiments To accelerate the evaluation of the spray combustion characteristics of LTC, to identify the soot formation threshold in LTC, and to implement the LTC concept in real diesel engines, LTC is modeled and simulated However, since the physics of LTC is rather complex, it has been a challenge to precisely compute LTC regimes by applying the available diesel combustion models and considering all spatial and temporal characteristics as well as local properties of LTC In this paper, LTC regimes in a constant-volume chamber with n-Heptane fuel were simulated using the ECFM3Z model implemented in a commercial STAR-CD code The simulations were performed for different ambient gas O2 concentrations, ambient gas temperatures and injection pressures The simulation results showed very good agreement with available experimental data, including similar trends in autoignition and flame evolution In the selected range of ambient temperatures and O2 concentrations, soot and NOx emissions were simultaneously reduced KEY WORDS : Low-temperature combustion, ECFM3Z model, STAR-CD, Autoignition, Soot, NOx INTRODUCTION rapidly at low cost However, the chemical kinetics involved in this concept not allow the use of classical diesel auto-ignition and combustion models based on oversimplified representations of combustion chemistry In these classical models, the time in the reactive zone is usually considered much smaller than the diffusion time of fuel and air towards the flame region This work analyzes the effect of various parameters such as the ambient gas O concentration, ambient gas temperature, and fuel injection pressure on the evolution of diesel flames and emission formation in low-temperature combustion regimes The results are obtained by simulating LTC conditions with the n-Heptane fuel and ECMF3Z model used by the STAR-CD code First, the features of various combustion models are compared with respect to LTC Then, the simulation results are presented in comparison with available experimental data International regulations ratified in recent years have imposed more stringent limits on pollutant emissions and fuel consumption in internal combustion engines To comply with these regulations and reduce diesel NOx and soot emissions, new combustion concepts and technologies are being developed aggressively (Workshop 2006; Kimura et al., 1999; Kawamoto et al., 2004; Pickett and Siebers, 2004a) As one technology, homogeneous charge compression ignition (HCCI) and conventional diesel-based Low-Temperature Combustion (LTC) concepts show great potential in reducing NOx and soot emissions simultaneously The LTC concept is a better candidate because it allows easier auto-ignition control and it can be applied to conventional diesel engines with minimal design modifications However, the differences in chemistry and combustion between this concept and conventional diesel combustion must be investigated to determine their effects on spray combustion characteristics as well as emissions (Beatrice et al., 2007) LTC processes are investigated by computer modeling and simulation, which provide better understanding of the combustion process of new combustion concepts Different low-temperature combustion regimes can be evaluated 2 ANALYSIS FORMULATION To mitigate the formation of NOx, diesel combustion must occur at low temperatures (Yu and Shahed, 1981), but low combustion temperature can generally lead to soot formation The soot, however, can be avoided by initiating combustion at an equivalence ratio below and flame temperature under 1800K (Kamimoto and Bae, 1988; Akihama et al., 2001; Kitamura et al., 2003) Diesel diffusion flames can *Corresponding author e-mail: sngkim@chonnam.ac.kr 265 266 U B AZIMOV, K S KIM, D S JEONG and Y G LEE have complete combustion at temperatures in the range of 1500~1600 K, where NOx formation is very low (Pickett, 2005) Therefore, there is a trend towards the development of low-temperature combustion strategies for diesel engines The initial premixed burn of classical diesel combustion is an example of this type of low-temperature combustion, and if the mixture is lean enough, soot will not form during the low-temperature combustion reaction Low-temperature combustion in diesel engines consists of fuel injection in which the fuel is allowed to vaporize and mix with the ambient gas before combustion occurs A high level of Exhaust Gas Recirculation (EGR) is usually used to reduce the combustion temperature, and heat release is controlled by the chemical reaction kinetics of the mixture (Aceves and Flowers, 2004) This introduces new variables due to the factors that are not present in traditional diffusion-burn diesel combustion, where combustion starts in a cetane number-based time delay after the start of the fuel injection As shown in Figure 1, after the fuel injection, fuel evaporation occurs as the hot air is entrained into the fuel jet and mixes During the fuel evaporation, chemistry becomes active and entrainment continues until ignition occurs Once ignition occurs, it is assumed that no more air mixes into the core of the fuel jet because the oxygen is consumed in the outer layers of the jet Consequently, nonsooting and low NOx combustion is realized at equivalence ratios below 2.0 and flame temperatures less than 1800 K In LTC, with the increase of EGR, the auto-ignition delay period is increased and fuel-air premixing is improved Although liquid fuel penetrates much further into the chamber, the higher energy released from premixed reactions contributes to the intense evaporation of liquid fuel (Higgins 2000; Idicheria and Pickett, 2005) Since the fuel and air are very well mixed, the amount of oxygen around the fuel molecules is sufficient to prevent pyrolysis and soot formation throughout the jet cross-section Inaccurate predictions of alternative diesel combustion regimes often originate from the fact that many numerical approaches use the Magnussen eddy break-up concept (Magnussen and Hjertager, 1976), in which the complexity of the combustion chemical reactions is eliminated with a fast chemistry limit As the diesel combustion progresses, there is a full spectrum of important chemical and turbulence time scales ranging between the limits of slow, distributed chemistry and turbulent, mixing-controlled, fast chemistry Both mixing and chemical time scales are crucial to the diesel modeling In LTC, introducing finiterate chemistry is important for accurately predicting pollutant formation To improve the accuracy of predictions while modeling diesel spray combustion, unified combustion models have been built to account for all types combustion modes simultaneously Abraham (1985) suggested replacing the controlling time scale in the Magnussen model by the slowest time scale of the mixing time and the chemical time Kong et al et al et al Figure Conceptual scheme of low-temperature combustion (1995) proposed an extended characteristic-time model based on Abraham (1985) which accounts for chemical and turbulence time scales simultaneously This model was combined with the Shell ignition model to simulate the overall combustion processes in a diesel engine In this combined model, the initiation of combustion relies on laminar chemistry, and turbulence starts to have an influence on combustion only after combustion events have already been observed, similar to the Magnussen model Even if premixed and non-premixed combustions are taken into account in this model, the non-mixture of the species within a computational cell is only represented by the mixing time-scale, which does not account for the mixing history Consequently, the transition between chemically controlled and mixing-controlled combustion needs to be monitored by an empirical function This model does not account for flame propagation combustion These modifications improve the eddy break-up model only to a minor extent because only the time scales from the limiting ends of the diesel combustion time scale spectrum are included A two-zone flamelet combustion model was developed by Chen (2000) Based on the classical flamelet model by Peters 1986, this model (in which the reactions occur in wrinkled turbulent flames that can be considered as a collection of laminar flamelets) suggests that each cell et al et al EVALUATION OF LOW-TEMPERATURE DIESEL COMBUSTION REGIMES WITH n-HEPTANE FUEL is divided by the flame front into two zones: the unburned zone and the burned zone The unburned zone consists of air, fuel vapor, and residual gases, and the burned zone contains combustion products The unburned zone is further divided into two regions: the segregated region and the fully mixed region The combustion is decoupled as two sequential events: mixing and burning However, the use of flamelet models requires a separation between chemistry and turbulence time scales, even though the chemical reactions are still in the fast chemistry limit Another approach called EPFM (Eulerian Particle Flamelet Model) was developed by Hasse (2000) This model is an extension of the RIF (Representative Interactive Flamelet) concept by Pitsch (1995) for a single representative flamelet The EPFM model can be used to solve multiple unsteady flamelets in a flamelet code, and simultaneously solve the Navier-Stokes equations in a CFD code The CFD code solves the three-dimensional equations for the flow, turbulence, enthalpy, mixture fraction and its variance The flamelet parameters are calculated from the turbulence and mixture field, and are then passed to the flamelet code The EPFM model assumes the introduction of different marker particles, which are associated with different flamelet histories depending on the path a particle takes through the turbulent flow field This model allows the representation of autoignition and diffusion flames, but represents the mixing and combustion, which are very local phenomena, in an averaged way since the flamelets are based on the averaged properties over all parts of the domain This model does not account for flame propagation and computational cost increases with the number of flamelets involved The CMC (Conditional Moment Closure) approach, which was independently developed by Klimenko(1990, 1993), and Bilger (1993), is considered to one of the more advanced models for turbulent reacting flows In this approach, the mixture fraction Z is not represented solely by its mean value and fluctuations like in most models; instead, the Z-space is discretized, and combustion and mixing processes are solved for different values of Z The main concept behind CMC is to find how the reactive scalars (e.g temperature, species mass fractions) depend on the mixture fraction CMC can be applied for infinitely fast and finite rate chemistry The CMC model calculates conditional moments at a fixed location within the flow field using modeled transport equations for the conditional moments of the reactive scalars with no assumptions on the small-scale structure of the reaction zones or on the relative timescales of chemistry and turbulence This approach is very promising but computational cost still remains unacceptable for industrial applications More recently, a new, flame surface density approach was proposed to model auto-ignition and diffusion flames It considers the dimension of mixing, represented by the mean mixture fraction and its fluctuation, and the dimension of progress of reaction, represented by the mean proet al et al 267 gress variable and its fluctuation (Pope, 1988; Candel and Poinsot, 1990; Bray , 2005) This approach is based on the Coherent Flame Model (CFM), which describes the rate of fuel consumption per unit volume as the product of the flame surface density (i.e the flame surface per unit volume) and the local flame speed at which it consumes the mixture This approach supposes that the chemical reaction of fuel oxidation occurs in a very thin layer This layer separates the burned and unburned gases and propagates toward the fresh mixture of fuel, oxygen and dilutant The CFM model was extended first to the ECFM model, and was specifically adapted to model combustion with perfectly or partially mixed mixtures and to simulate the combustion processes in direct injection spark ignition engines (Colin , 2003) Then ECFM was adapted to account for unmixed or diffusion combustion, and the three-zone description of the mixing state was added This new combustion model, called ECFM3Z (3-Zones Extended Coherent Flame Model), can therefore be seen as a simplified CMCtype model, which discretizes the mixture fraction space by only three points Therefore, this model was selected to be the most appropriate for simulating LTC regimes because it can reflect the real physics of LTC, it relies on flamelet libraries, and it is less computationally demanding than the CMC and EPFM models et al et al ECFM3Z MODEL CONCEPT The ECFM3Z model was briefly presented by Beard (2003) and was described in detail by Colin and Benkenida (2004) This model describes the unburned/burned gas zones based on the flame surface density equation In order to account for diffusion flames and mixing processes, each computational cell is split into three mixing zones: a pure or unmixed fuel zone, a pure air plus EGR zone or unmixed air and EGR zone, and a mixed zone, containing fuel, air and EGR This structure can account for the three main combustion modes: auto-ignition, flame propagation and diffusion flame as encountered in LTC It is based on two dimensions: the mixing state description and the reaction progress description The mixing state description is represented by the Probability Density Function (PDF) of the mixture fraction P( Z )=a δ ( Z )+bδ ( Z – ZM )+cδ ( Z – ) (1) M where Z is the average value of the mixture fraction in the mixed zone The first δ function corresponds to the unmixed air region, the second one to the mixed region and the third one to the unmixed fuel region In this structure, space is discretized by only three points The mixing model can reflect the transference of unmixed fuel and unmixed air into the mixed region The reaction progress description is represented by the progress variable et al Z Y˜ uFu m -u =1− c˜ =1− m Y˜ TFu (2) 268 U B AZIMOV, K S KIM, D S JEONG and Y G LEE where Y˜ uFu is the mass fraction of the fuel present in the unburned gases, and Y˜ TFu is the mass fraction of fuel before the onset of combustion (fuel tracer) Y˜ TFu is constant in space and time for perfectly mixed charges In practical applications Y˜ TFu varies in space and time because of the imperfect mixing of the charge In addition, a transport equation is solved to obtain the Favre average mass densities of the chemical species of the fuel, O2, N2, NO, CO2, CO, H2, H2O, O, H, N, OH and of the soot inside the computational cell containing the three mixing zones A detailed description and specific features of the ECFM3Z model are given in Colin and Benkenida, 2004; Colin , 2005; Reveille , 2006; Knop and Jay, 2006; Priesching , 2007; Shi , 2007 et al et al et al et al SIMULATION PROCEDURE The simulation was conducted using the STAR-CD commercial CFD code in a three-dimensional computation grid The ECFM3Z model with appropriate adjustments was incorporated into STAR-CD The computational grid assumes a cylinder-shaped constant volume chamber of 80 mm in diameter and 80 mm in length The discretization of space (number of cells) and time (time steps) are set after the Courant number (STAR-CD Methodology, 2006) In addition, the complete spray combustion duration was adjusted to match that of the experiment The mesh resolution was set to achieve good agreement between the simulation results and experimental results for the penetration of nonreacting and reacting fuel jets The fuel was injected with spray characteristics adjusted according to the spray characteristics assumed in the experiments In the spray model, atomization proceeded according to the Reitz-Diwakar model and the fuel droplets were formed according to the Reitz-Diwakar breakup model This atomization model assumed that the liquid emerges from the nozzle as a jet, waves form on the jet’s surface, and then the waves are amplified and the liquid is eventually broken up into droplets by aerodynamic forces caused by the high relative velocity between the liquid and the gas (Reitz, 1987) To apply this model, a semi-cone angle must be known and given as part of the input data Based on this angle, the initial droplet velocity is determined This angle was determined from experiments performed using the same common-rail spray characteristics and ambient gas conditions as those mentioned in this paper (Jeong, 2003) The autoignition in the present simulation was controlled by the double-delay autoignition model This autoignition model was developed to consider the effect of cool flames, which are characterized by a weak increase in temperature after an initial delay, followed by a slowing of the reaction rates until the second delay After this second delay, the reaction rate increases rapidly, and the main autoignition takes over This model makes use of pre-computed tables containing the results of complex chemistry calculations of the autoignition of n-heptane (Curran , 1998; Subramanian, 2007) The tables give values for the two delays and these delays are functions of pressure, temperature, equivalence ratio and EGR For emission simulation, the 3-step Zeldovich model and ERC model were used for NOx and soot emission calculations, respectively The simulation conditions are listed in Tables and The ambient gas temperature, ambient gas content, ambient gas pressure, fuel injection pressure, injection duration and single-hole injector orifice parameters correspond to those of the experiment et al RESULTS AND DISCUSSION 5.1 Evaluation Approach for LTC Regimes The present paper numerically evaluates the LTC regime of DI diesel combustion Since LTC differs from conventional diesel combustion, it is necessary to use a model that is universally applicable to both conventional and alternative diesel combustion applications Soot and NOx emissions computed for particular conditions, and the entire combustion event was evaluated to understand the physics of the combustion as well as the relations among the operating parameters For this purpose, the parameter called “Combustion Factor” was introduced: Y˜ Fb (3) ψ =1− Y˜ TFu where, Y˜ Fb is the mass fraction of fuel in the burnt gases, and Y˜ TFu is the fuel tracer This parameter is considered as an indicator of the combustion mode (premixed vs diffusion) in the complex LTC process It is extremely difficult to differentiate the diesel LTC process into certain modes because premixed, partially premixed and diffusion modes occur simultaneously However, it might be possible to map the combustion event and see which combustion mode prevails and Table Simulation conditions Fuel Ambient gas temperature [K] Ambient gas pressure [MPa] Rail pressure [MPa] Injection duration [ms] Injected fuel mass Nozzle hole n-Heptane 820, 870, 920 90, 135 1.2 8.4 mg, 10.3 mg dn=0.163 mm, ln/dn=5.52 Table Ambient gas content Molecular percentage N2 CO2 O2 21.0 79.0 16.0 81.0 3.0 12.0 81.0 7.0 MW 28.84 29.12 29.6 EVALUATION OF LOW-TEMPERATURE DIESEL COMBUSTION REGIMES WITH n-HEPTANE FUEL how it would change with the change of a certain operating parameter The combustion factor varies from to If this parameter approaches 1, the combustion is considered premixed, and if it approaches 0, the combustion is considered as diffusion flame mode To validate this approach, the original ECFM3Z model was invoked (Colin and Benkenida, 2004) As mentioned earlier, the ECFM3Z model consists of zones that allow computation of all combustion modes, autoignition, premixed flame and diffusion flame In the ECFM3Z model, the transport equations are solved to obtain the Favre average mass densities of chemical species as well as the fuel In the mixed zone, the fuel is divided into two parts: the fuel present in the fresh gases and the fuel present in the burned gases This division is necessary because the fuel in the fresh gases will be consumed by autoignition and the premixed flame, while the fuel in the burned zone will be consumed by the diffusion flame During the combustion event, if there is any fuel in the burned gases, this fuel will be consumed and post-oxidized by the source of the Magnussen EBU model or diffusion combustion model Therefore, it would be possible to evaluate the extent of diffusion combustion within the entire combustion event and to estimate the magnitude of soot formation, accepting the fact that soot formation can be avoided during the premixed mode To evaluate the entire range of the combustion regimes, the combustion factor was normalized using the expression below to obtain results for all the conditions under the same scale: ψ – ψ minψ n= 1–ψ (4) 5.2 Auto-ignition and Flame Development As mentioned earlier, the double-delay autoignition model was used within the ECFM3Z model to simulate the autoignitions of various LTC regimes Figure shows the autoignition delay mapping for all conditions mentioned in this paper This figure shows that the first autoignition delay moderately changes from that of the main autoignition Also, it is seen that with the decrease of O2 concentration or the increase of equivalence ratio and decrease of the ambient gas temperature, the autoignition delay slightly increases According to the low-temperature reaction mechanism, the first autoignition of hydrocarbon fuel is largely associated with the decomposition of the ketohydroperoxide species at temperatures between 800 and 850 K, and the end of the first autoignition occurs when the temperature reaches NTC zones The start of the first autoignition is determined by the time needed for the air/ fuel mixture to reach the decomposition temperature The figure shows that the first autoignition delay periods at 920 K and 870 K are almost similar, because the ambient gas temperatures are high enough to immediately initiate the decomposition of the mixture However, the first autoignition delay at 820 K is noticeably different, probably due to longer time required for the air/fuel mixture to reach 269 Figure Autoignition delay mapping the decomposition temperature In addition, probably at 820 K, there was no significant decomposition of fuel during evaporation, but some portion of fuel did decompose during evaporation for the high temperature case This assumption is in agreement with other presented work (Curran , 1998; Wang and Rutland, 2005) After the first autoignition, there is a period of very slight temperature increase due to “cool flame” chemistry This period is an ignition delay between the initial fuel decomposition and very rapid temperature rise The time interval between the first autoignition and the main autoignition is much greater at 820 K for similar O2 conditions compared to those of the other two cases This difference is due to both the retarded first autoignition and the further retarded reaction progress with the decrease of O2 concentration In the ECFM3Z model, defining the occurrences of the first and main autoignitions is straightforward because their computed values are automatically stored in the postprocessing file Figure compares spray combustion simulation results for various O2 concentrations and ambient gas temperatures with experimental data The results presented in Figure indicate good agreement with data from the experiments in terms of spray and combustion development The liquid fuel pattern, as well as the spatial distribution of the flame, matches well the pattern obtained from experiments Figure does not provide any information on the start of autoignition, but only depicts the comparison in spray flame development between simulation and experiment Nevertheless, based on Figures and 3, the effect of variation of charge composition showing the longer autoignition delay periods and decreased flame temperatures was evident for the diluted charge at lower O2 concentration et al 5.3 Flame Temperature The ambient gas temperature has a small influence on flame temperature and NOx formation but has great effect on the fuel/air equivalence ratio The ambient gas temper- 270 U B AZIMOV, K S KIM, D S JEONG and Y G LEE Figure Spray combustion development at P -90 MPa inj ature can only have an effect on autoignition initiation because of the reaction of fuel with oxygen in the hightemperature environment And then, as the flame propagates, the combustion is controlled by the O2 concentration and injection pressure Figure 4(a), (b), (c) shows a similar pattern in the temperature evolution for various combustion regimes with a change in ambient gas temperature Although the autoignition delay becomes longer with decreasing ambient gas temperature, the maximum values of the flame temperature at conditions with similar O2 content are nearly the same Note that the flame temperature is slightly increased by an increase in injection pressure, especially during the initial stage of flame propagation and temperature increase This may be explained by better evaporation and mixing at higher injection pressures (Gill , 2005) As the fuel is well-mixed and distributed, higher energy is released, and flame temperature is increased et al 5.4 NOx The three-step Zeldovich mechanism was used to compute the NOx emissions, which are generally believed to depend on only the flame temperature Figure 5(a), (b), (c) shows that with the decrease of O2 concentration, NOx emissions are gradually reduced, especially for the conditions of O212%, which corresponds to about EGR-60%, where NOx is reduced to almost zero These results are in agreement with experimental and numerical data, indicating that NOx formation can be avoided at reduced flame temperatures with decreased oxygen concentration in the ambient gas (Heywood, 1989; Abd-Alla, 2002; Egnell, 2000; Wagner et al., 2003; Alriksson and Denbratt, 2006) Also, the effect of higher injection pressure on NOx formation, as on the flame temperature discussed previously, is indicated A similar trend of increased NOx formation with higher injection pressures was shown by Henein , 2006 et al 5.5 Equivalence Ratio Akihama (2001) have shown that in addition to the notable decrease in NOx emissions, soot formation can also be avoided by producing combustion at flame temperatures less than 1800 K Kamimoto and Bae (1988) proposed that soot formation could be avoided by producing combustion at equivalence ratios below Non-sooting combustion has also been demonstrated at higher temperatures and higher equivalence ratios by entraining sufficient oxygen into the jet NOx formation occurs at high temperatures, but NOx is reduced to N2 under fuel-rich conditions and thus, NOx emissions are decreased with higher equivalence ratios In the case of low-temperature combustion, soot emission reduction appears to be related to an increase in ignition delay, which is due to the reduced O2 concentration, which provides more time for mixing before combustion and a possible decrease in the equivalence ratio of the igniting fuel-ambient gas mixture Figure shows equivalence ratio distribution at the time of main ignition for each combustion regime The deeper the fuel penetrates into the chamber, the more diluted and mixed it becomes as O2 concentration and ambient temperature decrease The equivalence ratio at the jet's leading part, the zone where ignition is supposed to occur and where the premixed burn occurs, is about The maximum equivalence ratio value corresponds to et al EVALUATION OF LOW-TEMPERATURE DIESEL COMBUSTION REGIMES WITH n-HEPTANE FUEL Figure Flame temperature variation with the change of O2 concentration, ambient gas temperature and injection pressure the core of the injected fuel and is about 6, and the stoichiometric values are along the jet periphery, which is in agreement with the experimental and numerical results obtained by other researchers for similar conditions (Idicheria and Pickett, 2007) 5.6 Soot The computed soot data were compared with experimental data of flame luminosity This direct comparison may not provide quantitative information about soot formation because soot is determined by directly solving the transport equations, but flame luminosity is experimentally related to soot concentration and local flame temperature It is accepted that the flame luminosity can be interpreted as a qualitative indicator of in-cylinder soot formation (Siebers , 2002; Mueller and Martin, 2002; Choi , 2004; Kim , 2007) et al et al et al 271 Figure NOx variation with the change of O2 concentration, ambient gas temperature and injection pressure In Figure 7(a) and (b), both the soot curves and the flame luminosity curves have a similar pattern After soot formation and flame luminosity reaches their peaks, soot oxidation will dominate and flame luminosity will decrease The soot oxidation process is clearly seen to be slower than the soot formation process in both (a) and (b) A similar trend was observed for other cases for different O2 concentrations and ambient gas temperatures, as in Figures 8(a), (b) and 9(a), (b) In general, higher flame luminosity peaks were found for lower injection pressures Higher luminosity peaks with higher injection pressure for the case of O2-21% and ambient temperature 920 K in Figure are probably due to a higher local equivalence ratio at higher ambient gas temperatures Although a higher injection pressure is believed to contribute to better mixing (Pickett and Siebers, 2004b), there is probably still not enough time to ensure sufficient 272 U B AZIMOV, K S KIM, D S JEONG and Y G LEE Figure Equivalence ratio at the time of ignition at Pinj-90 MPa Figure Comparison of calculated soot and measured flame luminosity at Tamb-870 K Figure Comparison of calculated soot and measured flame luminosity at Tamb-920 K mixing of the charge to decrease local equivalence ratio and prevent soot formation However, looking at cases with O2-16% and O2-12%, a similar trend can be observed where soot formation or flame luminosity peaks are always lower at higher injection pressures because of a longer ignition delay 5.7 Combustion Factor Based on the results mentioned above and applying the Figure Comparison of calculated soot and measured flame luminosity at Tamb-820 K combustion factor concept described earlier, the LTC regimes are evaluated to determine the best scenario of combustion in terms of soot and NOx emissions Note that EVALUATION OF LOW-TEMPERATURE DIESEL COMBUSTION REGIMES WITH n-HEPTANE FUEL Figure 10 Combustion factor at Tamb-920 K 273 Figure 13 Computed equivalence ratio and the fraction of fuel in burned gases at Tamb-870 K Figure 11 Combustion factor at Tamb-870 K Figure 14 Pressure rise comparison for different simulation conditions Figure 12 Combustion factor at Tamb-820 K the resulting curves of the combustion factor are a relative representation of one regime against the other within the framework of the conditions considered in this paper, solely mapping from the worst-case and the best-case scenario According to a number of experimental and numerical research results, simultaneous soot and NOx reduction can be achieved in premixed combustion with EGR and fuellean mixtures (Kamimoto and Bae, 1988; Dec, 1997; Kimura , 1999; Akihama , 2001) Figures 10, 11 and 12 show the results for combustion progress At the instant of ignition, combustion starts as premixed mode To confirm the assumptions stated above regarding the et al et al Figure 15 Apparent heat release rate for different simulation conditions combustion factor, the quantity of fuel present in the burned gases and the values of the equivalence ratio at the instances when the soot formation is at its maximum were determined For brevity, only the case with ambient temperature of 870 K is presented in Figure 13 As mentioned earlier, the fuel in the burned gases is consumed by the diffusion flame This fact can be utilized to determine the 390 H YANG et al appropriate for the mode used at low vehicle speeds of the dual mode, because these systems have good fuel economies at low vehicle speeds The high efficiencies of outputs and at high vehicle speeds provide merit for their use in dual-mode powertrains CONCLUSION Figure 15 Fuel economies at constant vehicle speeds The optimal planetary gear ratios, the maximum fuel economy, and acceleration performance for the output split systems acting in the single-mode powertrain are shown in Table For comparison, there is the reference power split system, which is the input split system used in the Prius This input split system is rated as the most superior power split system for the single-mode powertrain The X sign means that the input split system cannot follow the velocity profile of the FTP72 cycle or accelerate up to the vehicle speed of 100 km/h Outputs 2, 3, and cannot follow the FTP72 cycle and moreover, have bad acceleration performance Outputs 1, 5, and have better system performance than the other systems, but when compared with the reference input split system, they have no merits of fuel economy and acceleration performance Therefore, it can be concluded that output split systems are inappropriate for the single-mode hybrid powertrain 4.4 Fuel Economy of Output split Systems at Constant Vehicle Speeds The optimal planetary gear ratios and the maximum fuel economies at constant vehicle speeds are shown in Table Fuel economies are tested at the constant vehicle speeds of 20, 40, 60, 80, 100, and 120 km/h These results are represented in Figure 13 Outputs and have the worst fuel economies at all vehicle speeds, and therefore they can be used for neither the single- nor dual-mode powertrain Outputs and are There are six possible output split systems that use one planetary gear set and two motors In this study, we analyzed the characteristics, advantages, and disadvantages of these systems and evaluated the fuel economies of these systems in the FTP72 driving cycle and acceleration performance Output split systems provide no advantages when used for the single-mode powertrain, because the performances of all output split systems are worse than that of the input split system, known for its overall superiority However, outputs 1, 2, 5, and show good characteristics at low or high vehicle speeds, and this finding can be verified by the fuel economies at constant vehicle speeds If these systems are used to construct dual-mode powertrains in combination with other power-split systems, these dualmode powertrains are capable of good performance These subjects will be studied in future works REFERENCES Benford, H L and Leising, M B (1981) The lever analogy a new tool in transmission analysis SAE Paper No 810102 Conlon, B (2005) Comparative analysis of single and combined hybrid electrically variable transmission operating modes SAE Paper No 2005-01-1162 Fussner, D and Singh, Y (2002) Development of single stage input coupled split power transmission arrangements and their characteristics SAE Paper No 2002-010590 Jung, K H., Kim, H E and Kim, J K (2005) Performance and transmission efficiency analysis of 2-mode hydro mechanical transmission Trans Korean Society of Automotive Engineers 13, 1, 90−98 Yang, H., Cho, H., Kim, N., Lim, W and Cha, S (2007) Analysis of planetary gear hybrid powertrain system part 1: Input split system Int J Automotive Technology 8, 6, 771−780 Copyright © 2009 KSAE 1229−9138/2009/046−15 International Journal of Automotive Technology, Vol 10, No 3, pp 391−404 (2009) DOI 10.1007/s12239−009−0045−x EFFECT OF DROWSINESS ON DRIVING PERFORMANCE VARIABLES OF COMMERCIAL VEHICLE DRIVERS A MORTAZAVI , A ESKANDARIAN and R A SAYED 1) 2)* 3) California Center for Innovative Transportation (CCIT), University of California, Berkeley, USA Center for Intelligent Systems Research, School of Engineering and Applied Science, The George Washington University, 20101 Academic Way, Ashburn, VA 20147, USA Mechanical Engineering Deparment, NWFP-UET, Peshavar Pakistan 1) 2) 3) (Received 14 March 2008; Revised 25 October 2008) ABSTRACT−Driver drowsiness is a major safety concern, especially among commercial vehicle drivers, and is responsible for thousands of accidents and numerous fatalities every year The design of a drowsiness detection system is based on identifying suitable driver-related and/or vehicle-related variables that are correlated to the driver’s level of drowsiness Among different candidates, vehicle control variables seem to be more promising since they are unobtrusive, easy to implement, and cost effective This paper focuses on in-depth analysis of different driver-vehicle control variables, e.g., steering angle, lane keeping, etc that are correlated with the level of drowsiness The goal is to find relationships and to characterize the effect of a driver’s drowsiness on measurable vehicle or driving variables and set up a framework for developing a drowsiness detection system Several commercial drivers were tested in a simulated environment and different variables were recorded This study shows that drowsiness has a major impact on lane keeping and steering control behavior The correlation of the number and type of accidents with the level of drowsiness was also examined Significant patterns in lateral position variations and steering corrections were observed, and two phases of drowsiness-related degradation in steering control were identified The two steering degradation phases examined are suitable features for use in drowsiness detection systems KEY WORDS : Drowsiness, Commercial drivers, Drowsiness detection, Steering control, Lane keeping, Vehicle safety INTRODUCTION is fatally injured (NTSB, 1990) Other researchers have also addressed the relationship between impairment due to sleepiness exacerbated by alcohol and driving incidents, i.e lane drifting (Horne and Reyner, 1995) Based on NHTSA General Estimates System (GES) statistics (Knipling and Wierwille, 1994), although the frequency of drowsiness-related crashes involving passenger vehicles is greater than that of combination-unit trucks, the number of involvements per vehicle life cycle for trucks is about four times greater because of their very high exposure level, as well as the greater likelihood of night driving Moreover, truck crashes are more severe in terms of injury and property damage (Wang and Knipling, 1994a) In general, the factors that influence driver fatigue/drowsiness include daytime sleepiness, amount of night sleep, difficult work schedules, hours of work, driver’s age and experience, cumulative sleep debt, presence of a sleep disorder, and time of day (Hamelin 1987; Wylie , 1996; Gander and James, 1998; McCartt , 2000) Truck drivers are even more at risk because of the long hours of continuous wakefulness, irregular driving schedules, night shifts, sleep disruption or fragmented sleep because of split off-duty time, and other factors associated with the The National Highway Traffic Safety Administration (NHTSA) recognizes driver drowsiness as one of the major causes of single and multiple car accidents in the United States (Knipling and Wang, 1994) Every year, 100,000 crashes involve driver fatigue, resulting in more than 40,000 injuries The Fatality Analysis Reporting System (FARS) reports 1,544 fatalities each year due to driver drowsiness-related accidents More than 3% of drowsiness related crashes (a total of 3,300 crashes and 84 fatalities) involved drivers of combination-unit trucks Based on police reports, drowsiness accounts for 1% to 3% of all U.S motor vehicle crashes (Lyznicki, Doege , 1998) Police report studies are likely to substantially underestimate this proportion as the drivers involved in fatigue accidents not admit their state of drowsiness, and police may not investigate fatigue issues because of a lack of time and knowledge According to other researchers (Wang and Knipling, 1994b), fatigue has been involved in 15% of single vehicle fatal truck crashes Drowsiness is the most frequent contributor to crashes in which a truck driver et al * Corresponding author et al et al e-mail: eska@gwu.edu 391 392 A MORTAZAVI, A ESKANDARIAN and R A SAYED profession Long-haul truck drivers have been shown to get less sleep than is required to be alert (Mitler , 1997) In a separate study, it was found that drivers tend to continue driving, even though they are aware of the risk factors related to drowsiness (Nordbakke and Sagberg, 2006) A driver’s drowsiness can be evaluated using two types of phenomena: psychophysiological and driving performance measures Psychophysiological assessments include the measurement of brain waves by electroencephalogram (EEG) (Akerstedt and Gillberg 1990; Huang , 1996) and eye activity, including eyelid closures and blink rate (Skipper , 1984; Dingus , 1985; Ueno , 1994; Ogawa and Shimotani, 1997) PERCLOS (PERcent eyelid CLOSure) is one of the most widely accepted metrics for the detection and evaluation of drowsiness (Dinges , 1998; Grace , 1998) Unlike methods based on sensing driver psychophysiological phenomena, techniques based on monitoring driver controls or driving performance variables during driving have the advantage of being nonintrusive to the drivers There are numerous valuable studies on the effect of drivers’ inattention and drowsiness on driving performance (Dingus , 1997; Hanowski , 1999), especially for truck drivers (Klauer , 2003; Barr , 2005) These researchers used field data or naturalistic data for their analysis Because of the risks of involving drowsy drivers in dangerous scenarios, some researchers perform the experiments in a simulated environment (Chung , 2004; Gillberg and Kecklund, 1996; Pilutti and Ulsoy, 1997; Sayed and Eskandarian, 2001; Sayed , 2001a, 2001b) Previous experiments have shown that some driving performance variables degrade with drowsiness and, therefore, can be a sign of a driver’s level of drowsiness Other studies in the literature also indicate that variables related to vehicle lane position (Skipper , 1984; Dingus , 1985; Pilutti and Ulsoy, 1997) correlate well with drowsiness indicators (i.e PERCLOS) The correlation of driver steering activity with drowsiness has also been examined previously (Elling and Sherman, 1994; Dingus , 1985) Several researchers have also studied relationships between drowsiness and drivers’ braking, acceleration or speeding strategies (Dingus , 1985; Wierwille , 1994) The Center for Intelligent Systems Research (CISR) performed a series of experiments on drivers in a simulated environment to develop a drowsiness detection algorithm, which is based on an Artificial Neural Network (ANN) that learns a driver’s steering characteristics (Sayed and Eskandarian, 2001; Sayed , 2001a; 2001b) Their study showed that steering activity is a good candidate, among other variables, for evaluating a driver’s drowsiness Those experiments were conducted only on regular drivers in a passenger car simulator Because of the difference between the dynamics of trucks as compared to cars, and the professional skill level of commercial drivers, the effect of drowsiness on truck driving performance was not clear from that study et al et al et al et al et al et al et al et al et al et al The current study has the following objectives: • Explore driving performance variables for commercial truck drivers in a simulated environment to determine which variables are affected by drowsiness, • Gain a deeper understanding of driver control behavior at the onset of the drowsiness, • Conduct a thorough analysis of the driving performance variables affected by drowsiness to identify how the variables, specifically steering wheel control behavior, are degraded Experiments were conducted in a truck-driving simulator The level of drowsiness was measured by using PERCLOS and a subjective rating system (described below) through video observation The statistical analysis was used to distinguish the driving performance variables that were significantly affected by drowsiness and sleep deprivation This paper examines the effect of drowsiness on different variables, specifically steering wheel angle, and shows that drowsiness degrades steering control in two phases The authors aim to use the results of this paper to develop a new robust drowsiness detection system for commercial vehicles et al et al et al et al et al et al et al et al et al EXPERIMENT 2.1 Study Population A total of 13 commercial drivers from a variety of truck driving professions, ranging in age from 23 to 55 years (mean age=41 years with standard deviation of 9.1), including two females, completed the study The solicitation of professional truck drivers was based on their availability in the area and no other demographic variables like age and gender This study population does not address or consider any gender- or age-related issues Furthermore, the study of 13 subjects may not be statistically significant to extrapolate the findings to the general commercial truck driver population However, there are some important general technical characteristics associated with the drowsiness state of the drivers regarding their steering control actions These characteristics can be observed, even in the limited number of subjects used herein These discoveries are elaborated in the results and conclusion sections However, the results should be interpreted with caution because of the statistical limitations All the participants were required to have a valid commercial driver’s license and were screened for any susceptibility to simulator sickness 2.2 Driving Simulator Laboratory Experiments were performed at the CISR (Center for intelligent Systems and Research) Truck Driving Simulator Laboratory (TDSL) The TDSL is a fully interactive, high resolution, fixed base driving simulator, which has a fullsize truck cabin This truck-driving simulator was developed in partnership with the Modeling, Simulation and Driv- EFFECT OF DROWSINESS ON DRIVING PERFORMANCE VARIABLES OF COMMERCIAL VEHICLE DRIVERS ing Simulators (MSIS) research unit of the French National Institute for Transport and Safety Research (INRETS) Figure illustrates the various elements of this driving simulator TDSL includes: • Seven networked PCs used for real-time computation of detailed virtual environments, vehicle dynamics, traffic and sound effects • Sophisticated vehicle dynamics and traffic models with intelligent autonomous vehicle models • A five channel immersive projection systems with a 135 degree forward field-of view and two operational sideview mirrors • A multi-channel sound system • A subwoofer bolted to the truck cabin to generate realistic vibrations during driving • A fully instrumented truck cabin with all original accessories, controls, pedals, and an operational instrument panel • An electronic gearbox system with eight forward gear ratios, representative of a real manual transmission • An advanced steering feedback system using a fully dedicated computing unit, a Pulse-Width-Modulation (PWM) controller and a DC motor to generate a realistic truck steering feel • A state-of-the-art head-mounted eye closure measuring system • Four infrared digital cameras recording the driver’s face, hands, and feet motions and projected driving scene during simulation Additional description and verification of the simulator is provided in Eskandarian (2006) et al Figure Schematic diagram of the CISR truck simulator lab 393 2.3 Driving Scenario This experiment required a long and monotonous section of interstate highway to induce drivers’ boredom and drowsiness The simulated driving scenario was developed based on real highway data representing an 84-kilometer (km) section of Interstate 70 from Topeka to Junction City, Kansas The geometric design and terrain information was extracted from paper drawings (courtesy of Kansas DOT) Although the alignment, traffic, and other environmental conditions on this highway are very consistent with that expected to induce monotony and drowsiness, the setup was not based on any accident data, but rather on the availability of the design data Based on the highway design, the posted speed limit was 105 km/h (65 mph) The traffic vehicles surrounding the driven truck were also able to intelligently adjust their speed to keep a safe distance from other vehicles The drivers were asked to keep a safe distance from the vehicle in front of them during the test Traffic volume in the driver’s direction surrounding the driven vehicle was low enough to minimize traffic incidents Figure shows a sample of a virtual driving scene The 84-km highway scenario was divided into five equal consecutive segments, each 16.8 km long This allowed for investigation of the drivers’ behavior, and the evaluation and analysis of their drowsiness state on smaller intervals Analysis of the data from each segment enabled the comparison of steering wheel angle data of different sessions or of different level of drowsiness 2.4 Testing Procedure and Experimental Design The experimental protocol consisted of testing truck drivers when alert and drowsy Two driving sessions were conducted: a morning session and a night session Subjects were instructed to have at least eight hours of sleep the night before their scheduled testing time (8:30am~9:30am) Thus, the amount of sleep deprivation associated with the morning session was 1~2 hours and the night session was 18~19 hours All participants completed a practice training session to become familiar with the simulator They were observed for any possible unusual reactions, e.g simulation sickness, unusual driving behavior, or other unrealistic driving beha- Figure Sample scene of the driving scenario 394 A MORTAZAVI, A ESKANDARIAN and R A SAYED viors attributed to the simulated driving environment For the morning session, subjects drove the simulator for one full length of the 84-km scenario, hereafter referred to as a lap During this session, subjects were fresh and were expected to experience no fatigue due to sleep deprivation Morning sessions were only one lap long to avoid induction of monotony After completing the morning session, they were allowed to carry on with their daily life activity Subjects were asked to limit their caffeine intake and not sleep during that day The drivers were not monitored during this period They then reported back to the laboratory the same evening for the night session For the night session, subjects were picked up from home and arrived at the laboratory two hours before the start of the experiment The night experimental sessions were conducted between 1:30 AM and 5:00 AM or until the driver was too drowsy or fatigued to continue driving Most of the drivers stopped before 5:00 AM In this session, subjects were sleep deprived and were susceptible to falling asleep during driving The steering wheel angle values along the road were dependent upon road curvature To compare steering wheel angles (variable) for different sessions, the data should be examined under similar road geometry, i.e repeating laps For example, the effect of road curvature on mean steering wheel angle should the same after driving one lap and after driving three laps Therefore, the road curvature effect was similar for all the sessions, enabling analysis of the effect of morning and night factors on driving performance variables On the other hand, it was not possible to design an extremely long night session scenario because of the simulator software and hardware limitations As a result, the same 84-km driving scenario was continuously repeated, depending on each driver’s level of drowsiness during the night session Therefore, the total length of driving time was different for each driver during the night session A few drivers had higher tolerance for drowsiness, and did not experience doze-off periods during the experiment In general, the repetition of the scenario may have some drawbacks because the drivers could get used to the cycle However, this was not an issue in this experiment because the initial length of one lap was sufficiently large to prevent memorization of the road geometry and features After finishing the night session, each subject was questioned about his/her level of drowsiness The drowsiness-related crashes were also confirmed The participating truck drivers had very intense work schedules and were generally available for only two days Therefore, a counterbalanced experiment design was not feasible The experiment schedule thus consisted of a practice session on one day and a morning and a night session on the other day 2.5 Recorded Variables Four types of data were recorded during the experiment: (1) The driver’s inputs to the vehicle were recorded at 20 Hz from the steering wheel, brake, throttle, clutch, and gearshift (2) Vehicle (truck) kinematics and information on the surrounding traffic vehicles were recorded at 20 Hz (3) Digital videos of the driver’s face, hand and foot position, and the roadway scene were recorded from inside the cabin (4) Eye closure data were recorded at 60 Hz using the head-mounted eye tracking and measuring system Table shows a list of all dependent variables for which data were collected and measured during the driving simulator experiments For statistical analysis, the data were reduced by calculating the mean and standard deviation values within each segment for acceleration, lateral displacement, lateral position, throttle opening, speed and steering wheel angle DROWSINESS ASSESSMENT The sleep-wake cycle and the amount of wakefulness are the primary factors affecting driver drowsiness The likelihood of drowsiness onset and dozing off is higher during sleep cycle or night sessions The drowsiness of the tested drivers was identified and validated in two ways: i) Subjective Drowsiness Rating, and ii) Eye Closure Measures (PERCLOS) In Subjective Drowsiness Rating (SDR), Drowsiness was rated via visual inspection of each driver’s face, and determined based on a subjective judgment of facial attributes and eye closures Observations were made both during experiments (by real-time video surveillance cameras) and after the experiment (video recordings.) The subjective assessment of drowsiness level was based on a newly devised five-level rating scale: • SDR 0: alert • SDR 1: questionable, some primary signs of fatigue and drowsiness were detected, i.e sighing • SDR 2: moderately drowsy, the eye closure was slower and longer • SDR 3: very drowsy, the driver experienced doze-off • SDR 4: extremely drowsy, the driver was completely asleep The Eye Closure Measures (PERCLOS) indicates the intervals of time the eyes were closed PERCLOS (PERcent eyelid CLOSure) is the percentage of time the eye is more than 80% closed RESULTS Truck-driving simulator experiments were conducted for the study population of thirteen commercial drivers A statistical analysis of the resulting data was performed to identify the driver performance variables that differed in the morning versus night sessions, and to identify any EFFECT OF DROWSINESS ON DRIVING PERFORMANCE VARIABLES OF COMMERCIAL VEHICLE DRIVERS 395 Table List of analyzed variables (*normal font: measured data by simulator; **italic font: statistical measure) Variables Description Vehicle ACC* Vehicle acceleration (m.s−2) related ACC_MEAN** Mean vehicle acceleration (m.s−2) variables SPEED Instantaneous speed (m.s−1) SPEED_MEAN Mean speed (m.s−1) SPEED_STDEV Speed standard deviation (m.s−1) LATERALP Lateral displacement from road centerline (m) LATERALD Lateral displacement from center of driving lane (m) LATERALD_MEAN Mean lateral displacement (from center of driving lane) (m) LATERALD_STDEVStandard deviation of lateral displacement (from center of driving lane) (m) LANEDEP Lane departure, total driving time when the vehicle departed 0.5 m from the lane boundaries GAS Percentage of full throttle opening (%) GAS_MEAN Mean percentage of full throttle opening (%) GAS_STDEV Standard deviation percentage of full throttle opening (%) BRAKE_TAP Number of brake pedal taps/application T_CROSS Total driving time on or beyond the left/right shoulder or center line (s) T_CROSS_L Total driving time on or beyond the left shoulder line (s) T_CROSS_C Total driving time on or beyond the dashed line between travel lanes without using the turning signal (s) T_CROSS_R Total driving time on or beyond the right shoulder line (s) N_CROSS Total number of left/right shoulder or center line crossings N_CROSS_L Total number of left shoulder line crossings N_CROSS_C Total number of the dashed line between travel lanes crossings without using the turning signal N_CROSS_R Total number of the right shoulder line crossings CRASH_RD_L Number of left run-off-road crashes CRASH_RD_R Number of right run-off-road crashes CRASH_V Number of crashes with another vehicle or barrier STEER Steering wheel angle STEER_MEAN Mean steering wheel angle (degree) STEER_STDEV Standard deviation of steering wheel (degree) STEER_P Steering wheel angle signal power, average of squared steering values CSR Proportion of time steering wheel angle is constant over the total time CSR_MEAN Mean CSR Driver PERCLOS Proportion of time that the eyes are closed over a certain period of time related SDR Subjective drowsiness rating (0-4) variables SEVD Severity of drowsiness, the total time SDR ≥ divided by total driving time (0-1) correlation between these variables and drowsiness If sufficiently strong correlations exist, then the appropriate variables can potentially be used for a drowsy driver detection system All the data were checked and verified before use; portions of the experiment where the data recording was problematic or unrealistic were discarded Data recording problems mostly stemmed from the eye tracking system, and were caused by the system’s shortcomings and drivers’ sudden or abrupt movements These difficulties included the following: • Eyeglasses: Although the system worked satisfactorily with the majority of eyeglasses, it did not perform well with glasses that exhibited strong glare as the reflections caused erroneous recordings • Drivers, while scratching their faces or rubbing their eyes (particularly during night sessions), misaligned the reflecting mirror and caused the system to lose tracking of the eye In addition, the data from the intervals when the driver was already off the road (i.e right after run-off-road crashes) or when the unexpected movement of traffic forced the 396 A MORTAZAVI, A ESKANDARIAN and R A SAYED driver to perform unnatural maneuvers (i.e sudden lane change), were excluded from the data analysis These are considered steering maneuvers not directly caused or associated with a driver’s drowsy state First, variables representing the drowsiness level, such as SDR and PERCLOS, were analyzed Next, the analysis of the driver performance variables was conducted in two phases The effects of drowsiness caused by sleep deprivation and time of day were assessed in two phases In the first phase, the general trend of each variable was studied by analyzing total mean and total standard deviation to check whether there was any significant change in a variable between morning and night session using ANOVA and Tukey’s test (between treatments degrees of freedom=1) In this phase, each dependent variable was analyzed separately without considering the correlation between the variables The second phase involved examination of the correlation of the variables with the driver’s level of drowsiness in each of the five driving segments For each of the five segments, the correlation between drowsiness levels, measured by subjective rating (SDR) or PERCLOS, and driving performance variables were developed Because of the repetition of the laps (section 2.4), for each segment the variables were measured repeatedly For each driver, the measured variables corresponded to different levels of drowsiness, hence providing data to correlate the level of drowsiness and the driving performance variables The variations of particular variables were also studied during selected drowsy periods to establish any correspondence between the variables and drowsiness This was accomplished by splitting the desired variable signal into smaller windows and analyzing the effect of drowsiness on the variation of statistical variables with drowsiness 4.1 Drowsiness Analysis Table presents a summary of drowsiness analysis for the drivers during morning and night sessions The table includes the subjective ratings and PERCLOS data Generally, for an average of 0.75 hours of morning driving the average SDR was 0.15, while for an average of 2.13 hours of night driving, the average SDR was 2.25 The average PERCLOS was 7% for morning sessions and 22% for night sessions Both SDR and PERCLOS showed a significant difference between day and night conditions (SDR: df (effect), df (error) F=1,24 F=134.27, P3 divided by the driving time (Table 2) Figure shows average SEVD and PERCLOS for each driver during morning and night sessions The figure indicates that the severity of drowsiness and PERCLOS values are higher for night sessions, and differ from subject to subject 4.2 Driving Performance Variables The following is the analysis of vehicle state variables and their relationship, or lack thereof, to the drivers’ drowsy state 4.2.1 Driving performance ANOVAs: Morning and night effect Table displays the driving performance variables’ means, standard deviations and the summary of ANOVAs for morning and night sessions (effect of level of sleep deprivation and time of day) The alpha level was set to 0.05 The ANOVAs and Tukey’s comparison tests of morning and night data showed a significant difference for four variables: standard deviation of steering wheel angle (F1,24 =19.62, P