Multisensor Measurement of Train Driver Mental Fatigue: From Simulation to Reality

Motivation

• Automation shifts train drivers from manual control toward prolonged system monitoring, which can reduce active engagement and contribute to task-induced mental fatigue.In GoA2 operation, automated acceleration and braking leave drivers responsible for supervision and intervention during unexpected events.Source: §1 · pp. 2–3

Safety relevance

• Mental fatigue can impair sustained attention, situation awareness, action monitoring, error detection, and inhibitory control, creating risks for safety-critical railway performance.In railway operations, these risks include distracted driving and passing stop signals at danger (SPADs).Source: §1 · pp. 3–4

Measurement challenge

• Objective MF measurement remains challenging because MF is subjective and multidimensional, requiring indirect multimodal assessment through self-reports, physiological signals, behavioral signals, and performance indicators.No generally accepted standard specifies how to quantify MF or select and combine suitable sensor systems.Source: §1 · pp. 4–5

Study contribution

• The study evaluates multiple subjective, physiological, and behavioral MF indicators in a high-fidelity simulator and real-world rail operation to assess sensitivity, feasibility, and robustness across controlled and operational conditions.It addresses the limited evaluation of the full range of operationally viable sensors within a single rail study.Source: §1 · pp. 5–6

Operational constraints

• Operational railway conditions constrain MF assessment through lighting changes, cab vibration, sensor-placement limits, comfort requirements, and privacy concerns [42].These conditions can interfere with vision-based and physiological recordings and limit the feasibility of intrusive or burdensome systems.Source: §1 · pp. 5–6

• The study evaluates whether multisensor approaches for mental-fatigue assessment are sufficiently robust, feasible, and interpretable in real-world rail environments.Source: §1 · p. 6
• The work contributes to evaluating multisensor approaches for mental-fatigue assessment in the railway context.Source: §1 · p. 6
• The findings support further development of sensor-based mental-fatigue detection in train drivers.Source: §1 · p. 7

2.1. Mental Fatigue

• Mental fatigue has no universally accepted definition and develops gradually during sustained task engagement, while drowsiness can vary within seconds and MF accumulates more slowly.Source: §2.1 · p. 7
• Passive mental fatigue is particularly relevant to train driving because extended monitoring, monotony, and cognitive underload can reduce engagement, increase mind wandering, slow hazard responses, and raise collision risk.GoA2 operation is associated with lower workload but higher task-induced fatigue than GoA3, linked to monotony and reduced cognitive demand.Source: §2.1 · pp. 7–8

2.2. Traditional Mental Fatigue Monitoring and Its Limitations

• Conventional vigilance devices require simple motor acknowledgments, yet drivers can automate these responses while vigilance and situation awareness degrade.This limitation motivates direct, objective measures less susceptible to behavioral adaptation and training effects.Source: §2.2 · pp. 8–9

2.2.1. Eye Tracking Measures

• Multimodal non-invasive measures are promising for MF monitoring, with eye-blink dynamics offering a practical operational candidate because they can be captured simply and non-intrusively.Blink metrics detected drowsiness-related impairment under severe sleep deprivation and matched outcomes including out-of-lane events and early drive terminations. EEG markers such as frontal theta and parietal alpha or beta remain sensitive to task and context, with validity degrading outside controlled laboratories.Source: §2.2 · p. 9, §2.2.1 · p. 9, §2.2.2 · p. 10

2.2.2. Physiological Measures

• Physiological monitoring spans EEG, autonomic, cardiac, and respiration measures, but no single signal reliably discriminates fatigue across all task types, while cardiac measures support simple continuous wearable acquisition.EEG studies often use small samples and simulations, and neuroergonomic markers are strongly influenced by task characteristics, context, and confounds; recent HRV and respiration approaches target near-real-time, non-invasive monitoring.Source: §2.2.2 · pp. 10–11
• Autonomic and respiratory measures may offer a practical approach for monitoring, but further validation remains necessary.EDA reflects sympathetic nervous system activation and showed no clear sensitivity to fatigue-related changes.Source: §2.2.2 · pp. 11–12

2.2.3. Behavioral Measures

• Behavioral measures are task-specific and may remain intact despite reduced underlying vigilance, making them insufficient alone for complete or early fatigue detection.This motivates complementary physiological measures that track internal state continuously and independently of discrete task responses.Source: §2.2.3 · p. 12
• Because indicators differ in sensitivity, robustness, intrusiveness, and feasibility, recent work advocates multisensor evaluation focused on interpretable, usable measures in real-world rail environments.The present study therefore examines multiple fatigue-related indicators across simulator and real-world train driving, emphasizing feasibility, robustness, and practical suitability for rail ergonomics.Source: §2.2.3 · p. 13
• The study examines n-back-induced mental fatigue through subjective indicators, compares behavioral and physiological measure sensitivity, assesses ecological robustness, and evaluates drivers’ perceived comfort and usability.Source: §2.2.3 · pp. 13–14

3.1. Study design, Mental Fatigue Induction and Task Description

• Each session began with a 15-minute baseline Drive 1, followed by a one-hour auditory n-back MF induction task, and then a second 15-minute drive.Source: §3.1 · p. 15

3.3. Subjective Fatigue

• Subjective sleepiness was measured with the nine-point Karolinska Sleepiness Scale at four points spanning the drives and n-back task.This enabled within-participant tracking before and after each protocol stage.Source: §3.3 · p. 17

3.3.1. Electroencephalography

• The multisensor battery included EEG, blink-duration eye tracking, heart-rate variability, breathing rate, and electrodermal activity.EEG was recorded at 500 Hz in both settings, while cardiovascular, respiratory, and electrodermal signals used dedicated wearable sensors.Source: §3.3.1 · p. 19, §3.3.2 · p. 19, §3.3.3 · p. 19

3.4. Behavioral Indicator

• Driving behavior was assessed using reaction times to matched trackside targets in both environments and missed dead man’s handle activations in the simulator.Targets were designed to have comparable size, color, contrast, and track distance across settings.Source: §3.1 · p. 15, §3.4 · pp. 19–20

3.5.1. Simulator study

• The simulator study used the high-fidelity RailSET cab simulator, whereas the real-world study was conducted on a closed railway track in Saxony.Simulator driving involved automatic speed regulation and dead man’s handle monitoring, while real-world sessions used GOA1 operational tasks including speed regulation and level-crossing management.Source: §3.5.1 · p. 20, §3.5.2 · pp. 20–21

3.5.2. Real-world study

• The real-world procedure replicated the simulator sequence, with the train stopped during the one-hour n-back task and other train traffic excluded for safety.All physiological sensors were fitted before the session at Schwarzenberg station.Source: §3.5.2 · pp. 20–21

3.6. Participants

• The simulator study included 14 professional train drivers, and a separate real-world study was conducted with six professional train drivers.Source: §3.6 · p. 21

3.7.1. Physiological Data

• The study measured EEG using 26 shared locations, applied sampling and ocular-artifact preprocessing, detected ECG R-peaks, and included RMSSD, blinks, and breathing rate.EEG was sampled from 500 to 256 Hz and re-referenced to average reference; ocular artifacts were corrected using independent-component analysis. RMSSD, blinks, and breathing rate were among the physiological measures.Source: §3.7.1 · pp. 22–23

3.7.2. Behavioral Data

• Behavioral measures comprised reaction times to visual targets and missed dead man’s handle events extracted from simulator logs.Reaction time was defined from target visibility to signal-horn response.Source: §3.7.2 · p. 23

3.8. Data Analysis

• Data analysis combined descriptive results with Bonferroni-corrected paired Wilcoxon signed-rank tests per sensor and environment, while Table 1 documented available data by condition and sensor.Real-world sensor failures were attributed to USB connectivity problems from train vibration, Bluetooth saturation, and high data rates from eye-tracking and camera systems.Source: §3.8 · p. 24, Table 1 · p. 24

4. Results

• No paired Wilcoxon signed-rank tests were significant, with descriptive sensor values reported across drives and environments in Figures 3 and 4.Source: §4 · p. 24, Figure 3 · p. 28, Figure 4 · p. 29

4.1. Subjective Fatigue

• KSS increased from Drive 1 to Drive 2 in both environments, from 3.21 to 5.50 in the simulator and from 2.17 to 3.33 in the real environment.Source: §4.1 · p. 25

4.2. EEG

• EEG changes were inconsistent or stable across environments, and none of the frontal theta, parietal alpha, or parietal beta changes matched the hypothesized fatigue direction.Source: §4.2 · pp. 25–26

4.3. Heart Rate Variability (RMSSD)

• HRV was higher during Drive 2 than Drive 1 in both environments: simulator z-scored RMSSD rose from -0.35 (SD = 0.55) to 0.32 (SD = 0.50), while real-environment means rose from -0.31 (SD = 0.40) to 0.32 (SD = 0.42).Source: §4.3 · p. 26

4.4. Breathing Rate

• Breathing rate followed a comparable cross-environment pattern, with higher values during Drive 1 and lower values during Drive 2.Source: §4.4 · p. 26

4.5. Electrodermal Activity

• Electrodermal activity showed nearly unchanged simulator SCR counts but higher real-world SCR peaks during Drive 2, whereas blink duration remained nearly identical between drives.Source: §4.5 · p. 26, §4.6 · p. 27

4.7. Reaction Time

• Reaction times were higher in Drive 1 than Drive 2, especially in the real environment, while simulator dead-man-handle reminders differed only slightly and overlapped considerably.Source: §4.7 · p. 27, §4.8 · p. 27

4.9. Sensor comfort

• Sensor comfort was favorable overall, with median ratings of 4; EDA and the breathing belt were most comfortable, while eye tracking was least comfortable.Source: §4.9 · p. 30

5.1. RQ1: Fatigue Induction

• KSS ratings increased from Drive 1 to Drive 2 in both environments, supporting the hypothesis that the induction task increased self-reported fatigue.Simulator participants began with higher sleepiness, while lower real-world baseline fatigue may partly explain that group’s smaller absolute increase.Source: §5.1 · p. 31

5.2. RQ2: Indicator Sensitivity

• Overall, indicator sensitivity was mixed and diverged from hypotheses in most cases, with autonomic measures providing the clearest support for fatigue-related change.The discussion therefore distinguishes robust HRV and breathing-rate patterns from null or contrary findings in EDA, EEG, blink duration, and behavioral measures.Source: §5.2 · p. 31, §5.2.1 · p. 32

5.2.1. Heart Rate Variability and Breathing Rate

• HRV and breathing rate changed consistently across both environments: RMSSD increased and breathing rate decreased from Drive 1 to Drive 2, matching reduced-arousal expectations.Their convergence across independent autonomic measures and operational settings supports prioritizing them for future train-driver monitoring systems.Source: §5.2.1 · p. 32

5.2.2. Electrodermal Activity

• EDA did not support the fatigue hypothesis: SCR peaks increased slightly in the real-world environment, remained essentially unchanged in the simulator, and showed high inter-individual variability.The simulator’s absence of change suggests EDA was not sensitive to the fatigue manipulation under the tested conditions.Source: §5.2.2 · pp. 32–33

5.2.3. EEG

• Neither EEG hypothesis was supported: frontal theta at F3 and F4 decreased rather than increased between drives, highlighting context sensitivity under operational constraints.The authors treat this non-replication as meaningful because laboratory classification performance may not transfer directly to realistic driving conditions.Source: §5.2.3 · p. 33

5.2.4. Blink Duration

• Blink duration remained below individual baseline and was nearly identical between drives, indicating no sensitivity to the fatigue manipulation.The authors suggest sustained visual engagement or measuring blink duration without a composite index such as PERCLOS may explain the null result.Source: §5.2.4 · pp. 33–34

5.2.5. Reaction Time and Dead Man’s Handle

• Reaction times were faster in Drive 2 across both environments rather than slower, likely reflecting practice effects that could mask fatigue-related slowing.The especially large real-world improvement supports greater novelty there, while a longer familiarization phase is recommended.Source: §5.2.5 · pp. 34–35
• Dead man’s handle miss rates were very low and showed no meaningful change in the simulator, suggesting limited sensitivity once responses became highly automated.The authors identify this as a broader limitation of vigilance-monitoring devices.Source: §5.2.5 · p. 35

RQ3: Ecological Robustness

• HRV and breathing rate showed the same directional pattern in the technically constrained real-world sample as in the controlled simulator.This consistency is presented as a meaningful form of robustness under real-world conditions.Source: §5.3 · p. 35
• The sensor system was deployable in both contexts, and consistent patterns emerged for at least some indicators despite operational constraints.Real-world data collection introduced substantial technical challenges, making these findings primarily a feasibility result.Source: §5.3 · p. 35
• Future field studies need purpose-built data-acquisition pipelines and systematic robustness testing rather than direct transfer of laboratory sensor setups.These practical lessons follow from the substantial technical challenges encountered during operational rail data collection.Source: §5.3 · p. 35
• Differences in sample size, participant composition, collection timing, and operational demands limit direct comparison between environments.The real-world study included n=6, and small samples preclude inferential statistics for most measures; reported patterns should therefore remain descriptive and exploratory pending replication.Source: §5.3 · pp. 35–36

RQ4: Practical Feasibility

• Most sensors received favorable ratings, with a median of 4 out of 5; EDA was rated M=4.25, SD=0.79.These ratings indicate generally favorable perceived comfort and usability.Source: §5.4 · p. 36
• The breathing belt was the most comfortable sensor and also showed the clearest and most robust fatigue-related patterns.This convergence supports prioritizing chest- and abdomen-worn autonomic sensors for future operational monitoring systems on diagnostic and practical grounds.Source: §5.4 · p. 36
• Eye tracking was the least comfortable sensor, and its blink-duration output showed no fatigue sensitivity in the data.This combination weighs against prioritizing eye tracking for operational fatigue monitoring.Source: §5.4 · p. 36

Returning to the Ironies of Automation

• Reaction times improved and dead man’s handle misses stayed flat across the session, even though subjective sleepiness rose.These results show that overt task performance can remain stable as the driver’s underlying state shifts.Source: §5.5 · pp. 36–37
• Autonomic measures may offer a more direct route to detecting vigilance erosion than behavioral monitoring, including traditional dead man’s handle systems.Unlike behavioral output, autonomic measures do not depend on the driver actively producing a correct response.Source: §5.5 · p. 37

5.6. Adequacy of the Fatigue Induction

• The induction’s limited intensity may explain many null results, because physiological changes were expected within only 15 minutes after the task.Subjective fatigue increased, but the magnitude was modest in the real-world environment.Source: §5.6 · p. 37
• The n-back task induced fatigue through sustained effortful processing, whereas GoA2 monitoring fatigue arises mainly from prolonged inactivity and underload.These fatigue types may differ in both subjective experience and physiological signatures.Source: §5.6 · p. 37
• A longer induction designed to resemble a real monitoring shift could target operational mental fatigue more directly and produce stronger physiological effects.Greater ecological validity would require matching the duration and character of an actual monitoring shift.Source: §5.6 · p. 37

Monitoring priorities

• HRV and breathing rate are the most promising starting point for fatigue monitoring in rail operations, showing consistent directional changes across environments.Both measures were technically robust in wearable form and rated among the most comfortable sensors by drivers.Source: §5.7 · p. 37

Measurement limitations

• EEG, EDA, blink duration, and behavioral measures were not sensitive to the fatigue manipulation under the studied conditions.EEG also faces substantial practical challenges in field deployment, although methodological factors such as induction type and measurement duration may partly explain its insensitivity.Source: §5.7 · p. 38

Future validation

• Future work should prioritize ecologically valid fatigue induction paradigms, larger samples, and purpose-built data acquisition infrastructure for operational rail environments.These priorities address the need for stronger validation under realistic operating conditions.Source: §5.7 · p. 38

Operational acceptance

• Sensor comfort and wearability require systematic investigation because long-term acceptance by train drivers is a prerequisite for operational monitoring.The study addressed comfort through Likert ratings, but sustained acceptance remains an important implementation condition.Source: §5.7 · p. 38

• Subjective sleepiness increased in both settings, indicating that the fatigue induction procedure was effective at least at the self-report level.Source: §6 · p. 38
• Heart rate variability and breathing rate showed the most consistent, theoretically coherent patterns across simulator and real-world environments and were rated comfortable by drivers.These findings support their pragmatic use in rail driver-state monitoring and wearable sensing systems.Source: §6 · pp. 38–39
• EEG, electrodermal activity, blink duration, and behavioral measures did not show clear fatigue-related changes, including the expected frontal theta increase.This underscores the context-sensitivity of neurophysiological fatigue markers and cautions against transferring them uncritically across operational domains.Source: §6 · p. 39
• Reaction times and dead man’s handle performance improved or remained flat despite rising subjective fatigue, illustrating the Ironies of Automation.Drivers can sustain overt task performance through practiced, automated responses even as their underlying state shifts, motivating autonomic monitoring.Source: §6 · p. 39
• Real-world data collection exposed vibration, sensor-connectivity, and concurrent high-frequency acquisition challenges, requiring larger samples and purpose-built infrastructure in future work.Future studies should also use ecologically valid fatigue induction paradigms that better approximate passive monitoring demands in GoA2 operation.Source: §6 · p. 39

• Claude Sonnet 5 (Version July 2026) was used for spell-checking and wording refinement.Source: §8 · p. 40
• The authors reviewed and edited the content after using the tool.Source: §8 · p. 40
• The authors take full responsibility for the published article’s content.Source: §8 · p. 40