Asif Faisal Chowdhury
CV
Menu

ManuscriptFirst author and doctoral researcher

Measuring excavator operator behavior

94.5% mean balanced accuracy

94.5% mean balanced accuracy from wearable joystick signals.

All projects

My contribution

Analyzed joystick-mounted inertial measurements and separate gaze measures to study experience-related behavior.

Measure behavior from recorded signals

Joystick-mounted inertial measurement units (IMUs) capture control motion. The classification analysis uses 36 usable trials, with 19 experienced and 17 inexperienced operators. A separate gaze analysis uses 40 participants. These branches answer complementary questions and retain separate sample definitions.

Classification and coordination

Logistic regression reports 94.5% mean balanced accuracy with a fold standard deviation of 9.0 percentage points. The protocol repeats stratified five-fold cross-validation twenty times and refits preprocessing inside each training fold. Leave-one-out performance is a separate condition.

Experienced operators show 43.2% simultaneous joystick movement versus 20.4% for inexperienced operators, measured as a share of active time. This descriptive contrast is not the classifier’s score and does not prescribe an ideal overlap percentage for training.

Visual behavior

Eight of sixteen eligible gaze/head metrics differ after correction. Pupil variability has reported group medians of 0.153 and 0.216 mm, with Cliff’s delta −0.88 and a 95% confidence interval from −1.00 to −0.69. These are aggregate associations. Participant trajectories and distributions cannot be reconstructed from them.

System architecture

Select a step, or a block in the drawing, to see its input, computation, output and source.

Redrawn from manuscript

Two sensor branches with separate samples

Joystick IMUs feed classification and coordination analysis. Eye tracking feeds a separate gaze comparison. The branches use different samples and are not combined into one classifier.

  • Signal processing and statistics
  • Classifier
  • Reported result
Two sensor branches with separate samplesA simulator session feeds two separate branches. The IMU branch (36 usable trials) runs joystick IMUs to motion features, then four classifiers to an experience class, and separately a coordination analysis to simultaneous movement. The gaze branch (40 participants) runs an eye tracker to sixteen eligible metrics, then corrected comparisons to group differences.SimulatorsessionJoystickIMUsMotionfeaturesFourclassifiersExperienceclassCoordinationanalysisSimultaneousmovementEye tracker,gaze and headEligible gazemetricsCorrectedcomparisonsGroupdifferences
  1. Simulator sessionInput
    Input
    One excavator simulator task.
    Computation
    None.
    Output
    Recorded sensor streams.
    Source
    Wearable-sensor manuscript, Sections 3 and 4
  2. Joystick IMUsInput
    Input
    Joystick-mounted inertial measurement units.
    Computation
    Records control motion.
    Output
    IMU signals from 36 usable trials.
    Source
    Wearable-sensor manuscript, Sections 3 and 4
  3. Motion featuresDeterministic
    Input
    IMU signals.
    Computation
    Feature extraction; preprocessing is refitted inside each training fold.
    Output
    Per-trial features.
    Source
    Wearable-sensor manuscript, Sections 3 and 4
  4. Four classifiersModel
    Input
    Per-trial features.
    Computation
    Logistic regression, LDA, radial-basis SVM and random forest under twenty repeats of stratified five-fold cross-validation.
    Output
    Experienced or inexperienced label.
    Source
    Wearable-sensor manuscript, Sections 3 and 4
  5. Experience classOutput
    Input
    Classifier predictions.
    Computation
    Mean balanced accuracy with fold SD.
    Output
    Classification result.
    Source
    Wearable-sensor manuscript, Sections 3 and 4
  6. Coordination analysisDeterministic
    Input
    IMU activity.
    Computation
    Both-active time as a share of active time.
    Output
    Group mean simultaneous movement.
    Source
    Wearable-sensor manuscript, Sections 3 and 4
  7. Simultaneous movementOutput
    Input
    Coordination measure.
    Computation
    Descriptive group comparison.
    Output
    Not a training target and not the classifier score.
    Source
    Wearable-sensor manuscript, Sections 3 and 4
  8. Eye tracker, gaze and headInput
    Input
    Eye tracking during the task.
    Computation
    Records gaze and head movement.
    Output
    Signals from 40 participants.
    Source
    Wearable-sensor manuscript, Sections 3 and 4
  9. Eligible gaze metricsDeterministic
    Input
    Gaze and head signals.
    Computation
    Computes sixteen eligible gaze and head metrics.
    Output
    Per-participant metrics.
    Source
    Wearable-sensor manuscript, Sections 3 and 4
  10. Corrected comparisonsDeterministic
    Input
    Per-participant metrics.
    Computation
    Group comparisons with Cliff's delta and corrected q values.
    Output
    Effect sizes with 95% CIs.
    Source
    Wearable-sensor manuscript, Sections 3 and 4
  11. Group differencesOutput
    Input
    Corrected comparisons.
    Computation
    None.
    Output
    Associations only; pupil variation alone does not establish cognitive load.
    Source
    Wearable-sensor manuscript, Sections 3 and 4

Adapted from Skill Classification and Behavioral Analysis of Excavator Operators Using Wearable Sensors, Table 4; Section 4.2.1; Figure 6; Table 5.

No combined IMU and gaze classifier result is reported.

Animated pulses show conceptual signal flow only. They do not show measured execution time or a live inference run.

Findings

Reported results, grouped by sample. Each figure keeps its own scale and links to its source and method.

Usable IMU trials

36 trials. 19 experienced, 17 inexperienced; wearable classification and coordination branch.

Mean balanced accuracy, repeated five-fold validation

Reported in source

Highest reported value: Logistic regression, 94.5%. Whiskers show one standard deviation, not a confidence interval.

Data table
Mean balanced accuracy, repeated five-fold validation. Usable 36-trial IMU sample; twenty repeats of stratified five-fold cross-validation.
ConditionValueUncertainty
Logistic regression94.5%SD 9.0 percentage points
Linear discriminant analysis92.5%SD 11.6 percentage points
Radial-basis SVM92.2%SD 10.7 percentage points
Random forest89.0%SD 12.2 percentage points

Denominator: Usable 36-trial IMU sample; twenty repeats of stratified five-fold cross-validation.

Adapted from Skill Classification and Behavioral Analysis of Excavator Operators Using Wearable Sensors, Table 4; Section 4.2.1; Figure 6; Table 5.

Source and method

Mean simultaneous joystick movement

Reported in source

Highest reported value: Experienced, 43.2%.

Data table
Mean simultaneous joystick movement. Both-active time as a share of active time; reported group means.
ConditionValue
Experienced43.2%
Inexperienced20.4%

Denominator: Both-active time as a share of active time; reported group means.

Adapted from Skill Classification and Behavioral Analysis of Excavator Operators Using Wearable Sensors, Table 4; Section 4.2.1; Figure 6; Table 5.

Source and method

Separate gaze analysis

40 participants. Gaze/head comparisons; not the IMU classifier cohort.

Median pupil variability

Reported in source

Highest reported value: Inexperienced, 0.216 mm.

Data table
Median pupil variability. Separate gaze analysis; reported experience-group medians.
ConditionValue
Experienced0.153 mm
Inexperienced0.216 mm

Denominator: Separate gaze analysis; reported experience-group medians.

Adapted from Skill Classification and Behavioral Analysis of Excavator Operators Using Wearable Sensors, Table 4; Section 4.2.1; Figure 6; Table 5.

Source and method

Cliff's delta: experienced versus inexperienced

Reported in source

All 8 reported effects are negative, from −0.88 to −0.44, and every 95% CI excludes zero.

Data table
Cliff's delta: experienced versus inexperienced. 40 participants, 20 per group; q values Benjamini-Hochberg corrected across all 16 eligible metrics.
ConditionValueUncertaintyCorrected significance
Pupil variability−0.8895% CI [−1.00, −0.69]q < 0.0001
Mean saccade amplitude−0.8795% CI [−1.00, −0.65]q < 0.0001
Vertical gaze spread−0.7895% CI [−0.96, −0.55]q = 0.0001
Mean saccade duration−0.6995% CI [−0.94, −0.37]q = 0.0007
Stationary gaze entropy−0.6995% CI [−0.91, −0.41]q = 0.0007
Head rotation rate−0.6695% CI [−0.89, −0.38]q = 0.001
Median fixation duration−0.4995% CI [−0.76, −0.17]q = 0.0177
2D gaze dispersion−0.4495% CI [−0.73, −0.10]q = 0.0359

Denominator: 40 participants, 20 per group; q values Benjamini-Hochberg corrected across all 16 eligible metrics.

Adapted from Skill Classification and Behavioral Analysis of Excavator Operators Using Wearable Sensors, Table 4; Section 4.2.1; Figure 6; Table 5.

Source and method

Evaluation Scope & Methodology

Samples, denominators, uncertainty, sources and limits

Approach

Repeated stratified five-fold cross-validation with twenty repeats and training-fold preprocessing; separate corrected gaze comparisons.

Cohorts and samples

CohortSampleDefinition
Usable IMU trials36 trials19 experienced, 17 inexperienced; wearable classification and coordination branch.
Separate gaze analysis40 participantsGaze/head comparisons; not the IMU classifier cohort.

Reported measures

MeasureUnitCohortDenominatorUncertaintyEvidence
Mean balanced accuracy, repeated five-fold validationPercentUsable IMU trialsUsable 36-trial IMU sample; twenty repeats of stratified five-fold cross-validationLogistic regression: SD 9.0 percentage points; Linear discriminant analysis: SD 11.6 percentage points; Radial-basis SVM: SD 10.7 percentage points; Random forest: SD 12.2 percentage pointsReported in source
Mean simultaneous joystick movementPercentUsable IMU trialsBoth-active time as a share of active time; reported group meansNone reportedReported in source
Median pupil variabilityMillimetresSeparate gaze analysisSeparate gaze analysis; reported experience-group mediansNone reportedReported in source
Cliff's delta: experienced versus inexperiencedCliff's delta from −1 to 1Separate gaze analysis40 participants, 20 per group; q values Benjamini-Hochberg corrected across all 16 eligible metricsPupil variability: 95% CI [−1.00, −0.69]; Mean saccade amplitude: 95% CI [−1.00, −0.65]; Vertical gaze spread: 95% CI [−0.96, −0.55]; Mean saccade duration: 95% CI [−0.94, −0.37]; Stationary gaze entropy: 95% CI [−0.91, −0.41]; Head rotation rate: 95% CI [−0.89, −0.38]; Median fixation duration: 95% CI [−0.76, −0.17]; 2D gaze dispersion: 95% CI [−0.73, −0.10]Reported in source

Uncertainty definitions

  • SD. Standard deviation as reported by the source. It describes spread across evaluation folds or runs. It is not a confidence interval.
  • 95% CI. 95% confidence interval as reported by the source.

Protocol

  • Keep the N=36 IMU and N=40 gaze branches separate.
  • Classifier uncertainty is fold standard deviation, not a confidence interval.
  • Leave-one-out accuracy belongs to a separate validation condition.

Limitations

  • One simulator and task with a small sample.
  • The gaze sample is separate from the usable IMU sample.
  • No combined IMU–gaze classifier result or independent field validation is established.
  • Descriptive group differences are not validated training targets.
  • Pupil variability alone does not establish cognitive load.

Sources

  • Skill Classification and Behavioral Analysis of Excavator Operators Using Wearable Sensors, Table 4; Section 4.2.1; Figure 6; Table 5 (Supplied manuscript reviewed November 2026). Public release pending

Research records

LSU Industrial Training & Assessment Center · September 2024 to present

Energy Assessment Team Leader & Researcher

I lead industrial energy assessments: utility-bill analysis, on-site measurement, energy-balance calculations, and client technical reports. Totals are team assessment outcomes.

Recommended annual electricity savings

8,208,899 kWh/year

16 assessments with recommendations

Implemented annual electricity savings

5,950,754 kWh/year

11 assessments with implementation reports

Field assessments

Open to a conversation.

AI engineering roles where systems meet real operations.

Download CV