Today's presentation

• Introduction
• Objectives
• Methodology
• Results
• Future studies

Today's presentation

• Introduction
• Objectives
• Methodology
• Results
• Future studies

Introduction

Typical railway problems

• Asset degradation
• Maintenance planning
• Maintenance cost

• Continuous monitoring
• Increased safety
• Optimal planning
• Lower cost

Introduction

There are two main ways of quantifying track quality:

1. Monitor track irregularities (more common)​​
2. Monitor the dynamic response of the wagon to the track excitations

Introduction

There are two main ways of quantifying track quality:

1. Monitor track irregularities (more common)​​
2. Monitor the dynamic response of the wagon to the track excitations

Introduction

Figure: Instrumented railway vehicle (IRV)

Today's presentation

• Introduction
• Objectives
• Methodology
• Results
• Future studies

Today's presentation

• Introduction
• Objectives
• Methodology
• Results
• Future studies

Methodology

Methodology

EDA - Conclusions

• Outliers appeared likely due numerical errors caused by impact loads in the wheel rail contact interface
• More severe track irregularities mask the effect that track curvature has on the measured variables. Therefore, this information is not that helpful and is removed.
• Suspension sensors will likely be the most important
• Accelerometers positioned in the trailing bogie have lower correlation with vertical irregularities (bad sensor placemente position for this application)
• Triaxial accelerometer has low univariate correlation with the target variables, but it is important considering all variables simultaneously
• A baseline model results in                    , which is far below the desired objective

Methodology

Filtered dataset

• Umbalanced dataset
• Many small irrregulaties
• Only class FRA4 is used (railway classification)

Methodology

Methodology

Feature selection

After feature engineering

380 columns!

After feature selection

30 columns

Today's presentation

• Introduction
• Objectives
• Methodology
• Results
• Future studies

Methodology

Results

Better RMSE than 0.66 mm from Urda's research!

Conclusions

1. The current instrumentation of the IRV is capable of obtaining vertical track irregularities. Not likely for lateral track irregularities;
2. Sensor placement in the trailing bogie will likely not help in measuring vertical track irregularities;
3. The optimal model resulted in a RMSE metric of 0.523 mm, which is better than similar research publications;
4. Post processing with a low pass filter at the desired wavelength increased the RMSE from 0.523 mm to 0.421 mm;

Today's presentation

• Introduction
• Objectives
• Methodology
• Results
• Future studies

Future studies

1. Study model robustness with velocity and GPS error
2. Apply the methodology proposed here to obtain lateral irregularities using the virtual sensors
3. Degradation forecasting using track geometry data and ML model outputs to estimate optimal date for maintenance
4. Apply methodology on real measured data