CHAPTER 5

Conclusions and Way Forward

5.1 PATH Track Geometry Comparison Analysis

To analyze the extent of changes over time, a track geometry comparison analysis was conducted for data collected between 2013 and 2018 as well as between 2013 and 2020. Critical areas in the 2013 data were identified by selecting locations with high RQ results. These RQ values were collected while riding the PA5 car on the PATH network during Phase 2 of the project. The comparative analysis focused on these critical areas during each designated year.

For the 2013 versus 2018 geometry data comparison analysis, MxV Rail concluded that several areas of interest characterized by larger track geometry deviations in the data presented differences between 2013 and 2018. The observed differences could potentially result in a different dynamic response from the PA5 passenger car operating over the track deviations during each respective year and therefore, could change the RQ output.

During the analysis of the 2013 and 2018 track geometry data, PATH was in the process of commissioning a newly acquired track geometry measurement system. Once PATH completed the commissioning, MxV Rail requested and received recent track geometry data collected in late 2020 from the new track geometry system. MxV Rail then compared the 2020 data to the 2013 data in the critical areas and determined the compared data segments had similar median and IQR values, but significant differences were found in the extreme values (maximums and minimums). Track locations with high geometry deviations are expected to generate adverse RQ. Multiple areas with assumed adverse RQ had a poor match between 2013 track geometry data and the data collected in 2020.

Geometry analysis findings were presented to TCRP, and geometry changes over a 7-year span were discussed. The track geometry variations between 2013 and 2020 were deemed acceptable by TCRP, and MxV Rail engineers were greenlit to proceed with the use of the 2013 PA5 car RQ and geometry data to conduct the simulation work and evaluate the viability of PBTG technology for transit systems.

5.2 PBTG NN Development and Viability for Transit System

NN models were developed for the following PA5 car accelerations and wheel forces:

• Driver cab maximum lateral and vertical accelerations
• Car body trailing maximum lateral and vertical accelerations
• Maximum L/V ratio – left and right lead axle wheels of the lead truck
• Minimum vertical force – left and right lead axle wheels of the lead truck

The model prediction confidence achieved the following ranges:

• Car body lateral and vertical accelerations: 60 to 70 percent.
• Maximum wheel L/V ratio: 73 percent (left wheel); 60 percent (right wheel).
• Minimum vertical wheel force: 16 percent (left wheel); 18 percent (right wheel)

Using track segments with more RQ issues, the NN technique still showed merit for predicted car body accelerations and L/V ratios, but the model results did not reach 80 percent or higher, MxV Rail’s standard for confidence. The performance showed that the models cannot be deployed with confidence for track inspection using the PATH system. Based on MxV Rail’s experience, the observed model performance is due to:

• An insufficient amount of degraded track geometry in PATH track segments and the corresponding adverse vehicle responses to this insufficiency. More degraded geometry data carrying predictive information is needed for the model to better learn how to recognize the specific track and operating condition patterns that are more likely to generate unwanted vehicle responses that lead to RQ issues.
• More track distance and a wider range of track geometry and vehicle response data are needed to develop more accurate models at an acceptable confidence level of performance.

The Phase 2 PA5 NUCARS model was updated so the track geometry parameters used for the simulation would be integrated in the simulation results and synchronized with the dynamic response. The vehicle response outputs and track geometry measurements were aligned with the corresponding track locations. The data used to develop the acceleration models was derived mostly from revenue service test data from the PA5 car collected on PATH track in 2013. Only a few events from the NUCARS runs were included in the training data. These events were generated at 10 mph lower than the allowable track speed.

For the wheel-rail forces, only NUCARS output was used because no instrumented wheelsets were deployed during the 2013 revenue service testing. The PA5 NUCARS model developed during Phase 2 was validated for the car body accelerations using only test results from on-track testing.

After analyzing common railroad RQ and safety standards and evaluating the viability of current PBTG for RQ, it was determined that, without further development, the current PBTG could not be adapted directly for transit systems to predict frequency-weighted rms accelerations based on prescribed criteria and generate track maintenance reports recommending RQ-related remedial actions. Significant changes to the core PBTG technology are needed to be able to accomplish these critical tasks.

5.3 Suggested Way Forward

Phase 3 efforts showed the machine-learning technique still has merit for developing models that can be used for transit systems provided adequate amounts of degraded track geometry and vehicle response data are available. These efforts also showed the PBTG system, as designed, cannot be directly adapted to assess ride quality for passenger cars using RQ criteria, most notably the widely used ISO 2631-1997 criteria.

Takeaways from Phase 3 NN model development approach include:

• ISO 2631-1997 criteria for RQ assessment provides comfort levels calculated from frequency-weighted accelerations. The segment-based approach the current PBTG and NN models were developed with is inadequate for that type of assessment, and a point-by-point approach to develop the NN models would be more suitable. The differences between the two approaches include:
  • Segment-based approach aligns the track geometry data with the measured vehicle response based on a defined segment length of track geometry and using the variable statistics for geometry and vehicle extreme responses (maximums or minimums).
  • Point-by-point approach aligning the track geometry data with the raw measured vehicle response based on one-foot increments.

Takeaways from Phase 3 machine-learning application development in indicate:

• Most individual transit systems are not big networks and may not have sufficient and varied track geometry degradation with patterns that are easily recognizable by NN models.
• Transit agencies with similar track and rolling stock can collaboratively and anonymously create and share data sets of degraded track conditions and RQ responses. Data collaboration can lead to:
  • A data repository to help build more accurate models using machine-learning techniques.
  • Use of data by researchers, Transit Agencies, algorithm and application developers, and other stakeholders.
  • Acceleration of the development of robust models that can be deployed with confidence for RQ assessment of passenger cars

Future considerations for machine-learning model development for PBTG™-like technology designed for RQ assessment based on MxV Rail’s experience include:

• Conducting simultaneous over-the-road testing with a track geometry car and an instrumented passenger car in the same test.
• Collecting geometry data over various tracks consisting of varied track and operating conditions.
• Placing emphasis on passenger cars that are more prone to generating RQ issues when building optimum models.
• Developing a machine-learning application that is “universal” is not currently possible because of the wide variation in vehicle dimensions, masses, and suspension characteristics within the North American transit industry.
• Using multibody system models that have been developed and validated with on-track test results with different geometry and operating conditions to expand the data to be used for model development.