Changing the landscape of rail through advanced asset health monitoring systems – A novel method to increase the resilience of track infrastructure

Adequate maintenance of railway track infrastructure is vital to its efficient, reliable and safe operation. Fundamental to maintenance is the ability to assess the condition of infrastructure. Difficulties in doing this are compounded by the length and accessibility of the track, especially in remote regions or busy metro environments. The dynamic response of rolling stock on any given section of track is an indication of the track’s condition and the characteristics of the rolling stock. By measuring and analysing this dynamic response it is possible to extract meaningful information about track infrastructure.

Method

This project aimed to deliver a tool for advanced asset health monitoring of rail track infrastructure. The tool developed predicts the motion of the rolling stock during operation, with track geometry measurements and the speed of the rolling stock as inputs. These inputs are measured using instrumented revenue vehicles (IRVs) developed by the Institute of Railway Technology at Monash University. Predictive models were developed for three classes of wagons, a passenger wagon in Indonesia, a freight wagon in Australia and a faster regional passenger wagon in Australia.

The project addressed challenges arising from a lack of consensus amongst operators on the standards to be used. As an example, chord length of various track geometry measurements varies to suit the rolling stock in use. These disagreements require track geometries to be measured multiple times under various configurations, especially when a variety of rolling stock classes operate on it. This is time-consuming and impacts operational efficiency and maintenance cost. First principle mathematical models were developed to convert track geometry measurements from one chord length to another. The accuracy and reliability of the mathematical models was investigated to understand the limitation of the method.

The main focus of the project was developing a vehicle dynamics predictive model. The in-service wagon identified for modelling was instrumented in the initial stage. Track geometry parameters and operational speed were measured using the IRV, and the data analysed and processed to develop predictive models. Two distinct types of models were developed: a first principle mathematical model and a data-driven model from a self-learning algorithm (machine learning).

Some assumptions and approximations were made in developing a first principles mathematical model. These assumptions are valid for the passenger wagon, but the predictive performance of these models decreases for complex rolling stock systems, i.e. freight wagons. For more complex systems, machine learning was used to determine the relationship between the inputs and the dynamic response of the wagon, from which the predictive models were developed. Researchers used machine learning to develop two types of predictive models for assessing track health and operational risk, based on the dynamic behaviour of the wagon:

• A model that classifies every 50-metre section of track between class one and four, with class one denoting a section where measured response is highest, and class four denoting the lowest.
• A regression model that predicts the time series dynamic response of the wagon.

Findings

Both models were extremely useful in quantifying operational risks from track condition. Innovative sampling strategies were used to develop a reliable predictive model from datasets without the desired level of distribution. Optimisation schemes were employed in determining the structure of the machine learning predictive tool. In addition to developing predictive models capable of evaluating dynamic performance, the tool was used to assess the risk to passengers. The predictive tool was trained to predict jerk, which is a measure of the shock sensed by passengers in various orientations. Large jerk events can lead to passengers toppling. The ability to preempt jerk allows operators to adjust operational parameters to minimise the risk of injury.

The outcomes demonstrated that measured data from IRVs can be used to derive business value and quantitatively evaluate performance and risk. Predictive modelling also facilitates migration from corrective maintenance to predictive maintenance. Project outcomes will not just improve the safety and comfort of passengers, but also promote effective maintenance.

Conclusion

With developed tools, maintenance strategies can be shifted from conditional to predictive. This shift could potentially maximise asset life and minimise maintenance cost, and allow for optimised operations and more informed operational decisions. The tool can also be used to evaluate the effectiveness of speed restriction to improve operational safety. Further refinement in the model from permanently instrumented revenue vehicles will improve the predictive capability of the model.