July 2016

Nowcasting and Forecasting tecniques for railway switches maintenance

A large proportion of the costs of the European railway Infrastructure is related to maintenance and the current situation of railway maintenance .

The increasing usage of the railway infrastructure, due to a growing frequency of passenger and freight trains on it, and environmental and safety regulations, the increasing of maintenance requirements cannot be met without a substantial shift in maintenance strategies.

In this prospective, the new “Proactive/Predictive Maintenance”  approach will involve/have an impact all the main railway stakeholder on long-term preservation of assets (expressed in RAMS requirements) minimizing life cycle costs, while for the railway operators, the proactive maintenance will improve the railway infrastructure availability and reliability.

Switches and Crossings (S&C) are fundamental infrastructure assets that allow efficient routing of trains on the network. Assessing their current status and ensuring their proper functionality is obviously a key requirement to guarantee the railway infrastructure availability for the railway operators. For these reasons, Ansaldo STS tries to discover in the Mantis project, references concerning railway S&C condition monitoring based on a wide range of methodologies from the world of statistics, data mining, time series analysis, machine learning, and filtering.

Given the examined literature, it seems clear that the use of the word “nowcasting” can be associated with: a shorter timeframe respect to “forecasting”, a different approach or algorithm for performing the estimation of the value of interest, the fact that the data used for the estimation is imprecise, uncertain, incomplete or is only indirectly related to the phenomenon of interest, and, finally, with the purpose of providing an alert for a sudden event or a possible anomaly.

A reasonable simple definition could be adopted in the framework of the Mantis project for differentiating “forecasting” and “nowcasting” processes:

Forecasting:

The process of exploiting past and present data to make deductions about the future.

Nowcasting:

The process of exploiting past and present uncertain or incomplete data to make deductions about the present.

Many of the failures are not detected until the asset is being operated by the interlocking system when trying to lock the train route. This means that for some failures, nowcasting cannot be done before operating the unit. Frequent test procedures could solve this issue but it has other disadvantages like the introduction of increased wear, cost and reduced inherit capacity. Another solution for this problem is to introduce additional condition monitoring systems for detection of different states of critical components in the S&Cs.

Estimating the remaining component lifetime – a key activity of of MANTIS

A key activity of the MANTIS WP4 work on analysis and decision making functionalities has focus on identification and development of methods for forecasting the remaining useful life of an asset by modelling it’s deterioration and by extrapolating this into the future. When the extrapolated deterioration reaches a certain threshold that marks the end-of-life of the asset, and that point in time can be used as the forecasted remaining useful life (RUL). Simultaneously, the extrapolated deterioration gives an indication of the expected wear and tear, which can be compared with the observed wearing out to monitor the behavior of an asset.

Based on the predicted remaining lifetime, suitable maintenance tasks can be planned, in order to prevent unscheduled system down time.

 

Overall working strategy

The approach followed by MANTIS to address the estimation challenge is as follows:

  • Develop algorithms and train models to estimate the RUL or predict the wear and tear of an asset.
  • Monitored data (from embedded sensors or from quality assurance checks) is streamed to the cloud based storage system, where it will be investigated using time-series analysis or temporal data mining approaches.
  • These methods will be used to discover revealing relationships between parameters to identify significant patterns, trends and anomalies.
  • Based on this analysis of the historical data, a diagnostic model will be made self-learning, comparing predicted and actual data.
  • Flexible modelling techniques will be used on diverse information, such as condition parameters, data patterns, sparse data and/or existing expertise to learn more complex relationships between the different parameters or train models which allow prediction of future performance in terms of expected failures.

 

Current status of work

An overview of relevant deterioration models and methods to estimate the remaining useful life (RUL) has been made. Also, for each use case, the present status with respect to available data, deterioration models and methods for estimation of RUL has been collected. In particular, for each of the use cases, the following information is relevant:

  • Description of data typically available
  • Description of relevant deterioration processes
  • Description of models used for estimation of RUL

 

Next steps

During the next 6 months, the general models and methods will be analysed and compared with the present status for the individual use case (available data, deterioration process, RUL estimation models). Following this analysis, specific models and methods will be selected and unified as much as possible, and thereafter applied and validated on the use cases during the second half of the MANTIS project.

This will be achieved by using a bottom-top-bottom approach:

  • Collecting input for individual use cases
  • Seeking for commonalities and unification into a generic approach
  • Applying the generic approach on the individual use cases

Maintenance data analytics; How to deal with free text

Problem description

For analyzing maintenance problems, such as determining the root cause, in general three sources of information are available:

  • sensor data
  • machine generated logging messages
  • service and/or maintenance engineer reports

In contrast to sensor data and machine generated logging messages, of which the format is determined at design time, the service and maintenance reports have a free format, making it hard to analyze automatically.

Problem reporting
Problem reporting

Possible solutions

Before going into possible solutions, a better description of the problem is needed. Free format is this case means that a report can contain

  • unstructured text
  • ambiguous descriptions e.g. mixing commercial and technical identification
  • multiple languages e.g. using local language
  • spelling errors, ad hoc abbreviations, missing data etc.

For easy analysis it would be good if it is possible to bring free text back into the predefined format realm. Ways to solve this are

  • using predefined forms with e.g. dropdown lists for selecting options
  • using intelligent text editors that can recognize the vocabulary needed for describing the specific problems
  • combinations of the above
  • less stringent is specifying rules people have to abide by when entering maintenance logs (typo’s are still possible then)

As it is impossible to predict all problems that will be encountered during the lifetime of a machine, there still needs to be the option to enter free text, but it should be clear that this is only to be used when other options do not cover the problem at hand.

Conclusion

It is clear that free text will always be used in reporting, but in order to facilitate data analysis of service and/or maintenance reports computer aid or rules can make life a lot easier.