AusRAIL, Operations & Maintenance, Rollingstock & Manufacturing

A risk-based approach to enable predictive maintenance

predictive maintenance

Using what is already there has allowed Infinitive to deploy an advanced predictive maintenance solution based on existing data.

Knowing in advance the risk of a service affecting failure is one of the holy grails of rollingstock asset maintenance managers and is a key step on the path to predictive maintenance. Being able to do that without adding an additional array of sensors and equipment could be thought of as near impossible.

But this is exactly the solution that UK consultants Infinitive Group have developed, and the numbers are already speaking for themselves.

The technological solution has been applied on the London Underground and other mainline routes, where a service affecting failure has significant consequences, said Richard Johnson, managing director – Australia for Infinitive Group.

“You can imagine in the London Underground if you have a failure in a tube tunnel that whole line has to stop. It is the same on a mainline in the UK, a failure there affects many different operators and Network Rail imposes a minimum fine of £500 ($903) per minute for any interruptions on the mainline. If you can look at the data and predict what’s going to fail ahead of time, then that’s quite a significant benefit to the service itself.”

However, reducing service affecting failures without installing more hardware is no easy task. Indeed, it is only through recent developments in technology that Infinitive were able to deliver a solution using the existing data captured, but not always used, by rollingstock.

“What we found is that rollingstock that’s been in service for the past 20 years have had a significant number of sensors already fitted to the vehicles,” said Johnson.“What’s been required is just a technology convergence.”

In addition to the existing sensors, advances in processing power and cloud storage, along with the emergence of new technologies such as machine learning that can analyse vast amounts of data, are enabling new ways of understanding what components on a train are reaching the end of their life or are under additional strain. With just the sensors already installed on multiple lines of rollingstock in London, already seven billion data samples are recorded each hour across the fleet.

“Manually it’s impossible to get across that kind of data so you need the ability to process and store it and then you also need machine learning to tease the information out,” said Johnson.

While the data is already there, without the tools to store, process, and analyse it, this resource was being underutilised.

One example of how the data can now be understood through the application of Infinitive’s solution was in doors. Uploading data from the wayside data store to the cloud and analysing it using machine learning, alerts could be generated to identify doors that were being obstructed during closing and opening. This was caused by issues such as dirt and debris build up. These alerts would identify which doors required servicing. Similar scenarios could be run for compressor units and other systems to enable predictive maintenance prior to failure.

“Prior to this technology, the machine learning and processing power, you’d have to write manual algorithms for each use case which is very time consuming,” said Johnson. “It has been done and you can do it, but machine learning allows you to explore in many different areas that you wouldn’t even realise to pick out trends and anomalies.”

The data frames for analysis are assembled in sets according to their inspection points e.g. door locations, each frame also contains the inspection aspects e.g. fault severities and other associated fields. Where there are known influencers, like the vehicle being common to multiple door locations, this is also included in the data. The model then works through the analytics job highlighting anomalies. Post processing contextual information, such as known events, can be included and the model re-run. This allows human domain knowledge to be worked into the model.

GETTING TO A POINT OF CHANGE
Throughout the implementation of the technology and the approach to predictive maintenance, the issues of establishing this solution were not so much technical as they were process-based.

“The data was being collected and was being stored but it was in an older database format, so we brought that into a database that was also tailor-made for machine learning and was quite scalable. The ability to do that was more a process than a technical issue where you have to organise access to the data, and once the data is accessed you can start running the analysis on it.”

Rather than a need for new gadgets there was a need for expert human intervention to make the most of the existing data. Anomalies needed to be removed from the dataset so that only the real data was being fed into the algorithm.

“That improves the accuracy of the machine learning quite significantly because if you just throw all the data in unprocessed your analysis is not as accurate,” said Johnson.

With the system now established, the alerts needed to be actioned to have tangible benefits. By partnering closely with its operators, Infinitive was able to propose changes to ways of working that have typically improved the Mean Distance Between Service Affecting Failures by 20 per cent. In addition, savings of £80 million over five years are expected to be realised.

As Johnson points out, when the technology is applied alternatives can be suggested to ensure that services continue without causing disruptions.

“The outcome can be multilayered and in one example that we did with diesel trains, we had a look at turbocharger performance and at the end, we were able to go back to the customer and say, ‘Based on our analysis, if you run the train at this load under this temperature condition this turbocharger has an 80 per cent chance of failure.’ Then, although the operator cannot book that train in for servicing today, they can put it on a route that’s not as hilly and lower load and then book it in later,” said Johnson.

“This risk-based maintenance approach allows the operator to make better decisions.”