From the opening of the first line from Tokyo to Yokohama in 1872 through to the shinkansen bullet trains of today, the evolution of the Japanese railway has been a triumph of ambition and a pioneering achievement in the fields of technology and engineering.
Passenger rail services in Japan have become a byword for efficiency, and the network spanning more than 27,000km remains one of the most utilised, punctual and least subsidised in the world.
Ensuring that operations and maintenance (O&M) work is carried out in a timely fashion is central to this success. In recent years, however, operators such as East Japan Railway Company, or JR East, have faced multiple challenges including aging infrastructure, a dearth of new train maintenance specialists due to Japan’s decreasing population, and spiralling costs alongside shrinking budgets.
To help improve train efficiency and safety for the six billion passengers that use JR East services every year, the company turned to Palo Alto Research Centre (PARC), an open innovation company based in Silicon Valley focused on predictive analytics using the industrial internet of things (IIoT).
“Many of Japan’s capital-intense rail assets were deployed decades ago and JR East would obviously like to extract the maximum life out of them without compromising on safety,” says PARC strategic execution director Ajay Raghavan. “The shrinking population means that rail revenues are declining and so there is also pressure on O&M teams to be lean, plus a lot of more experienced technicians are beginning to retire.
“In light of these factors, JR East realised that its existing time-based maintenance (TBM) practices were not necessarily the best solution going forward, and approached PARC six years ago in search of a more effective and sustainable long-term solution.”
Window of opportunity: time vs condition-based maintenance
Traditional time-based maintenance (TBM), and reactive ‘fail and fix’ or planned maintenance practices, can be costly, prone to human error, or lead to downtime or accidents. Train companies are understandably moving away from TBM in increasing numbers in favour of condition-based maintenance (CBM), in which servicing of machinery is performed when the need arises.
Raghavan uses an analogy of a car in order to illustrate the concept of CBM. A car needs an oil change every six to 12 months or 5–10,000 miles. Newer models have sensors that tell the owner when the service is due based on average usage time, but this may not take into account factors such as heavy loads or frequent use. The service schedule may be too conservative, for example.
The worst that can happen, of course, is that the car stops working, but extrapolate that out to a large complex asset such as a train and the problem immediately becomes much more significant.
“The traffic in the Tokyo subway system is crazy, especially during peak hours, and if a train stops working the entire system clogs up, or in the most extreme cases accidents can result in loss of life with operators liable,” says Raghavan. “TBM may have worked for assets that have similar usage patterns, but not necessarily for those with a significant geographical or climate spread, or those stretched beyond their original design life; in these instances TBM practices are no longer cutting it.
“Train companies are therefore moving away from TBM to this more predictive paradigm where O&M teams may get a week or even a month’s notice before assets fail, allowing them to prepare schedules and resources to avoid taking a critical train out of service and disrupting the system.
“Predictive analytics is also extremely valuable in terms of long-term planning. Providing operators with data on the remaining useful life of an asset one or even six months before it has to be retired enables them to plan ahead for a major capital outlay, for example.”
The MOXI™ technology suite explained
JR East had collected a significant amount of lab and field data, but were struggling to understand it.
“They were getting 70%–80% success but not enough for the operations and maintenance teams to be sufficiently confident with the capabilities,” says Raghavan. “They were also sceptical about CBM.”
PARC deployed its MOXI™ technology suite, which utilises model-based algorithms that enable 90% or higher accuracy and low false alarm rates, and in some cases require only minimal data sets.
“The first system we explored was train doors which, in the UK, can cause as much as 30% of all delays,” explains Raghavan. “There is a lot of them and they are used every time someone uses the subway or other major transit trains. Doors are supposed to work 100% autonomously and they do a reasonable job most of the time, despite being subjected to quite enormous loads, but it only takes one to fail to take the entire train out of operation.
“So this was of huge importance to JR East because the company prides itself on being on time, all the time. Within a matter of five months our PARC team was able to demonstrate first proof of concept of our predictive CBM working for the train door system with 95% accuracy or higher in classifying the different types of faults and, thanks to our algorithms, very low to minimum false alarm rates.”
Dashboards were also developed to enable JR East engineers to visualise and better understand the obtained data. PARC and its Asian consultancy partner NRI are now working with JR East Technical Centre teams to test and implement these solutions on the heavily used Yamaman rail line.
“We have also built a computerised analytics solution that automatically analyses the images taken by Solus trains that operate on tracks overnight,” adds Raghavan. “Our algorithms make sense of them and identify different types of faults.”
Bridge to the future: optical radar and analytics
PARC’s MOXI™ technology suite has also been deployed in smart factories, as well as critical aerospace and energy systems. According to Raghavan, its success is in a large part down to its ability to obtain extremely high-accuracy analytics.
“We do not count on black box machine learning approaches, but instead apply a first principles, physics-based approach in order to understand how a system operates, functions and fails,” he explains. “We do not necessarily rely on large amounts of labelled data sets, which is what most contemporary machine learning algorithms need for success. In CBM, getting labelled data sets is a significant challenge because asset failures tend not to be that frequent – obtaining even modest amounts can take months.
“However, as more assets come online we are getting more data and our predictive analytics technology is very well positioned to take advantage of that.”
PARC’s global footprint extends to Australia, where it is currently embedding fibre optic sensors into rail bridges owned by the Rail Track Corporation of Victoria (VicTrack) and using optical radar and analytics to make sense of the resulting data.
“A lot of bridges were built in the 1950 and 1960s, and designed with a 100-year lifespan, but they are beginning to show their age,” says Raghavan. “So far Victoria has managed to avoid the sort of bridge failures we have seen elsewhere in the world, including in the US, where it is a major concern.
“The state of Victoria is being proactive and, like JR East, has adopted a predictive maintenance strategy to manage these assets.”