Machine Learning at the Core of Automated Inspection and Predictive Maintenance
Inspection and maintenance are key ways to ensure the reliability of railway assets throughout their life cycle. Equipment that fails in-service entails costs higher than planned replacement, in the form of safety risk, disruptions to the network, delays and service cancellations.
Reactive or corrective maintenance - correcting equipment after a fault has occurred - should be minimised, and more proactive forms of maintenance, preventing the failure to occur, should be the norm.
However, this is a difficult task. Preventive or planned maintenance, which is based on general reliability and life expectancy statistics available for an asset class, does not guarantee that the inspections carried out identify the specific piece of equipment which is going to fail next.
In contrast, predictive maintenance, which uses real-time asset condition monitoring, collecting data on equipment during operations to identify issues and predict the moment when the equipment is going to fail, is a promising area for the railway industry. It is recognised that not all asset types warrant this approach but the ability to determine the right moment when a piece of equipment should be repaired or replaced helps avoid incidents and downtime, while keeping maintenance frequency as low as possible.
Predictive maintenance would not be possible without the many sensors deployed by industries to capture real time information on equipment health and condition (vibration analysis, remote visual monitoring, infrared and acoustic analysis, temperature and pressure monitoring). Predictive maintenance would also not be possible without the means of making sense of all this data and turn it into predictive insight: this is where Machine Learning can come in.
Predictive machine learning software can learn which combination of factors lead to certain types of failure, finding complex underlying relationships, so when these factors are identified again, failure can be predicted. Combined with data mining techniques, models and tools can be developed that can assess asset health and predict remaining useful live.
Computer vision, or machine vision, involves using digital imaging systems to automatically extract, analyse and understanding information from images and other forms of multidimensional data. These techniques are gaining popularity in areas where repetitive visual inspections are used, for example facial recognition at airport security and social media apps. Applying machine vision technologies to railway inspection offers a way for the industry to reap huge benefits. Human inspectors are generally used to carry out visual and/or measured inspections of critical assets. This approach is time intensive, costly and, given its reliance on humans, can be inconsistent and prone to error. Making use of machine vison technology, able to capture detailed recordings of asset condition, provides on opportunity to improve safety for both staff and passengers, while driving down overall cost. Imaging techniques such as line and area scanning, thermal imaging and colour identification can be used together to identify fault irregularities leading to improved safety, reliability, and service.
There are several railway companies that are beginning to embrace these techniques to revolutionise how work is being carried out in these areas.
Machine vision for infrastructure inspection
In 2012 Network Rail introduced the plain line pattern recognition (PLPR) system to replace visual track inspection. The system uses a network of line scan, 3D and thermal imaging cameras to capture images of the railhead. The machine vision software can identify defects, that are then validated by an inspector.
The pattern recognition system is used alongside track geometry measurement and positioning systems to cross check the status of track components and their potential effects on rail support and relative position. The image below indicates a variance in track geometry that can be in part attributed to a clamped rail defect.
This system has improved safety for track workers, increased train path availability and has improved the quality and efficiency of track inspections.
There is a clear case for extending this technology to other infrastructure assets. The applications are most readily suited to assets where there are more homogenous and visible components with defined geometry relationships such as overhead line systems.
Predictive maintenance for rolling stock
Train manufacturers, like Hitachi and CRRC, are pushing for a digital revolution in train maintenance. Hitachi’s new IEP trains are fitted with sensors that collect data from over 48,000 signals, real-time, to feed into a system that monitors condition and provides information to the relevant technical teams to support their decision making. CRRC have developed a machine-learning based prognostic and health management system (PHM) that has been applied to the high-speed railway network in China. This system can streamline operations and maintenance decisions making by combining data on physical characteristics, fault diagnostics, health assessments from various sources.
CRRC’s system has a novel way of dealing with challenge of noise in the data without using huge amounts of computing power, improving accuracy and efficiency. By determining the probability density function (PDF), a way of calculating the relative likelihood of a variable falling between a particular range, it is possible to determine if the reading is a characteristic fault or not. The PDF values are compared with reference values and the Hellinger distance, that is the similarity between two probability distributions, calculated. The Hellinger distance then feeds into a Bayesian inference calculation, which is modified with each new piece of data it received to create a single figure to show the likelihood of the component being faulty.
Since introducing this system equipment maintenance costs have gone down by over 25% and the reported failure rate have decreased by around 75%. As these fleets spend more time in service machine learning techniques can be applied to historical data gathered to identify new combinations of signatures and trends that are precursors to failure.
While these techniques are being embraced by some in rail, there are still huge opportunities to transform the way many aspects of inspection and maintenance are carried out, that are, in many cases, manual, repetitive and based on traditional, time-interval based methods.
Rolls Royce FLARE – a pair of ‘snake’ robots which are flexible enough to travel through an engine, like an endoscope, before collaborating to carrying out patch repairs to damaged thermal barrier coatings.
Radical innovations are being enabled by the use of machine learning in the world of robotics, improving their ability to navigate, adapt to surroundings, manipulate objects and control motion. The powerful combination of machine learning with robotics makes it possible to create ‘find and fix’ systems, that require limited human interaction. Examples of this can be found in the marine and automotive industries, as shown by the Rolls Royce FLARE, and are starting the be seen in the rail industry.
The motivation to employ machine learning techniques for automated inspection and predictive maintenance should be strong in the rail industry, where we have assets with long lifespans, that we try to keep in service for as long as possible. Improving maintenance and inspection can reduce operating costs and keep assets going for longer.
To fully exploit these emerging technologies, the rail industry needs to rethink its approach to data gathering and storage. A broad scope and high resolution are needed for fault trends and signatures to be identified. Another key challenge related to railway maintenance is the operation and management of the complex, interwoven schedules and plans that involve people, assets and access. This is a rich subject area in which exploiting machine learning applications and has the potential to resolve conflicts, achieve higher utilisation and ultimately provide a better service to the customer.
In the next article in this series, Justin Willett, Professional Lead Operation and Performance, and Giulia Lorenzini, Senior Partnerships and Grants Manager, will discuss the application of machine learning to planning, scheduling and timetabling.