Since 2016, 4Tel Pty Ltd of Newcastle, Australia, has been investing in the development of artificial intelligence for application in the rail industry generally. As a part of this activity, 4Tel has a research and development contract with the University of Newcastle Robotics Laboratory known as NUBots, where 4Tel is their Platinum Sponsor. The work is being conducted under Project HORUS, which seeks to develop an Advanced Driver Advisory System (ADAS) using real-time sensors and software to assist a driver in the safe operation of a locomotive. As the technical basis to this work, 4Tel has selectively applied modern autonomous car technology to achieve very sophisticated artificial, intelligence based, ADAS functionality.
The HORUS System
For safe and efficient operations, a locomotive needs to know exactly where it is, recognise the objects around it, and continuously monitor the authorised route for normal operations. The HORUS system is being developed to fulfil this role by integrating multiple sensor data in real-time to allow a comparison with the previous record of any given track section using neural network processing in an on-board processor. HORUS then provides the functionality to apply software processes to conduct the computationally intensive algorithms for localisation, object detection, awareness, dynamics and route monitoring.
The total system also requires a central data centre that collects as-run video, or sensor data more generally, that is then used to update the HORUS stylised track reference record called a “Master Sequence” for the route travelled. The overall process uses ‘deep learning’ techniques to ‘learn’ any objects or route changes and update the Master Sequence. On-board a locomotive, the differences allow the locomotive’s HORUS processor to recognise both normal and abnormal operations for any given track location.
Deep Learning Processing Pipeline
Deep Learning plays a vital role through the entire HORUS processing pipeline. In this way, computers are not simply programmed on what to expect along a route, but also with processes for learning route knowledge enabling the ‘artificial intelligence’ system to become increasingly ‘smarter’ based on experience. The HORUS processing pipeline involves:
- Detection. The detection process uses multiple sensors to understand the environment around the locomotive;
- Localisation. The localisation process uses what’s measured to create a detailed localisation map of where the locomotive is situated with regard to its surroundings;
- Awareness. The awareness process interprets the real-time virtual 3D environment around the locomotive for hazards or differences as relevant to operations;
- Dynamics. The dynamics process calculates how to drive smoothly and within movement authorities given the track infrastructure condition, the weather environmental conditions, and the dynamic characteristics of the train
- Monitoring. The monitoring process continuously checks that the train is operating within authorised parameters of the locomotive’s movement authority, that all alarms have been actioned, and that the locomotive is otherwise operating normally.
HORUS ADAS Capabilities
The HORUS processing pipeline allows many operational ADAS capabilities, including;
- location definition to a specific rail line or place along a rail line
- Signal Lamp Post recognition and reading of displayed aspect
- Speed sign recognition and reading of speed
- Detection of an unexpected object or hazard on or about the rail track
- Monitoring of rail infrastructure
- Unexpected rail geometry
Notwithstanding that the HORUS technology has been designed to evolve to more automation in the future, it is not currently configured to operate as a fully automated self-driving system. HORUS is in-effect currently configured as a very comprehensive driver advisory system with the purpose of moving routine route monitoring tasks to a computerised assistant.
However, the underlying technology being used by HORUS will allow evolution to autonomous operations on selected lines as more confidence is obtained in the technology. This is not simply a technical issue. The safety case needs to be made with extensive data and operational experience, which will come over time. Technology itself will also evolve, so an autonomous future is inevitable once the safety achieved with the technology exceeds that provided by a human driver.
Business Benefits
The technology delivers some important business benefits:
- Reduced Infrastructure.
Users can identify their operating track and location with very high confidence, meaning that track-based location methods using track-circuit detection are no longer needed in all situations.
- Improved Safety.
Train operating safety is improved because HORUS does not get tired or distracted. It performs its set monitoring and train protection functions 24/7 relentlessly. - Continuous Improvement.
With continuous machine learning, HORUS also improves in capability the more times a route is operated. The track and route data files are updated with every train movement. Such continuous learning does mean that HORUS needs a suitably scaled back-office computer centre to support network operations. - Shared Route Knowledge.
HORUS can share any hazard or track difference with any other locomotive via communications, without that other locomotive having to operate the track first. All route knowledge is shareable in real time without further learning. This is an extremely efficient capability and differs significantly from human learning. - Separation of Train Control from Infrastructure.
Because all track identification, hazard detection and operations monitoring is conducted on-board the locomotive using its AI Agent, a locomotive does not need to be restricted to a specific track area if it is loaded with the correct route data file for a given route. This method of on-board management allows a train to operate as per its movement authority controlled entirely by an above-rail process. In this manner, a movement authority has the potential to operate across different types of infrastructure. In essence, below-rail train control by track-side signalling is replaced by above-rail, on-board control as per a digital movement authority issued by the below-rail access provider.
While autonomous trains will certainly eventuate, probably on closed vertically integrated, access-limited systems first, the technology is currently focussed on improving safety. Detailed safety assessments based on extensive operational experience and data is needed to evolve to autonomous operations. Nevertheless, it is inevitable that autonomous trains are coming. AI and machine learning represent a cost-effective way for the rail industry to immediately obtain a safety benefit from current car-based AI technology while assisting in the development of corporate knowledge in the important field of AI and machine learning on rail generally. Operational data will be collected and compared to computerised models over time to establish the maturity of the technology for more advanced automated roles, but those roles and any safety case will be addressed in a future evolution of the technology.
– Derel Wust – Managing Director
