Capitalising on artificial intelligence for maintenance scheduling
By Angela Tam
If Hollywood is to be believed, robots as good-looking as Jude Law will one day populate the world much as they do in the film "Artificial Intelligence", acting and thinking like humans and, well, being glamorous like movie stars. In today's world, though, artificial intelligence (AI) is a far more prosaic concept which has yet to be combined with the creations of robotics experts and the artists at Madame Tussaud's.
AI does, however, provide effective solutions for practical problems. Its range of applications is not always obvious, which is why an AI engine developed by the City University of Hong Kong (CityU) for the MTR Corporation has won it the American Association of Artificial Intelligence's Innovative Application of Artificial Intelligence Award 2005 in the "deployed applications" category.
What CityU researchers developed for MTR using AI is a scheduling tool that streamlines the management and allocation of engineering resources for the MTR network, which comprises the five urban lines, the Airport Express and the Disneyland Resort Line, which starts taking passengers to the theme park this month.
Managing the complex and sometimes competing demands for engineering resources on a busy subway network is a time-consuming and challenging task. Everyday there are numerous requests for various maintenance and installation work which has to be carried out during the few short hours at night, when the network is closed, according to strict safety requirements. Each request was formally submitted in paper forms until 2003, when the submission procedure was web-enabled. These submissions number between 100 and 200 per week, with requests for routine maintenance interspersed with those calling for emergency repairs. In the past, they were reviewed and prioritised by all the managers sitting down together regularly. On a particularly eventful day, they might find themselves with several high priority requests that required the deployment of the same resources, plus more routine requests which would have to be met another time.
According to MTR infrastructure engineering manager Ir Richard Keefe, the corporation had explored the use of a traditional programming technique to manage the scheduling because it is good for data processing and management information analysis.
"However, resource planning and allocation with complex constraints cannot be handled effectively by traditional programming," Ir Keefe said.
After considering a number of vendors, the CityU team headed by Dr Andy Chun of the Department of Computer Science was finally selected to develop an AI engine for MTR.
According to Dr Chun, the team initially looked to other railway systems around the world to see how such scheduling tools were built, but discovered that no other network had attempted to use an AI engine for such a purpose. The team therefore developed the algorithm from scratch.
The AI engine streamlines the whole scheduling process starting from the form submission stage. Previously, the system could only check whether the input on forms submitted via the web was typed correctly. The AI engine does a lot more. Domain knowledge combined with AI enable it to determine whether:
- a request makes sense logically
- there are enough equipment to meet the request
- the request clashes with other requests
- there is a conflict with other work; for example, when work on the track is required when electrified work will be carried out at the same time
If there is a conflict or problem allocating the resources, the staff submitting the form will be advised to change his request, to increase its chance of acceptance.
"Due to the large volume, complexity and stringent rules for engineering works carried out every night, longer time is needed for manual scheduling and resource allocation. With the AI engine, the efficiency of the possession resource planning process has been improved by about 50% in terms of time," Ir Keefe said.
The AI engine does more. In addition to reviewing the requests in terms of resources availability and potential conflict, it also puts its access to the database to good use by automatically combining work procedures to save time, labour and resources. For example, it can deploy a works wagon, along with its loco, supervisor and driver, for two unrelated work requests, thus maximising use of limited resources.
Safety being paramount, the AI engine also simulates the entire night-time operation of the network, ensuring that all works are carried out with the required separation. This is an aspect of the works in which the engine really excels: the MTR's safety regulations are contained in a ream-thick manual that takes considerable time to leaf through. Experienced staff, of course, are well-versed in the requirements, but it makes sense to have that and other information available electronically regardless of whether those staff are on duty or not.
"There are aspects of the network that only a few people are familiar with," Dr Chun said. "For example, there are several sidings in Admiralty, of which one has to be left empty so trains can pass through. The engine puts the corporation's most precious knowledge onto the computer so it becomes available at all times."
The engine took a team of 15 programmers about one year to develop using C Sharp with .NET Framework. Spread across different countries, the team adopted what Dr Chun called an "internet-style" of distributed development to produce the engine. For Dr Chun, the most challenging aspects of the project were learning MTR's operation and finding out how its knowledge could be turned into a computer-generated form; as well as how the algorithm could be optimised to reply to a request within seconds.
To ensure its reliability, the engine was put through four months of testing using several hundred test cases derived from historical data. MTR began using it in July 2004. Using the system requires minimal training. The AI engine is purely a web service with a simple, service-oriented architecture. According to Ir Keefe, a graphical user interface was developed so that users can interactively make jobs or resource changes on the computer and see the effects of the changes. The AI engine then performs all the backend computation work and displays the results on screen.
Dr Chun said the AI engine had other applications.
"The same algorithm can be adopted by utility companies, which tend to receive many maintenance requests within a short time," he said. "I believe our AI technology can easily be applied to many different types of organisations in Hong Kong, to help them streamline operations and maximise performance and revenue."
MTR, as the first railway network in the world to adopt AI for resource management, has also built up a new, marketable area of expertise.
Said Ir Keefe: "In terms of system functionalities, the system can be used by other railway operators in the world for possession resource planning, but each railway has its unique geographical and operational features and requirements. MTR has not designed the system as a product for use in other markets, but we are willing to work with other railway operators on how to adapt the system to meet their specific needs."
Ir Keefe would not divulge the cost of the system, but said cost-effectiveness was a definite advantage.
Allocating resources for routine maintenance as well as high priority work used to be a complex, time-consuming task
Dr Chun (left) and Ir Keefe
The AI engine at work