Rio Tinto AI tool aids defect elimination in rail operations

Rio Tinto has built a tool called ReconAI to help it identify “chronic issues” affecting assets such as autonomous locomotives, and prioritise fixes.

Rio Tinto's Milan Pareek.

Senior data scientist and superintendent engineering Milan Pareek told AWS Summit Sydney that the miner wanted to maximise its use of data to improve asset reliability and availability.

ReconAI targets an engineering process called defect elimination - where asset health data is used to determine a problematic asset and put it in line for corrective action.

“Defect elimination projects are typically mechanical or electrical engineering projects, or there might be changes to maintenance tactics as well,” Pareek said.

Specifically, ReconAI takes over three of the six steps of the process - “classifying all of the faults using data from all different systems so that the engineers don’t have to do it”, and then measuring the effectiveness of corrective actions.

In doing so, Rio Tinto is removing repetitive - albeit specialist - administrative work that would otherwise be performed by engineers, giving them more time to spend on corrective action.

“That empowers the engineers to spend more time in the field actually delivering projects, working with ops and attending to breakdowns and such,” Pareek said.

The miner has tested this out for its locomotive fleet, numbered in the “hundreds”, which autonomously haul ore from pit to port on the company’s automated rail freight network.

“With such a large, remote and autonomous operation, it’s really critical for us to minimise the downtime with our assets in the field because any fault that we experience can often require operational staff to go out and actually recover the train, which increases the time to fix the fault,” Pareek said.

The company has access to a lot of data about asset equipment faults. While this includes traditional telemetry - logs, alarms, events and so on, there’s also a lot of free text data captured in SAP and another “extensively” used application called HOSDI.

“This free text data contains information that operational staff have entered when they’ve actually recovered a train, for example, or if they’ve gone out into the field to fix a fault,” Pareek said.

“SAP often contains data that’s been entered by maintainers in the workshops in terms of how they’ve diagnosed an issue, how they’ve troubleshooted it, what they’ve replaced and so on.”

ReconAI is designed to give maintenance and reliability engineers working onsite “a pristine view of all the faults that have occurred with each of our different types of assets”, and to “classify these faults to a causing equipment or a component or a subsystem of the asset, and then also to a failure mode.”

The classifications are granular - dealing with hundreds of systems and subsystems on a locomotive as just one asset type, for example - hence the historical requirement that engineers do them.

In the space of a year, ReconAI has proven effective at classifying faults based on qualitative and quantitative data, with the tool in production to prioritise defect elimination projects.

At a high level, the miner uses Palantir Foundry on AWS to collect all its rail, SAP and other related data.

“We replicate this data to S3 and then we don’t need any further data engineering because we make use of Amazon Glue and Athena to natively or directly read this data. 

“The Recon AI backend application runs in EC2 and it uses the data from each fault event as well as all of the contextual maintenance history and other data associated with the assets involved to go through a retrieval augmented generation (RAG) workflow … to classify each fault to a causing equipment and failure mode. 

“We call Anthropic Claude models - predominately Sonnet - through [Amazon] Bedrock, and we also make use of many Bedrock knowledge bases that are synced to SharePoints as well, which the engineers put documents in.

“Users access the classifications through Foundry, through the frontend Recon application.”

Pareek noted that Recon shows users “the full transparent reasoning of the entire agentic workflow from the very beginning of the workflow all the way through to the end, including what’s been retrieved and its justifications or its classifications throughout the whole process.”

“That’s critical for trust,” he said.

“Even with users who are as sceptical and harsh and critical of AI as engineers, and even for complex domains where there’s quite a lot of jargon, acronyms and business complexity, the users do actually trust it if you provide them the full transparent reasoning and bring them along for the journey.”

Pareek said a proof-of-concept started this time last year achieved 80 percent classification accuracy “on a relatively granular classification of the causing equipment or the components involved - not as deep as we go now, but still quite granular and good enough for the engineers.”

Through iterative experiments, the company managed to “push that performance up all the way to 96 percent in production for classifying locomotive faults.”

Those iterations saw the miner - among other things - adopt a new model in the form of Claude 3.5 Sonnet v2, and introduce new data as well as a new agent to the process.

The company also made changes where misclassifications occurred, and had engineers create additional documentation that could be used by the tool where gaps were identified.

Pareek said that in addition to being accepted by reliability and maintenance engineers, the tool is also contributing to improvements in data capture.

“For example, the maintenance supervisors are getting feedback on how to better capture data in SAP, so that it not only helps this process but also helps all the other foundational processes as well,” Pareek said.

Rio Tinto is in the process of applying ReconAI to more rail asset classes beyond locomotives, including ore cars, signalling, and ports conveyors.

While most effort on the technology side had focused on improving accuracy, the company has been monitoring its “cost per classification” and intends to make some optimisations around batch inferencing and prompt caching to reduce its costs.

Ry Crozier attended AWS Summit in Sydney as a guest of AWS.