BRED: Minimizing ATM Downtime With Machine Learning

Global business information provider IHS Markit predicts that the business value of AI in banking will reach $300 billion by 2030. To be sure, there is no shortage of use cases when it comes to data science, machine learning, and AI in the banking and financial services sector. Instead, the larger challenge is often choosing the use cases that will prove to be most successful as a data-driven initiative based on potential impact.

We spoke to members of The Data Factory — including Bertrand Ring, Head of Data Service — at BRED to understand how they took a previously untapped wealth of available data on the bank’s more than 600 ATMs and used it to provide concrete value to the business: minimizing downtime by monitoring repair status as well as predicting possible failure before it occurs.

Challenges & Solution

The Data Factory at BRED built a system in production that monitors the status of its more than 600 ATMS, delivering outage information in a meaningful way for business teams to take action (regarding maintenance allocation and activity, etc.).

• The data: The model uses outage data coming from the ATMs themselves, including the state of its different components (like the network connection, card reader, etc.). Through initial analysis in Dataiku, the team found that this data was quite imbalanced: for example, only 20 percent of outages lasted longer than 15 minutes, while another 20 percent lasted for less than four seconds and 40 percent less than two minutes.
• The model: Leveraging a clustering algorithm in Dataiku, the team found that a large group of ATMs (more than 70 percent) rarely have problems. Other clusters allowed the team to identify specific technical defects (e.g., with the card reader or with network connection) with certain ATMs. But ultimately, the project uses a mix of models — clustering, survival analysis, and time series — in its final solution.
• The solution: The solution that The Data Factory at BRED delivered alerts the business team when maintenance needs to be sent to repair an ATM based on thresholds – for example, unavailability that lasts a few seconds doesn’t create false alerts. They also delivered dashboards for predictive maintenance of ATMs.  

Keys to Success 

There were a few key elements integral to La BRED’s success in building their predictive ATM maintenance solution:

1. Center of excellence model: In order to bootstrap its ability to innovate and work quickly using data, BRED created The Data Factory in 2018, and in 2019, it became a center of excellence for data science within the company. This unit is responsible for not only underlying data architecture and administration of data tools, but also for developing and driving data science projects in collaboration with lines of business, with a focus on production.
2. Clear business question and scope: The Data Factory at BRED is successful because when tackling a use case, they start with one clear business question or objective with a relatively small scope, and they keep the project focused and contained. In this example, their objective was to limit downtime of ATMs by identifying those most likely to break down and providing a way for the business to monitor ATM downtime (before, during, and after maintenance occurs).
3. Focus on deliverables: In addition to choosing a clear business problem, the team also was successful because they took time upfront to understand the needs of the business when it came to final deliverables. In this example, they identified three key needs: statistical analysis of the data, a predictive layer, and dashboards for final consumption by business teams. Importantly, they presented initial results of their project for feedback before delivering a final solution to ensure that deliverables aligned with expectations.
4. Gain time with the right tools: Upon creation, The Data Factory at BRED was using a host of different tools and solutions strung together to take a data project from raw data to production. In late 2018, they turned to Dataiku, which allows them to conduct quicker data analysis, pre-processing, and hyper-parameter tuning plus an easy way to operationalize and productionalize data projects while still remaining flexible enough to allow for coding or importing external libraries.

Results and Next Steps

In addition to building a solution that makes it easier for business teams to maintain ATM uptime overall, thanks to Dataiku, The Data Factory at BRED was also able to separate the global project into sub-projects, offering more granularity on ATM maintenance according to specific geography, maintenance team, and more.

Next, The Data Factory at BRED will work on optimizing the model’s performance as well as testing new Dataiku features for opportunities to introduce more automation. Because the project allowed the team to help the entire enterprise understand the value of data science, they will also continue to tackle new use cases from other lines of business for even more impact.