Ensuring data speed and resiliency in a mobile IoT application

Control is a telemetry company that has carved a niche in race car data connectivity. They have showcased their expertise from iconic race events like Daytona to the legendary Le Mans and Nurburgring. Their primary focus is ensuring that data from racing cars is transmitted flawlessly and seamlessly to engineers situated around the globe, regardless of distance or geographical challenges.

The opportunity

Ensuring optimal network connectivity in racing telemetry

Control supplies racing teams and manufacturers with race-winning telemetry solutions. These cellular devices contain three modems, allowing the device to be attached to up to three mobile networks simultaneously. The device can automatically switch between these modems up to 15 times per second during the race to optimize connectivity.

The problem was that when a device arrived at a new racetrack, the modems automatically connected to the first available network, not the best available network. This could lead to performance degradation of up to 23% compared to an optimal condition.

Initially, this problem was solved manually. A Control engineer would log in to an online portal, review the historical signal quality for the location, and set a priority network for the device. The problem with manual configurations:

• They take significant engineering resources, about an hour per session to configure 30 cars
• Devices can be missed, leading to support calls and a scramble to reconfigure
• Customers attending unsupported test days can suffer needless performance degradation, leading to support calls and manual reconfiguration
• A sudden change in network conditions, such as an operator outage, requires manual remediation mid-session

To improve their product and reduce operational costs, Control needed:

1. A way to automate the configuration of their IoT devices
2. A way to monitor, test, optimize and edit the configuration while the device is in operation

The solution

Machine learning for automatic network performance detection and optimization

Control partnered with Quix to build and deploy an ML model that could automatically detect network performance in real time and optimize connectivity by automatically updating the device configuration.

However, machine learning models are rarely simple and often take months or even years to deploy. A Dotscience report revealed that 64.4% of the organizations surveyed take 7–18 months to develop ML and AI models from idea to production.

Control needed to build and execute this project a whole lot faster. So they turned to Quix’s production-ready stream processing platform to help them quickly build, train, test and deploy a fast and resilient ML system.

Quix’s solution called for these steps:

• Connect Quix to Control’s existing Azure Event Hub infrastructure to receive data streams
• Transform raw messages into a structured data schema with standard business semantics
• Contextualize data, so each car is associated with just one data stream
• Store high-quality data from actual vehicles and train ML models on that data
• Deploy the models into a stream processing pipeline to rank the best connectivity
• Automatically update a configuration table in real time to optimize connectivity

Because Control’s devices operate in various locations, Quix’s solution created 82 machine learning models — one for each venue and transmission mode — to tailor the solution to each environment. The models rank all available networks, weigh multiple variables and deliver a single recommendation for the best network at any point.

This didn’t just work for real-time data. The ML models were also able to predict real-time network performance using historical data. And when new venues are added, the ML script alerts engineers that these also must be added to the set of models.

This result is populated to live streams to enable automation — continuously ranking the networks to maximize speed and connectivity.

The outcome

Significant operational savings with enhanced performance reliability

Control’s use of Quix has enabled the company to update each car’s connectivity automatically. Previously, a Control engineer had to configure and test each car for each session manually. Now, the vehicles configure themselves.

This took a massive operational burden off Control engineers’ shoulders while eliminating the risk that a customer would needlessly suffer up to 23% performance degradation.

Control’s journey to developing and deploying ML models seemed, at first, like an impossibility. The company’s small team leaned heavily toward software engineers rather than data scientists, and they found it to be too costly and time consuming to build on Microsoft Azure.

In partnership with Quix, Control stood up, trained, tested, and deployed 82 ML models in just two weeks. In addition, Quix provided Control with a more resilient data pipeline by replicating and sharding the data and the processing. This gave Control confidence that their new data product would perform consistently without adding resources to monitor and maintain the system.