ecoexperts – Tunnel Sensor Fusion: Implementation of Machine Learning Models in Critical Infrastructure

ecoexperts is a leading infrastructure automation company specializing in the development and implementation of state-of-the-art control systems for critical infrastructure. With a focus on tunnels, traffic control and energy farms, ecoexperts ensures continuous availability, protection against cyber attacks and future-proof scalability. Since its foundation in Austria in 2011, the company has successfully implemented major infrastructure projects and is a key partner of ASFINAG, the Austrian government’s road infrastructure operator. ecoexperts stands for sustainability, innovation and customized solutions to create efficient and secure infrastructures.

A modern highway tunnel contains a large number of different sensors, such as cameras for video detection, microphones, counting loops, CO measuring devices, etc. (in this case, there are over 350 different sensors over a distance of more than two kilometers). Each of these sensors detects specific events in the tunnel, such as slow vehicles, stationary vehicles or accidents. Due to various factors, the individual sensors report a large number of false-positive events (events without an actual problem in the tunnel). The tunnel control center staff, in turn, must respond to each of these messages and take appropriate action (close the tunnel, slow the speed, etc.).

Our goal was to use a machine learning model to significantly reduce the number of false positive reports without reducing the number of the true positive ones. The challenge was to appropriately combine the information from the different sensors and deliver a more accurate event notification for the control center.

To work out the data and system architecture, we started with a data thinking workshop. In this format, we first analyze the business processes and data in terms of potential use cases and then create a detailed project plan, with necessary steps, required resources and expected timeframe.

After coming up with the basic data architecture, we implemented data pipelines to convert the raw sensor data into a suitable target format for a machine learning model.

In the next step, we tested and evaluated different machine learning models. Choosing the right evaluation metric played a crucial role in model selection, as Accuracy, for example, weights false-positive reports equally to false-negative ones.

Git was used for version control of the code and MLFlow was used for versioning of models, artifacts, and metrics. APIs for the models were implemented via fastAPI and via MLFlow the production-ready models were eventually made available to the API.