Onboard Data Acquisition: Data is collected and pre-processed onboard the vehicle, prioritizing data collection and transmission to minimize communication overhead. This approach focuses on capturing valuable information for higher-level CORVETTE services and novelty. Contextualized data for advanced CORVETTE services is generated using Statistical Machine Learning (ML) techniques, such as Autoencoders and Gaussian Mixture Models, along with preprocessing methods. The primary goal is to ensure robust and stable ML methods, enabling reliable detection even with variations in input data. Privacy preservation is achieved through ML-based data preprocessing.
Backend Data Processing: The central CORVETTE backend efficiently manages the Onboard Devices, handling device management and aggregating all collected data for higher-level services. The project aims to achieve a flexible and scalable backend, implementing routing mechanisms for automatic provision of specific services for incoming Onboard events and context-sensitive data processing. Multimodal ML methods will be designed to automatically assign incoming Onboard data to situation classes. The backend will support individual Onboard ML model training, where existing Automated Machine Learning Frameworks will be extended for automated model generation, with a focus on adapting models for embedded ML on Onboard Devices (e.g., Pruning). We will provide suitable visualizations and Northbound interfaces for easy data access and seamless integration with higher-level systems.
Demonstration and Testing: The developed methods in CORVETTE will undergo thorough testing and demonstration through various implemented use cases, showcasing their adaptability across different scenarios. Agile and iterative method development, based on real vehicle data, will ensure continuous evaluation. The project will culminate in a long-term test, where the methods and the CORVETTE Framework will be showcased.

The project adopts a scientific approach, with a primary focus on developing systematic methods for integrating and utilizing real-time vehicle data in higher-level services, both at the vehicle level and fleet level.
The following topics are investigated:
• ML-based anomaly detection and novelty detection in multimodal measurement time series
• Automatic ML-based contextualization and preprocessing of measurement time series
• Privacy-aware ML
• Methods for improved ML Robustness & Reliability
• SOA / Microservice Service Architectures
• Automated generation of ML models
• Methods for automated model transformation (Embedded ML)

The research contributes to the field of industrial applications of machine learning, specifically addressing domain-specific challenges posed by multimodal sensor data. This includes tackling issues related to robustness, privacy, and changepoint/novelty detection, tailored to the unique requirements of the industry.
The achievement of the goals becomes visible by:
• Robust and adaptive ML models for anomaly or novelty detection in vehicle sensor data.
• Situation-specific generation of ML classification models
• Prototype CORVETTE onboard measurement hardware
• Prototype CORVETTE Backend
• Demonstration of the CORVETTE concept in the context of two defined use cases (Automated model generation, OnDevice data acquisition / recognition, Automated classification of incoming OnDevice measurement time series on fleet level, Higher-level service integration)
In the project's initial phase, the software infrastructure was successfully established, enabling effortless integration of new applications within Docker containers. This modular approach ensures easy addition of functionalities. Applications can be conveniently downloaded and updated directly on onboard devices in the Docker environment.
A unified software system remains consistently available, accessible at system startup regardless of the number of vehicles or onboard devices. Several implemented use case applications are currently operational on the devices. These applications include:
• Digit Dashboard Detection – QR code approach to identify regions of interest (ROIs)
• Weather and Tunnel Detection
• Anomaly Detection – Monitoring acceleration data
• Up/Download of recorded data to a backend where data analysis algorithms are applied
• Retraining of ML models in the backend which can be updated on the edge device in runtime