Data evaluation based on machine learning

Fully autonomous systems (e.g. vehicles, robots and drones) will no longer be able to return the system control to humans in case of detected problems. Accordingly, they need a kind of intelligence to carry out independent analysis in unknown environments/situations, to plan actions and to act safely according to decisions made.

Automatic detection of new and relevant data sets based on machine learning

Fully autonomous systems (e.g. vehicles, robots and drones) will no longer be able to return the system control to humans in case of detected problems. Accordingly, they need a kind of intelligence to carry out independent analysis in unknown environments/situations, to plan actions and to act safely according to decisions made.

Challenge

Autonomous systems need a precise model of their environment for independent planning and safe actions. The safeguarding of the sensors and environmental detection is crucial. In the development of environmental detection systems, artificial intelligence (AI)-based software systems, which rely on "machine learning" (ML), are now dominating. However, the amount of data required for the training of ML-based systems is immense. For example, 6.62 billion test kilometers are required to secure autonomous driving functions of vehicles, 1 which makes data storage and processing extremely complex. Currently available methods manually filter out the relevant data from the total amount of data, which is a time-consuming and cost-intensive process.

Our solution

This patent application presents a novel data reduction method for the training of ML-based functions (Fig. 1), developed by Prof. A. Rausch (TU Clausthal). With the data from an existing data pool, a so-called semantic discriminator is trained. It is based on an autoencoder network and can evaluate whether the next incoming data stream is already known or whether it represents a substantial extension of the existing data pool. In this way, the originally very large data pool can be significantly reduced. However, if substantial new data is fed into the auto-encoder, this is detected by a differ during decoding and the data pool is expanded.

Festoxidbrennstoffzelle in SchnittdarstellungFig. 1: Architecture and process for continuous monitoring of data sets based on a dependability cage approach with ML-based semantic novelty determination (adapted from patent application).

Advantages

  • Memory volume (data pool) significantly reduced
  • Data processing accelerated/facilitated
  • Efficient recognition of new and relevant data sets
  • Substantial expansion of data bases
  • Time, energy and cost savings

Applications

  • Vehicles
  • Robots
  • Drones

Development status

The functionality was successfully tested.

Patent situation

European patent application filed.

Patent owner

Technical University Clausthal

Contact

Dr. Mirza Mackovic
Patent Manager Technology
E-Mail: This email address is being protected from spambots. You need JavaScript enabled to view it.
Tel.: +49 551 30724 153
Reference: CPA-2250-T270

Tags: Software und Algorithmen

Search

A subsidiary of

Georg-August-Universität Göttingen