Our scope of services
Artificial intelligence methods, in particular machine learning, allow computer systems to improve themselves automatically on the basis of sample data. Explicit programming is no longer required, not only allowing the performance of already existing functions to be optimized but also enabling access to entirely new activities.
At Fraunhofer ISST, application-specific AI and ML pipelines (i.e. interlinked processing steps) are researched and developed for the areas of healthcare, logistics and data management. Depending on the application, these can be based on different data sources such as biosignals (such as measurements using EKG or 3D acceleration sensors), audio, images, videos, texts or a combination of several data sources. We support our partners along the entire pipeline, starting with the identification and feasibility assessment of ML application scenarios, through the preprocessing of the (raw) data to the selection and training of suitable models and their evaluation based on application-specific performance criteria. Depending on the requirements, classical learning methods (such as Support Vector Machines, Decision Tree) are used as well as Deep Learning approaches.
The services offered by Fraunhofer ISST include identification and feasibility assessment of machine learning (ML) deployment scenarios, data pre-processing, selection and hyperparameter optimization of suitable ML models and their evaluation.
Identification and feasibility assessment of ML application scenarios
- Identification of application-specific optimization potentials as well as new areas of responsibility through the use of AI or ML
- Feasibility assessment of the use of AI or ML based on the available data
Preliminary processing of data
- Clean up the raw data
- Feature calculation based on biosignal data (e.g. 3D acceleration, EKG, audio) from the time and frequency domains
- Use of feature selection and extraction methods
Training and evaluation of ML models
- Conception of AI/ML-based applications
- Selection from different learning approaches, e.g. classic classification methods, deep learning, association analysis and clusters
- Hyper-parameter optimization, evaluation based on application-specific performance metrics
- Use of windowing methods, time series analysis
- Use of techniques for dealing with imbalanced data, e.g. data augmentation, cost-sensitive classification
Artificial intelligence methods, in particular machine learning, can be used to support various industries or completely open up new areas of responsibility. Whether for automated quality control in logistics, the diagnosis of diseases, the real-time detection of critical situations in the health sector or for the extraction of information from documents, the possibilities are only limited by the availability of data.