Periodic Reporting for period 1 - DATAMINE4.0 (Advanced Data Modeling and Analysis Applied to next Generation Industry 4.0 settings and the Internet of Things)
Período documentado: 2018-01-22 hasta 2020-01-21
We considered the following applications: 1) food processing - detection of foreign matter in food with help of hyper-spectral imaging, 2) agriculture - forecast of the rice blast disease, and 3) human-robot collaborative application - object detection problem. These applications require very high prediction accuracy, and even a small improvement in the accuracy can have a high impact in a long run (for example, less food waste in food processing application).
1. We developed a novel machine learning model for anomaly detection in hyperspectral images for the application of the detection of foreign matter in food. The model uses the recent advances in the unsupervised machine learning modelling, in particular the use of autoencoders for the purpose of dimensionality reduction.This model was tested in a pre-industrial setting with an IRIS proprietary hyperspectral camera. The tests have shown an improvement over the methods traditionally used in chemometrics.
2.We trained a forecasting model for the detection of the rice blast disease. The model is based on a recurrent neural network, which is one of the state-of-the art models for time series data. Our approach has shown an advantage in several cases over the 'process-based' models traditionally used for this application.
3. We trained an object detection model for a robotic application. We used one of the state-of-the-art model called YOLOv3. The model is a trained on artificial images obtained by a simulator of a robotic application. The study confirmed usefulness of the YOLOv3 model for this kind of images.
The results were disseminated through one published conference paper, submission of two journal papers of which one is published and another one is under the revision process, and a poster presentation on a conference. All mentioned scientific papers were peer-reviewed.