Dipartimento di Informatica

Attenzione: Contattare i docenti (Idilio Drago e Elvio Amparore) prima di porre la propria candidatura all'azienda.

Industrial robotics offer many challenges in several digital areas that are highly relevant for the development of innovative functionalities and advanced solutions. These theses will be developed in collaboration with COMAU S.p.A., and is organized around multiple challenges/topics that are highly relevant in data analysis and AI-based applications: 

1 - Synthetic dataset generation

To ensure high accuracy of ML algorithms, one of the main difficulties is related with the lack of data to train the neural networks supporting the systems. Recently multiple algorithms have been proposed for producing synthetic datasets, such as generative neural networks, autoencoders, and others. These techniques are usually combined with classic image processing approaches, such as to vary lighting conditions, to change background colors or to create cluttered scenes. 

This topic aims at developing a data augmentation framework for robotics applications. The goal is to evaluate different approaches for artificial data generation. The best techniques will be incorporated in a proof-of-concept system to generate synthetic data from few real images or CAD. The application of the proof-of-concept system will produce extended and representative datasets to be used in the training of AI models for robotics applications. 

2 - Object detection and/or instance segmentation from 2D images/videos

To address the new challenges in industrial automation, object detection has been used to recognize and localize the objects in 2D images. This topic consists in starting from already designed and fully-trained neural networks (like ResNet or YOLO) or just fine tuning them to reach the best accuracy for a specific robotic-oriented task. The application sectors are logistics, food&beverage and agribots. Moreover, a look at the instance segmentation problem will also be considered, if the available data allows for this problem to be explored.

3 - Multi-modal anomaly detection on manufacturing chains 

Anomaly detection consists of finding elements that diverge from what is considered "normal". Anomaly detection supports assembly lines and manufacturing chains by providing means for detecting problems in products, which otherwise would be invisible to human experts. This thesis consists of the design and development of systems to support the decision of qualified personnel during the quality assurance test on the assembly lines and manufacturing chains, which is cumbersome and error-prone. Machine and deep learning algorithms – using multi-model data, such as images and time-series data –  allow the detection of defects even when not visible to naked eye. The student will research techniques to classify and segment the anomaly extent as well as to predict the risk of failure.

4 - Time-series analysis of bigdata

Time series analysis has the objective to highlight the underlying causes of trends or systemic patterns over time, e.g., to predict the likelihood of future events with time series forecasting. The aim of the thesis is the creation of tools able to analyze multi-dimensional time-series present in big datasets to extract valuable information from these series. 

The datasets consist of product data, sensor data, and the production data (e.g. cycle time). The challenges to be addressed in the thesis come from the multi-dimensional and high volume of the datasets, which challenge classic solutions for time-series analysis.