This project focuses on processing real-life digital signals (time-series and images), which often exhibit a non-stationary nature. We plan to utilize advanced signal processing techniques and artificial intelligence to analyze and classify such data. The project envisages the transformation of time-series into images (time-frequency representations providing simultaneous insight into signal characteristics in both domains). Special attention will be focused on the development of adaptive, data-driven, and computationally efficient representations – which implies the design of representations with reduced interference and high resolution.
One of the factors that significantly affects the classification accuracy of such representations using machine learning is the noise, which corrupts, in almost all cases, real-world data. Therefore, we plan to incorporate newly developed denoising techniques into the classifiers, both in the time/frequency domain and in the time-frequency domain.
Special attention will be paid to regularization so that the developed method does not adapt too much to the input data and learns a trivial function that relates the noise to the data. For this reason, the model types, optimization objectives, and, finally, optimization algorithms will be carefully selected, taking into account the original distribution contained in the data (before/after transformation of the input space/feature extraction). For the purpose of this research, it is envisaged to use the most modern deep learning architectures such as Visual transformers, DenseNet architecture, U-Net, text transformers such as BERT, but also traditional machine learning methods such as Random Forest, Support Vector Machine, and K-Nearest Neighbors. A major role in the selection of methods will be played by the amount of collected data, which, in the event that it is not sufficient, will be supplemented with data generated using the Stable Diffusion and Generative Rival Networks models.
