Outdoor sparse sensing and superresolution of acoustic fields
||Outdoor sparse sensing and superresolution of acoustic fields|
||Elias Zea <firstname.lastname@example.org>|
||Kungliga Tekniska högskolan|
||2023-11-04 – 2024-06-01|
In this project, we will train artificial neural networks for the application of sparse sensing and superresolution of acoustic fields outdoors. The development of models to predict outdoor noise and the associated sources can enormously impact the well-being of communities exposed to high sound levels.
The project's overarching goal is to quantify the impact of noise pollution caused by air and land transportation near airports by transferring knowledge from canonical cases of acoustic fields measured inside rooms. As a first step, we will train and deploy convolutional neural networks and autoencoders to improve the spatial resolution (i.e., the number of microphones) and investigate the modal structure of the fields. As a second step, we will: (i) directly deploy the trained networks on field measurements near Uppländs Väsby and (ii) fine-tune and deploy the networks on such measurements. The problem can be solved for small training datasets without GPU capabilities, but for larger sets—especially outdoor scenarios—the situation is highly computationally intensive. Moreover, we are interested in testing the performance of the networks against various hyperparameters (e.g., depth, input/output resolution, types of activation functions, training set size, etc.), for which the HPC capabilities offered by Berzelius will be vital. Additionally, as we want to test larger architectures (e.g., GANs), GPU acceleration will be critical to speed up the training and tuning of the hyperparameters. We will use the project storage to load, save and access the data. We expect the project to start as soon as possible and be finished by the Fall of 2023.