UC Davis

The advent of high-precision plasma diagnostics tools has created vast amounts of data to be processed and, as a result, increased demand for tools that automate data analysis. One common task in analyzing plasma data is processing spectrogram data to find structures that correspond to physical processes that have occurred, one such example being the toroidal Alfven eigenmode. Due to the extreme environments of the plasma, the data collected are often noisy, thus making automatic detection of structures in the data difficult.This thesis presents a pipeline that takes in raw data from a channel of an Electron Cyclotron Emission Imaging (ECEI) system, which then detects and processes any excited Alfven eigenmodes. Each detected mode is tagged and individually filtered out in the pipeline. The main feature of the pipeline is a deep denoising model. The deep denoising model is created to combat the noise in the spectrograms and enhance the detected signal. A deep denoising architecture is chosen because of its strong ability to denoise images without blurring, unlike traditional denoising models. The architecture used is a Multi-Wavelet Convolutional Neural Network which was then trained using synthetic data. Synthetic data were used due to limited availability of clean measured ECEI data. When applied to real ECEI data, the denoising model achieves significant background denoising, at the cost of occasional hallucinated or spurious structures. The pipeline, in total, has nine steps to process and filter the raw ECEI data. The steps of the pipeline are: time domain low-pass filtering, short-time Fourier transform, spectrogram denoising, thresholding, morphological thinning, trace clustering, trace separation, filter mask creation, mask filtering, and inverse short-time Fourier transform. The combination of the operations in the pipeline allows for the automatic detection and filtering of signals in the spectrogram with a strong rejection of noise.