Deep neural networks for time series classification

• Main Author: Cheng, Wen Xin
• Other Authors: Ponnuthurai Nagaratnam Suganthan
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/170362
• Institution: Nanyang Technological University

Recently, deep neural networks are getting popular in various classification problems. Random Vector Functional Links (RVFL) are proposed for many tasks, such as classification, forecasting and visual tracking. Also, many Convolutional Neural Networks are being proposed for other tasks such as image classification (CNN), image segmentation, object detection, speech recognition, biomedical tasks and time series classification. However, deep neural networks are not extensively explored in time series classification. RVFL is not able to extract useful local features in time series and performs poorly in time series classification. Deep CNNs often have many trainable weights and require a lot of training samples for training. Deep CNNs often contain multiple pooling layers that scale down the extracted features' length. Therefore, deep convolutional neural networks are unsuitable for short time series signals. Also, since stock market trends rapidly change, models must be retrained to learn new patterns. Therefore, this thesis focuses on enhancing deep neural networks for time series classification. Firstly, we compensate for RVFL's inability to extract meaningful local patterns by extracting features from trained Residual Networks. We proposed feature selection for edRVFL, which picks useful features and filters out redundant features for direct links. Next, we propose adaptive scaling in U-Net to allow the network to adapt to different lengths of time series signals from different time series classification tasks. Then we propose Deep Ensemble Randomized Convolutional Network for EEG-based emotion recognition, which is effective in classifying such signals while requiring less computation effort. Finally, propose Dynamic Ensemble and Online Learning for RVFL ensembles to adapt to the rapidly changing trends stock market.