Deep learning based sensory data analytics

• Main Author: Lee, Jing Yang
• Other Authors: Mao Kezhi
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Computer science and engineering::Information systems; Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/141926
• Institution: Nanyang Technological University

This FYP project constitutes developing and evaluating deep learning models for 2 primary tasks – Remaining Useful Life (RUL) prediction and News Popularity prediction. Remaining Useful Life (RUL) prediction of industrial systems/components helps to reduce the risk of system failure as well as facilitates efficient and flexible maintenance strategies. In this project, an architecture comprising a Dilated Convolutional Neural Network, which utilises non-causal dilations, combined with a Long Short-Term Memory Net-work: DCNN-LSTM model for RUL prediction is proposed. This model was validated on the publicly available NASA turbofan dataset and its performance was benchmarked against previously proposed models, showing the improvement by our proposed model. Next, DCNN-LSTM model was used in a transfer learning setting where both the issues of model retraining and limited availability of experimental data were addressed. The results showed a significant reduction in training time compared to the time required for retraining models from scratch, while achieving similar performance. Moreover, the proposed method also achieved at par performance by utilizing a much smaller amount of data. A Self-Attention model was also developed and benchmarked on the same RUL prediction dataset, outperforming previously proposed models. Furthermore, a GAN was also implemented to augment existing datasets with synthetic data. Online news popularity prediction is extremely relevant to news organizations. Accurate news popularity prediction would allow news organizations to gauge the popularity of their news items prior to release. This would allow them to boost readership by modifying the news items in order to maximize the predicted popularity. In this project, an hierarchal multitask deep learning model which leverages the natural language understanding capability of pretrained language models for concurrent topic detection, sentiment analysis and news popularity detection based on news headlines and titles is implemented. The News Popularity in Multiple Social Media Platforms dataset provided by UCI is used in this project. We evaluate the performance of our model on all 3 tasks, demonstrating the effectiveness of hierarchal multitask training as well as the effect of different language models (BERT, RoBERTa and ALBERT) on overall performance of the model.