Neural Ordinary Differential Equations Auto-Encoder for Fault Detection in Process Systems

With the advent of Industry 4.0, there has been a paradigm shift in the operations of the manufacturing and industrial sectors. The fourth Industrial revolution has compelled the industries to integrate Machine Learning with its core processes. Over the years, data-driven approaches have been a key contributor to the smooth functioning of a process plant. In the proposed work, Neural ordinary differential equations Auto-encoder being a special class of neural networks are utilized for fault detection. The proposed methodology consists of a neural network coupled with an ordinary differential equation solver. The neural network is used to parameterize the derivatives of the hidden states and results in a continuous transformation of the states. For time series processes, the methodology proves to be of greater use because of its inherent ability of extrapolating the values for a particular time step. The proposed methodology was validated for the Tennessee Estman process and yielded better results when compared to the different deep learning-based models used for time series predictions in LSTM and its variants, and machine learning framework such as Dynamic Principal Component Analysis (DPCA).