Energy Out-of-distribution Based Fault Detection of Multivariate Time-series Data

A major challenge faced by the chemical process industry is carrying out operations safely and safely. The proposed work entails a fault detection approach for a multivariate time series dataset by utilizing the energy scores instead of the traditional approach. This work proposes a loss function which utilizes the concept of in-distribution and out of the distribution of data. Energy scores are more theoretically aligned with the probability density of the inputs and can be used as a scoring function. For a pre-trained neural network, energy can be utilized as a scoring function and can also be used as a trainable cost function. The concept of out-of-distribution is similar to that of any outlier identification method. Similarly, for energy out of distribution, an energy value which falls below a certain threshold can be considered an outlier and is addressed as out-of-distribution. The values within the range are in-distribution. Higher energy values imply a lower likelihood of occurrence and vice versa. The proposed approach is compared with different deep learning approaches like Auto-encoders (AEs), LSTMs and LSTM-AEs that are traditionally used for anomaly detection and utilize the softmax scores. The Proposed methodology is also compared with some state-of-the-art fault detection methods, such as the PCA and DPCA and returns encouraging results. Energy based out of distribution is coupled with various deep learning methods to identify faulty and normal points. When teamed with the Auto-encoder network, energy-based scoring proved to be of significant dominance compared to other methods. The study was validated for the benchmark Tennessee Eastman data for fault detection.