Second-order component analysis for fault detection

Process monitoring based on neural networks is getting more and attention. Compared with classical networks, high-order have natural advantages in dealing heteroscedastic data. However, might bring the risk of overfitting, which learning both key information from original data noises or anomalies. Orthogonal constraints can greatly reduce correlations between extracted features, thereby reducing overfitting risk. This paper proposes a novel fault detection method called second-order component analysis (SCA). SCA rules out heteroscedasticity process by optimizing autoencoder orthogonal constraints. In order to deal this constrained optimization problem, geometric conjugate gradient algorithm adopted paper, performs combination Stiefel manifold Euclidean manifold. Extensive experiments Tennessee-Eastman benchmark show that outperforms compared state-of-the-art methods missed rate (MDR) false alarm (FAR).