Multifractal analysis and stacked autoencoder-based feature learning method for multivariate processes monitoring

Deep learning has become a popular tool for fault detection in industrial processes to learn complex nonlinear features. However, the features extracted from most traditional deep networks usually ignore the geometric characters and singularities of the process data. The representative features cannot be extracted effectively, which may lead to the inaccurate modeling and is not beneficial for the process monitoring. Thus, this paper proposes a feature learning method based on multifractal analysis and stacked autoencoder (MF-SAE) for fault detection in complex multivariate processes. MF-SAE can learn high-level multifractal features from the raw data in an unsupervised way. Multifractal analysis is first introduced to extract the multi-scale self-similar characteristics from industrial process data, in which sigmoid function is added for preprocessing the process data. Owe to the redundant information existing in the multi-scale feature, SAE is then utilized to learn key feature from the extracted multifractal feature. The learned hidden feature and residual feature are provided to construct the monitoring statistics. The fault detection performance of MF-SAE is tested on the Tennessee Eastman (TE) process.