An adaptive spatiotemporal feature learning approach for fault diagnosis in complex systems

The machine fault diagnosis is being considered in a larger-scale complex system with numerous measurements from diverse subsystems or components, where the collected data is with disparate characteristics and needs more prevailing methods for data preprocessing, feature extraction and selection. This work presents a novel diagnosis framework that combines the spatiotemporal pattern network (STPN) approach with convolutional neural networks (CNN) to build a hybrid ST-CNN scheme. The proposed framework is tested on two data sets for diagnosing unseen operating conditions and fault severities respectively, to evaluate its generalization ability, which is essential for the application in machine fault diagnosis as not all of the aforementioned scenarios have sufficient labeled data to train a model. The results show that the proposed ST-CNN framework outperforms or is comparable to shallow methods (support vector machine and random forest) and 1D CNN. Through visualizing the activations, it is verified that the spatial features can elevate the diagnosis accuracy, and more general features are determined by the proposed approach to form an adaptive classifier for diverse operating conditions and different fault severities.