Incremental Learning based Broad Canonical Correlation Analysis Network for Industrial Process Monitoring

This paper proposes an incremental learning based broad canonical correlation analysis network (ILBCCAN) to address the time-consuming problem of deep learning models in training process. It uses canonical correlation window, mapping window, and enhancement window to extract the correlation and nonlinear structure between variables, thus improves the feature extraction capability. Moreover, it continues to maintain the advantage of the broad learning system in terms of time overhead by using a pseudo-inverse strategy to determine the network parameters, which greatly reduces the training time. In addition, the proposed ILBCCAN model has an incremental learning capability. In scenarios where high accuracy is required or where the network structure needs to be adjusted online for real-time updating, the number of node parameters is flexibly determined by dynamically adding enhancement nodes within the incremental window, which avoids the need to retrain the entire network from scratch for achieving the desired level. Finally, the ILBCCAN model is used for fault monitoring. Case studies show that the proposed model outperforms other fault monitoring methods based on Tennessee Eastman benchmark.