A multi-objective optimization based deep feature multi-subspace partitioning method for process monitoring

Deep neural networks (DNNs) have shown advantages in dealing with complex nonlinear problems and have been applied in process monitoring. However, traditional DNN based process monitoring methods face the following problems. Firstly, only utilizing the features of the last layer to characterize industrial processes may lead to information loss. Secondly, due to the existence of irrelevant features, using all the deep features to establish a monitoring model may lead to the degradation of monitoring performance. Therefore, this paper proposes a multi-objective optimization based deep feature multi-subspace partitioning method for process monitoring. Firstly, a DNN model is established based on normal data, and all the deep features that can represent process information are extracted. Next, the residual is combined with the deep features to form a multi-level information representation vector. Then, aiming at the detection accuracy and detection diversity of the subspaces, the optimal multi-subspace partition can be solved with multi-objective optimization algorithm and the validation data (including normal data and fault data). Finally, the detection results of multiple subspaces are integrated through a fusion strategy to realize process monitoring. Applications in two industrial processes demonstrate the effectiveness and advantages of MOO-DFMSP.