DRE-xLSTM: Distributed Reversible Embedding Extended Long Short-Term Memory (xLSTM) for Complex Industrial Data Imputation

The large amount of sensor data poses a challenge for imputation in complex industrial processes. The prominent method of xLSTM can achieve highly accurate data imputation in time series, however, xLSTM is incapable of distributed imputation. To solve the problem in a distributed environment, a distributed reversible embedding xLSTM imputation method (DRE-xLSTM) is proposed. First, a reversible embedded xLSTM method is designed to scale the data and fuse complex features using reversible normalization and multilayer linear modules. Second, a searchable prototype federated learning method is developed to accomplish the distributed learning using searchable cell states in a reversibly embedded xLSTM for delivery. The proposed DRE-xLSTM produces more accurate imputation results than comparative methods on the QINGHE dataset.