Incremental learning using generative-rehearsal strategy for fault detection and classification

In this study, we propose a novel pseudorehearsal method for modeling fault detection and classification. As manufacturing processes become increasingly advanced, it is often necessary to model the architecture when the data change over time. Particularly, learning with the addition of new fault types is called class incremental learning. Although learning systems must acquire new information from new data, this includes problems that can lead to catastrophic forgetting and class imbalance, wherein the number of instances in a particular class is greater than those in the other classes. Classification performance degrades when the existing model is trained under such conditions. Therefore, we propose a generative-rehearsal strategy that combines a pseudorehearsal strategy with independent generative models for each fault type. This method overcomes catastrophic forgetting and enables incremental learning with unbalanced data. The performance of the proposed method was superior to that of existing incremental and nonincremental methods while being memory efficient.