A Two-Stage Privacy-Preserving Domain Adaptation for Industrial Time-Series Prediction

Domain adaptation (DA), as an emerging computational intelligence technology, is crucial for industrial time-series prediction, since various operating environments and tasks of industrial equipment will lead to variants in the distribution of monitoring data. Recently, many DA methods have been proposed to adapt to cross-domain industrial scenarios with data distribution shifting. However, due to privacy preserving concerns, data owners are reluctant to share their data, resulting in inaccessible source data. The artificial intelligence (AI) model trained by the inaccessible source data can be only applied as a blackbox. This makes it difficult to transfer source domain knowledge to the target domain. To solve this issue, we propose a two-stage source-free domain adaptation method for unsupervised knowledge transfer for industrial time-series prediction. At the first stage, an adversarial training method is proposed to improve the model's ability to represent data in the target domain, which may significantly differ from the source distribution. At the second stage, an unsupervised feature alignment method based on mean-teacher is proposed to align the target domain data with source knowledge. Additionally, we defined two contrastive loss functions to strengthen the consistent representation of target data. Experiments conducted on datasets N-CMAPSS and FEMOTO-ST demonstrate the effectiveness of the proposed method.