Cross-domain health conditions identification based on joint distribution modeling of fused prototypes
International design engineering technical conferences and computers and information in engineering conference, 87295, V002T02A003
*Corresponding author
Cite
@inproceedings{10.1115/DETC2023-114851,
author = {Xie, Tingli and Huang, Xufeng and Choi, Seung-Kyum},
title = "{Cross-domain health conditions identification based on joint distribution modeling of fused prototypes}",
booktitle = {International design engineering technical conferences and computers and information in engineering conference},
volume = {87295},
year = {2023},
pages = {V002T02A003},
doi = {10.1115/DETC2023-114851},
url = {https://doi.org/10.1115/DETC2023-114851},
}
Abstract
Abstract Intelligent diagnostic models with deep learning networks have been widely used for health conditions identification of complex mechanical systems. However, the distributional shift and data scarcity phenomenon inevitably exist in practical and real-world scenarios, which further affect the generalization and representation capability of neural networks. Recently, meta-learning has provided an effective way to solve domain shifts and extremely limited data problems. By introducing non-parametric similarity measures into meta-knowledge learning, metric-based meta-learning can seek embedding spaces to compare class prototypes to identify unseen testing samples. This article proposes a novel meta-neural network based on joint distribution modeling of fused prototypes (IFJPN) to alleviate distribution discrepancy in embedding spaces between different domains. First, multi-source time-series data are transformed into images for modality alignment and information aggregation. Second, a feature embedding module with multiple convolutions is developed to obtain fused prototypes from converted images. Then, the Brownian distance covariance is used for efficient joint distribution modeling of fused prototypes, which improves the discriminative capability of cross-domain health conditions with extremely limited data. Extensive experiments on a benchmark dataset in different cross-domain scenarios under two limited data settings show that the proposed IFJPN has more robust recognition and generalization capability than typical transfer-learning or meta-learning methods.