A complex process fault diagnosis method based on manifold distribution adaptation☆
Introduction
With the continuous development of complex industrial processes, industrial computers have been increasingly applied to the actual production process, and a large number of production process data reflecting the system’s operating state are collected and saved. Therefore, data-driven process monitoring methods have attracted more and more attention. However, with the continuous change of the production process, the monitoring model established with the data of the old process (source domain) has poor performance when applied to the new process (target domain). One of the important reasons for this phenomenon is that traditional data-driven process monitoring methods have strict assumptions that training data and test data must obey independent and identical distribution. However, the actual production process data is difficult to satisfy this harsh assumption, and marking new process data is expensive and time-consuming. Domain adaptation is a transfer learning method specially used to solve the different distribution of source domain data and target domain data, which is capable of transfer knowledge from different but related domain to facilitate learning of target domain tasks and has been widely used in image processing (Wang et al., 2019, Luo et al., 2018, Luo et al., 2017), target detection (Yu et al., 2019, Yan et al., 2018, Brust and Denzler, 2018), and text categorization (Han and Eisenstein, 2019, Chen et al., 2018). Therefore, we consider using domain adaptation to solve the problem of different distributions of old process data and new process data in complex industrial processes.
Section snippets
Related work
Domain adaptation is a special means to reduce the distribution differences between domains. At present, the research on domain adaptation has produced a lot of research results, which can be roughly divided into the following two types.
Manifold distribution adaptation
In this section, we present the Manifold Distribution Adaptation (MDA) approach in detail.
Experiments and evaluations
In this section, we will conduct extensive experiments on public large-scale transfer learning datasets and actual industrial process datasets to evaluate the performance of MDA.
Conclusion
Due to the different distribution of new process data and old process data in complex industrial processes, the performance of the old monitoring model is degraded when it is directly applied to the new process. In addition, it is difficult to establish monitoring model directly for new process because the new process data are few and unlabeled, therefore, a manifold distribution adaptation (MDA) approach is proposed in this paper. Specifically, we first map the old process data and the new
References (30)
- et al.
Learning with augmented features for supervised and semisupervised heterogeneous domain adaptation
IEEE Trans. Pattern Anal. Mach. Intell.
(2014) - et al.
Transferring knowledge fragments for learning distance metric from a heterogeneous domain
IEEE Trans. Pattern Anal. Mach. Intell.
(2018) - Baktashmotlagh, Mahsa, Harandi, Mehrtash T., Lovell, Brian C., Salzmann, Mathieu, 2014. Domain adaptation on the...
- et al.
Not just a matter of semantics: the relationship between visual similarity and semantic similarity
(2018) - et al.
Knowledge-based transfer learning explanation
(2018) - Dai, W., Yang, Q., Xue, G.R., et al., 2007. Boosting for transfer learning. In: ICML, pp....
- Fernando, B., Habrard, A., Sebban, M., Tuytelaars, T., 2013. Unsupervised visual domain adaptation using subspace...
- Gong, B., Shi, Y., Sha, F., Grauman, K., 2012. Geodesic flow kernel for unsupervised domain adaptation. In: CVPR, pp....
- et al.
Domain adaptation for object recognition: An unsupervised approach
- et al.
Domain adaptation for object recognition: An unsupervised approach
Grassmann discriminant analysis: a unifying view on subspace-based learning
Unsupervised domain adaptation of contextualized embeddings: A case study in early modern english
An unsupervised domain adaptation approach for cross-domain visual classification
Adaptation regularization: A general framework for transfer learning
IEEE Trans. Knowl. Data Eng.
Cited by (9)
Multiscale domain adaption models and their application in fault transfer diagnosis of planetary gearboxes
2021, Engineering Applications of Artificial IntelligenceCitation Excerpt :Guo et al. (2019) proposed a new intelligent method named deep convolutional transfer learning network (DCTLN), which can confuse domains using MMD loss and a domain discriminator. Wang and Zhao (2020) proposed Manifold Distribution Adaptation (MDA) to solve data transfer problems in practical engineering. The above classical DDA methods calculate the domain confusion losses by using only one TF.
A new transferable bearing fault diagnosis method with adaptive manifold probability distribution under different working conditions
2021, Measurement: Journal of the International Measurement ConfederationCitation Excerpt :As a classical nonlinear dimensionality reduction method, manifold learning attempts to explore the data structure characteristics embedded in the high-dimensional observation space. This unsupervised manifold learning method is also widely utilized in machinery fault diagnosis [30–34]. The results verify that manifold learning can effectively capture the inherent manifold features associated with mechanical fault information.
Temporal Attention Source-Free Adaptation for Chemical Processes Fault Diagnosis
2024, IEEE Transactions on Industrial InformaticsDeep transfer learning based on dynamic domain adaptation for remaining useful life prediction under different working conditions
2023, Journal of Intelligent ManufacturingProduction quality prediction of cross-specification products using dynamic deep transfer learning network
2023, Journal of Intelligent Manufacturing
- ☆
No author associated with this paper has disclosed any potential or pertinent conflicts which may be perceived to have impending conflict with this work. For full disclosure statements refer to https://doi.org/10.1016/j.engappai.2019.103267.