Statistical transfer learning for modeling, monitoring, and prognosis of manufacturing systems

Amirhossein Fallahdizcheh

doi:10.25820/etd.007175

Thanks to the advances in data collection and storage systems, data-rich environments are widely available. Examples of these data acquisition systems are GM's OnStar, Amazon Go, and Siemens MindSphere. Despite the easy-to-access features, the cost of generating or collecting data in many processes/systems is still very high due to collection conditions. Moreover, data is very scarce in most start-up processes. These factors result in the data scarcity issue. Fortunately, researchers have proposed feasible solutions for dealing with this issue by leveraging the resources in data rich environment to benefit the understanding of data-scarce processes. This idea is known as transfer learning. The key to facilitating transfer learning is to find/create a shared domain that can serve as the bridge for transferring knowledge from the data rich source processes to the data scarce target process. Transfer learning has been applied in various scenarios and achieved success in information sharing and process learning. However, many new challenges arise due to the more and more complicated data structures and collection environments. First, the heterogeneity across different processes widely exists in transfer learning. These heterogeneities represent the unique features of different systems, and the transfer learning is expected to identify shared similarities from heterogeneous data. Second, incomplete/imbalance data is an intrinsic problem in transfer learning. The data availability in the target process is typically smaller than in source processes. The data incompleteness should be handled in an appropriate way, otherwise, the extracted similarity will be biased toward the data rich systems. In addition, computational efficiency is also a critical concern in modeling these incomplete data, especially when large volume of data involved in real-time monitoring tasks. Finally, heterogeneous sampling rates are commonly observed when dealing with multiple sources of data. In this case, methods such as interpolation were traditionally used. Unfortunately, these methods treat each source of data independently thus ignoring the correlation among different processes. On the other hand, the incorporation of such correlation requires significant computational complexity in modeling, optimization, and prediction. It is critical to reaching a balance between model complexity and performance when dealing with heterogeneous sampling rates in transfer learning. To address these challenges, three tasks are investigated in this thesis under different applications of transfer learning. Transfer learning among heterogeneous degradation processes. Data-driven statistical models are proposed to facilitate source selection and transfer learning among heterogeneous degradation processes. The effects of negative transfer is studied to demonstrate the significance of screening informative sources in the process of transfer learning. Transfer learning for monitoring incomplete profile data. The multi-output Gaussian process is investigated under the incomplete data, where the traditional maximum likelihood estimator cannot provide parameter estimations. A special covariance structure is proposed to facilitate the parameter estimation using incomplete data, and the performance in monitoring incomplete profiles validates the superiority and practicability of the proposed method. Efficient monitoring of incomplete profile data via variational inference based multi-output Gaussian process. A variational inference-based solution is proposed to reduce the computational complexity of the multi-output Gaussian process. A novel convolutional structure is proposed to tie the correlation between the variational inference and the multi-output Gaussian process. The proposed method is applied to incomplete profile data monitoring and demonstrates significant computational complexity reduction. Efficient modeling and prognosis for data transfer with heterogeneous sampling rates. A novel multivariate functional principal component analysis is proposed to deal with heterogeneous sampling rates in different data streams, where the cross-process correlation is captured to improve the data interpolation and prediction performance. The proposed method is applied in transfer learning of prognosis of degradation signals and achieves remarkable improvement in prognosis accuracy.

Statistical transfer learning for modeling, monitoring, and prognosis of manufacturing systems

Files and links (1)

Abstract

Details

Metrics