Difference between revisions of "Transfer Learning for Machine Diagnostics and Prognostics"
YuantaoFan (Talk | contribs) |
YuantaoFan (Talk | contribs) |
||
| (15 intermediate revisions by one user not shown) | |||
| Line 1: | Line 1: | ||
{{StudentProjectTemplate | {{StudentProjectTemplate | ||
| − | |Summary= | + | |Summary=Develop a deep adversarial neural networks (DANN) based method to predict failures and estimate machine health, under transfer learning settings. |
|Keywords=Transfer Learning, Domain adaptation, Domain Adversarial Neural Networks, Fault diagnosis, Prognostics | |Keywords=Transfer Learning, Domain adaptation, Domain Adversarial Neural Networks, Fault diagnosis, Prognostics | ||
|TimeFrame=Fall 2019 | |TimeFrame=Fall 2019 | ||
| Line 18: | Line 18: | ||
[8] Long, Mingsheng, et al. "Unsupervised domain adaptation with residual transfer networks." Advances in Neural Information Processing Systems. 2016. | [8] Long, Mingsheng, et al. "Unsupervised domain adaptation with residual transfer networks." Advances in Neural Information Processing Systems. 2016. | ||
| + | |||
| + | [9] Fan, Yuantao, Sławomir Nowaczyk, and Thorsteinn Rögnvaldsson. "Transfer learning for Remaining Useful Life Prediction Based on Consensus Self-Organizing Models." arXiv preprint arXiv:1909.07053 (2019). | ||
|Prerequisites=Artificial Intelligence and Learning Systems courses; good knowledge of machine learning and neural networks; programming skills for implementing machine learning algorithms | |Prerequisites=Artificial Intelligence and Learning Systems courses; good knowledge of machine learning and neural networks; programming skills for implementing machine learning algorithms | ||
| − | |Supervisor=Yuantao Fan, | + | |Supervisor=Yuantao Fan, Mohammad Ghaith Altarabichi, Sepideh Pashami, |
|Level=Master | |Level=Master | ||
| − | |Status= | + | |Status=Open |
}} | }} | ||
| − | Transfer Learning (TL) [1, 2] refers to methods for transferring knowledge learned in one setting (the source domain) to another setting (the target domain) and it is needed in the field of machine diagnostics and prognostics. The current industrial approach for developing machine diagnostic and prognostic methods usually relies on data acquired in experiments under controlled conditions prior to deployment of the equipment. Detecting or predicting failures and estimating machine health in this paradigm assumes that future field data will have a very similar distribution to the experiment data. However, many machines were operated under dynamic environmental conditions and were operated in a variety of ways. This makes the assumption that pre-deployment data and post-deployment data follow very similar | + | Transfer Learning (TL) [1, 2] refers to methods for transferring knowledge learned in one setting (the source domain) to another setting (the target domain) and it is needed in the field of machine diagnostics and prognostics. The current industrial approach for developing machine diagnostic and prognostic methods usually relies on data acquired in experiments under controlled conditions prior to deployment of the equipment. Detecting or predicting failures and estimating machine health in this paradigm assumes that future field data will have a very similar distribution to the experiment data. However, many machines were operated under dynamic environmental conditions and were operated in a variety of ways. This makes the assumption that pre-deployment data and post-deployment data follow a very similar distribution is unlikely to hold. Moreover, since industrial systems are not allowed to run until failure (for safety reasons), collecting data that has comprehensive coverage on various usages patterns and deterioration patterns is very challenging. |
| + | |||
| + | Therefore, machinery diagnosis and prognosis need to: (i) adapt to more complex scenarios where unseen degradation patterns and new operating conditions are present (in the testing data); (ii) learn to utilize and transfer knowledge gained from operations of similar equipment/assets. A suitable solution is to perform TL, using feature-representation based TL or domain adaptation methods, for transferring knowledge between tasks, e.g. [5], and/or dealing with new conditions/faults, e.g. [9]. Feature-representation based TL and domain adaptation methods aim at discovering meaningful common structures between the source and the target domain, finding transformations that project the source data and the target data into a common latent feature space, which has predictive qualities for solving the target task. The discrepancy of marginal distributions between the source and the target data in the latent feature space is expected to be reduced at the same time. | ||
| − | + | Recently, Domain Adversarial Neural Networks (DANN) [3, 7, 8] has been applied for fault diagnosis and machine prognosis [4, 5, 6, 9]. The idea is to train a deep neural network for extracting domain-invariant features that has predictive power for the classification/regression task. DANN includes a deep feature extractor and a label predictor, which is a standard architecture for performing supervised learning. The unsupervised domain adaptation task is carried out by a domain classifier, which backpropagates gradients for making features domain-invariant. The main objective of this work is to develop a DANN based method, e.g. designing network structure and cost functions, to perform machine diagnostics and prognostics, using multivariate time series data. The proposed method will be evaluated using both simulated data and real data coming from heaving duty vehicles. | |
Latest revision as of 10:39, 14 October 2019
| Title | Transfer Learning for Machine Diagnostics and Prognostics |
|---|---|
| Summary | Develop a deep adversarial neural networks (DANN) based method to predict failures and estimate machine health, under transfer learning settings. |
| Keywords | Transfer Learning, Domain adaptation, Domain Adversarial Neural Networks, Fault diagnosis, Prognostics |
| TimeFrame | Fall 2019 |
| References | [[References::[1] Pan, Sinno Jialin, and Qiang Yang. "A survey on transfer learning." IEEE Transactions on knowledge and data engineering 22.10 (2009): 1345-1359.
[2] Weiss, Karl, Taghi M. Khoshgoftaar, and DingDing Wang. "A survey of transfer learning." Journal of Big data 3.1 (2016): 9. [3] Ganin, Yaroslav, and Victor Lempitsky. "Unsupervised domain adaptation by backpropagation." arXiv preprint arXiv:1409.7495 (2014). [4] W. Lu, B. Liang, Y. Cheng, D. Meng, J. Yang, T. Zhang, Deep model based domain adaptation for fault diagnosis, IEEE Transactions on Industrial Electronics 64 (2017) 2296–2305. [5] Wang, Qin, Gabriel Michau, and Olga Fink. "Domain Adaptive Transfer Learning for Fault Diagnosis." arXiv preprint arXiv:1905.06004 (2019). [6] Da Costa, Paulo R. de O., et al. "Remaining useful lifetime prediction via deep domain adaptation." arXiv preprint arXiv:1907.07480 (2019). [7] Akuzawa, Kei, Yusuke Iwasawa, and Yutaka Matsuo. "Adversarial Invariant Feature Learning with Accuracy Constraint for Domain Generalization." arXiv preprint arXiv:1904.12543 (2019). [8] Long, Mingsheng, et al. "Unsupervised domain adaptation with residual transfer networks." Advances in Neural Information Processing Systems. 2016. [9] Fan, Yuantao, Sławomir Nowaczyk, and Thorsteinn Rögnvaldsson. "Transfer learning for Remaining Useful Life Prediction Based on Consensus Self-Organizing Models." arXiv preprint arXiv:1909.07053 (2019).]] |
| Prerequisites | Artificial Intelligence and Learning Systems courses; good knowledge of machine learning and neural networks; programming skills for implementing machine learning algorithms |
| Author | |
| Supervisor | Yuantao Fan, Mohammad Ghaith Altarabichi, Sepideh Pashami |
| Level | Master |
| Status | Open |
Transfer Learning (TL) [1, 2] refers to methods for transferring knowledge learned in one setting (the source domain) to another setting (the target domain) and it is needed in the field of machine diagnostics and prognostics. The current industrial approach for developing machine diagnostic and prognostic methods usually relies on data acquired in experiments under controlled conditions prior to deployment of the equipment. Detecting or predicting failures and estimating machine health in this paradigm assumes that future field data will have a very similar distribution to the experiment data. However, many machines were operated under dynamic environmental conditions and were operated in a variety of ways. This makes the assumption that pre-deployment data and post-deployment data follow a very similar distribution is unlikely to hold. Moreover, since industrial systems are not allowed to run until failure (for safety reasons), collecting data that has comprehensive coverage on various usages patterns and deterioration patterns is very challenging.
Therefore, machinery diagnosis and prognosis need to: (i) adapt to more complex scenarios where unseen degradation patterns and new operating conditions are present (in the testing data); (ii) learn to utilize and transfer knowledge gained from operations of similar equipment/assets. A suitable solution is to perform TL, using feature-representation based TL or domain adaptation methods, for transferring knowledge between tasks, e.g. [5], and/or dealing with new conditions/faults, e.g. [9]. Feature-representation based TL and domain adaptation methods aim at discovering meaningful common structures between the source and the target domain, finding transformations that project the source data and the target data into a common latent feature space, which has predictive qualities for solving the target task. The discrepancy of marginal distributions between the source and the target data in the latent feature space is expected to be reduced at the same time.
Recently, Domain Adversarial Neural Networks (DANN) [3, 7, 8] has been applied for fault diagnosis and machine prognosis [4, 5, 6, 9]. The idea is to train a deep neural network for extracting domain-invariant features that has predictive power for the classification/regression task. DANN includes a deep feature extractor and a label predictor, which is a standard architecture for performing supervised learning. The unsupervised domain adaptation task is carried out by a domain classifier, which backpropagates gradients for making features domain-invariant. The main objective of this work is to develop a DANN based method, e.g. designing network structure and cost functions, to perform machine diagnostics and prognostics, using multivariate time series data. The proposed method will be evaluated using both simulated data and real data coming from heaving duty vehicles.
