Trainee Project
Title:
Joint learning to compensate missing features in machine prognostics
Dates:
2023/03/01 - 2023/09/30
Supervisor(s):
Description:
The context of this internship is the industry of the future and more particularly the contribution of digitalization and Artificial Intelligence to prognostic and health management (PHM).
PHM deals with equipment maintenance which is a critical and costly activity: studies show that depending on the industry, between 15 and 70 percent of total production costs come from maintenance activities [3]. Nevertheless, maintenance has always been a major factor in a company's ability to compete in terms of performance and to deliver high-quality products at a reasonable price. To this end, health and prognosis management (HPM) has received a lot of attention in recent years due to its ability to manage maintenance in a more optimal manner.
Prognostic process providing the remaining useful life (RUL) estimation, on top of diagnosis and fault detection, is the core process in PHM, as it provides PHM the ability to anticipate fault and provide relevant information on failure time to maintenance decision-makers. The remaining useful
life (RUL) of equipment is defined as the time from the present to when it no longer performs its intended function. Accurate and reliable RUL prognostic provides decision makers with valuable information to adopt an appropriate maintenance strategy to maximize equipment utilization and
avoid costly breakdowns. Numerous prognostic algorithms for RUL estimation have been reported in the literature. Nevertheless, a recent trend in the literature is the use of Deep learning approaches for prognostic.
Deep learning (DL) approaches have multiple potential benefits that have been explored in various fields. But for prognostics, this is not the case due to the scarcity of representative data in particular applications, and this can be caused by several factors such as equipment reliability by design, preventive maintenance, unrepresentative degradation by sensors values, and the high cost of testing to obtain run to failure data. This limits the research in this field even though these types of applications will have a strong impact on the industrial world.
To overcome these challenges, we should be able to leverage multiple data sets, which is not yet fully attained. One of the reasons for this is that even data sets collected from similar machines may have different input features, different operating conditions, and different failure modes, making it
challenging to exploit all the knowledge available. In this internship, the student will work on a specific sub-problem of the latter. We are interested
in the case where we do not have the same features between the support and target data sets, with the possibility of the target features being less representative of the degradation than the support ones, thereby raising the question of whether the knowledge can be compensated by other datasets.
To explore this possibility and build a model that can generalize well, domain adaptation, transfer learning, and joint learning approaches are to be explored to make the best use of multiple small data sets.
To perform the experiments, two public data sets (CMAPSS [4] and N-CMAPSS [1]) will be used to develop and evaluate the proposed approach, as they are well known benchmarks in this field. In addition, the base DL model to be used is the one proposed in [2] which is based on LSTM cells (this
is open to change depending on the progress of the work).
References:
[1] Manuel Arias Chao et al. "Aircraft engine run-to-failure dataset under real flight conditions for prognostics and diagnostics". In: Data 6.1 (2021), p. 5.
[2] Alaaeddine Chaoub et al. "Learning representations with end-to-end models for improved remaining useful life prognostic". In: European Conference of the Prognostics and Health Management Society. 2021.
[3] Christian Krupitzer et al. A Survey on Predictive Maintenance for Industry 4.0. 2020. arXiv:2002.08224 [cs.LG].
[4] Abhinav Saxena et al. "Damage propagation modeling for aircraft engine run-to-failure simulation". In: 2008 international conference on prognostics and health management. IEEE. 2008,pp. 1-9.
PHM deals with equipment maintenance which is a critical and costly activity: studies show that depending on the industry, between 15 and 70 percent of total production costs come from maintenance activities [3]. Nevertheless, maintenance has always been a major factor in a company's ability to compete in terms of performance and to deliver high-quality products at a reasonable price. To this end, health and prognosis management (HPM) has received a lot of attention in recent years due to its ability to manage maintenance in a more optimal manner.
Prognostic process providing the remaining useful life (RUL) estimation, on top of diagnosis and fault detection, is the core process in PHM, as it provides PHM the ability to anticipate fault and provide relevant information on failure time to maintenance decision-makers. The remaining useful
life (RUL) of equipment is defined as the time from the present to when it no longer performs its intended function. Accurate and reliable RUL prognostic provides decision makers with valuable information to adopt an appropriate maintenance strategy to maximize equipment utilization and
avoid costly breakdowns. Numerous prognostic algorithms for RUL estimation have been reported in the literature. Nevertheless, a recent trend in the literature is the use of Deep learning approaches for prognostic.
Deep learning (DL) approaches have multiple potential benefits that have been explored in various fields. But for prognostics, this is not the case due to the scarcity of representative data in particular applications, and this can be caused by several factors such as equipment reliability by design, preventive maintenance, unrepresentative degradation by sensors values, and the high cost of testing to obtain run to failure data. This limits the research in this field even though these types of applications will have a strong impact on the industrial world.
To overcome these challenges, we should be able to leverage multiple data sets, which is not yet fully attained. One of the reasons for this is that even data sets collected from similar machines may have different input features, different operating conditions, and different failure modes, making it
challenging to exploit all the knowledge available. In this internship, the student will work on a specific sub-problem of the latter. We are interested
in the case where we do not have the same features between the support and target data sets, with the possibility of the target features being less representative of the degradation than the support ones, thereby raising the question of whether the knowledge can be compensated by other datasets.
To explore this possibility and build a model that can generalize well, domain adaptation, transfer learning, and joint learning approaches are to be explored to make the best use of multiple small data sets.
To perform the experiments, two public data sets (CMAPSS [4] and N-CMAPSS [1]) will be used to develop and evaluate the proposed approach, as they are well known benchmarks in this field. In addition, the base DL model to be used is the one proposed in [2] which is based on LSTM cells (this
is open to change depending on the progress of the work).
References:
[1] Manuel Arias Chao et al. "Aircraft engine run-to-failure dataset under real flight conditions for prognostics and diagnostics". In: Data 6.1 (2021), p. 5.
[2] Alaaeddine Chaoub et al. "Learning representations with end-to-end models for improved remaining useful life prognostic". In: European Conference of the Prognostics and Health Management Society. 2021.
[3] Christian Krupitzer et al. A Survey on Predictive Maintenance for Industry 4.0. 2020. arXiv:2002.08224 [cs.LG].
[4] Abhinav Saxena et al. "Damage propagation modeling for aircraft engine run-to-failure simulation". In: 2008 international conference on prognostics and health management. IEEE. 2008,pp. 1-9.
Keywords:
Predictive Mainetance, Prognostic, Deep Learning, Federated learning
Conditions:
Location: Vandoeuvre-Lès-Nancy (France)
Duration: 5 or 6 months
Expected starting date: february-march 2023
Payment: 550¬/month
Duration: 5 or 6 months
Expected starting date: february-march 2023
Payment: 550¬/month
Department(s):
| Modeling and Control of Industrial Systems |
