Ph. D. Project
Title:
Contribution to the failure prediction of industrial products using a hybrid approach combining artificial intelligence and physical models - Application to Schneider Electric products
Dates:
2024/10/01 - 2027/09/30
Supervisor(s):
Other supervisor(s):
Cathignol Augustin (augustin.cathignol@se.com)
Description:
I. Context and objective:
This thesis is a CIFRE thesis proposal which is part of the industrial context of Schneider Electric with the objective of improving the performance of the products sold (e.g. circuit breakers, transformers, inverters, batteries, etc.) through the deployment of a predictive maintenance. This thesis is built on a synergy between Schneider Electric France (the industrial) and the Nancy Automation Research Center (CRAN-UMR CNRS 7039; the academic) of the University of Lorraine. The CRAN is a scientific actor recognized both nationally and internationally on the issues of prognosis and decision support in (predictive) maintenance within the framework of the PHM (Prognostics and Health Management) communities within the federations. GdR MACS, IEEE, PHM society, IFAC and CIRP. As part of this thesis, we will focus more particularly on the heart of the failure prediction algorithm. Indeed, the objective of this thesis is to develop forecasting methods making it possible to combine in an optimized manner the physical knowledge of experts and the most advanced AI techniques. This physical knowledge can take various forms: very structured (equations, models) or less structured (observations, expert opinions), quantitative or qualitative, etc.
II. Scientific challenges and expected contributions:
Overall, the developed prognosis approaches can be classified into two groups: model-based approaches and data-driven approaches for predicting failure or residual life of a system. Model-based approaches consist of physically modeling the system, including the associated failure degradation mechanisms. These approaches are sometimes very difficult to use for real applications due to the complexity of the model to be produced in order to best approach the complexity of the real system, the consideration of all of its degradation modes but also influencing factors, updating the model, etc. In addition, these approaches have low performance in a "real-time" use situation. Conversely, with the rapid development of machine learning and artificial intelligence (AI) techniques, data-driven approaches such as Gaussian process regression (GPR), vector machines support (VSM), artificial neural networks (ANN), offer new perspectives in prognosis. Compared to model-based approaches, data-driven approaches are non-parametric and do not initially take into account a priori knowledge about the system. These approaches can solve complex problems, however their performance strongly depends on the quantity and quality of the available data. An original direction to resolve these limitations while retaining the advantages of a data-driven approach is to move towards hybrid approaches, based both on the physical knowledge of systems and on artificial intelligence techniques, thus using in synergy models and data. These hybridizations should be considered as promising solutions to improve the performance and calculation time of purely data-driven approaches. For example, PINNs (Physics-Informed Neural Networks) constitute a new (2019) class of neural networks that combine machine learning and physical laws. To date, existing hybrid approaches are mainly proposed for specific systems (e.g., production system, machine tools) with specific types of physical knowledge. They are therefore not directly applicable to systems/products developed by Schneider Electric. Faced with this observation, our aim in this thesis is to found a (generic) methodology to support hybridization which makes it possible to effectively combine the most advanced AI techniques and the physical knowledge of the experts that the Schneider Electric group possess. not only on its products, but also on the lines that manufacture them, and the data that is generated. One of the major challenges is therefore the ability to integrate various forms of knowledge into the fault prediction method. Theoretical work will consist of both evaluating existing techniques (e.g. PINNs, Bayesian approaches and digital twin construction techniques) and creating new ones in order to build a methodology adapted to different types of data and physical and professional knowledge.
The major scientific originalities of the thesis in connection with the previous issues are, therefore:
- The formalization of heterogeneous knowledge (physical, professional) of experts on the degradation/failure process;
- The development of a methodology to support a hybridization of forecasts making it possible to effectively combine available data and formalized expert knowledge;
- A proposal for metrics making it possible to evaluate the performance and robustness of this hybridization of prognoses with regard to a failure prediction objective with a certain degree of confidence;
- Consideration for this methodology of contextual elements adapted to the needs and requirements of Schneider Electric with regard to the efficient deployment of predictive maintenance on its products.
Throughout the thesis the doctoral student will rely on very concrete and varied cases from Schneider Electric for which are available, on the one hand data acquired on the systems during endurance campaigns up to to failure as well as physical knowledge and business knowledge provided by the company's various internal experts. The systems on which this methodology will be applied are varied (circuit breakers, transformers, inverters, batteries, industrial machines) and will therefore make it possible to test and improve the generalizability of the method.
This thesis is a CIFRE thesis proposal which is part of the industrial context of Schneider Electric with the objective of improving the performance of the products sold (e.g. circuit breakers, transformers, inverters, batteries, etc.) through the deployment of a predictive maintenance. This thesis is built on a synergy between Schneider Electric France (the industrial) and the Nancy Automation Research Center (CRAN-UMR CNRS 7039; the academic) of the University of Lorraine. The CRAN is a scientific actor recognized both nationally and internationally on the issues of prognosis and decision support in (predictive) maintenance within the framework of the PHM (Prognostics and Health Management) communities within the federations. GdR MACS, IEEE, PHM society, IFAC and CIRP. As part of this thesis, we will focus more particularly on the heart of the failure prediction algorithm. Indeed, the objective of this thesis is to develop forecasting methods making it possible to combine in an optimized manner the physical knowledge of experts and the most advanced AI techniques. This physical knowledge can take various forms: very structured (equations, models) or less structured (observations, expert opinions), quantitative or qualitative, etc.
II. Scientific challenges and expected contributions:
Overall, the developed prognosis approaches can be classified into two groups: model-based approaches and data-driven approaches for predicting failure or residual life of a system. Model-based approaches consist of physically modeling the system, including the associated failure degradation mechanisms. These approaches are sometimes very difficult to use for real applications due to the complexity of the model to be produced in order to best approach the complexity of the real system, the consideration of all of its degradation modes but also influencing factors, updating the model, etc. In addition, these approaches have low performance in a "real-time" use situation. Conversely, with the rapid development of machine learning and artificial intelligence (AI) techniques, data-driven approaches such as Gaussian process regression (GPR), vector machines support (VSM), artificial neural networks (ANN), offer new perspectives in prognosis. Compared to model-based approaches, data-driven approaches are non-parametric and do not initially take into account a priori knowledge about the system. These approaches can solve complex problems, however their performance strongly depends on the quantity and quality of the available data. An original direction to resolve these limitations while retaining the advantages of a data-driven approach is to move towards hybrid approaches, based both on the physical knowledge of systems and on artificial intelligence techniques, thus using in synergy models and data. These hybridizations should be considered as promising solutions to improve the performance and calculation time of purely data-driven approaches. For example, PINNs (Physics-Informed Neural Networks) constitute a new (2019) class of neural networks that combine machine learning and physical laws. To date, existing hybrid approaches are mainly proposed for specific systems (e.g., production system, machine tools) with specific types of physical knowledge. They are therefore not directly applicable to systems/products developed by Schneider Electric. Faced with this observation, our aim in this thesis is to found a (generic) methodology to support hybridization which makes it possible to effectively combine the most advanced AI techniques and the physical knowledge of the experts that the Schneider Electric group possess. not only on its products, but also on the lines that manufacture them, and the data that is generated. One of the major challenges is therefore the ability to integrate various forms of knowledge into the fault prediction method. Theoretical work will consist of both evaluating existing techniques (e.g. PINNs, Bayesian approaches and digital twin construction techniques) and creating new ones in order to build a methodology adapted to different types of data and physical and professional knowledge.
The major scientific originalities of the thesis in connection with the previous issues are, therefore:
- The formalization of heterogeneous knowledge (physical, professional) of experts on the degradation/failure process;
- The development of a methodology to support a hybridization of forecasts making it possible to effectively combine available data and formalized expert knowledge;
- A proposal for metrics making it possible to evaluate the performance and robustness of this hybridization of prognoses with regard to a failure prediction objective with a certain degree of confidence;
- Consideration for this methodology of contextual elements adapted to the needs and requirements of Schneider Electric with regard to the efficient deployment of predictive maintenance on its products.
Throughout the thesis the doctoral student will rely on very concrete and varied cases from Schneider Electric for which are available, on the one hand data acquired on the systems during endurance campaigns up to to failure as well as physical knowledge and business knowledge provided by the company's various internal experts. The systems on which this methodology will be applied are varied (circuit breakers, transformers, inverters, batteries, industrial machines) and will therefore make it possible to test and improve the generalizability of the method.
Keywords:
Prognostics, PHM, AI, Data-driven models, hybrid approaches
Department(s):
| Modeling and Control of Industrial Systems |
Publications: