Ph. D. Project
Data-driven nonlinear system identification and health state prognostics using deep learning. Application to predictive maintenance of business jet aircraft
2020/11/20 - 2023/12/04
Other supervisor(s):
The successful candidate will carry out research in the field of data-driven model-based predictive maintenance. The proposed PhD
project is part of a collaboration with Dassault Aviation whose final objective is to develop a function for monitoring the condition of
business jet aircraft systems by analyzing data recorded in flight.
The aim of the PhD is to design novel methods that lie at the crossovers of data-driven nonlinear dynamical model identification
approach and deep learning for robust predictive maintenance of business jet aircraft. The research will have both theoretical and
applied components.
The data-driven modeling and predictive maintenance of complex systems is undergoing a revolution, driven by the rise of big data,
advanced algorithms in machine learning and optimization, and modern computational hardware. Increasingly, these model-based
predictive maintenance strategies are aided by data-driven techniques that characterize the input-output dynamics of a system of
interest from measurements alone, without relying on first principles modeling. This is known as system identification, which has a long
and rich history in control theory. However, with increasingly powerful data-driven techniques, such as those coming from the machine
learning community, nonlinear system identification has encountered renewed interest.
Amongst the different models and associated techniques, deep learning methods will be conside- red and investigated. Deep learning
has indeed emerged in the domain of Artificial Intelligence as an efficient data-driven method for data/event classification and
prediction for dynamical nonlinear systems under variable operating conditions. The PhD will therefore explore the links bet- ween the
classical system identification approach [1, 2] and deep learning methods [3] for such systems.
Predictive maintenance is closely based on prognostics of failure [4] and prediction of the remaining useful life (RUL) of critical
components [5]. Prediction of RUL becomes a difficult problem in absence of exact deterioration knowledge. The availability of
degradation data base must be leveraged for efficient RUL predictions.
Recently, several works have highlighted the utility of Recurrent Neural Networks (RNNs) and Long short Term Memory (LSTM) neural
networks for identification of dynamic systems and prediction of futuristic events such as faults, failures and remaining useful life of
industrial systems [6, 7]. These recent works would be referred to study, explore and be the basis for developing novel Deep LSTM-
based solutions for identification/prediction using the observed data.
Among other aspects, the research will focus on improving the generalization capability of the method in the sense that the developed
methodology/algorithms should demonstrate excellent generalizable capacity in face of unknown/unseen asset properties or
The developed algorithms should be verifiable and should propose efficient means of quantifying the accuracy of output results.
Suitable procedures should be developed to analyze the prediction accuracy of novel algorithms. Another major focus area would be
unsupervised type of learning wherein it is expected that developed algorithms are able to learn relevant representations of
anomalies/health symptoms in self-supervised or unsupervised manner leading to better generalizations and accuracy.
The aim is the PhD is therefore to propose solutions for the different research challenges formulated above. The proposed solutions will
be implemented and tested by exploiting routine operation data
coming from several business jet aircraft.
[1] Ljung, L. (1999). System identification. Theory for the user. Prentice Hall.
[2] Garnier, H. & Wang, L. (Eds.) (2008). Identification of continuous-time models from sampled data. Springer-Verlag.
[3] Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press.
[4] Jha, M. S., Dauphin-Tanguy G., & Ould-Bouamama B. (2016), Particle filter based hybrid pro- gnostics for health monitoring of
uncertain systems in bond graph framework. Mechanical Systems and Signal Processing 75, 301-329.
[5] Jha, M. S., Bressel, M., Ould-Bouamama, B., & Dauphin-Tanguy, G. (2016). Particle filter based hybrid prognostics of proton exchange
membrane fuel cell in bond graph framework. Computers & Chemical Engineering, 95, 216-230.
[6] Deutsch, J., & He, D. (2017). Using deep learning-based approach to predict remaining useful life of rotating components. IEEE
Transactions on Systems, Man, and Cybernetics : Systems, 48(1), 11-20. [7] Li, X., Ding, Q., & Sun, J. Q. (2018). Remaining useful life
estimation in prognostics using deep convolution neural networks. Reliability Engineering & System Safety, 172, 1-11.
Deep learning, system identification, predictive maintenance, business jet aircraft
Control Identification Diagnosis