Rafael Gouriveau

Degradations analysis and aging modeling for health assessment and prognostics of PEMFC

Applying prognostics to Proton Exchange Membrane Fuel Cell (PEMFC) stacks is a good solution to help taking actions extending their lifetime. However, it requires a great understanding of the degradation mechanisms and failures occurring within the stack. This task is not simple when applied to a PEMFC due to the different levels (stack - cells - components), the different scales and the multiple causes that lead to degradation. To overcome this problem, this work proposes a methodology dedicated to the setting of a framework and a modeling of the aging for prognostics. This methodology is based on a deep literature review and degradation analyses of PEMFC stacks. This analysis allows defining a proper vocabulary dedicated to PEMFC׳s prognostics and health management and a clear limited framework to perform prognostics. Then the degradations review is used to select critical components within the stack, and to define their critical failure mechanisms thanks the proposal of new fault trees. The impact of these critical components and mechanisms on the power loss during aging is included to the model for prognostics. This model is finally validated on four datasets with different mission profiles both for health assessment and prognostics.

Machine health condition prediction via online dynamic fuzzy neural networks

Machine health condition (MHC) prediction is useful for preventing unexpected failures and minimizing overall maintenance costs in condition-based maintenance. The neural network (NN)-based data-driven method has been considered to be promising for MHC prediction due to the adaptability, nonlinearity and universal approximation capability of NNs. This paper presents an online MHC prediction approach using online dynamic fuzzy NNs (OD-FNNs) with structure and parameters learning. To meet the requirement of real-time application, the original OD-FNN is simplified based on an extreme learning machine technique as follows: (1) initial fuzzy rules are randomly generated without the knowledge of training data; (2) fuzzy rules are added and pruned uniformly by fired strength-based criteria; (3) antecedent parameters are fixed after generation so that only consequent parameters are updated online. The modified OD-FNN is particularly suitable for MHC prediction since: (1) fuzzy rules can evolve as new training datum arrives, which enables us to cope with non-stationary processes in MHC; (2) learning mechanisms applied are simple and efficient for real-time implementation. The validity and superiority of the proposed MHC prediction approach has been evaluated by real-world monitoring data from the accelerated bearing life.

A New Multivariate Approach for Prognostics Based on Extreme Learning Machine and Fuzzy Clustering

Prognostics is a core process of prognostics and health management (PHM) discipline, that estimates the remaining useful life (RUL) of a degrading machinery to optimize its service delivery potential. However, machinery operates in a dynamic environment and the acquired condition monitoring data are usually noisy and subject to a high level of uncertainty/unpredictability, which complicates prognostics. The complexity further increases, when there is absence of prior knowledge about ground truth (or failure definition). For such issues, data-driven prognostics can be a valuable solution without deep understanding of system physics. This paper contributes a new data-driven prognostics approach namely, an “enhanced multivariate degradation modeling,” which enables modeling degrading states of machinery without assuming a homogeneous pattern. In brief, a predictability scheme is introduced to reduce the dimensionality of the data. Following that, the proposed prognostics model is achieved by integrating two new algorithms namely, the summation wavelet-extreme learning machine and subtractive-maximum entropy fuzzy clustering to show evolution of machine degradation by simultaneous predictions and discrete state estimation. The prognostics model is equipped with a dynamic failure threshold assignment procedure to estimate RUL in a realistic manner. To validate the proposition, a case study is performed on turbofan engines data from PHM challenge 2008 (NASA), and results are compared with recent publications.

Prognostics in switching systems: Evidential markovian classification of real-time neuro-fuzzy predictions

Condition-based maintenance is nowadays considered as a key-process in maintenance strategies and prognostics appears to be a very promising activity as it should permit to not engage inopportune spending. Various approaches have been developed and data-driven methods are increasingly applied. The training step of these methods generally requires huge datasets since a lot of methods rely on probability theory and/or on artificial neural networks. This step is thus time-consuming and generally made in batch mode which can be restrictive in practical application when few data are available. A method for prognostics is proposed to face up this problem of lack of information and missing prior knowledge. The approach is based on the integration of three complementary modules and aims at predicting the failure mode early while the system can switch between several functioning modes. The three modules are: 1) observation selection based on information theory and Choquet Integral, 2) prediction relying on an evolving real-time neuro-fuzzy system and 3) classification into one of the possible functioning modes using an evidential Markovian classifier based on Dempster-Shafer theory. Experiments concern the prediction of an engine health based on more than twenty observations.

Defining and applying prediction performance metrics on a recurrent NARX time series model

Nonlinear autoregressive moving average with exogenous inputs (NARMAX) models have been successfully demonstrated for modeling the input-output behavior of many complex systems. This paper deals with the proposition of a scheme to provide time series prediction. The approach is based on a recurrent NARX model obtained by linear combination of a recurrent neural network (RNN) output and the real data output. Some prediction metrics are also proposed to assess the quality of predictions. This metrics enable to compare different prediction schemes and provide an objective way to measure how changes in training or prediction model (neural network architecture) affect the quality of predictions. Results show that the proposed NARX approach consistently outperforms the prediction obtained by the RNN neural network.

Remaining Useful Life Estimation by Classification of Predictions Based on a Neuro-Fuzzy System and Theory of Belief Functions

Various approaches for prognostics have been developed, and data-driven methods are increasingly applied. The training step of these methods generally requires huge datasets to build a model of the degradation signal, and estimate the limit under which the degradation signal should stay. Applicability and accuracy of these methods are thereby closely related to the amount of available data, and even sometimes requires the user to make assumptions on the dynamics of health states evolution. Following that, the aim of this paper is to propose a method for prognostics and remaining useful life estimation that starts from scratch, without any prior knowledge. Assuming that remaining useful life can be seen as the time between the current time and the instant where the degradation is above an acceptable limit, the proposition is based on a classification of prediction strategy (CPS) that relies on two factors. First, it relies on the use of an evolving real-time neuro-fuzzy system that forecasts observations in time. Secondly, it relies on the use of an evidential Markovian classifier based on Dempster-Shafer theory that enables classifying observations into the possible functioning modes. This approach has the advantage to cope with a lack of data using an evolving system, and theory of belief functions. Also, one of the main assets is the possibility to train the prognostic system without setting any threshold. The whole proposition is illustrated and assessed by using the CMAPPS turbofan dataset. RUL estimates are shown to be very close to actual values, and the approach appears to accurately estimate the failure instants, even with few learning data.

Joint Particle Filters Prognostics for Proton Exchange Membrane Fuel Cell Power Prediction at Constant Current Solicitation

Proton Exchange Membrane Fuel Cells (PEMFC) are promising energy converters, but still suffer from a short life duration. Applying Prognostics and Health Management seems to be a great solution to overcome that issue. But developing prognostics to anticipate and try to avoid failures is a critical challenge. To tackle this problem, a hybrid prognostics approach is proposed. It aims at predicting the power aging of a PEMFC stack working at a constant operating condition and a constant current solicitation. The main difficulties to overcome are the lack of adapted modeling of the aging for prognostics, and the occurrence of disturbances creating recovery phenomena through aging. Consequently, this work proposes a new empirical model for power aging that takes into account these recoveries based on different features extracted from the data. These models are used in a joint particle filter framework directly initialized by an automatic parameter estimate process. When sufficient data are available, the prognostics can give accurate behavior predictions compared to experimentation. Remaining useful life estimates can be given with an error smaller than 5% for a horizon of 500 hours on a life duration of 1750 hours, which is clearly long enough for decision making.

Novel failure prognostics approach with dynamic thresholds for machine degradation

Estimating remaining useful life (RUL) of critical machinery is a challenging task. It is achieved through essential steps of data acquisition, data pre-processing and prognostics modeling. To estimate RUL of a degrading machinery, prognostics modeling phase requires precise knowledge about failure threshold (FT) (or failure definition). Practically, degrading machinery can have different levels (states) of degradation before failure, and prognostics can be quite complicated or even impossible when there is absence of prior knowledge about actual states of degrading machinery or FT. In this paper a novel approach is proposed to improve failure prognostics. In brief, the proposed prognostics model integrates two new algorithms, namely, a Summation Wavelet Extreme Learning Machine (SWELM) and Subtractive-Maximum Entropy Fuzzy Clustering (S-MEFC) to predict degrading behavior, automatically identify the states of degrading machinery, and to dynamically assign FT. Indeed, for practical reasons there is no interest in assuming FT for RUL estimation. The effectiveness of the approach is judged by applying it to real dataset in order to estimate future breakdown of a real machinery.

Reducing arbitrary choices in model building for prognostics: An approach by applying parsimony principle on an evolving neuro-fuzzy system

Abstract Failure prognostics requires an efficient prediction tool to be built. This task is as difficult as, in many cases, very few knowledge or previous experiences on the degradation process are available. Following that, practitioners are used to adopt a “trial and error” approach, and to make some assumptions when developing a prediction model: choice of an architecture, initialization of parameters, learning algorithms… This is the problem addressed in this paper: how to systematize the building of a prognostics system and reduce the influence of arbitrary human intervention? The proposition is based on the use of a neuro-fuzzy predictor whose structure is partially determined, on one side, thanks to its evolving capability, and on the other side, thanks to parsimony principle. The aim of the approach is to automatically generate a suitable prediction system that reaches a compromise between complexity and accuracy capability. The whole proposition is illustrated on a real-world prognostics problem concerning the prediction of an engine health.

Improving accuracy of long-term prognostics of PEMFC stack to estimate remaining useful life

Proton Exchange Membrane Fuel cells (PEMFC) are energy systems that facilitate electrochemical reactions to create electrical energy from chemical energy of hydrogen. PEMFC are promising source of renewable energy that can operate on low temperature and have the advantages of high power density and low pollutant emissions. However, PEMFC technology is still in the developing phase, and its large-scale industrial deployment requires increasing the life span of fuel cells and decreasing their exploitation costs. In this context, Prognostics and Health Management of fuel cells is an emerging field, which aims at identifying degradation at early stages and estimating the Remaining Useful Life (RUL) for life cycle management. Indeed, due to prognostics capability, the accurate estimates of RUL enables safe operation of the equipment and timely decisions to prolong its life span. This paper contributes data-driven prognostics of PEMFC by an ensemble of constraint based Summation Wavelet-Extreme Learning Machine (SW-ELM) algorithm to improve accuracy and robustness of long-term prognostics. The SW-ELM is used for ensemble modeling due to its enhanced applicability for real applications as compared to conventional data-driven algorithms. The proposed prognostics model is validated on run-to-failure data of PEMFC stack, which had the life span of 1750 hours. The results confirm capability of the prognostics model to achieve accurate RUL estimates.