Fleet PHM for Critical Systems: Bi-level Deep Learning Approach for Fault Detection

Abstract

Data-driven approaches are highly relying on the representativeness of the dataset used for training the algorithms. For Prognostics and Health Management (PHM) applications, a lack of representativeness will result in detecting new operating conditions (that have not yet been observed during the period used for training) as faults. This is particularly a challenge for PHM in critical systems, for which long and consistent datasets with all operating conditions are generally not available.   Among the many data-driven approaches applied to PHM, Deep Learning has recently brought promising results, enabling an automation of traditional PHM tasks, including signal processing, feature engineering and signal monitoring. Yet, as the parameters to optimize in a Deep Neural Networks are numerous, the training requires also huge datasets: another quality that critical system datasets are often missing.   When a fleet of systems is monitored, however, a solution to compose more representative and bigger training datasets is to combine condition monitoring data from systems with similar operating conditions. Identifying similar operating conditions would, traditionally, require comparing the distances and the distributions of multi-dimensional time series, a computationally intensive task worsened by the curse of dimensionality.   In this paper, we propose to use a deep neural network, designed for measuring similarities between the training and the testing datasets: a Hierarchical Extreme Learning Machine (HELM). HELM have demonstrated excellent abilities to jointly learn features and monitor deviations from the training data. Training first this network on individual systems, HELM can be used to identify other systems with similar operating conditions. Afterwards, the same network can be trained again with this representative dataset composed of condition monitoring data of several systems, to monitor more efficiently the health of the individual systems. Any deviation in the algorithm output would signify that the system is not anymore in operating conditions seen in the training fleet, and is likely experiencing a fault.   The novelty of the proposed approach lies in the usage of the same architecture twice in a bi-level framework: first, for selecting the representative datasets from a fleet of systems and second, using the selected datasets to train the health monitoring algorithm. This approach achieves good performances on both tasks. Learning from the fleet attenuates the impact of changing operating conditions (\eg summer/winter trends), and improves the quality of the fault detection.   The approach is tested on a fleet of 112 power plants, some of which experienced a stator vane failure.