Ensemble Learning Based Convolutional Neural Networks for Remaining Useful Life Prediction of Aircraft Engines

Remaining useful life (RUL) prediction is an essential task of Prognostics and Health Management (PHM) of aircraft engines performed utilizing the huge data collected from multiple sensors attached to them to ensure their safe operation. While many studies have been reported on RUL prediction for aircraft engines, only a few of them focus on ensemble learning of CNN models for RUL prediction. The success of ensemble learning which is a combination of several base models developed using either same or different machine learning or deep learning algorithms, critically depends on the diversity among the base models generated. This paper proposes a new data-driven approach for RUL prediction of aircraft engines using ensemble learning based convolution neural networks (CNN) by investigating various steps to generate more diverse base models. The main objectives and contributions of this paper are as follows: Explore various CNN model architectures for RUL prediction - After data preprocessing and exploratory data analysis, two different CNN approaches, namely 2D CNN and 1D CNN with multiple channels, are investigated employing time window approach for time-series input preparation for better feature extraction by CNN. Each approach is experimented with multiple architectures to achieve the best possible outcome. Investigate engine specific RUL target function - For RUL prediction of turbofan engines using the C-MAPSS dataset, typically two RUL target functions, namely linear and piecewise linear, are used to determine RUL target values. In the piecewise linear RUL target function approach, which yields better performance in the reported studies, the RUL target value is assigned based on a piecewise linear degradation model which assumes a constant (and maximum) RUL target value in the early phase before linearly degrading the RUL targets. In the literature, this maximum RUL target value was chosen same for all the engines by taking a value of 125 or 130 without providing proper rationale. In this study, we adopt an approach based on the widely known health index to determine an engine specific initial (and maximum) RUL target value that can be used with the piecewise linear degradation model to determine RUL target values. Investigate hyperparameter optimization of CNN models to generate diverse base models - For the purpose of developing a high performance ensemble CNN model for RUL prediction, hyperparameter optimization of CNN models is performed to optimally determine the network structure (such as # of filters, filter size, stride, padding, # of convolutional, pooling, and dense layers, activation functions etc.) as well as the hyperparameters that determine the network training process (such as optimization method, learning rate, momentum, batch size etc.). Investigate ensemble learning to select and combine diverse CNN models for RUL prediction – In order to develop model combiners, diverse CNN models as base learners are selected using multiple performance measures such as RMSE, score function, MAE, and R2 score, and employing the non-negative least squares method, random forest regression, and extreme learning machine (ELM) to train model combiners. Evaluate the above proposed approach using the C-MAPSS dataset - To show the effectiveness of the proposed approach, various evaluations for RUL prediction using the popular C-MAPSS dataset (including all the four sub-datasets denoted as FD001, FD002, FD003 and FD004) are carried out and the results will be compared against the state-of-the-art results on the same dataset. Major emphasis of this proposed approach is on the generation of diverse CNN base models by carrying out various steps as explained above, and it is expected the results of this proposed approach will contribute towards enhancing the RUL prediction performance especially on the sub-datasets FD002 and FD004 which are challenging for the existing state-of-the-art RUL prediction techniques.