Online prognostic strategies with deep leaning models for fleets of engineering assets

Abstract

The accurate prediction of remaining useful life for fleets of engineering assets is an increasingly important task in prognostics and health management (PHM). This is because the accurate prediction of time to failure allows for improved planning, scheduling and decision-making of maintenance tasks for fleets of engineering assets. Accurate prediction of remaining useful life therefore has high potential to increase the reliability, availability, production output, profitability and safety, and to decrease downtime, unnecessary maintenance and operating costs for fleets of engineering assets. This work proposes general and convenient prognostics strategies with data-driven deep learning models for online remaining useful life prediction for fleets of engineering assets, where historical run-to-failure condition monitoring measurements with trendable exponential degradation trajectories are available. The modeling of long-term sequence information in condition monitoring measurements has previously been shown to be very challenging and crucial for effective data-driven prognostics. Long short-term memory (LSTM) and gated recurrent unit (GRU) recurrent neural network (RNN) deep learning models are currently the state-of-the-art sequence modeling techniques and can effectively model long-term sequence information. These gated recurrent neural networks have however to date not been comprehensively investigated and compared for data-driven prognostics for fleets of engineering assets. In this work we investigate data sets which include a general asset degradation data set, turbofan engine degradation data set and turbofan engine degradation benchmarking data sets. The investigated data sets all simulate the exponential degradation trajectories and condition monitoring measurements for fleets of engineering assets that were run to failure, where each asset had either univariate or multivariate condition monitoring sensor measurements performed at fixed time intervals over its lifetime. The data sets investigated include training set examples and testing set examples. The training set examples represent historical (previously seen) engineering assets with condition monitoring measurements that were run to failure. The testing set examples represent future (completely unseen) general engineering assets with condition monitoring sensor measurements that were run to failure. The objective and challenge is therefore to propose a prognostics strategy and train a model on the condition monitoring measurements of the training set examples offline. The proposed prognostics strategy and trained model must then predict the remaining useful life from the condition monitoring measurements of the testing set examples fully online. The turbofan engine degradation benchmarking data sets allow for simple and effective prognostics performance comparisons between publications with different prognostics strategies and models. The proposed general prognostics strategies for this work include a prognostics classification strategy and prognostics regression strategy. The proposed prognostics classification strategy is to structure the remaining useful life modeling problem as a sequence-to-sequence classification deep learning problem, where the input sequence is the univariate or multivariate condition monitoring measurement time series and the target sequence is the remaining useful life classification time series. The remaining useful life classification time series for each individual training and testing set example consists of remaining useful life classes with different degradation levels that are based on its linearly decreasing remaining useful life time series with an applied threshold. The prognostics regression strategy is to structure the remaining useful life modeling problem as a sequence-to-sequence regression deep learning problem, where the input sequence is the univariate or multivariate condition monitoring measurement time series and the target sequence is the remaining useful life regression time series. The remaining useful life regression time series for each individual training and testing set example consists of remaining useful life values that are based on its linearly decreasing remaining useful life time series with an applied threshold. The sequence-to-sequence classification and regression deep learning models then learns and generalizes the mapping between the condition monitoring measurement time series and the remaining useful life classification and regression time series for all the training set examples offline. The trained sequence-to-sequence classification and regression deep learning models is then used to predict the remaining useful life classification and regression time series from the condition monitoring measurement time series for all the testing set examples fully online. The proposed sequence-to-sequence deep learning classification and regression model architectures that are investigated and compared for the proposed prognostics classification and regression strategies include a feedforward neural network (FNN), simple recurrent neural network (S-RNN), long short-term memory recurrent neural network (LSTM-RNN) and gated recurrent unit recurrent neural network (GRU-RNN). The sequence-to-sequence deep learning classification and regression model architectures are trained on the training sets of the investigated data sets with the new and effective Adam algorithm. The deep learning models are also regularized with a combination of the early stopping, weight decay and dropout regularization techniques in order to reduce overfitting and improve generalization and prediction performance. The prognostics classification and regression strategies were successfully applied on the investigated data sets with the FNN, S-RNN, LSTM-RNN and GRU-RNN classification and regression model architectures. The LSTM-RNN and GRU-RNN models drastically outperformed the FNN and S-RNN models as expected. The GRU-RNN models slightly outperformed the LSTMRNN models and the S-RNN models significantly outperformed the FNN models on average. The prognostics regression strategy and LSTM-RNN and GRU-RNN regression models achieved very competitive results when compared with other state-of-the-art publications on the turbofan engine degradation benchmarking data sets. The prognostics regression strategy and GRU-RNN regression model achieved a very competitive benchmarking score of 589 on the PHM08 turbofan degradation benchmarking data set.