JavaScript is disabled for your browser. Some features of this site may not work without it.

Louw, Carel Johannes

Date:
2018

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.

Dissertation (MEng)--University of Pretoria, 2018.