You are on page 1of 7

Optimal RUL Estimation: A State-of-Art Digital Twin Application

Mohamad Danish Anis, University of Toronto


Sharareh Taghipour, PhD, Ryerson University
Chi-Guhn Lee, PhD, University of Toronto

Key Words: Prognostics, RUL Estimation, Deep Learning, Long-short Term Memory, IIoT, Autoencoder

SUMMARY & CONCLUSIONS technologies are constantly challenging the traditional practices
in many industries worldwide. The onset of what is seen by
A real world Industrial IoT set up has paved way for
many as the fourth industrial revolution (Industry 4.0) finds its
simultaneous monitoring of several sensors at their unique
basis in a new generation of virtual reality and big-data driven
sampling rates. This has realized the need for artificial
models [1]. Digital twin technology has been recognized as a
intelligence tools for robust data processing. However, the large
core component of Industry 4.0 with its ability to provide a
size of input data requires real time monitoring and
comprehensive functional mirror of a component, product or
synchronization for online analysis. As the star concept behind
system, aiming to replicate the physical asset's performance in
the Industry 4.0 wave, a digital twin is a virtual, multi-scale and
virtual space with real-time synchronization [2].
probabilistic simulation to mirror the performance of its
As the industry grows to cope with the increasing
physical counterpart and serve the product lifecycle in a virtual
production demands, machinery reliability and cost efficiency
space. Evidently, a digital twin can proactively identify
of maintenance practices have been areas of intensive research.
potential issues with its corresponding real twin. Thus, it is best
Owing to a system’s requirement of continued availability and
suited for enabling a physics-based and data-driven model
reduced manual dependency, precise self-prediction and self-
fusion to estimate the remaining useful life (RUL) of the
assessment are mainly a digital twin’s focus. As a part of its
components. Traditional RUL prediction approaches have
prognostic application, a digital twin can proactively identify
assumed either an exponential or linear degradation trend with
propagating anomalies with its physical counterpart. Thus,
a fixed curve shape to build a Health Index (HI) model. Such
RUL prediction of specific components can be facilitated by
an assumption may not be useful for multi-sensor systems or
fusing pre-processed real time sensory data flow into a digital
cases where sensor data is available intermittently. A common
twin model.
constraint in the industry is irregular sensor data collection. The
Data-driven prognostics approaches, given their simplicity,
resulting asynchronous time series of the sporadic data needs to
are mostly preferred to make RUL predictions by directly
be an accurate representation of the component’s HI when
analyzing equipment behavior from condition-monitoring data
constructing a degradation model. In this paper, we extend the
[3]. Complex sequential data requires machine learning based
Long-Short Term Memory (LSTM) Recurrent Neural Network
deep learning (DL) techniques utilized in offline computing
(RNN) technique to generate RUL prediction within a digital
resources for accurate non-linear processing [1]. They are
twin framework as a means of synchronization with changing
trained using a combination of machine learning algorithms and
operational states. More specifically, we first use LSTM
data analytic techniques to create a model of a specific target,
encoder-decoder (LSTM-ED) to train a multilayered neural
for instance a flight-critical component or a high-speed rotor
network and reconstruct the sensor data time series
shaft that can derive an actionable outcome from the model.
corresponding to a healthy state. The resulting reconstruction
In this paper, we estimate RUL using an encoder-decoder
error can be used to capture patterns in input data time series
based LSTM-RNN network. An LSTM-ED architecture can
and estimate HI of training and testing sets. Using a time lag to
capture the inter-relationship of a multivariate time series data
record similarity between the HI curves, a weighted average of
leading to better prediction. The proposed model learns the
the final RUL estimation is obtained. The described empirical
normal/healthy behavior of the system, retains the healthy
approach is evaluated on publicly available engine degradation
sequence information and reconstructs the input time series
dataset with run-to-failure information. Results indicate a high
corresponding to the healthy state. Input sensor data in the form
RUL estimation accuracy with greater error reduction rate. This
of multivariate time series is mapped to a fixed dimensional
demonstrates wide applicability of the discussed methodology
vector representation with an LSTM-encoder. The LSTM-
to various industries where event data is scarce for the
decoder uses these vector representations to generate the target
application of only data-driven techniques.
sequences. The LSTM-ED memory cell stores the
1 INTRODUCTION healthy/normal instances of data which are used to reconstruct
the input time series. This new time series is re-constructed
State of art advancements and evolution in digital

 
 






Authorized licensed use limited to: University of Massachusetts Amherst. Downloaded on August 02,2020 at 10:24:03 UTC from IEEE Xplore. Restrictions apply.
without being trained on any faulty data. This leads to a high ability to add or remove information to the cell state. The output
reconstruction error when the machine starts to degrade with value is kept in a binary form, 0 or 1, for each number in the
every faulty time instance. The resultant HI is finally used to cell state Ct-1, where 1 represents any information being
predict the RUL. All figures discussed in this paper are included allowed to pass and 0 represents nothing can pass. The forget
in the Appendix section at the end. gate decides which information must be discarded from the cell
This methodology can overcome some of the constraints state. Only conditional information is stored in the cell state and
observed earlier with DL based prognostics: this is decided by the sigmoid layer (input gate layer) and the
tanh layer (new candidate values). The old cell state, Ct-1, is thus
1.1 Practical Constraints
updated with information from the new cell state Ct, forget gate,
x Most case studies assume a linear or exponential input gate and new memory. Lastly, the output gate decides
degradation trend instead of utilizing unsupervised which information will be converted from the cell state into the
methods when building HI prediction models. While this current hidden layer data (to be passed to the next cell).
approach has its benefits when the data doesn’t have As the flow of information is regulated in and out of the
explicit health parameters or labels, it doesn’t consider the cell, the memory cell can preserve its state for longer duration
time taken for the machine to degrade and is slow to (i.e. learning long term dependencies) and influence future
converge to accurate RUL predictions. predictions. Several impactful studies have been conducted in
x As is common in an IIoT device, several sensors recording the field of RUL prognostics that confirm the same. Ref. [6]
data may be dependent on one another such that the output used LSTM to extract important degradation features from
from one may affect the other after some time. Such several operating conditions in a pre-training procedure and
temporal dependencies can be captured by LSTM-based showed enhanced performance compared to RNN and GRU-
models as their architecture is able to retain information. LSTM. Ref. [6] & [7] proposed novel LSTM-based DL
methods to capture degradation and predict RUL. Some of the
2 LITERATURE REVIEW
more recent works on LSTM-based RUL estimation have been
2.1 Deep Learning based Prognostics looking to build a HI instead of relying on degradation trends
for RUL prediction by using LSTM encoder-decoder.
DL is one of the most active sub-fields of machine learning
research. Significant advantages that DL offers in predictive 2.3 LSTM Encoder-Decoder (ED)
models are accuracy and increased processing power, thus,
LSTM-ED have successfully been used for both anomaly
reducing the computational burden [4]. A bulk of literature
detection and sequence-to-sequence learning tasks like natural
studies how deep neural networks can extract hierarchical
language generation and reconstruction, parsing and image
representations from input data and perform a layer-wise non-
captioning [8]. Autoencoders or ED re-construct their inputs
linear transformation into outputs. This allows learning
and learn a lower dimensional representation of input data in
complex concepts from simpler information fed as inputs.
the hidden layer. While the encoder maps a multivariate input
Furthermore, the model parameters don’t have to be trained
sequence to a fixed dimensional vector representation, the
individually. Some studies have used DL to train and re-train
decoder produces target sequences using this vector
their models for both diagnostic and prognostic applications.
representation.
Variants of neural networks are the recent DL techniques
used frequently in the literature for RUL predictions. LSTM The main rationale behind using LSTM-ED was to reduce
networks in particular are widely being explored for the dependency of the traditional prognostic approaches upon
prognostics and have proven to be efficient in supervised and the assumption that the degradation trend must follow a specific
semi-supervised studies. curve., as was followed in many famous case studies [10]-[11].
Autoencoders capture system degradation without relying on
2.2 Long-short Term Memory Network domain knowledge or the assumption regarding degradation
A variant of the traditional recurrent neural network curve. The additional memory cells in LSTM-ED can
(RNN), the LSTM model was initially proposed to address the reconstruct the input time series, store the healthy data
vanishing/exploding gradient problem in RNNs and further information, detect anomalies, reconstruct the error and obtain
improved to avoid any long-term dependency issues [5]. LSTM an unsupervised HI for accurate RUL estimation. Given the
models have since achieved great success in sequence learning novelty of this approach, not a lot of studies exist that make use
tasks such as speech recognition and machine translation where of autoencoders to make accurate RUL predictions.
hidden patterns are to be discovered. The architecture of LSTM 3 MODEL OVERVIEW
uses a well-designed memory cell with four different gates to
replace the activation function of the hidden state in RNN [5]. LSTM-ED is trained to reconstruct the input time series in
Let Ct be any given LSTM cell state and Ct-1 be the the form of embeddings corresponding to the healthy state of
previous state where input is fed. From Ct-1 to Ct, the cell state the system. The rationale behind using autoencoders for this
converts input information from end to end and stores it in a cell purpose is because of their proven ability to retain underlying
memory. Within the memory cell, the model has three gates patterns from time series representation while filtering out
composed out of a sigmoid neural net layer that provide the noise. As the learned model reconstructs subsequences that

Authorized licensed use limited to: University of Massachusetts Amherst. Downloaded on August 02,2020 at 10:24:03 UTC from IEEE Xplore. Restrictions apply.
belong only to the healthy state, the resulting high dimensional final hidden state %& . The decoder on the other hand
reconstruction error for unhealthy subsequences gives the HI at uses %& as its initial hidden state to reconstruct the input time
those specific time instances. The difference in embeddings for series by performing gated transformations. The detailed
normal and faulty machines tends to provide the degree of mathematics behind this transformation and the non-linear
degradation and subsequently the RUL. mapping of fixed dimensional time series between encoder and
We consider a scenario where RUL of a system is to be decoder can be found in ref. [9].
()
estimated at a certain operational instance with multi-sensor The reconstruction sequence (%. ) and error (/ )
historical end-of-life data available. For this given scenario, let obtained using LSTM-ED can mathematically be explained as
the multivariate input time series data be represented by equations (4) and (5) below:
(4)
equation (1) below: %. = 0 3 + 5 (4)
() () () () () () () ()
 = [ ,  ,  , … 
] (1) / = 6% + %7 6 (5)
()
where  and p is the dimension of the sensor data Assuming % is a target healthy state at the beginning,
() ()
(p no. of sensors at a time t),  : is a set of train instances % and %7 are inputs for training and inference respectively,
() (4)
of the system,  are data points on every time step till failure, 3 is the hidden state of the decoder. 0 is the weight and b
and Tn is the time of failure. The LSTM model takes each is the bias. The LSTM-ED model is trained to minimize the loss
() function ( 9 ) by squaring reconstruction error as shown in
sequence of sensor measurements  and learns how to equation (6) below.
model the whole sequence with respect to target RUL [5]. Let
() () 
the time sequence of the ith sensor measurement be given by 9 = :6% + %7 6 (6)
equation (2): As no additional inputs are provided, enough information
() () () () ()
required to reconstruct the time series is stored by the decoder.
 = [ ,  ,  , …  ] (2)
If the resulting reconstructed time series is smooth, it means that
where  = (1,2, … ),  = (1,2 …  ). Let the sensor readings the embeddings learned can retain the necessary degradation
()
be z-normalized over all U instances such that  is the sensor patterns from the input.
reading at time t for sensor i at instance u be transformed to
()
3.2 Obtaining HI and RUL Calculation
 
where ! and " are the mean and standard deviation The obtained encoder function or the embedding with

for the ith sensor data. Equation (3) below denotes a important patterns can be used to highlight any differences
() between the healthy and faulty data. As the deterioration
subsequence of length l from the original input time series 
progresses over time, the healthy embeddings are used to train
represented by: each subsequence of the original input series to provide
()
 (, #) = $ = [% , % , … %& ] (3) prediction error. This difference between actual and predicted
()
Such that 1 '  ' * + # + 1. As seen in many previous values is calculated to be the target normalized HI for that
()
case studies like [11], correlated multi-sensor data require an specific data point and is represented by ; in equation (7):
unsupervised pre-training stage for degradation feature ()
() ()
<> <
extraction before the DL model can perform any supervised ; = () () (7)
<> <?
fine-tuning. Many times, the sensor data is inconsistent and () ()
requires a clustering algorithm for regime partitioning. This where /@ and / are the maximum and minimum values of
allows for a potentially more accurate RUL prediction. PCA has reconstruction errors for u instances. Prediction error exists
also previously been used in multi-sensor prognostics to reduce because the model is only trained on healthy data. With
data dimensionality and preparing the validation set. deterioration, the visible error gives machine HI. The HI can
Referring to equation (2), every failed instance have a binary value, with 0 being poor health and 1 being
()
corresponds to the total operating life and in equation (3), healthy or normal. Most studies estimating HI use some scaling
the input time series has been segmented into fixed length or normalizing procedure. The obtained HI curves for train and
subsequences or windows. The encoder of our LSTM model test instances are later matched for similarity, as shown in
can be trained in an unsupervised fashion to estimate health at equation (8), by varying the Euclidean distance between the
()
the end of the fixed length  (, #). The embeddings for these two.
subsequences at the end of their fixed length retain important The actual RUL prediction of the ground truth or test data
degradation related information and provides with the is performed by the algorithm counting the number of train
reconstruction error for healthy and faulty states. The following instances left after the last cycle of the test instance. If Hu* and
sections explain this process in detail. Hu are the train and test instance HI curves, then their similarity
3.1 Encoder-Decoder based Embeddings can be quantified by equation (8) below, where t is the time lag
introduced to scale the inequality between the curves, d is the
The RHS of equation (3) are the time series data points that Euclidean distance and is an arbitrary parameter proportional
an LSTM encoder iterates through till they reach a fixed to similarity. Finally, the weighted average RUL for the testing

Authorized licensed use limited to: University of Massachusetts Amherst. Downloaded on August 02,2020 at 10:24:03 UTC from IEEE Xplore. Restrictions apply.
instance can be given by equation (9) below, where u and u* length l = 20 can be selected to train the model with respect to
are the train and test instances respectively, A(B , , ) is the healthy data. RUL parameters are selected in the degradation
train and test instance similarity +C  (B , , ) is the Euclidian model through a 5-cross validation process in the code. Other
B
distance similarity between train and test instances, D( ) is the parameters like LSTM units hidden in the encoder and decoder
weighted RUL. can be specified as c = 30, similarity threshold A (B , , ) =
F  (B ,,) 0.0005, the alpha parameter that decides the length of RUL
A(B , , ) = / E H (8) B
( D( ) ) plot, N =0.87 and maximum time lag allowed O = 40.
G
B B
: I( ,,).K D (B ) (,) The LSTM network built here has 100 units in its first layer.
D( ) = : B (9) As training this model involves running several epochs, it
I( ,,)
is always a good practice to use ‘Model Checkpoint’, ‘Call
The statistical formula basically gives the number of
remaining cycles for train instance after the last cycle of test
instance. So, the RUL is calculated iteratively to the previous
B
train instance at a given time t ( D( ) (, ) ). Thus, the HI curve
for a test instance is compared with the HI curve of failed (train)
instance to estimate the RUL of the test instance.
4 EXPERIMENTAL EVALUATIONS
4.1 Data Description
The proposed model is used to calculate the RUL on
NASA’s famous C-MAPSS dataset for turbofan engine
performance degradation and prognostics, generated from ref.
[10]. The data set is a multivariate time series representing
turbofan engine usage from beginning till the end of its life. The
data is from a fleet of engines of the same type. Each entry (row)
in the data set reflects an operational cycle of a specific engine
identified by engine id and cycle time. There are multiple
entries per engine to represent different reporting times. Other
columns represent different features like 3 operational settings
and 21 sensors.
The C-MAPSS data set is divided in 4 subsets each for
training, test and RUL estimation (FD001, FD002, FD003, Figure 1: Data Visualization - Time (Operating Cycles) vs.
FD004) and each subset has data from 100 engines. A subset Sensor Reading Amplitude
can have a different operational condition and consists of Back’ and ‘Early Stopping’ so the model doesn’t overfit and
different number of engines. In the training set, every engine training disruptions don’t interrupt progress. The model learns
has been operated for a specific number of cycles and the last by running iterations through every training HI points and
cycle signifies failure. For instance, if a randomly selected during each iteration, the last output from the previous iteration
engine has 192 distinct time series events (operating from 1st to is appended at the end of the current input. The process stops
192nd cycle), the RUL will start with 192 and go down to 1, when the last output returns a 0 value, meaning termination.
since the last cycle signifies failure. A better visualization of the Figure 2 shows the RUL estimation results of a random fleet of
representative sensor raw data showing degradation in cycles of engines taken from each of the test files. The blue line reflects
10 randomly selected engines can be depicted in figure 1. the actual RUL while the orange line is the predicted RUL. The
There are also some cases where, due to error in data actual RUL values are taken from the ground truth or actual
collection or sensor mounting, one cannot see any visualization RUL within the C-MAPSS dataset.
at all. Such data will eventually impact the RUL plots. While This study is looking at the RUL prediction for 100 engines
there is some visible noise, we can say that not all engines in terms of their cycles. Figure 2 shows our predictions for 6
display degradation at the same time and there is no necessary randomly selected engines using their unique 4 digit IDs out of
correlation between the data as the engine approach failure. The the entire dataset (All 100 can’t be displayed due to space
goal of the model is to identify RUL at the point where the time restrictions). Although we can observe that one of the engine
series for an engine effectively ends. RUL clearly has a larger error as compared to others, this could
4.2 Model Learning and RUL Estimation Results either be understood as an outlier with error in data collection
(indicated earlier with reference to figure 1), or the RUL was
Training and testing datasets were randomly split in an simply difficult to be calculated due to absence of a clear trend
80:20 ratio respectively (A 70:30 split would not have made a in HI for that particular engine. But overall, for the entire
significant difference, and this is up to the discretion of the dataset, we generally observe good RUL trends with no major
analyst). Since we assumed that the initial readings of the re-construction errors. This is also reflective in the model
dataset belong to normal or healthy operation, subsequences of

Authorized licensed use limited to: University of Massachusetts Amherst. Downloaded on August 02,2020 at 10:24:03 UTC from IEEE Xplore. Restrictions apply.
evaluation results in section 4.3.

Figure 2: RUL of Random set of Engines from test files

- Train - Train
- Test - Test

- Train
- Test

Figure 3: Robustness of LSTM-ED Model with Overall Performance Metrics - MAE, R2, Model Loss and Model Accuracy

Authorized licensed use limited to: University of Massachusetts Amherst. Downloaded on August 02,2020 at 10:24:03 UTC from IEEE Xplore. Restrictions apply.
inference can be explored to work with plain operational data
4.3 Performance Metrics
instead to further optimize the area of deep learning based
The literature proposes several performance metrics to prognostics.
evaluate the performance of prognostic models. Robustness of REFERENCES
the proposed LSTM-ED model is evaluated on metrics such as
1. Tao F, Cheng J, Qi Q, Zhang M, Zhang H, Sui F. Digital
Model Accuracy, Mean Absolute Error (MAE), Coefficient of
2 twin-driven product design, manufacturing and service
Determination (R ) and Model Loss. If N is the total number of
with big data. The International Journal of Advanced
test instances, Model Accuracy can be expressed as equation
Manufacturing Technology. 2018 Feb 1;94(9-12):3563-76.
(10) below:
RR S B
2. Boschert S, Rosen R. Digital twin—the simulation aspect.
Q= : BV T(U( ) ) (10) In Mechatronic Futures 2016 (pp. 59-74). Springer, Cham.
S
B B 3. Anis, MD. Towards Remaining Useful Life Prediction in
where error metric TWU( ) X = 1 if the test set error, U( )
Rotating Machine Fault Prognosis: An Exponential
[+O , O ] and O , O are the time lag with null values. Intuitively,
B Degradation Model. 2018 (CMD). IEEE.
TWU( ) X = 0 if there exists a time lag or if O , O have values
4. Zhang J, Wang P, Yan R, Gao RX. Long short-term
greater than 0. Mean Absolute Error is expressed as equation memory for machine remaining life prediction. Journal of
(11) below: manufacturing systems. 2018 1;48:78-86.
 ( B )
YQZ = :S B \U \ (11) 5. Graves A., Schmidhuber J. Framewise phoneme
S  V
( B ) classification with bidirectional LSTM networks. In
A prediction is false positive if U < +O and false
( B ) Proceedings. 2005. IEEE International Joint Conference on
negative if U < O . R tells us how predictable the future
2
Neural Networks, 2005. (Vol. 4, pp. 2047-2052). IEEE.
variable/value is given the previous data points while Model
6. Wu Y, Yuan M, Dong S, Lin L, Liu Y. Remaining useful
Loss is just a measure of a model’s prediction on a single
life estimation of engineered systems using vanilla LSTM
example. Both R2 and Model Loss can be calculated using
neural networks. Neurocomputing. 2018 31;275:167-79.
inbuilt functions in Python’s Keras library.
7. Hinchi AZ, Tkiouat M. Rolling element bearing remaining
Graphical results of each of these performance metrics are
useful life estimation based on a convolutional LSTM.
shown in figure 3. Presented therein is the overall performance
Procedia Computer Science. 2018 1;127:123-32.
of the model on the entire dataset (not a specific engine). While
8. Malhotra P, Ramakrishnan A, Anand G, Vig L, Agarwal P,
there can be observed some minor under fitting towards the end
Shroff G. LSTM-based encoder-decoder for multi-sensor
for each of the three statistical metrics, we can see a high
anomaly detection. arXiv:1607.00148. 2016.
accuracy displayed by the model. We can re-affirm our choice
9. Malhotra P, TV V, Ramakrishnan A, Anand G, Vig L,
of LSTM-ED over LSTM or other DL variants for RUL
Agarwal P, Shroff G. Multi-sensor prognostics using an
predictions when these evaluation results are compared against
unsupervised health index based on lstm encoder-decoder.
the ones observed in previous studies [6][8][9].
arXiv preprint arXiv:1608.06154. 2016.
5 CONCLUSION 10. Mosallam A, Medjaher K, Zerhouni N. Data-driven
prognostic method based on Bayesian approaches for direct
In this paper, we propose a deep learning model that forms
remaining useful life prediction. Journal of Intelligent
the underlying analytics behind developing a digital twin for
Manufacturing. 2016 Oct 1;27(5):1037-48.
prognostic applications. The objective achieved in this work
11. Ellefsen AL, Bjørlykhaug E, Æsøy V, Ushakov S, Zhang
was to demonstrate accurate RUL estimation using an LSTM-
H. RUL predictions for turbofan engine degradation using
ED architecture. The main advantage of the proposed
semi-supervised deep architecture. Reliability Engineering
methodology is that machine health index is obtained without
& System Safety. 2019 1;183:240-5.
the need of making assumptions about the degradation trends in
the dataset, as it has been observed in most of the literature. BIOGRAPHIES
Instead, the model relies on reconstructing input time series data
Mohamad Danish Anis, PhD Student
with training corresponding to healthy data. Naturally, there
Center for Maintenance Optimization & Reliability
exists larger reconstruction error when the reconstructed time
Engineering Laboratory (C-MORE Lab)
series is compared against the degradation data on every time
Department of Mechanical & Industrial Engineering,
instance. Thus, the obtained health index curves for training and
University of Toronto
testing data can be used to accurately estimate the machine
5 King’s College Rd,
RUL. The proposed model displays higher accuracy based on
Toronto, ON M5S 3G8, Canada
the evaluation metrics as compared to previous models.
While the advantages of the proposed method have been e-mail: mdanish.anis@mail.utoronto.ca
highlighted in the paper, a major limitation that needs to be
Danish Anis joined the University of Toronto’s Department of
addressed in the future work is the heavy reliance of such
MIE in September 2018 as a full time PhD student at the Center
reconstruction models on labelled data with run-to-failure
for Maintenance Optimization and Reliability Engineering (C-
history. Models like reinforcement learning and Bayesian
MORE Lab). He received his BEng in Mechanical Engineering

Authorized licensed use limited to: University of Massachusetts Amherst. Downloaded on August 02,2020 at 10:24:03 UTC from IEEE Xplore. Restrictions apply.
and MSc in Reliability Engineering and Asset Management in Iran. Her research interests are in the area of engineering,
2017, both from The University of Manchester, UK. His PhD maintenance optimization and their applications in evidence-
under Prof. CG Lee and Prof. Sharareh Taghipour' co- based decision making. Dr. Taghipour is a member if IISE, SRE
supervision is focused on developing a digital twin, a state-of- and IEEE.
art predictive maintenance tool, to facilitate Remaining Useful
Chi-Guhn Lee, PhD, Peng, Professor
Life (RUL) prediction using data driven and physics based
Center for Maintenance Optimization & Reliability
models. Preliminary research is being conducted on analysis of
Engineering Laboratory
asynchronous multivariate time series data using statistical and
Department of Mechanical & Industrial Engineering
neural network methods.
University of Toronto
Sharareh Taghipour, PhD, LEL, Associate Professor 5 King’s College Rd,
Reliability, Risk and Maintenance Research Laboratory Toronto, ON M5S 3G8, Canada
Department of Mechanical and Industrial Engineering,
e-mail: cglee@mie.utoronto.ca
Ryerson University
Toronto, ON M5B 2K3, Canada Chi-Guhn Lee is currently a Professor at the Department of
Mechanical & Industrial Engineering and Director of C-MORE
e-mail: sharareh@ryerson.ca
Lab at the University of Toronto. He received his Ph.D. in the
Sharareh Taghipour is currently an Associate Professor in the area of Industrial & Operations Engineering from the University
Department of Mechanical and Industrial Engineering, Canada of Michigan, Ann Arbor, and joined the University of Toronto
Research Chair in Physical Asset Management and Director of faculty in 2001. Professor Lee has done both theoretical and
RRMR lab at Ryerson University. She obtained her PhD in applied research in dynamic optimization under uncertainty.
Industrial Engineering from the University of Toronto and her Applications of his interest have been diverse from finance and
BSc in Mathematics and Computer Science as well as her MASc supply chain to reinforcement learning and deep reinforcement
in Industrial Engineering from Sharif University of Technology, learning.

Authorized licensed use limited to: University of Massachusetts Amherst. Downloaded on August 02,2020 at 10:24:03 UTC from IEEE Xplore. Restrictions apply.

You might also like