Vidyaya Amrutham Ashnuthe

B.N.M. Institute Of Technology

PREDICTION OF REMAINING USEFUL LIFE OF
AIRCRAFT ENGINE
USN Name
1BG16EC051 Karthik G Sarode
1BG16EC064 Mohit Kumar Pandey
1BG16EC084 Rajesh Kumar Jha
1BG14EC021 Chaitra Ilkal

Project Guide
Mr. Sudarshan
Assistant Professor
Dept. of ECE
BNMIT, Bengaluru

Reference
Remaining Useful Life
Prediction for a Machine with
Multiple Dependent Features
Based on Bayesian Dynamic
Linear Model and Copulas
Date added to IEEE Access,
August 4, 2017,
Literature Survey
Description
In this paper, a multivariate
degradation modeling approach based
on the Bayesian dynamic linear model
(BDLM) is proposed to calculate the
RULs of degradation features, and the
Copula function is employed to capture
the dependence among RUL
distributions.
Limitation Or Gap
The current model just selects two
typical BDLMs, linear growth
model and seasonal factors model.
In the future, more types of BDLM
should be investigated.
No scope for non-linearity in
degradation process.

Reference
Battery remaining useful life
prediction algorithm based on
support vector regression and
unscented particle filter
Date Added to IEEE Xplore: 15
August 2016
Literature Survey
Description
In this paper, an improved method is
proposed by using support vector
regression-unscented particle filter
(SVR-UPF), which increases the
accuracy of the RUL prediction results.
Firstly, an exponential model is
adopted to approximately express the
degeneration of battery capacity.
Secondly, a novel SVR-UPF method is
presented to solve the degeneracy
phenomenon of the UPF algorithm,
and then it is applied to predict the
battery RUL.
Limitation Or Gap
The algorithm uses a non-linear
exponential model, which may
show a certain deviation at
extreme values .Future works
include changing the regression
algorithm in re-sampling process
and finding a better degeneration
model to imitate the evolution of
capacity.
Reference
A New Method Based on
Stochastic Process Models for
Machine Remaining Useful Life
Prediction
IEEE TRANSACTIONS ON
INSTRUMENTATION AND MEASUREMENT,
VOL. 65, NO. 12, DECEMBER 2016
Description
This paper proposes a new method
based on stochastic process
models for machine RUL
prediction. First, a new stochastic
process model is constructed
considering the multiple variability
sources of machine stochastic
degradation processes
simultaneously. Then the Kalman
particle filtering algorithm is used
to estimate the system states and
predict the RUL.
Limitations / Gap
The limitation is that the model
parameters need to be initialized
using large numbers of available
measurements that are hard to be
acquired in real applications.
MLE is used to initialize the
parameters.
Accelerated degradation alters
degradation mechanism.
 Run ML
Procure data & Suggest class Store the data
algorithm to
pre process based on RUL for future use
find RUL

RUL Classes,

➢ CLASS 2 : RUL from 0 to 10, most economic valuable

➢ CLASS 1 : RUL from 11 to 40,
➢ CLASS 0 : RUL above 40,
▪ The data for this project has been taken from NASA's Turbofan Engine Degradation Simulation, wherein they used
the simulated run to failure events from different no's of engines. The data package comes with 4 datasets each for
training and testing the model and one set containing RUL(Remaining useful life) values.
 SENSORS USED
Sensor No. Symbol Description Units
1 T2 Total temperature at fan inlet °R

2 T24 Total temperature at LPC outlet °R

3 T30 Total temperature at HPC outlet °R

4 T50 Total temperature at LPT outlet °R

5 P2 Pressure at fan inlet psia
6 P15 Total pressure in bypass-duct psia

7 P30 Total pressure at HPC outlet psia

8 Nf Physical fan speed rpm
9 Nc Physical core speed rpm
10 Epr Engine Pressure ratio −
11 Ps30 Static pressure at HPC outlet psia

12 Phi Ratio of fuel flow to Ps30 pps/psi

13 NRf Corrected fan speed rpm
14 fNR Corrected core speed rpm
15 BPRht Bypass ratio −
16 carBN Burner fuel–air ratio −
17 BleedP Bleed enthalpy −
18 f-dmd Demanded fan speed rpm

19 CNfR-dmd Demanded corrected fan speed rpm

20 W31 HPT coolant bleed lbm/s

21 W32 LPT coolant bleed lbm/s
 Flow Chart / Approach

Data
Data collection Data Analysis Data
Cleaning/wrangling
Visualization

Supervised
Learning
Applying ML Splitting Feature
Algorithms Training Dataset Engineering
Unsupervised
Learning

Data Standardisation Dummy

Variable
 Supervised Unsupervised
Learning Learning

Random Neural Network

SVM
Forest

Selecting the
model

Comparing

Optimized Output
Data visualization of the various sensor outputs plot of all the sensor data
• OLS is the method used to find the simple linear regression of a set of data.

• In statistics, ordinary least squares are a type of linear least squares method for estimating the
unknown parameters in a linear regression model.

• The OLS R- Squared produces only 65% accuracy.

• Hypothesis testing is used to assess the plausibility of a hypothesis by
using sample data A statistical hypothesis is a hypothesis that is testable
on the basis of observing a process that is modeled via a set of random
variables.

• A statistical hypotheses test performed when the test statistic is chi-

squared distributed under the null hypothesis.

• Before training the model we first know about the importance of the
feature such as which feature is more useful.

• Below figure show the chi2 test scores of the different features
used in this project, the scores are in descending order the variables
which is more important at top and so on.
• Regression algorithm similar to SVM
• In SVR we try to fit the error within a certain threshold

We see value of R-square which is almost 1 that

implies that the accuracy of the model is 100%
and it is not good for the model.

We observe the fig in which the x(Actual value of

X_test i.e y_test) and y(predicted value of X_test)
variables are fitted linearly and this is the perfect
case of Over fitting.

So, we did not use this model for further

prediction or operation.

SVR Result
➢ Random forests or random decision forests are an Ensemble learning method Classification /
regression by constructing a multitude of decision trees at training time and outputting the class that is
the mode of the classes (classification) or mean prediction (regression) of the individual trees.

➢ Random decision forests correct for decision trees' habit of Over Fitting.
❑ Figure shows the relation between the x (Actual value
of X_test i.e y_test) and y (predicted value of X_test).

❑ The R-squared value is .55 and mean square value is

2123.71.

❑ In Regression model the accuracy of the model is

determined by their R-square value greater the R-
square value higher is the accuracy and lower the
mean square error better is the result.

❑ By viewing the result of random forest, we did not

satisfy because the accuracy is not good as we want.

Random Forest Result

SGD is efficient discriminative learning for linear classifiers and convex loss functions.

The word ‘stochastic‘ means a system or a process that is linked with a random probability.

Gradient Descent can be described as an iterative method which is used to find the values of the parameters of a
function that minimizes the cost function as much as possible.

Thus in SGD, we find out the gradient of the cost function of a single example at each iteration instead of the sum
of the gradient of the cost function of all the examples.

Path taken by
Stochastic
Gradient Descent
In this sequential model we use the different number of epochs (200, 300 and 2000) and batch size (100, 200 and 900)
and checks the performance of the model. In the model two activation function Sigmoidal and Rectifier function is
used, the learning rate is 0.001 and the optimizer is Stochastic Gradient Descent (SGD).

Figure shows the Error (mean absolute error)

and Loss (mean square error) at each
iterations or epoch.

Viewing of both figures we observe that with

increment of number of epochs the mean
absolute error and mean square error is
reduces this implies that the losses and error
is very low in this model which increases the
performance of the model.

Loss and Error

Fig showing the difference between predicted and true remaining useful life, here the x-axis (i.e unit_id of the
engines) is common for both True and predicted RUL.

By viewing this plot, we say that the performance of the model is good and it is observed that the accuracy of
the model is very good if the remaining useful life of any engine exceeds the value of 200 life cycles.

Comparison between true and predicted RUL

 True RUL Predicted Differenc
Figure showing the result of our model here we store the RUL e
data into the .csv or excel file in which there are 3 column first
259 [270.0] [11.0]
column is the True value of the RUL 2nd column is the
Predicted value of the RUL and 3rd column is the difference 258 [205.0] [-53.0]
between the Predicted value and True value. 257 [203.0] [-54.0]
256 [190.0] [-66.0]

We get a huge list of RULs, the sample output is shown here. 255 [190.0] [-65.0]
We see that the accuracy is better using SGD than other 254 [198.0] [-56.0]
algorithms. 253 [186.0] [-67.0]
252 [227.0] [-25.0]
251 [195.0] [-56.0]
250 [192.0] [-58.0]
Final Result
A breakthrough in machine learning would be worth
ten Microsofts

-Bill Gates