HYBRID PHYSICS-BASED AND DATA-DRIVEN PHM

H. Hanachi, W. Yu, I.Y. Kim and C.K. Mechefske

Queen’s University, Kingston, ON, Canada, K7L 3N6
{houman.hanachi; wennian.yu; kimiy; chris.mechefske}@queensu.ca

ABSTRACT
To address the core needs and competitive demand within the field of prognostics and health
management (PHM), an innovative research direction is described and demonstrated. The new
methodology is a hybrid combination of physics-based and data-driven PHM. The process of
combining intelligent sensor based monitoring and diagnostics to enhance performance and
reliability throughout a system’s life is the basis of data-drive PHM. Use of physics-based models
to predict system response is the basis of physics-based PHM. Both approaches have distinct
advantages and disadvantages when applied to specialized integrated electrical, mechanical, and
software components. Individually, they fall short of delivering broad capability and value across
the entire range of existing and future intelligent devices. While there are many methods and
algorithms for health management, such as Fuzzy Logic Predictions, Bayesian Belief Networks,
Artificial Neural Networks, and Autoregressive Moving Average to name a few, Adaptive Neuro-
Fuzzy Inference Systems (ANFIS) are one of the most effective approaches that can be used for
both data-driven and physics-based PHM diagnostics and prognostics. These ANFISs can be
trained by applying sensor data as boundary conditions in physics-based computer aided analysis
over the entire operating envelopes of a system. Kalman filter and particle filter are systematic
sequential fusion frameworks to combine the diagnostic results from the measurement signals and
the degradation models. Hybrid PHM systems that utilize both physics-based and data-driven
approaches are a relatively unexplored multi-disciplinary methodology. By using this hybrid
approach an established ANFIS model can interpolate between input parameters and desired
output information to optimize system function as well as facilitate ongoing self-training using
new input data. Successful integration of physics-based and data-driven models and its application
will be a truly transformative breakthrough in intelligent system development. This paper shows
the results of preliminary methodology development and application on gas turbine engines and
gearboxes.

1. INTRODUCTION
From the earliest days of human existence, we have witnessed degradation and failure of our tools
during use. It took some time until we could compile our experience to form a basic knowledge of
failure prediction in machinery. A classical approach is to employ the laws of physics to make
fault models with two main objectives:
1) To quantify the fault symptom as a function of fault magnitude.
2) To predict growth of the fault with operating time.
The former, also referred to as a measurement model, establishes instantaneous causality
relationships between the root causes of the faults and the symptoms manifested in the
measurements of a faulty system. For instance, principles of continuum mechanics can be used to
calculate natural frequencies of a structure in the presence of a crack with known dimensions.
When fault detection is the objective, an inverse model is needed to receive the fault symptoms
and step backward to infer the component faults leading to such symptoms. In this regard, a
physics-based diagnostic approach is the inverse of physics-based measurement modeling. It
requires a thorough understanding of the system under study and advanced modeling tools for the
analysis. There is no consideration of time in this modeling and all relations are considered
simultaneously. This instantaneous causality relationship and the inference process between the
faults and the symptoms are demonstrated in Fig. 1.
Another methodology is to include the operating time as an input for modeling. This is in fact the
second approach referred to above. Here, the knowledge of the physics of failure is used for
developing a function of time to predict the fault magnitude at some time in the future. For
instance, the Paris Law can be applied to calculate crack length as a function of loading cycles.
However, fault growth phenomena are generally complicated resulting in available prediction
models needing to include a great deal of uncertainty.
Information compiled from multiple sources is believed to be more reliable than that from a single
source. This can be proven theoretically, while it has been verified in numerous applications such
as sensor fusion [1]. This philosophy has inspired the development of hybrid data-model fusion
frameworks, where the faults are both inferred from their observable symptoms and predicted by
fault growth models. The results are sequentially fused together to achieve a higher accuracy.
In this paper, the structure of the hybrid framework for data-model fusion PHM is laid out in the
next section, and two applications of this framework in machinery fault detection and prediction
are presented in later chapters. A conclusive summary is provided the end of the paper.
Component Faults

System Symptoms
CAUSALITY

Component System System

Fault Growth
Performance Performance Measurements

INFERENCE

Fig. 1. Causality relation of faults and symptoms, and inference process to detect the faults

2. HYBRD FRAMEWORK
To describe this concept, assume a two-shaft gearbox with a spall on the flank of a gear tooth. The
fault is parametrized with a vector 𝑥 in terms of the dimensions of the spall. The gearbox is
operating at condition 𝑢 (the load and the shaft speed) and there are some measurements 𝑦
available from the system (a vibration transducer signal). The measurement signal is influenced
by both the fault parameter and the operating conditions. We can build a measurement model based
on vibration principals to formulate the relationship between the observed measurements and the
operating condition. In discrete form, at a time step 𝑘,
𝑦𝑘 = 𝐺𝑘 (𝑥𝑘 , 𝑢𝑘 ) + 𝜀𝑘 , (1)
where 𝜀𝑘 is the error due to model inaccuracy and the measurement noise. We know that when
operating the gearbox under loaded conditions, the spall grows as a function of time. In discrete
form, given the operating condition has not changed between consecutive time steps, the fault level
at time step 𝑘 depends on its level at the previous time step and the loading condition in the
meantime.
𝑥𝑘 = 𝐹𝑘 (𝑥𝑘−1 , 𝑢𝑘 ) + 𝜏𝑘 , (2)

where 𝜏𝑘 is the prediction error, also known as the process noise, caused by prediction uncertainty.
In Eq. 2, we have assumed that the fault growth is a memoryless process, where the state of the
next step depends on the immediate previous step and not the older steps. This type of process is
called a first order Markov process. With this approach, there are two sources of information for
estimating the fault through the measurement model and the prediction model as in Eqs. 1 and 2.
This concept is not restricted to only a gearbox as in our example, but it is extendable to any system
experiencing degradation.
Fusion of the results from the predictions and the measurements is through a systematic sequential
process. Starting from the prediction model, at time step 𝑘, Eq. 2 predicts the fault based on its
value at the previous time step and the operating condition. The prediction includes an uncertainty
characterized with the distribution of the prediction error 𝜏𝑘 . At the same time step, new
measurements from the system are obtained. The predicted fault (along with its distribution) will
be updated with respect to likelihood of the new measurement. Figure 2 shows the process of
hybrid fault detection for a first order Markovian system. Depending on the design of the prediction
model and the measurement model, different data-model fusion techniques have been applied for
fault detection. Here, we elaborate on two techniques widely used in this field.

𝑢𝑘 SYSTEM 𝑦𝑘

Data-model fusion:
... 𝑥𝑘−1 𝑥 ′ = 𝐹𝑘 (𝑥𝑘−1 , 𝑢𝑘 ) + 𝜏𝑘 𝑥𝑘 ...
൜ 𝑘
𝑦𝑘 ′ = 𝐺𝑘 (𝑥𝑘 ′, 𝑢𝑘 ) + 𝜀𝑘
𝑥𝑘 ~𝑓(𝑥𝑘 | 𝑦𝑘 ′ = 𝑦𝑘 )

Fig. 2. Sequential data-model fusion framework for fault detection.

2.1 Kalman Filter:

If the prediction and the measurement models are linear and their errors have Gaussian distribution,
the optimal solution can be found in closed form with a fusion framework known as a Kalman
filter (KF), where the estimated fault will have the highest accuracy achievable with available
information. Linear models for Eqs. 2 and 1 can be written in matrix form,
𝑥𝑘 = 𝐴𝑘 𝑥𝑘−1 + 𝐵𝑘 𝑢𝑘 + 𝜏𝑘 , (3)

𝑦𝑘 = 𝐶𝑘 𝑥𝑘 + 𝜀𝑘 , (4)

where, 𝐴𝑘 and 𝐵𝑘 are the fault transition model and the input model matrices respectively, and 𝜏𝑘
is a zero-mean process noise vector with normal distribution. 𝐶𝑘 is the matrix of measurement
model with measurement noise vector 𝜀𝑘 . The system fault optimally estimated by KF is
𝑥𝑘 = 𝑥̃𝑘 + 𝐾𝑘 (𝑦𝑘 − 𝐶𝑘 𝑥̃𝑘 ), (5)

where, 𝑥̃𝑘 is the priori fault, and 𝐾𝑘 is the optimal gain as per the following,
𝑥̃𝑘 = 𝐴𝑘 𝑥𝑘−1 + 𝐵𝑘 𝑢𝑘 , (6)

𝐾𝑘 = [cov(𝑥𝑘 − 𝑥̃𝑘 )] 𝐶𝑘𝑇 [cov(𝑦𝑘 − 𝐶𝑘 𝑥̃𝑘 )]−1 . (7)

Equation 3 to 7 constitute the data-model fusion framework with generic KF for linear systems.
When the fault transition model and the input models are nonlinear, modified versions of KF such
as extended KF and unscented KF can be used. The KF and its variants have been extensively used
for fault detection in machines. For instance, in a comparative study, Simon applied linear,
extended and unscented KFs for fault estimation in a turbofan engine with nine fault modes across
the gas path. Comparison of the three fault detection frameworks showed that both extended and
unscented KFs outperform the linear KF. The computational effort of using extended and
unscented KFs is higher than the linear KF by one order and two orders of magnitude respectively.
Simon concluded that nonlinearity of GT dynamics can be acceptably addressed by using an
extended KF for fault detection, and an unscented KF is unnecessary, given its massive
computational expense [2].

2.2 Particle Filter:

A simulation-based approach for data-model fusion is known as the particle filter (PF). It can be
applied for fault detection on nonlinear systems with non-Gaussian noise structure. PF is a
numerical simulation for sequential Bayesian state estimation. In this approach, to represent the
distribution of the fault state, a set of fault scenarios, referred to as particles, along with their
corresponding weights {𝑥𝑘𝑖 , 𝜔𝑘𝑖 }𝑖=1∶𝑚 are considered at each time step as shown in Fig. 3. A delta
Dirac function can be used to construct a discrete distribution for the fault state
𝑥𝑘 ~ ∑𝑚 𝑖 𝑖
𝑖=1 𝜔𝑘 𝛿(𝑥𝑘 − 𝑥𝑘 ). (8)

However, the density function can be regularized with kernel density 𝐾ℎ , resulting in a continuous
density function [3], shown as the green curve in Fig. 3.
𝑥𝑘 ~ ∑𝑚 𝑖 𝑖
𝑖=1 𝜔𝑘 𝐾ℎ (𝑥𝑘 − 𝑥𝑘 ). (9)

{𝑥𝑘𝑖 , 𝜔𝑘𝑖 }𝑚
𝑖=1

𝜔𝑘𝑖

𝑥𝑘𝑖

Fig. 3. Fault probability distribution function in descrete and regularized form.

Using Eq. 2, the fault scenarios are propagated to the next time step as priori particles. Once a new
measurement from the system is available the distribution of the priories can be updated. At this
step, the updated weights of the particles are set proportional to the likelihood of the measurement
for each particle. As per the same sequence, diagnostic process through fusion of the prediction
result and measurement result is repeated for the next time steps.
Different versions of PF such as auxiliary PF and regularized PF have been developed and tested
for fault diagnostics. Regularized PF is one of the variants that has shown promising performance
while having relatively simple structure. Table 1 shows the sequential process of fault detection
with regularized PF for a system with prediction model and measurement model as in Eqs. 2
and 1.

Table 1: Sequential process of regularized PF

{(𝑥𝑘𝑖 , 𝜔𝑘𝑖 )}𝑖=1∶𝑚 = RPF [{(𝑥𝑘−1
𝑖 𝑖
, 𝜔𝑘−1 )}𝑖=1∶𝑚 , 𝑢𝑘 , 𝑦𝑘 ]
 Draw process noise: 𝑖
𝜏𝑘 ~𝑓𝜏 (𝜏𝑘 )
 Propagate priori: 𝑥𝑘𝑖 = 𝐹𝑘 (𝑥𝑘−1
𝑖
, 𝑢𝑘 ) + 𝜏𝑘𝑖
 Update the weights: 𝜔𝑘𝑖 ∝ 𝜔𝑘−1𝑖
𝑓𝜀 (𝑦𝑘 − 𝐺𝑘 (𝑥𝑘𝑖 , 𝑢𝑘 ))
 Construct posterior density: 𝑓𝑥 (𝑥𝑘 |𝑥𝑘−1 , 𝑢𝑘 , 𝑦𝑘 ) ≃ ∑𝑚 𝑖 𝑖
𝑖=1 𝜔𝑘 𝐾ℎ (𝑥𝑘 − 𝑥𝑘 )
 𝑖 𝑖
Resample equally likely particles: (𝑥𝑘 , 𝜔𝑘 = 1/𝑚)~𝑓𝑥 (𝑥𝑘 |𝑥𝑘−1 , 𝑢𝑘 , 𝑦𝑘 )

In the next sections, examples of hybrid data-model fusion techniques for fault detection are
presented.

3. GAS TURBINE DIAGNOSTICS

In this section, a case study for application of the data-model fusion technique is presented. The
hybrid framework is applied for multimode diagnosis in a single shaft SGT100 gas turbine used
for power generation. System inputs, outputs and internal faults are explained in table 2. The
objective is to use the measurable parameters and inputs of the gas turbine to detect four system
faults (compressor mass flow decrease, compressor efficiency decline, turbine mass flow increase,
and turbine efficiency decline). Two methods are examined and compared for this purpose:
measurement-based diagnostics and hybrid diagnostics.
The causality relationship between component faults and their system level symptoms, as shown
in Fig.1, is formulated in Eq. 1. Measurement-based diagnosis attempts to build the inverse of the
measurement model and infer the faults using the system measurements. We call this model the
diagnostic model.
𝑥𝑘 = 𝐺𝑘−1 (𝑦𝑘 , 𝑢𝑘 ) + 𝜀′𝑘 , (10)

The forward measurement model of the gas turbine can be constructed using a heat balance
approach [4]. The inverse model however, is not easy to make. As we see in table 2, the diagnostic
model should receive eight parameters (four input parameters and four system measurements), and
return four fault parameters. A practical approach is to obtain a trainable numerical modeling
platform and fit it to a bank of data generated with forward modeling. With high-dimensionality
of the input of the diagnostic model, numerical models such as multi-variate polynomials, artificial
neural networks and adaptive neuro-fuzzy inference system (ANFIS) have been practiced [5-7]. [5][6][7]
In a recent work, authors employed ANFIS structure and trained it with simulation results from a
high-fidelity gas turbine model under different fault scenarios [8]. ANFIS was introduced by Jang
in early 1990s [9]. It uses the adaptive capability of the neural network to adjust the fuzzy model
parameters such that it optimally fits the training data. In this work, we use an ANFIS structure
with a first-order Sugeno model for the fuzzy inference, and use two membership functions for
each input parameter.

Table 2: Measurable parameters and internal faults

vector symbol parameters
𝜌𝑊𝐶 : compressor mass flow decrease
𝜌𝜂𝐶 : compressor efficiency decline
𝑥: fault parameter
𝜌𝑊𝑇 : turbine mass flow increase
𝜌𝜂𝑇 : turbine efficiency decline
𝑃𝑊: output power
𝑁: shaft speed*
𝑢: system input 𝑇𝑎𝑚 : ambient temperature
𝑃𝑎𝑚 : ambient pressure
𝜙𝑎𝑚 : relative humidity
𝑊𝐹 : fuel mass flow
𝑃𝐶𝑜 : compressor discharge pressure
𝑦: system measurement
𝑇𝐶𝑜 : compressor discharge temperature
𝐸𝐺𝑇: exhaust gas temperature
*: the gas turbine in this study is set on a constant shaft speed (16500 RPM) for power generation.

The model is trained on 1000 data sets sampled from a total of 35,000 data sets representing a
four-year period of operation of the gas turbine. The training data is contaminated with noise of
different signal to noise ratios. The ANFIS-based diagnostic model is then run to predict faults
over the entire operating period using the noisy measurements. To quantify the diagnostic
accuracy, normalized root mean squared error is taken as the metric for each of the faults. Figure
4 partly shows the diagnostic results. The blue circles show the diagnostic results using the
instantaneous measurements with 𝑆𝑁𝑅𝑑𝐵 = 50 dB. The results appear to be highly scattered, which
translates to a large diagnostic error. The average diagnostic error for four faults at this condition
is 0.594.
In the second trial, we examine the introduced hybrid framework for fault detection. In the hybrid
framework, a fault transition model is required as in Eq. 2. We obtain and use an exponential model
for fault transition from the literature for the faults considered.
𝑥𝑘 = 𝑒 𝑏 ∆𝑡 𝑥𝑘−1 + 𝑎 (1 − 𝑒 𝑏 ∆𝑡 ) , (11)

where ∆𝑡 is the time duration between consecutive time steps. Data-model fusion is implemented
with the regularized PF according to table 1. The diagnostic results for the same time frame are
presented in Fig. 4 with black color. It is visible that the hybrid diagnostic results are much closer
to the actual system faults (the red curve) compared to model-based diagnostic results. The average
diagnostic error for four different faults at this condition is 0.063, which is 9.4 times smaller than
model-based error.
Diagnostic errors depend on the noise level in the measurement signals. The dependency of error
to the noise level is plotted in Fig. 5(a). Diagnostic error significantly improves with the hybrid
approach. The hybrid approach with a given signal to noise ratio yields an accuracy comparable
to the model-based method with about 20 dB cleaner measurement signal. For instance, the
diagnostic error of the hybrid method at 𝑆𝑁𝑅𝑑𝐵 = 40 dB is 0.184, whereas to achieve the same
accuracy with the model-based method, the 𝑆𝑁𝑅𝑑𝐵 should be over 60 dB.
Figure 5(b) shows the improvement in diagnostic errors with the hybrid method at different noise
levels compared to the model-based method. When the 𝑆𝑁𝑅𝑑𝐵 varies from 30 to 55 dB, the hybrid
method has the largest effect and reduces the diagnostic error down to 0.1. With larger 𝑆𝑁𝑅𝑑𝐵
values, the improvement effect gradually reduces, however it never reaches more than 0.5, which
is still over 50% reduction of error.

Fig. 4. Fault detection in multiple modes with model-based diagnostics and hybrid diagnositcs at
𝑆𝑁𝑅𝑑𝐵 =50 dB.

Fig. 5. Diagnostic error, (a) variation with noise level, and (b) improvement of error with hybrid
framework.
4. GEARBOX DIAGNOSTICS

4.1 Validation of Physical Model

A classic approach to assess the performance of a gear transmission system is to predict the
expected dynamic performance parameters (for use as baseline data) through dynamic analysis
based on a physical model of the gear system. A model-based approach is useful to solve remaining
useful life (RUL) estimation when there is not enough failure data available [10]. Normally, such
a model requires experimental verifications to ensure its feasibility [11]. Taking as an example, a
one-stage gearbox, which consists of a 1:1 speed ratio spur gear pair (20 teeth each) shown in Fig.
6(a), we demonstrate this process. A shallow spalling defect (rectangular-shaped in this case) was
implemented on one of the driving gear teeth as shown in Fig. 6(b). Two encoders were mounted
on the driving and driven shafts respectively to collect the angular displacements of the two gears.
An accelerometer was attached on the outer surface of the gearbox in the upright direction near
the damaged gear, in order to collect vibration data. During the experiment, the input speed of the
driving gear was 600 rpm, and the load applied on the gear system was 21 Nm.

(a) (b)
1

Fig. 6. Spur gear pair with surface defect: (a) photos of the experimental set-up (1-
accelerometer; 2- encoder on driving shaft; 3- encoder on driven shaft), (b) tooth surface defect.

The governing equations of motion (normally partial differential equations) of the gear system
are derived based on the physical model built in [12,13]:
[12][13]

𝑀𝑞̈ + 𝐶𝑞̇ + (𝐾𝑏 + 𝐾𝑔 (𝑡, 𝑞))𝑞 = 𝐹(𝑡, 𝑞) (12)

where 𝑞 is the degree of freedom vector of the gear pair. 𝑀, 𝐶, 𝐾𝑏 are the mass, damping and
bearing stiffness matrices, respectively. 𝐹 is the excitation resulting from the applied torques,
inertial effects produced by unsteady rigid-body rotations (gear eccentricities) and the tooth
surface defect. Detailed descriptions about these matrices and vectors are given in [12,13].
Figure 7 shows comparisons of the time-history acceleration, power spectrum density (PSD) and
spectrum after Hilbert envelope analysis on the resonance range 1700-2300 Hz between the
experimental measured data and simulation data. In the time-history acceleration plots, both the
experimental and simulation results are characterized by periodic impulses, which are due to the
periodic meshing of the defected tooth per rotation of the driving gear. In the PSD plots, there are
broad-band frequency components near 2 kHz, which is the primary resonance frequency of the
gear system. There are also multiple sidebands near the resonance frequency range, and the interval
of the sidebands is approximately 200 Hz, which is the mesh frequency of the gear system. In the
Hilbert spectra, there are a series of densely arrayed harmonics of 10 Hz at both the experimental
and simulation results, which correspond to the rotating frequency of the driving gear. The
amplitudes of these characteristic frequency components can be monitored to reflect the fault state
of the tooth surface.

(a) (b)

2000Hz
2016Hz
2200Hz
1814Hz
1800Hz
2227Hz

10Hz 200Hz
201Hz

10Hz 403Hz
400Hz

Fig. 7. Comparisons between the simulation results and experimental results: (a) experimental
results, (b) simulation results.

In general, the simulation results based on the physical model agree well with the experimental
results, which validates the feasibility of the proposed physical model in predicting the dynamic
performance of the gear system with the inclusion of the localized tooth surface defect. In fact, the
simulation accuracy and efficiency directly depend on the physical model built for the gear system.
Therefore, discrepancies between the simulation results and experimental results can be evaluated
for the further improvement of the physical model.

4.2 Performance Parameters

Measurements through sensors on monitored components are usually noisy multidimensional time
series signals. Thus it is essential to extract characteristic performance parameters from original
signals that represent the degradation evolution over time [11]. Taking the above-mentioned
gearbox as an example, statistical indicators such as the RMS, kurtosis (Kurt), Crest factor (C) of
the time-history acceleration, and the average of the amplitudes of the characteristic frequency
(rotating frequency) and its harmonics in the Hilbert spectrum (AH ) can all be monitored to reflect
the fault status of the tooth surface defect. These parameters constitute the instantaneous
measurable performance vector y of the gearbox:
𝑦 = [RMS, Kurt, 𝐶, AH ] (13)

The input speed 𝜔 and applied load 𝑇 on the gearbox are the two main control parameters in the
control input vector 𝑢:
𝑢 = [𝜔, 𝑇] (14)

For the proposed physical model, the physical status 𝑥 of a rectangular-shaped tooth surface defect
(dimensions of the spall) is:
𝑥 = [𝑤𝑠 , 𝑙𝑠 , 𝑑𝑠 ] (15)

where 𝑤𝑠 , 𝑙𝑠 and 𝑑𝑠 are the width, length in the direction of gear tooth roll, and depth of the
rectangular-shaped spall as shown in Fig. 8.

ls

ws
ds

Fig. 8. Dimensions of the rectangular-shaped spall.

Based on the physical model (Eq. 1), the performance vector 𝑦 can be modelled as a function of
the control vector 𝑢 and fault state 𝑥 of the gear system:
𝑦 = 𝐺(𝑥, 𝑢) (16)

Figure 9 shows the variations of each performance parameter in 𝑦 with the spall width 𝑤𝑠
increasing from 0.1 inch to 0.5 inch (2.5 mm to 12.7 mm) for different rotating speeds. It can be
found that most of the performance parameters in 𝑦 increase rapidly with the increase of the fault
state, which demonstrates these parameters can serve as degradation indicators for the assessment
of gear tooth surface condition. Besides, baseline data that correlates the gear performance vector
𝑦, control vector 𝑢 and fault state 𝑥 can be obtained through the dynamic simulation results of the
physical model built for the gear system. This database can be used to train a fault growth model
of the gear system, which will provide gear condition information and RUL prediction based on
given measurement input from the raw monitoring signals of the gear system.
It should be noted that, in this section, we assume that the spalling defect propagates only along
the tooth face width direction. In the future, a data-driven fault growth model will be developed
and fused with the proposed model-based prediction method for more general and accurate
gearbox diagnosis and prognosis.
 Fig. 9. Variations of the performance indicators with different input speeds and tooth surface
fault state.

5. CONCLUSION
In this work, a hybrid framework for fault diagnostics is presented for estimating the unmeasurable
faults from the system measurements. The hybrid framework uses two sources of information for
fault estimation; measurement signals, believed to include the fault symptoms, and fault growth
model(s), which predict the faults based on their value in the recent past. The hybrid framework
fuses together results from the measurement data and the model in a systematic manner. Different
fusion techniques have been developed and used for hybrid frameworks. Kalman filter and Particle
filter along with their variants are among the most popular techniques covering the range of linear
systems to highly nonlinear systems.
Examples of applications of a hybrid technique for fault detection in gas turbines and gearboxes
are presented. For the gas turbine, the results are compared to simple model based diagnostic
results. The results show that the hybrid diagnosis leads to up to ten times more accurate results,
compared to the results of simple model-based diagnostic approach.
For the gearbox, a physical model-based approach is introduced for the gear fault diagnosis and
prognosis. The dynamic model of a 1:1 speed ratio spur gear pair with a localized shallow spalling
defect was built, from which the simulation results were directly compared with the experimental
measured result. Several performance indicators were used to reflect the fault state of the gear
tooth defect. Baseline data that correlates the gear performance vector, control vector and fault
state vector can be obtained, which can be used for training fault growth model of the gearbox.
The developed hybrid framework has great potential for improving the diagnostic results in various
systems when the faults are inaccessible for direct measurement.
6. ACKNOWLEDGEMENTS
This work was financially supported by Fond de Recherche Nature et Technologies (FRQNT)
from the Quebec government in Canada, and the Natural Sciences and Engineering Research
Council (NSERC) of Canada.

7. REFERENCES
[1] R. R. Brooks and S. S. Iyengar, 1998, Multi-sensor fusion : fundamentals and applications with
software, Prentice Hall PTR.
[2] D. Simon, 2008, A comparison of filtering approaches for aircraft engine health estimation,
Aerosp. Sci. Technol., vol. 12, no. 4, pp. 276–284.
[3] N. Oudjane and C. Musso, 2000, Progressive correction for regularized particle filters, in
Proceedings of the Third International Conference on Information Fusion, p. THB2/10-THB2/17
vol.2.
[4] H. Hanachi, J. Liu, A. Banerjee, Y. Chen, and A. Koul, 2015, A Physics-Based Modeling
Approach for Performance Monitoring in Gas Turbine Engines, IEEE Trans. Reliab., vol. 64, no.
1, pp. 197–205.
[5] I. Loboda and Y. Feldshteyn, 2010, Polynomials and Neural Networks for Gas Turbine
Monitoring: A Comparative Study, in Volume 3: Controls, Diagnostics and Instrumentation; Cycle
Innovations; Marine, pp. 417–427.
[6] K. Salahshoor, M. S. Khoshro, and M. Kordestani, 2011, Fault detection and diagnosis of an
industrial steam turbine using a distributed configuration of adaptive neuro-fuzzy inference
systems, Simul. Model. Pract. Theory, vol. 19, no. 5, pp. 1280–1293.
[7] A. Cavarzere and M. Venturini, 2012, Application of Forecasting Methodologies to Predict
Gas Turbine Behavior Over Time, J. Eng. Gas Turbines Power, vol. 134, no. 1, p. 12401.
[8] H. Hanachi, J. Liu, and C. Mechefske, 2017, Multi-mode Diagnosis of a Gas Turbine Engine
using an Adaptive Neuro-Fuzzy System, Chinese J. Aeronaut., (accepted).
[9] J. S. R. Jang, 1993, ANFIS: adaptive-network-based fuzzy inference system, IEEE Trans. Syst.
Man. Cybern., vol. 23, no. 3, pp. 665–685.
[10] S. Zheng, K. Ristovski, and A. Farahat, 2017, Long Short-Term Memory Network for
Remaining Useful Life estimation, Progn. Heal.
[11] A. Mosallam, K. Medjaher, and N. Zerhouni, 2016, Data-driven prognostic method based on
Bayesian approaches for direct remaining useful life prediction, J. Intell. Manuf., vol. 27, no. 5,
pp. 1037–1048.
[12] W. Yu, C. K. Mechefske, and M. Timusk, 2017, The dynamic coupling behaviour of a
cylindrical geared rotor system subjected to gear eccentricities, Mech. Mach. Theory, vol. 107, pp.
105–122.
[13] P. Velex, 2012, On the Modelling of Spur and Helical Gear Dynamic Behaviour, in
Mechancial Engineering, Murat Gokc., InTech Open Access Publisher, pp. 75–106.
BIOGRAPHIES

Houman Hanachi received his B.Sc. and M.Sc. in mechanical engineering from
Sharif University of Technology, Tehran, Iran and his Ph.D. in mechanical
engineering from Carleton University, Ottawa, Canada. He worked as a
professional engineer in automotive industry for thirteen years and is currently a
postdoctoral research fellow at Queen’s University, Kingston, Canada. His
research focus is on diagnostics, prognostics and health management, and hybrid
techniques for state estimation and fault detection in complex systems such as gas
turbine engines.

Wennian Yu received his B.Sc. in mechanical engineering and automation from

Wuhan University of Technology, Wuhan, China; his M.Sc. in mechatronic
engineering from Chongqing University, Chongqing, China and his Ph.D. in
mechanical engineering from Queen’s University, Kingston, Canada. He is
currently a postdoctoral research fellow at Queen’s University under the
supervision of Dr. Mechefske. His research focus is on the dynamic modelling of
gear transmission system, and condition monitoring of gear system for
diagnostics, prognostics and health management.

Il Yong Kim received his B.Sc. in mechanical engineering from Korea

University and his M.Sc. and Ph.D. in mechanical engineering from KAIST
(Korean Advanced Institute of Science and Technology), South Korea. He
worked as a postdoc researcher at MIT, USA. He is currently an Associate
Professor in the Department of Mechanical and Materials Engineering at Queen’s
University, Kingston, Canada. His research expertise includes multi-disciplinary
design optimization and computational mechanics with automotive and aerospace
applications. He has published over 40 peer-reviewed journal articles, and he
currently works with various industrial partners including General Motors,
Magna, Bombardier, and General Dynamics.

Chris K Mechefske is a Professor in the Department of Mechanical and

Materials Engineering at Queen’s University in Kingston, Ontario, Canada. His
research interests include vibration based machine condition monitoring and fault
diagnostics, maintenance and reliability, machine dynamic analysis, and vibration
and noise reduction. He is a member of the editorial board of the Journal of
Condition Monitoring and Diagnostic Engineering Management; the
International Institute of Acoustics and Vibration (Director 2007-2009); and a
Fellow of the Canadian Society of Mechanical Engineers. Chris was the president
of the Canadian Machinery Vibration Association from 2003 to 2005.