Advanced Diagnostics and Prognostics For Gas Turbine Engine Risk Assessment
Advanced Diagnostics and Prognostics For Gas Turbine Engine Risk Assessment
Abstract- Real-time, integrated health monitoring of gas of enhanced diagnostic and prognostic algorithms that can
turbine engines that can detect, classify, and predict predict, within a specified confidence bound, time-to-failure
developing engine faults is critical to reducing operating and of critical engine components can provide many benefits
maintenance costs while optimizing the life of critical including:
engine components. Statistical-based anomaly detection
algorithms, fault pattern recognition techniques and • Improved safety associated with operating and
advanced probabilistic models for diagnosing structural, maintaining gas turbine engines.
performance and vibration related faults and degradation • Reduced overall life cycle costs of engines from
can now be developed for real-time monitoring installation to retirement.
environments. Integration and implementation of these • Ability to optimize maintenance intervals for specific
advanced technologies presents a great opportunity to engines or fleets of engines and prioritization of tasks to
significantly enhance current engine health monitoring be performed during the planned maintenance events.
capabilities and risk management practices. • Increased up-time/availability of all engines within a
fleet.
This paper describes some novel diagnostic and prognostic • Provides engineering justification for scheduling
technologies for dedicated, real-time sensor analysis, maintenance actions with corresponding economic
performance anomaly detection and diagnosis, vibration benefits clearly identifiable.
fault detection, and component prognostics. The
technologies have been developed for gas turbine engine The development of enhanced diagnostic and prognostic
health monitoring and prediction applications which strategies built upon existing engine condition monitoring
includes an array of intelligent algorithms for assessing the platforms can allow for continuous monitoring and prediction
total ‘health’ of an engine, both mechanically and of component failure rates and degraded engine performance.
thermodynamically. A block diagram illustrating how prognostic technologies can
be integrated within existing diagnostic system architectures
TABLE OF CONTENTS is shown in Figure 1.
1. INTRODUCTION
2. TECHNICAL APPROACH
3. ENHANCEMENTS TO ENGINE HEALTH
MONITORING AND DIAGNOSTIC S YSTEMS
4. COMPONENT PROGNOSTIC MODELING
5. CONCLUSIONS
6. REFERENCES
1. INTRODUCTION
From Figure 1, the integration of prognostic technologies consequential and fixed costs can then be used to optimize
within existing diagnostic systems begins with validated the maintenance intervals or alter operational plans.
sensor information currently measured on the engine being
feed directly into the diagnostic algorithms for fault 2. TECHNICAL APPROACH
detection/isolation and classification. The ability of an
enhanced diagnostic system to fuse information from
multiple diagnostic sources together to provide a more In order to evolve purely diagnostic health monitoring
confident diagnosis is emphasized along with a system’s systems to those that are capable of more robust diagnostics
ability to estimate confidence and severity levels and failure prognostics, a probabilistic framework is
associated with a particular diagnosis. In a parallel mode, advantageous. Certainly, a prognostic system output that
the validated sensor data and real-time current/past only reports a specific time-to-failure without having any
diagnostic information is utilized by the prognostic confidence (uncertainty) bound associated with the prediction
modules to predict future time-to-failure, failure rates would be unwise. This is true for simple prognostic
and/or degraded engine condition (i.e. vibration alarm approaches that only utilize historical reliability data (such as
limits, performance margins, etc.). The prognostic Weibull distributions) to the more advanced prognostic
modules will utilize physics-based, stochastic models modeling approaches that take design parameter and
taking into account randomness in operation profiles, operating condition uncertainties into account.
extreme operating events and component forcing. In The prognostic modeling approach implemented in this paper
addition, the diagnostic results will be combined with past takes advantage of the directly sensed parameters, fused and
history information to train real-time algorithms (such as a diagnostic EHM system results, as well as inspection and
neural networks or real-time probabilistic models) to historical reliability data to provide critical inputs for
continuously update the projections on remaining life. The producing accurate failure predictions. As shown in Figure
specific approaches and algorithms for determining these 2, this process begins with a comprehensive evaluation of the
component prognostic results are described in this paper. sensor data including multiple techniques that are fused
together for identifying incipient sensor failure modes. Once
Once predictions of time-to-failure or degraded condition the possibility of a sensor malfunction has been minimized,
are determined with associated confidence bounds, the multiple diagnostic algorithm outputs are combined with
prognostic failure distribution projections can be used in a diagnostic fusion techniques to enhance the fault
risk-based analysis to optimize the time for performing identification capabilities of the engine health monitoring
specific maintenance tasks. A process which examines the system. An example of diagnostic fusion might be
expected value between performing maintenance on an intelligently processing information from vibration data,
engine or component at the next opportunity (therefore performance parameters, and oil related measurements to
reducing risk but at a cost of doing the maintenance) versus detect bearing faults with higher certainty. Next, in order to
delaying maintenance action (potential continued increased obtain the prognostic projections on component failures or
risk but delaying maintenance cost) can be used for this unsatisfactory engine operation, different levels of prognostic
purpose. The difference in risk between the two strategies can be utilized. Approaches to be considered in
maintenance or operating scenarios and associated this paper include probabilistic, physics-based models that
take design parameter and operating condition uncertainties
into account.
2
IGTI/ASME Turbo Expo, Munich, Germany, May 2000
The sensor validation and diagnostic techniques fall in two because system response characteristics may be used to aid
categories; signal processing-based and physics-based. The in fault identification in highly dynamic situations.
signal processing techniques accomplish their tasks
(detecting and diagnosing sensor anomalies) independent of The more advanced detector schemes are rooted in the
a monitored system’s characteristics. Conversely, the engine’s physical characteristics. The fuzzy logic based
physics-based techniques accomplish more advanced tasks sensor analysis continuously assesses the “normal” bands
(sensor recovery and virtually sensing parameters) because associated with each sensor signal at the current operating
they are developed from a-priori knowledge of the system condition. When a signal goes outside these bands, while
characteristics. Signal auto-correlation and high-pass others remain within, an anomaly is detected associated with
filtering are two signal processing techniques used for those specific sensors. The neural network operates by
identifying grounding and intermittent spikes at high levels comparing the physical relationships between signals as
of sensitivity. The correlation matrix tracks the degree of determined from a gas path model of the engine’s
co-linearity between signals in real-time and can detect performance parameters. The neural network has the ability
clipping, noise and multiplexer cross-talk. This technique to recovery sensors that have failed or “virtually” sense
bridges signal processing and physics-based approaches critical diagnostic parameters that is not monitored. All of
these parallel algorithms may be combined in a data fusion
3
IGTI/ASME Turbo Expo, Munich, Germany, May 2000
4
IGTI/ASME Turbo Expo, Munich, Germany, May 2000
Pattern recognition of a fault pattern has been successfully The function Ö( ) is the standard normal cumulative
performed using an unsupervised neural network (Kohonen distribution. The notation â is the reliability index. If F and
Map) for clustering fault types in series with a standard back C are not normal or lognormal variables then the probability
propagation neural network to classify specific fault patterns of failure can be determined only by using a computer
in 2-D space [4]. The diagnostic predictions made by these algorithm as described in [4].
two techniques are then fused together using a Bayesian
inference algorithm that either increases or decreases the Two different fault measures have been developed from this
confidence of a particular diagnosis based on the outputs stochastic technique [1]. First, a cumulative sensitivity
agreement or disagreements. A discussion and example of index defined by the global non-dimensional variation from
the probabilistic fault diagnostic method is provided in this the initial state, at time 0, to the final state, at time t (over
paper. the time interval [0,t]). Next, an evolutionary sensitivity
index defined by the local non-dimensional variation from
As mentioned, the probabilistic fault identification process an intermediary state, at time ti, to another intermediary
utilizes known information on how measured parameters state, at time ti+1, over the time interval [ti,ti+1]. Both of
degrade over time and compares them with the current these indices are given below.
parameter distribution to track and identify degraded fault
conditions. The amount of each parameter’s shift from the Cumulative Sensitivity Index:
expected baseline operation at a particular operating
condition is continuously analyzed along with the associated βt − β0
statistical confidence level. The shifted distributions are C0 ,t = − (2)
β0
compared against the baseline parameter’s probability
distribution for all known faults. Comparing the measured
distributions against all the known fault distributions yields Evolutionary Sensitivity Index:
the confidence of a known performance degradation
problem. β ti+1 − β ti
Eti ,ti +1 = − (3)
β ti
A stochastic interpretation of the measured and known
parameter distributions can provide a powerful means of where β is the reliability index related to the Euclidean
calculating multiple fault contributions to a current engine distance between the current conditional distribution and a
operating condition and envisioning the most likely given fault distribution.
evolution of the fault. A manipulation of the First Order
Reliability Method (FORM) can be used to gage a fault’s Figure 6 illustrates the fault detection and classification
evolution in state vector space where the faults and the process in a two-dimensional parameter space. Starting at
current condition are not known with complete certainty. the origin, (representing initial normal engine operation) the
The conceptual framework for structural reliability and measured parameter distributions begin to shift as some type
more specifically the FORM algorithm is provided by of performance degradation begins to occur. After several
classical reliability theory [5]. missions when the anomaly level is reached an anomaly
detection warning is issued. As shown in Figure 6, the
In essence, the probabilistic fault diagnostic process current measured PDF moves to point T1 then to T2 as time
involves assigned PDF’s to performance error patterns progresses, ultimately approaching the 4% fault. The
associated to known faults in N-dimensional space. described situation in Figure 6 can be handled with
Similarly, the current error exists as a PDF in the space as mathematical accuracy by using the proposed probabilistic
well. The probability that the current condition (C), may be fault diagnostic calculations previously described.
attributed to a given fault (F) is determined by their joint
probability density function. If C and F can be assumed to
be normally distributed, the probability of association (Pa)
can be found using:
F −C (1)
p a = 2 Φ( − ) = 2Φ ( − β)
σ +σ
2 2
f c
where:
5
IGTI/ASME Turbo Expo, Munich, Germany, May 2000
Figure 6 Probabilistic Fault Diagnosis Process collected from accelerometers at specified locations on the
Figures 7 and 8 provide a further example of the evolution machine can be enhanced using neural-fuzzy diagnostic
of a performance error pattern. Three different performance techniques. Domain knowledge associated with particular
faults at a 2% level were identified using a gas path model vibration fault frequencies, fixed frequency ranges, per-rev
of an engine. In this example, and shown in Figure 7, the excitations, structural resonance’s, etc. are extracted from
PDF of the current error pattern initially evolves towards a the vibration spectrums and used to develop a knowledge
2% HP Compressor efficiency fault. This is also shown in base from which the fuzzy logic membership functions and
Figure 8 from the fact that from T=0 to T=3 the Euclidean rulebases are developed. This feature extraction process is
distance between the current PDF and the HPC fault gets illustrated in Figure 9. Non-vibration related data such as
smaller. However, as time goes by, the current condition the performance parameters, oil analysis data, etc. can also
evolves towards, and eventually past, the HPC fault. From be integrated (knowledge fusion) into the fuzzy expert
Figure 8, at T=5 the current PDF is closest to the 2% HP system to collaborate on a particular diagnosis.
Turbine Efficiency fault. Figure 7 illustrates that the
engine’s degradation has indeed evolved beyond association
with the HP Compressor fault to high association with the
HP Turbine fault. In this example, the final position rests at
9.98% association with the HPC fault and 22.8% association
with the HPT fault.
6
IGTI/ASME Turbo Expo, Munich, Germany, May 2000
7
IGTI/ASME Turbo Expo, Munich, Germany, May 2000
analyzed components such as bearings should have clearly associated with this steady stress due to variations in the
identified relationships between diagnosed fault severity and operating environment, temperatures, manufacturing
life consumption. tolerances, etc must be accounted for with a statistical
distribution computed by testing and field service data.
Example of a Stochastic Physics-Based Model Another factor critical in predicting turbine blade life is the
dynamic stress as a function of the uncertainty in harmonic
Sophisticated fracture mechanics and damage accumulation excitation and response characteristics. The strength or
analysis have shown that accelerated crack nucleation and resistance capability of the material must also be considered
micro-crack formation in components can occur due to start- as a function of the uncertainty in material properties.
ups and shutdowns, transient load swings, higher than
expected intermittent loads, or defective component
materials. More commonly, normal wear causes
configuration changes (loose fit of assembled parts, work
hardened surfaces, and reduced structural section areas) that
contribute to increased or unexpected dynamic loading
conditions. High cycle dynamic and transmission loads
cause micro-crack incubation and formation at material
grain boundaries in stress concentrated regions (especially
between hardened surfaces and softer subsurface material
interfaces, and at acute changes in component material
geometry). The majority of crack growth evolves in a sub-
critical propagation process of crack tip blunting, unstable
crack formation, and crack elongation. As super-critical
loading in the cracked material region is approached, growth
accelerates resulting in material dislocation and detachment.
Sub-critical crack evolution is highly dependent on a
component’s material, geometry, loading conditions, and the
particulars of the unique component crack growth cycle.
This kind of failure-mode knowledge is often times
overlooked in determining the potential usefulness of a
particular prognostic or diagnostic algorithm.
8
IGTI/ASME Turbo Expo, Munich, Germany, May 2000
calculate in a strict analytical sense. The result of the remaining life and predicting future usage associated with
simulation is also shown in Figure 12. critical components.
The damage due the low cycle fatigue may be given by a 6. REFERENCES
non-linear damage accumulation rule proposed by Gary
Halford at NASA Langley [7]:
[1] Bowerman and O’Connel, Forecasting and Time Series,
r1 Duxbury Press, 1993.
n
Damage = 1 (5)
Nf 1L [2] M.J. Roemer, and B. Atkinson, “Real-Time Engine
Health Monitoring and Diagnostics for Gas Turbine
The complete turbine blade prognostic model must further Engines,” Proceedings of the International Gas Turbine &
account for the other failure modes at other critical locations Aeroengine Congress and Exhibition, Orlando, 1997
on the blade, however that is outside the illustrative scope of
this paper. The net result, however, is the path that’s been [3] M.J. Roemer, and D.M. Ghiocel, ”A Probabilistic
taken to determine the current component life consumption Approach to the Diagnosis of Gas Turbine Engine Faults”
as shown in Figure 13. 53rd Machinery Prevention Technologies (MFPT)
Conference, Virginia Beach, VA, April 1999
9
IGTI/ASME Turbo Expo, Munich, Germany, May 2000
10