Modeling of Vibration Measurements For Gear Fault and Damage Detection On Aircraft
Modeling of Vibration Measurements For Gear Fault and Damage Detection On Aircraft
DETECTION ON AIRCRAFT
ABSTRACT
The Intelligent Health and Safety Monitoring activity has been supported by NASA’s
Computing, Information, and Communications Technology (CICT) Program for several years and
includes a unique combination of basic information technology, rotorcraft, and aviation safety
applications. One aspect of the program has been the evolution of a very close working
relationship between researchers at Ames and Glenn Research Centers, who share flight and
test-rig resources to address many challenges in developing effective damage detection and
prognosis technology. At both Centers, Army contributions are also essential, for both
computational- and flight-research support. This paper briefly reviews several kinds of modeling
efforts that are currently underway at the two Centers.
INTRODUCTION
The detection of gear faults, and particularly their precursor damage patterns, has proven
to be extremely difficult in dynamically operating aircraft. This is due to the complexity of the
rotating gear system components, as well as ongoing modulations of their vibration signals during
flight. Accurate and reliable detection of fault precursors in operating aircraft, therefore, depends
upon understanding both the physical alterations that take place in the rotating equipment, as well
as the ever-changing dynamic effects that must occur in the context of flight.
Researchers at NASA Glenn and NASA Ames are addressing these interrelated
technical problems of damage detection through a highly coordinated multidisciplinary program.
The joint effort is designed to take advantage of the unique in-house ground-test and flight-
research capabilities at the two Centers, and provides an ongoing forum for the exchange of
technical information and measurement methodology. The work at Glenn includes the
development of physics-based gear models, the formulation of metrics that reflect gear
degradation, and the development of techniques for non-intrusive monitoring of gear damage
propagation. Work at Ames includes the collection of vibration data under experimental flight
conditions, the evaluation of nonstationarity effects in the flight environment, first-principles
vibration signal simulation, flight dynamics modeling and studying its relationship to vibration
signal modulation. Both Centers are active in supporting multivariate signal analysis, particularly
with regard to the fusion of sensor data, such as oil debris, and the maturation of generic
methods for planetary signal decomposition.
Using multivariate methods, the statistical properties of flight data are being analyzed
specifically to understand the effects of speed, torque, fuel depletion, and other effects on the
time/frequency characteristics of measured signals. Since the variance of recorded time series
reflects the influences of both known and unknown deterministic processes, as well as random
noise, attention is focused on computational methods for “correcting” the signals for these effects.
The most commonly used method is continuous signal averaging, in either the time- or frequency-
domain, but this has shown to be inadequate if not dangerously misleading for flight applications
due to underlying non-stationarity dynamics.
Work reported at the AHS-2000 [1] and 2001 [2] Annual Forums that examined and
compared the three-dimensional responses of the AH-1 and OH-58C helicopters using Principal
Components Analysis (PCA), was expanded in FY-03 to include mesh frequency power
distributions [3]. This approach parsed the total power (i.e., RMS) in each principal axis, and was
combined with an exploratory state-space analysis defined by instantaneous engine torque,
torque derivative, and output shaft speed. The paper presented at the AHS-2003 Forum
demonstrated the highly complex, non-linear nature of the vibration response surface. It also
revealed that a single 2-hr. mission is, at least for the OH-58C, inadequate to capture a sufficient
number of points within the state-space to model the surface. If this observation were to hold for
heavier vehicles, the broad implication is that in-flight systems will need sufficient intelligence to
learn (i.e., adapt to) the specifics of each specimen vehicle over time.
During FY-04, the OH-58C flights discussed above are being re-flown with more stringent
recording of pilot observations and the addition of GPS information. These will be compared with
the earlier data, and multivariate techniques, such as factor-analysis, used to identify “hidden” or
latent variables that may be inferred from the response covariance matrix. These same
techniques will also be applied to legacy AH-1 data collected in FY-01, where aircraft attitudes
were recorded that allow for a more comprehensive state-space analysis.
Simple measures of vibration amplitude, such as peak-to-peak and RMS, are the most
elementary measures that could be used to monitor the health of machinery. In operational
aircraft, however, such simple measures of amplitude are insufficient to determine the health of
the transmission. Machinery on aircraft operates in a dynamic environment with constantly
changing forces producing nonstationary vibration signals. Several gear health metrics, based
upon a straightforward kinematic model of gear vibration, have been developed to detect
localized damage in gears [4-6]. This kinematic model is based on the idea that the vibration
produced by a healthy gear is the sum of a periodic signal containing components at the integer
harmonics of the gear mesh frequency plus random noise. These characteristics assume that all
gear teeth are identical and thus all gear mesh interactions are the same. To accommodate
speed variation in the gear, the frequencies are based upon rotation angle instead of time. The
idea is that in a gear with localized damage, the tooth-to-tooth interactions in the damaged area
differ from in the undamaged area, so the vibration signal deviates from the ideal periodic signal.
As an example, the FM4 metric is a measure of the kurtosis of the time synchronous
averaged vibration with the gear mesh periodic components removed. Based on the assumption
that the residual signal is Gaussian, the expected value is FM4=3 for an ideal gear vibration
signal. With localized damage, the value of FM4 would be expected to increase, presumably due
to the introduction of local phase and amplitude distortions resulting from the damage.
The FM4 metric has been used extensively to test for damage in gears [5-10]. An
example from NASA Glenn Research Center [11] reported the behavior of FM4 in fatigue damage
tests. Figures 1a and 1b show the distribution of FM4 for regions of the tests where there was no
damage, unknown damage and known damaged. Note that for the first gear in figure 1a, the
distributions match the expected behavior of FM4 according to the kinematic model with very
good separation between the damaged and undamaged states. In figure 1b the distributions are
not so well distinguished. Distributions of the FM4 metric from a gear in a good transmission
measured from flight are shown in figures 2a and 2b. Measurements from different flight
conditions are shown in each curve in figure 2a and measurements from different accelerometers
are shown in separate curves in figure 2b. The distributions from undamaged transmissions in
flight under conditions labeled B and C in figure 2a and from accelerometer number three in
figure 2b are resemble the distributions measured from a damaged gear as well as damaged
gear in a ground based test rig in figure 1.
Figure 1 Estimated Probability Density Function of FM4 from Spur gear in NASA Glenn Research
Center test rig. (from [11])
Although the model of periodic gear mesh vibration plus noise looks very much like the
measurements, the model does not appear to be sufficiently accurate to use as a sole basis for
distinguishing between good and damaged gears with vibration measurements.
1
10
Accel 1
Accel 2
Accel 3
Accel 4
0 Accel 5
10 Accel 6
Kinematic Model
Estimated PDF
−1
10
−2
10
−3
10
1.5 2 2.5 3 3.5 4 4.5 5
FM4
Figure 2 Estimated Probability Density Function of FM4 from gear in OH-58 Flight Test at NASA
Ames Research Center. (2a from[11])
The reliability and accuracy of detecting damage in rotating machinery is a direct function
of the accuracy of the physical sensor and corresponding diagnostic techniques being used.
Studies at NASA Glenn [10, 12] have shown that by integrating the signals and features from a
variety of sensors, the integrated system shows improved detection and decision-making
capability as compared to individual measurement methods. This is especially true for those
failure modes for bearings and gears that generate both wear debris in the oil and vibration
signatures when failure occurs. In this effort at Glenn Research Center, multivariate decision
fusion is being used to develop diagnostic tools to accurately and reliably detect bearing and gear
damage on-line.
During the coming year work will be completed on a general purpose diagnostic software
tool and evaluated experimentally with gear/bearing vibration and oil debris data from NASA
Glenn Test rigs, e.g., the 500-HP Helicopter Transmission Test Facility, Spur Gear Fatigue Rig,
Spiral Bevel Gear Fatigue Rig, etc. Multivariate decision fusion schemes are also being
incorporated for planetary fault detection using vibration, oil debris, and acoustic emission data
from test rigs and transport aircraft as they become available.
Cracks that develop in gears are a very important category of mechanical failure. Unlike
those that occur within individual gear tooth, fractures that expand through the gear rim may lead
to the catastrophic loss of the transmission or engine, and seriously compromise aircraft safety.
Predicting crack propagation in gears by means of numerical modeling is a recent development
[13, 14]. Knowledge gained through numerical modeling can be used to design gears with less
catastrophic failures modes, help with prognostic estimates of useful gear life and gain insight for
developing fault detection methods.
Crack propagation modeling in gears involves multiple methods combined together. The
general steps are 1) define the geometry of the gears, 2) generate a grid, 3) determine the
contact path, 4) initiate the crack, 5) remesh around the crack, 6) solve structural equations 7)
propagate crack and continue to loop through steps 5 to 7 until some end point is reached.
The simplest modeling is done with 2-dimentional finite elements and has been used to
study crack propagation in spur gears. At NASA Glenn, Lewicki [13, 15, 16] used this method
successfully within the framework of linear elastic fracture mechanics to study the effects of rim
thickness, initial crack location, and rotational speed of crack propagation Gear tooth coordinates
were modeled based upon the method of Hefeng [17] and the tooth load was placed at the
highest point of single tooth contact. This finite element modeling method utilized eight-node
quadrilateral and six-node triangular elements, as show in Figure 3. In the vicinity of the crack,
six-node elements were used to model the inverse square-root stress singularity. After each step
of the structural equation solver, the stress intensity factors were calculated and used to
determine the extent and direction of crack propagation. In order to validate the method, the
predicted crack propagation path was compared with the results of experiments also conducted
at Glenn Research Center. Figure 4 shows both the prediction and measurements from spur
gears with three different rim thicknesses. It is notable that the location of the physical crack
closely follows the predicted crack in all three cases; of particular interest is the agreement of
prediction with experiment with regard to the location and direction of the crack through the tooth
or through the rim.
Simulated
crack
Crack tip Quarter- trajectory
point
Initial
rosette crack
region
(a) (c)
(b)
Figure 3 Grids from 2-D finite element crack propagation. (from [15])
Figure 4 Comparison of predicted crack propagation with experiment (from [15])
For more complex gears such as a spiral bevel gear, enhancements have been added to
the crack propagation modeling. Moving tooth loads through modeling are included in some
studies [16, 18, 19]. Boundary element [14, 18] and 3-dimentional finite element [20, 21] models
have been used for the structural equations.
To identify the unique vibration characteristics of a specimen gear system, the most
comprehensive approach is to compare vibration output with that of a physics-based
computational model that represents what a perfect gear system would produce under ideal
operating conditions. This is being pursued by first-principles, finite-difference modeling of the
gear component geometries under rotation by solving the three-dimensional elastodynamic partial
differential equations with the appropriate dynamic boundary conditions. Numerical simulations
of stresses over component geometries that define an overall geometry for gear teeth in mesh for
rotating systems such as helicopters are being carried out using first-principles modeling
approach. Using this technique, it is anticipated that nominal “reference” vibration signatures will
shortly become available. Following that, “ideal” damage signatures will be produced for various
problems of interest. Since extremely intensive computing is required, however, NASA Advanced
Supercomputing facilities at Ames Research Center will most probably be required.
A new three-dimensional elliptic grid generation methodology has been developed that is
based on new boundary constraints for elliptic grid generation boundary-value problems [22], thus
making it possible to generate gear grids during simulation without user intervention. Figure 5
shows a 2-dimentional section of an automatically generated grid for a gear. Since the new elliptic
grid generation methodology makes it possible to generate the grids automatically [23], it will
make finite difference solution of structural dynamics problems a viable technology to simulate
gear teeth crack propagation in time, as well as to simulate other asymmetric time-dependent
gear fault phenomena involving gear grids where individual teeth are subject to deformation,
wear, pitting, etc.
A Structural Dynamics Simulator and Animator (SDSA) have been developed and are
presently being improved with regard to prediction accuracy. In particular, work is being carried
out on the boundary condition formulation based on the theory of hyperbolic systems. Work is
also ongoing to improve the usefulness of the animation capability, which now not only displays
the predictions graphically in simulation tine, but also can also be interrupted and checked for
anomalies either in the grids or the physical quantities that are displayed. The evolving grids can
now be inspected in three-dimensional space by rotating, zooming or translating them on the
screen, at any point during the simulation.
On the whole, what has been learned is that successful damage detection methods must
account for the dynamical influences of real-time vehicle operations, which result in non-
stationary signals and highly nonlinear signal response surfaces. Attacking this broad problem is
unmistakably one of the most demanding technical challenges for computational sciences
research in the vehicle health-monitoring arena. Success, no doubt, will depend upon
maintaining a focused interdisciplinary research capability that spans classical statistics, signal
analysis, data fusion, control theory, and computer architecture.
ACKNOWLEDGMENTS
The authors would like to thank the Management of NASA’s Computing, Information,
and Communications Technology (CICT) Program, for providing the ongoing support for this
effort. We would also like to complement the Army Flight Projects Office (FPO) for the
coordination of all aircraft instrumentation and operations. Special thanks is also due for
California Signal Processing (Sigpro), for outstanding flight research support.
REFERENCES