27.1 Introduction

Back

Present helicopter power trains are significant contributors to both flight safety incidents and

maintenance costs. For example, for large and medium civil transport helicopters in the period 1956 to

1986, gearboxes were the principal cause of 22% of the accidents with potential loss of life

and aircraft (Astridge, 1989). To prevent such incidents, routine maintenance is scheduled

at a significant ratio of the total maintenance cost for the helicopter. Rapid and reliable

detection and diagnosis (isolation) of faults1 in helicopter gearboxes is therefore necessary to prevent

major breakdowns due to progression of undetected faults, and for enhancing personnel safety by

27-1

1Determining whether the overall machinery is healthy is referred to as fault detection, whereas diagnosis is analogous to

isolating the source of failure.

© 2005 by Taylor & Francis Group, LLC

preventing catastrophic failures. Fault detection and diagnosis is also necessary for reducing maintenance

costs by eliminating the need for routine disassembly of the gearbox, and for saving time during

inspection.

Detection and diagnosis of helicopter gearbox faults, like most rotating machinery, is based on oil

analysis and vibration monitoring. Oil analysis, which is used to detect the presence of metallic debris,

is performed by: (1) magnetic plugs (chip detectors), (2) oil filters, and (3) spectrometric oil analysis

(SOA) (Lurton, 1994). Among these, magnetic plugs are the most popular because of their in-flight

utility and their ability to quantify the severity of wear by measuring the rate of detected debris.

However, magnetic plugs can only collect ferromagnetic particles, and their capture efficiency may be

poor. Oil filters used in helicopter gearboxes range between 3 and 150 mm, with finer meshes of 3 mm

more common in the recent years. SOA, which is a ground-based technique, is useful for detecting

fine debris, typically below 10 mm in size, caused by wear conditions such as rubbing, cutting, and

corrosion wear, and fine surface fatigue such as micropitting. Therefore, development of a reliable

method of identifying common forms of fatigue such as spalling or wear that generate particles

greater than 10 mm in size is of particular interest. The effectiveness of SOA is also affected by the

level of filtering performed on the used oil, which often leaves the oil free of particles (Lurton, 1994).

The need for improving wear detection has motivated development and use of other analysis

methods such as image processing of oil filters, ferrography, thermography, and ultrasonic analysis

(Thornton, 1994).

The other method of gearbox fault detection and diagnosis, and by far the more popular one, is based

on vibration monitoring. The basic principle behind vibration monitoring is that, under normal

operating conditions, each component in the gearbox produces vibrations at specific frequencies related

to the component’s rotational frequency. In the case of a component fault, the vibration generated by the

faulty component is different from the normal vibration, and will be reflected at the component’s

rotational frequency and its harmonics. As such, monitoring the changes of vibration should

theoretically give an indication of the fault. In practice, however, changes in the measured vibration as a

Triplex

Bearing(TB)

Spiral Bevel

Pinion(SBP)

Spiral Bevel

Gear(SBG)

Mast Ball

Bearing(MBB)

Sun

Planet Gear(SG)

Bearing(PB)

Planet

Gear(PG)

Ring

Gear(RG)

Gear Roller

Bearing(GRB) Duplex

Bearing(DB)

Mast Roller

Bearing(MRB)

Pinion Roller

Bearing(PRB)

FIGURE 27.1 Layout of various components in a typical helicopter (OH-58A) gearbox.

27-2 Vibration and Shock Handbook

© 2005 by Taylor & Francis Group, LLC

result of component faults are not always distinct

due to the attenuation of vibration by the housing

and other components it travels through, as well

as the noise in the signal. To provide a more

tangible framework for the concepts in vibration

monitoring, the layout of a typical helicopter

gearbox is shown in Figure 27.1, with the location

of the accelerometers on its test stand shown in

Figure 27.2.

In order to enhance identification of vibration

changes by component faults, the raw vibration is

generally processed to obtain “features” that

characterize the vibration at the frequencies

associated with the gearbox components. Accordingly,

the main focus of research in vibration

monitoring has been the identification of individual

features that consistently reflect specific gearbox

faults (Dyer and Stewart, 1978; McFadden and

Smith, 1986; Mertaugh, 1986; Zakrajsek et al.,

1995). A typical set of vibration features obtained

from each vibration measurement is shown in

Figure 27.3 (Stewart Hughes Ltd., 1986). Among

them, envelope band and tone energies, cepstra,

and synchronous-time averaged signals, are associated

with various bearing frequencies and gear

mesh tones. Envelope band and tone energies,

which are the sum of the harmonics of the

bearings’ fundamental rolling-element frequencies

within a filtered bandwidth, could be used for early

detection of failure in rolling-element bearings

(Barkov and Barkova, 1995). Similarly, a cepstrum,

which is the power spectrum of the logarithm of a

power spectrum, is often used because of its

insensitivity to transmission path effects for

identification of families of uniformly spaced

sidebands in gearbox vibration spectra (Randall,

1982). Synchronous-time averaging is a signal

processing technique that isolates the fundamental

and harmonics of the gear meshing frequency, and

is a primary analysis technique for detection of

gear and shaft faults (McFadden and Smith, 1986).

The traditional approach to fault diagnosis has

relied on human expertise to relate vibration

features to faults. In this approach, a diagnostician first identifies abnormalities in vibration features,

then relates them to component faults, considering both the proximity of the accelerometer producing

the feature to various components and the information about the type of fault characterized by the

feature. Using this information, the diagnostician hypothesizes faults in specific components and then

verifies or discards the hypothesis by examining features from other accelerometers in the proximity of

the suspect component. The disadvantages of this approach arise from: (1) the difficulty in identifying

abnormality in features which are contaminated with noise, and (2) the tediousness of examining the

numerous features obtained from all of the accelerometers. Owing to the large number of features

FIGURE 27.2 Location and orientation of the accelerometers

on the OH-58A test stand.

Tape

Recorder

Digitization /

Processing

(1) Skewness

(2) Kurtosis

(3) Crest Factor

(4) Peak-to-Peak

(5) RMS

(6) WHT

(7) RFR

(8) TEO-G

(9) TEO-P

(10) TM1-G

(11) TM1-P

(12) CEP(1911)

(13) CEP(572)

(14) TON(1911)

(15) TON(572)

(16) BE

(17) BKV

(18) EB

(19) ET

STAT BBPS BRGA SGAV

(1) MF1

(2) FM4a

(3) FM4B

(4) FM4-B

(6) WCH

(5) ACH

(7) SCH

For each of the 5 Gears

FIGURE 27.3 A typical set of vibration features

extracted from each accelerometer.

Fault Diagnosis of Helicopter Gearboxes 27-3

© 2005 by Taylor & Francis Group, LLC

associated with a gearbox2, the diagnostician

cannot often pay equal attention to all the features

and is likely to ignore information that contradicts

the hypothesis.

The most efficient method for integration of

features is pattern classification. To this end,

artificial neural networks have been widely investigated

and shown to provide excellent results

(Solorzano et al., 1991; Chin et al., 1993; Kazlas

et al., 1993; Chin et al., 1995). Neural networks offer the following advantages in diagnosis: (1) as pattern

classifiers, they can efficiently cope with noise in features, (2) through training, they can form the

signatures of individual faults in the multidimensional space of features, and (3) they can process

vibration features in parallel, so diagnosis is not hindered by the enormity of feature space. The main

disadvantage of supervised neural networks, however, is their need for prior training, which requires a

comprehensive set of features during normal operation and at various fault instances. While training data

can be obtained experimentally through accelerated fatigue tests (Lewicki et al., 1992) or seeded fault

studies (Naval Command, Control, and Ocean Surveillance Center, 1995), their cost is considered too

high, limiting the utility of supervised neural networks in fault diagnosis. The lack of training data poses a

similar restriction for statistical pattern classifiers which need a priori statistics of the features.

In the absence of training data required by supervised neural networks, the monitoring system can be

formatted as in Figure 27.4 to take advantage of unsupervised pattern classification. The strategy depicted

in Figure 27.4 considers fault detection independent of abnormality scaling of features, so it can be

performed before or in parallel to this stage, based on the overall deviation of all of the features from their

normal state. An example of abnormality scaling based on unsupervised pattern classification is

described in the Section 27.2.

For fault diagnosis, the proposed system should have detailed information about the relation between

the individual features and component faults. One format for providing that information is expert

diagnostic software, which can be developed at two different levels. At one level, “shallow expert systems”

can be developed to compile a human diagnostician’s knowledge relating measurements to faults, often as

“if …then” rules, similar to those already developed for simpler rotating machinery (Liddle and Reilly,

1993). The main obstacle in such a development will be the design of a robust inference engine that can

resolve conflicting conclusions from the large pool of suspect features. At another level, “deep expert

systems” can be developed to represent the diagnostic knowledge derived from the physics of the process.

In deep expert systems, measurements need to be related to component faults by a model of the energy flow

via the structural connections between components and sensors. A suitable format for defining the relation

between components and sensors is the fuzzy set theory, which can address the approximate nature of

vibration modeling. An example of a deep expert diagnostic system for helicopter gearboxes is the

structure-based connectionist network (SBCN) (Jammu et al., 1998) described in Section 27.3, which

takes advantage of the integration capability of neural networks while avoiding the need for supervised

training. The salient feature of this connectionist network is that its weights can be determined a priori,

based on the proximity of components to various accelerometers and the type of faults characterized by

individual features. Although this system has been developed on a very approximate model of vibration

flow between the gearbox components and accelerometers, it has produced promising results when used

for fault diagnosis of helicopter gearboxes (Jammu et al., 1998).

A side benefit of a model of the structural connections between gearbox components and sensors is

its use in sensor location selection (Wang et al., 1999). An important issue in helicopter gearbox

diagnostics is the determination of the number of accelerometers to be used for monitoring and their

location on the gearbox housing. Accelerometers are generally located by experts, based on their

2The large number of frequencies (tones) associated with the various components of the gearbox necessitates a huge number

of features (often in excess of a hundred) to be obtained for vibration readings from each of the several accelerometers.

Raw

Vibration

Signal

Processing

Fault

Detection

Abnormality

Scaling Diagnostic

features

FIGURE 27.4 Generic structure of an unsupervised

diagnostic system.

27-4 Vibration and Shock Handbook

© 2005 by Taylor & Francis Group, LLC

proximity to gearbox components, orientation, and ease of mounting on the housing. However, this

approach often leads to too many accelerometers and too high a demand for on-line monitoring on the

on-board computer. Another problem with the extra accelerometers is the cost of extra mountings,

cabling, and signal conditioning equipment. Sensor location selection based on a gearbox model is

discussed in Section 27.4.

Навигация