Tracking the health of aircraft electrical generators

  • Image: fig 1 2012-01-2234 jan TU.jpg
  • Image: fig 3 2012-01-2234 jan TU.jpg
Image: fig 2 2012-01-2234 jan TU.jpg

Basic modules and overall architecture for diagnostic and prognostic assessment of aircraft electric power systems.

The U.S. Navy's P-3 aircraft uses a Bendix (later AlliedSignal and currently Honeywell) generator that was designed over 30 years ago. The P-3 generator is a salient 8-pole (8 rotor bars), 6000-rpm, three-phase brushless ac generator. The running speed of the generator is 5700 to 6300 rpm (line frequency of 380 Hz to 420 Hz). The rated voltage and power is 115-V ac and 60 kVA (20 kVA/phase), respectively. It has a 12-pole ac exciter and a three-phase, half-wave diode rectifier rotating with the exciter armature and main generator field assembly. A single-phase permanent magnet generator (PMG) furnishes control voltage and power for the voltage regulator.

The electrical and mechanical issues (due to being continuously operated beyond design point and less than optimal drive end bearing support) combine to cause premature failures of the P-3 generators. Repairs are time consuming and costly.

Currently, diagnostic/prognostic technologies are not implemented for P-3 generators and other electrical power systems. Although some time series data (such as phase voltage and current) is collected during ground testing, no time series data is collected in-flight for the generators.

Researchers at NAVAIR and Global Technology Connection have developed feature extraction and diagnostic algorithms to ultimately 1) differentiate between generator failure modes and normal aircraft operational modes; and 2) determine the degree of damage of a generator, providing maintenance personnel with information about the current health state and remaining life of generators so that timely action can be taken.

Health-monitoring algorithms

Researchers came up with basic modules of the proposed diagnostic and health-management system architecture based upon data-driven algorithms. Also, they showed how the architecture can provide inputs to the condition-based maintenance (CBM) module for maintenance execution.

The feature extraction unit takes raw sampled data from a generator and converts it to a form suitable for the diagnostic and prognostic modules. The diagnostician monitors continuously critical feature data and decides upon the existence of impending or incipient failure conditions. The detection and identification of an impending failure triggers the prognosticator. The prognosticator reports the remaining useful lifetime of the failing machine or component to the CBM module. The CBM module schedules the maintenance so that uptime is maximized while certain constraints are satisfied.

The prognostic architecture is based on three constructs: 1) a static “virtual sensor” that relates known measurements to material deterioration; 2) a predictor that attempts to project the current state of the damaged material into the future thus revealing the time evolution of the damage and allowing the estimation of the material's remaining useful lifetime; and 3) a Confidence Prediction Neural Network (CPNN), whose task is to account for uncertainty and manage/shrink the prediction bounds.

Feature extraction

Initial time-domain and frequency-domain feature extraction algorithms were developed to distinguish between healthy and common failure modes such as bearing failure. The time domain features and frequency domain features based upon Electrical Signature Analysis (ESA) were calculated from (demodulated) voltage and current. In this work, the time domain features were statistics of demodulated three-phase voltage vs. its phase angle, while the frequency domain features were magnitude deviation statistics of selected (inter) harmonics of demodulated three-phase voltage and exciter current.

Each extracted feature from a set of time and frequency-domain based features was evaluated separately by statistical comparison with the corresponding feature database. This feature database was determined from the feature construction, selection, and extraction analysis of healthy/low-hour generators operating at various loads and operating frequencies.

Statistical margins were defined to denote the range of healthy/low-hour generators. Values outside of this range indicate the presence of a degraded/high-hour generator. Consequently, the healthy operation range was based on separate asymmetrical calculations of high and low margins that were calculated by generalized higher moments of features.

Electrical signature analysis

ESA is the term used for the evaluation of the voltage and current waveforms. This provides an increased advantage to diagnostics as power-related, motor-related, and load-related signals can be quickly compared. A key consideration when using ESA is that voltage signatures relate to the upstream of the circuit being tested (toward power generation) and current signatures relate to the downstream of the circuit being tested (toward the motor and load).

ESA uses the machine being tested as a transducer, allowing the user to evaluate the electrical and mechanical condition from the control or switchgear. Typically, ESA is done in the frequency domain, relying upon FFT techniques for accurate analysis.

Frequency-domain-based ESA techniques have been developed to detect and track a seeded bearing failure in a P-3 generator, which was not detectable in the vibration signals using visual inspection alone (no large spikes in vibration data were observed).

Also, frequency-domain techniques have been used to identify the location/component of degradations in P-3 generators using available phase voltage and current waveforms. In this work, ESA was done both in the time-domain as well as the frequency-domain.

Time-domain based ESA was used to assess the general health of generators by analyzing the amplitude demodulated signal vs. phase angle of the complex signals, while frequency-domain based ESA was used to identify the location/component of degradations in P-3 generators using available phase voltage and current waveforms.

Fault classifier

The diagnostician, implemented as a multiple-input, multiple-output fuzzy neural network (FNN), serves as a nonlinear discriminator to classify impending faults. The fault classifier is trained to recognize generator faults from a vector of (inter) harmonic signatures corresponding to air gap, rotor, stator, rotating diodes, and bearing failures.

Air gap and rotor winding failures are detected from demodulated line frequency harmonics with the latter utilizing only even harmonics. Stator winding and rotating diode failures are detected from exciter current harmonics. The first ten harmonics of bearing fundamental train frequency (FTF), ball pass outer race (BPOR), ball pass inner race (BPIR), and the two times the ball spin frequency (2×BSF) were used to identify bearing failures.

The bearing fault signatures appeared as harmonics of FTF × RS, BPOR × RS, BPIR × RS, and 2 × BSF × RS in the demodulated three-phase voltage where RS is the running speed of the generator. The bearing parameters (rolling element diameter, pitch diameter, contact angle and number of rolling elements) needed to calculate FTF, BPOR, BPIR, and 2×BSF were obtained from a manufacturer's catalog given the particular manufacturer and size of the bearing.

A virtual sensor calculated a failure measure indirectly through a (neural network) mapping of features and operating condition. Consider, for example, the case of an electrical generator. No direct measurement of the degree of stator/rotor winding degradation, bearings damage, etc. occurring in a generator is possible when it is in an operational state. That is, there is no such device as a “fault meter” capable of providing direct measurements of the fault evolution. The fault dimensions take the form of a vector of integer state-of-health (SOH) values where the values range from 100 (healthy) to 0 (fault).

Results and conclusions

Time domain features based upon ESA that were previously developed to detect the degree of damage in P-3 generators have been augmented by frequency-domain analysis algorithms that can accurately differentiate between normal operational modes and generator failure modes.

The results showed that the high-accuracy frequency-domain analysis algorithms were able to accurately detect bearing and rotating diode failures for two P-3 generators at full load, half load, and no load operating conditions. The deviations in the magnitudes of the (inter) harmonics above the -80 dB threshold for phase voltages and the -40 dB threshold for exciter current were used to determine what type of failure mode the generator was experiencing.

Also, the general harmonics to detect air gap, bearing, rotating diode, stator winding, and rotor winding were determined. Physical inspections have not been performed on the two P-3 generators to verify the diagnostic results.

Further work is needed to determine how the reference time-domain feature and the magnitudes of key frequency harmonics change with respect to nonlinear, leading power factor, and lagging power factor loads at various load levels.

This article is based on SAE technical paper 2012-01-2234 by Freeman Rufus and Ash Thakker, Global Technology Connection Inc.; Sean Field, Naval Air Systems Command; and Nathan Kumbar, NAVAIR.

HTML for Linking to Page
Page URL
Rate It
4.17 Avg. Rating

Read More Articles On

Gowanda's "world's smallest" SMG1812 was designed specifically for applications that are magnetically sensitive and therefore require non-magnetic components.
Palmer Wahl Instrumentation Group's custom dial faces—its BiMetal and Direct Drive Dial thermometers, as well as its pressure gauges with screw-on bezels—may be customized, no matter the dial size.
W. L. Gore & Associates’ Shielded Twisted Pair Cables reduce weight while maintaining reliable signal integrity.
A*STAR Institute of Microelectronics (IME), GLOBALFOUNDRIES Singapore Pte. Ltd., and Masdar Institute of Science and Technology entered into a collaboration to develop and advance micro-electro-mechanical systems (MEMS) technologies for automotive, aerospace, consumer, health care, industrial, and mobility applications.

Related Items