当前位置:首页 >> 兵器/核科学 >>

Assessment of data and knowledge fusion strategies for prognostics and health management


Assessment of Data and Knowledge Fusion Strategies for Prognostics and Health Management’
Michael J. Roemer Gregory J. Kacpnyaski RolfF. Orsagh Impact Technologies, U C 125 Tech Park Drive Rochester, NY 14623 (716) 424-1990

mike.roemer@impact-tek.com

Abstract-Various data, feature and knowledge fusion strategies and architectures have been developed over the l a s t several years for improving upon the accuracy, robustness and overall effectiveness of anomaly, diagnostic and prognostic technologies. Fusion of relevant sensor data, maintenance database information, and outputs fkom various diagnostic and prognostic technologies has proven effective in reducing false alarm rates, increasing confidence levels in early fault detection, and predicting time to failure o r degraded condition requiring maintenance action.
The data fusion strategies discussed in this paper are principally probabilistic in nature and are used to aid in directly i d e n w g i t h specific component fault confidence bounds associated w identifications and predictions. Dempster-Shafkr fusion, Bayesian inference,fuzzy-logic inference, neural network fusion and simple weightinghotingare the algorithmic approaches that are discussed in this paper. D a t a h i o n architectures such as centralized fusion, autonomous fusion, and hybrid fusion are described in terms of i n a lgoal their applicability to fault diagnosis and prognosis. The f is to find the optimal combiition of m e a s u r e d system data, data fusion algorithms, and associated architectures for obtaining the highest overall predictiddetection confidence levels associated with a sp&c application. To achieve this goal, a set of metrics has been developed for gauging the performance and effectiveness of a Fusion strategy. Specifically, this paper will demonstratehow various metrics are used for assessing individual and fused vibration-based diagnostic algoritbms. Evaluation of the diagnostic stmtegies was performed using gearbox seeded-fault and accelerated failure data taken with the MDTB (Mechanical Diagnostic Test Bed) at the ARL Lab at Penn State University.

6. CONCLUSION 7. REFERENCES

1. INTRODUCTION
Once sensor signals have been validated, the objective of data fusion is to combine their respective information in the most diagnostically efficient method possible. Multi-sensor data fusion refers to intelligent processing of an array of 2 or more sensors that have cooperative, complimentary and competitive qualities. As long as the sensor m y does not contain independent sensors, arrays usually contain various levels of these three qualities. Cooperative sensors are those that work together to create a new piece of diagnostic information, while a complimentary array creates a more complete picture of a problem. Finally, a competitive array provides unrelated measurements of the same physical phenomena for improved reliability (Brooks, 97).

2. FUSIONAPPLICATIONAREAS
Within a health management systeq there are three main areas where h i o n technologies play a contributing role. These areas are shown in Figure 1. At the lowest level, data fusion can be used to combine information kom a multi-sensor data array to validate signals and create features. One example of data fusion i s combining a speed signal and a vibration signal to achieve time synchronousaveraged vibration features.

FUSION APPLICATION AREAS 3. F U S I O N A R ~ C T U R E S 4. FUSIONTFCHNIQUES 5. ASSESSINGTHEBENEFITS OFFUSION- M.ETRICS
CL7803-5599-201/$10.00 Q 2001 IEEE

1. 2.

blTRODUCTION

6-2979

Centralized Fusion

Detection

Figure 2 - Centralized Fusion Architeture The autonomous fusion architecture shown in Figure 3 quells most of the data management problems by placing feature extraction before the fusion process. The creation of features prior to the actual fusion process provides the sigmiicant advantage of reducing the dimensionality of the information to be processed The main undesirable effect of a pure autonomous fusion architecture is that the feature fusion may not be as accurate as in the case of raw data fusion because a signilkant portion of the raw

Figure 1 - Fusion Application Areas At a higher level (area 2), fusion may be used to combine features in intelligent ways so as to obtain the best possible diagnostic information. This would be the case if a feature related to particle count and sue in a bearing's lubrication oil were fused with a vibration feature such as kurtosis. The combined result would yield an improved level of confidence about the bearing's health. Finally, Knowledge Fusion (area 3) is used to incorporate experienced-based i n f o d o n such as legacy failure rates or physical model predictions with signal-based infomation. One of the main concerns in any fusion technique is the danger of producing a fused system result that is actually performing worse than the best individual tool. This is because poor estimates can drag down the better estimates. The solution to t h i s concern is to weigh the tools according to their capability and performance, which must be realized a priori. The degree of a priori knowledge is a function of the inherent understanding of the physical system and practical experience with the system. The ideal knowledge fusion process for a given application should be selected based on the characteristics of the a priori system information.

signalhasbeeneliminated.

Autonomous Fusion

... ...

3. FUSION AR-CTURES
Identifying the optimal fusion architecture and approach at each level is a vital factor in assuring that the realized system truly enhances health monitoring capabilities. A brief explanation of fusion architectures will be provided here. The centralized fusion architecture fuses multi-sensor data while it is still in its raw form as shown in Figure 2. In the fusion center of t h i s architecture, the data is aligned and correlated during the first stage. This m e a n s that the competitive or collaborative nature of the data is evaluated and acted upon immediately. Theoretically, this is the most accurate way to fuse data, however, it has the disadvantage of forcing the fusion processor to manipulate a large amount of data This is often impractical for real-time systems with a relatively large sensor network (W 97). Figure 3 - Autonomous Fusion Architecture

A hybrid fusion architecture takes the best of both and is often considered the most practical because raw data and extracted features can be fused in addition to the ability to "tap" into the raw data ifrequiredby the fusion center (Figure 4).

6-2980

Hvbrid Fusion

uses three different sources of information. A-priori probability of failure at time t, (PF~t)) , the probability of failure as determined data, and feature reliability fiom the diagnostic classifier (Cwktl) a r e must be taken to which is independent of time (&,,). C prevent division by zero.

...

...

...

I

il

id

Validation

Figure 4 - Hybrid Fusion Architecture

The Bayesian process is a common and w e l l established fusion technique, but also has some disadvantages. The knowledge required to generate the a priori probability distributions may not always be available, and instabilities in the process can occur if conflicting data is presented or the number of unknown propositions is large compared to the know propositions.
Dempster-Shajm Method

4. FUSION TJ3CHNIQUES
There are probably hundreds of techniques for performing data, feature or knowledge fusion Because of this fact, sorting through which technique is best can be a daunting and involved task. In addition, there are no hard and fast rules about what fusion r architectures work best for any particular techniques o application. The proceeding sections will describe some common fusion approaches such as Baysian Jnference, Dempster-Shafer combination, WeightingNoting, Neural Network Fusion and Fuzzy Logic Inference. In addition, a set of m d c s will be discussed for independently judging the performance and effectiveness of the h i o n techniques within a diagnostic system Bayesian Infwence Bayesian Inference can be used to detamine the probability that a dtagnosis is correct, given a piece of a priori information Analyticallythis process is described as follows:

The Dempster-Shafer method addresses some of the problems discussed above and specifically tackles the a priori probability issue by keeping track of an explicit probabilistic measure of the lack of infomtior~The disadvantage of this method is that the process can become impractical for time critical operations in large fusion problems. Hence, the proper choice of method should be based on the specific diagnostic/prognostic issues that are to be addressed

In the Dempster-Shafer approach, uncertainty in the conditional probability is considered. The Dempster-Shafer methodology hinges on the COIlSLSUCtion of a set, called the frame of discernment, which Contains every possible hypothesis. Every hypothesis has a belief denoted by a mass probabiity (m). Beliefs are combined in the following m e r .

Belitf(H, ) =

1- C m , ( A ) - m j ( B )

AnB=Hn

where:
p ( f , l q= The probability of fault (f) given a diagnostic output

The technique can be best explained through the use of the following example. Given: A diagnostic classifier detects Fault A with the following probability and associated uncertainty: P~=0.80+/-0.15 The a priori probability of Fault A Occurring (based on current conditions and a priori information) is the following: P g = 0.30 +/- 0 . 1 0

(0),p(o,I~;)=The probability that a diagnostic output (0)is
~ The probability of the fault associated with a fault (f), and p ( ) =
( f ) occurring.

Bayes’ theorem is only able to d y e discrete values of confidence fiom a diagnostic classifier (i.e. it observes it or it doesn’t). Hence, a modified method has been implemented that

6-2981

Therefore:

m(A) = 0.65 m(A,A’) = 0.30 m(B) = 0.20 m(B,B’) = 0.20

m(A’) = 0.05 mQ3’) = 0.60

membership functions are used to calculate the fused output by either taking the centroid,max height or midpoint of the combined
hCti0n.

A

I

I

B I B’ 0.13 I 0.39

I
I

B,B’ 0.13

An example of a feature fusion process utilizing fuzzy logic is shovm below in Figure 5 . In this example, features f “ an image

A,A’
And

m(A) + m(B) {True)= (0.13 + 0.13 +0.06)/(1-(0.01 +0.39)) = 0.53

are being combined to help determine if a “foreign” object is present in an original image. Image features such as tonal mean, midtones, kurtosis and many others are combined to give a single output that ranks the probability of an anomalous feature being present in the image.

This result is called the ‘‘belief’ and it is the fused probability lower bound. The uncertainty in this result is the following:

m(A,A’)

+ m(B,B’) = 0.06 / (140.01 + 0.39)) =
0.10 or +/- 0.05

(4)

Hence, the probability of Fault A having actually occurred given the diagnosticoutput and in-field experienceis 0.58 +/- 0.05.
WeightingNotingFusion

% Shadows,

Tonal Mean
# of modes

YO Midtones

YO Highlights
Standard Dev.
Kurtosis

Both the Bayesian and Dempster-Shafer techniques can be computationally intensive for real-time applications. A simple weighted average or voting technique is another approach that can be utilized. In both these approaches, weights are assigned based on a prior knowledge of the accuracy of diagnostic/prognostic techniques being used. The only condition is that the sum of the weights must be equal to one. Each confidence value is then multiplied by its respective weight and the results are summed for each moment in time. Weights can also change as a function of time.

Modified Outputs Figure 5 - Example of Fuzzy Logic Inference
Neural Nelwork Fusion

n=l

A well accepted application of artificial neural networks (ANNs) is data and feature fusion. For the purposes of fusion, a networks ability to combine information in real-time with the added capability of autonomous relearning (if necessary) makes it very attractive for many fusion applications.
Artificial neural networks (ANN) utilize a network of simple

Where i is the number of features, C is the confidence value, and W is the weight value for that feature. Although simple in implementation, choosing proper weights is of critical importance to highlightingthe proper features under various operating modes.
Fuzzy Logic Infwence

Fuzzy Logic Inference i s a fusion technique that utilizes the membership h c t i o n approach to scale and combine specilic input quantities to yield a fused output. The basis for the combined output comes fiom scaling the developed membership functions based on a set of rules developed in a rulebase. Once t h i s scaling is accomplished, the scaled membership functions are combined ‘on, maximumor by one of various methodologies such as sur“& “single best” techniques. Finally, the scaled and combined

processing units, each having a small amount of local memory. These units are connected by “communication” l i n k s ,which carry n i t s operate only on their local data, which is numerical data.The u received as input to the units via the connections. M o s t ANN’S have some sort of training rule by which the weights of connections are adjusted based on some optimization criterion. Hence, ANN’s learn fiom examples and exhibit certain capability for generalization beyond the braining data (examples). ANN’S represent a branch of the artificial intelligence techniques t h a t have been increasingly accepted for data fusion and automated diagnostics in a wide range of aerospace applications. Their abilities to fuse features, recognize patterns, and to leam fiom samples have made ANN’Sattractive for fitsing large data sets kom complex systems.

6-2982

The ANN structure is called its architecture, which is an expression of the number of processing units and of the connections among these units. Most processing units are arranged in lqers (a layer is a collection of the units aligned for the same computational sequence), and the ANN is often referenced by the number of layers and the number of units in each layer.

processing of a large number of training patterns. In the supervised m e t h d a ‘’teacher” is present during the training phase who tells the network how well it is performing or what the correct behavior should be, i.e., it is used for specifying what target outputs should result from an input pattern. A representative application of neural network fusion would be to combine individual features from different feature extraction algorithms to give a single representative feature. An example of this type of neural network fusion will be given in the following section
Results

._._...I

-+ +
4

. _ . . I _ . . . . _ ”

Input layer

Hidden layer

Output layer

Figure 6 - Simple Neural Network Fusion Architecture Each & d connection line in Figure 6 represents a numerical value called the weight, representjng the connecting strength n i t s . Each circle is a unit and it between the two inter-connected u performs three sequential computations: the first is to multiply the weight by the output of the unit on the other end of the connection; the second is to sum the weighted outputs fiom all mnnections; and the third is to apply the weighted sum to a function (usually nonlinear and bounded) called an activation function. One of the most common activation functions is called the sigmoid function and t h e b w f ( x ) (0 to 1 input) and bipolar g(x) (-1 to 1 input) versions are given below. They are useful because the simple form of the derivative reduces the compuhtiod burden during training.

The fusion techniques previously described have been implemented on various vibration features extracted from a data set developed during a series of transitional run-to-faihu-e tests on an industrial gearbox at Penn State ARL.In these tests, the torque was cycled from 100% to 300% load starling a t approximately 93 hours. The drivegear experienced multiple broken teeth and the test was stopped at approximately 114 hours. The data collected during the test was processed by many feature extraction algorithm techniques t h a t resulted in 26 vibration features calculated from a single accelerometer attached to the gearbox housing. The features ranged in complexity &om a simple RMS level to a measure of the residual signal (gearmesh and sidebands removed) from the time synchronous averaged waveform. More information on these vibmtion features may be found in Byington, 19971. Figures 7 and 8 show plots of two of these features, Kurtosis and NA4, respectively. The green, smoother s the ‘‘puud truth severity” or the line in each of these plots i probabiltity of failure as determined ikom visual inspections discussed next.

The functional value of the weighted sum is called the output (or threshold) of the unit This sequence of computation is carried out for each unit and for each layer u n t i l the outputs layer of the ANN is reached. Training a neural network for a fusion application involves the process of adjusting the weights and evaluating the activations of the numerous interconnections between the input and output layers. There are two fundamental types of leaming methods used for feature h i m applications: Zmrupemked and supervised. In the unsupervised method, learning is autonomous; networks discover properties of the d a t a set a n d learn to reflect these properties in its output, ie., it is used to group similar input patterns to facilitate

105

110

5

Figure 7 - Kurtosis Feature

6-2983

0.5

0.4 -

0.3

~

0 . 20.1

-1

___

95

100 105 Time (Hours)

110

115

1

Figure 8 - NA4 Feature Borescopic inspections of the pinion and drivegear for this particuler test run were performed to bound the time period in which the gear had experienced no damage and when a single tooth had failed. These inspection results, coupled with the evidence of which features were best a t detecting tooth cracking prior to failure features (as determined tom the diagmotic metrics discussed l a t e r ) , was the a-priori information used to implement the Bayesian Inference, Weightinfloting, Neural Network, and Dempster Shafer lsion processes. The seven best vibration feature as determined by a consistent set of metrics described in the next section, were assigned weights of 0.9, average performing features were weighted 0.7, and low performers 0.5 for use in the voting scheme. These weights are directly related to the feature reliability j n the Baysian Inference fusion. Similarly, the best features were assigned the uncertainty values of (0.05), average performm (0.10) and low performers (0.15), for the Dempster Shafer combinatioa The prior probability of Mure required for the Neural Network, Bayesian Inference and Dempster Shafer fusion were built upon the experiental evidence that a drive gear crack will form in a mean time of 108 hours with a variance of 2 hours. Seven of the 26 total vibration features calculated are shown in Figure 9. Note that some of the features have little correlation to the actual tooth failure as de&ed by the ground tmth inspection data. The results of the Dempster-Shafer, Bayesain and Weighted fusion techniques on all 26 features is shown in Figure 10. All three approaches increase in their probability of failure estimates a t around 108 hours (index 269). Clearly, the voting fusion is most succeptible to false alarms, the Baysian Inference suggests a probability of failure increase early on but isn't capable of producing a high d d e n c e level. Finally, the Dempster-Shafer combination provides the same early detection, achieves a higher confidence level, but is more sensitive tbroughout the failure transition region overall.

300

350

Figure 9 - Top Seven Vibration Features

0

50

100

150

MO

250

300

360

Figure 10 - Fusion of all Features Next, the same fusion algorithms were applied to just the best seven features as defined by the metrics. As one can see, the fusion of these seven features produces more accurate and stable h a t the Dempsterresults, which are shown in Figure 11. Note t Shafer combination can now retaiu a high con6dence level with more robustness throughout the critical failure transition region.

6-2984

5. ASSESSING THE BENEFITS OF FUSION - METRICS
I
0.8 O.)i

Bavesian

Uempster Schaefer

--

0'71
0.6

"I

I

I

h
150 200

50

100

250

300

393

Based on the technical results fiom the vibration feature fusion demonstrated in the previous section, it is now worthwhile to assess the benefits of these fusion processes utilizing a consistent set of mathematical ground rules (i.e. metrics). Through implementation of fusion techniques at difFerent points in the health management intiormation flow (data - features knowledge), as was show in Figure 1, we expect to obtain some technical and economic benefits. To illustrate a measure of the magnitude of these benefits, some performance metrics associated with an algorithm fault detection and diagnostic capabilities were developed, along with effectiveness metrics to capture system implementation and cost issues. The metrics developed are shown in Figure 13 and are briefly discussed next. In this paper, we will present only a few results of an exhaustive study that was performed on the benefits of combining individual vibration features to aid in the detection of cracked gear teetk Specifm regarding the individual vibration features that were fused for improved fault detection can be found in [Orsagh and Roemer, 20001. This paper also provides the detail on how the p e r f o m c e and effectiveness metric scores were developed and calculated The results fiom this paper show how the detection ability of different algorithm afect the total ownership cost of implementing a particular fusion technology.

Figure 11 - Fusion of 7 best features Finally, a simple back propagation neural network was trained on four of the top seven features previously fused (RMS, Kurtosis, NA4, and MSA). In order to t r a i n this supervised neural network, r u t h " was the probability of failure as defined by the ''ground t required as a-priori information as described earlier. The network aukmatically adjusts its weights and thresholds (not to be conhsed with the feature weights) based on the relationships it sees between the probability of failure curve and the correlated feature maguitudes. Figure 12 shows the results of the neural network after being trained by these d a t a sets. The difference between the n e d network output and the "ground truth" probability of failure curve is due to error that still exists after the h i s error. Once network parameters have optimized to minimize t trained, the neural network fusion architecture can be used to intelligently fuse these same features for a diixent test under similar operating conditions.

1

08

Figure 13 - Technology Performance Metrics
06

04

02

0 50 100

Cycling Load

150

I

250

I

300

I

340

200

Crack Formed

Failure

Figure 12 - Fusion with a Neural Network

The benefits achieved through accurate detection and fault isolation by implementing fusion techniques are weighed against the costs associated with false alarms, inaccurate diagnoses, licensing, and resource requirements of implementing and operating specific fusion techniques. The simplified cost function shown below states the Technical Value provided by a diagnostic or detection technology for a particular failure mode. The value of a fusion-based diagnostic tool in a particular application is the summation of the benefits it provides over all the failure modes that it can diagnose less the implementation cost, operation and maintenance cost, and

6-2985

consequential cost of incorrect assessments as stated in “Total Value” equation layout as savings-costs.
Technicalvalue= P,*(D*a + I * B ) - ( l - P,)

* (P,
where:

* ( # I

- 4 *e>

(7)

the fault severity approaches the detection level. The Stability c measures the range of confidencevalues that occur over the fault transition by integrating the peak to peak difference at each point in the transition. In addition, diagnostic systems should detect anomalies over the full range of operating (duty) conditions such as loads, speeds, etc. The Detection Oun, Sensitivitv Metric measures the difference between the outputs of a diagnostic tool under various duty conditions.
w

Pf= Probability (time-based)of Occurrence for a failuremode D = Overall Detection Confidencemetric score a = Savingsrealized by detecting a fault prior to failure I = Overall isolation confidencemetric score p = Savings realized through automated isolation of a fault PD= False positive detectionmetric score Q = Cost associatedwith a false positive detection PI =False positive isolation metric score 6 = Cost associatedwith a false positive isolation
TotalVulue=
CTechnicuIVulue, - A - 0 - ( 1 - % ) * 6
FmIumW&a

A diagnostic tool that incorrectly reports anomalies is unacceptable because it reduces availability and increases maintenance costs for the equipment. The Fake Positive ConfidenceMetric measures the frequency and upper confidence limit associated with false anomaly detection by a diagnostic tool. Calculation of the false confidence metric is based on the fdse positive function as defined in reference [Orsagh and Roemer, 20001. In an operational environment, sensor data is sometimes contaminated with noise that may interfere with the operation of diagnostic algorithms. The robustness of an algorithm to noisy d a t a is measured by the Noise Sensitivitv Mm’c. Acquisition and implementation costs of the diagnostic algorithm may have a significant effect on the overall system’s cost effectiveness. The Implementation Cost Metric simply measures the cost of acquiring and implementing a diagnostic system on a single application. If the diagnostic system is applied to several pieces of equipment, any shared costs are divided among them. Operation and maintenance costs may also play a significant role in determining whether a diagnostic system is cost effective. The O&M Cost Metric measures the annual cost incurred to keep the diagnostic system running. These costs may include manual data collection, inspections, laboratory testing, data archival, re-licensing fees and repairs. The ability of the diagnostic algorithms or system to be run w i t h specified time requirements and on traditional computer platforms with common operating systems is important when considering implementation on multiple platforms. Therefore, a metric that takes into account computational effort as well as static and dyllarmc memory allocation requirements is necessary. The C ~ n ~ ~ uResource ter Mehic computes a score based on the normalized addition of CPU time to run (in terms of floating point operations), static and dynamic memory requirements for RAM and static source code space, and static and dynamic hard disk storagerequirements. Computer requirements may be a s@cant issue in some applicationssuch as aircraft. Complex systems are generally more susceptible to unexpected behavior due to unforeseen events. The System Complaitv Metric measures the complexity of diagnostic systems in terms of the number of source lines of code (SLOCs) and the number of inputs required

where:
A = Acquisition and ImplementationCost 0 = Life Cycle Operation and Maintenance Cost P, = Computer Resource Requirementscore 6 = Cost of a standardcomputer system

Table 1 shows the resulting scores for each of the individual diagnosticalgorithm evaluated as well as for the Dempster-Shafer fusion technique. For all of the metrics, a low score indicates an undesirableresult, and a high score indicates a desirableresult. For example, a high Computer resource requirement score is awarded to algorithms that use a small portion of the computer’s resources. A brief discussion of the metric developed is given next for completeness, but refer to [Orsagh and Roemer, 20001 for details on the exact calculation.

e t r i cm e a s u r e s an algorithms ability to The Detection Threshold M identify anondous operation associated with incipient f d t s with a specified confidence level. Confidence levels of 67% and 95%, corresponding to one and two standard deviations, are used to calculate the detection threshold metric. An assessment of the detection cdidence level over the entire severity range for 0 to 1 i s achieved with an Overall Detection Confidence metric. An algorithm that detects an incipient fault with high confidence will receive a high Overall Confidence score, while an algorithm that does not report a fault until it becomes very severewould receive a low score.
A confidence level that fluctuates Wildly is diflicult to interpret

and therefore undesirable. For example, a diagnostic tool t h a t produces a B o o l e a n result of either no fault or fault may flicker as

6-2986

Calculation of the Detection Technical Value, Overall Performance and Overall Effectiveness are based on weighting factors described in Tables 2 and 3. These factors were used to calculate the results shown in Table 1.

6. CONCLUSION
This paper provides an in-depth discussion about many aspects of fusion including where fusion should exist within a health management system the Herent types of fusion architectures, and a number of different fusion techniques. These fusion techniques were applied to vibration features extracted during a transitional failure test associated with an industrial gearbox The results yielded conclusive evidence that fusion can be very valuable in the dmgnostic process if chosen judiciously. Finally, an approach was presented for assessing the performance and effectiveness of the vibration feature extraction algorithms and the fusion techniques themselves.

[l] Hall, D., and Llinas, J., “An Introduction to Multisensor Data Fusion”, Proceecllngs of the IEEE, January 1997.
Benefit of Detection Cost of Std. Computer
$50000 $2000

[2] Leferve, E., and Colot, O., “A classification method based on the Dempster-Shafer’s theory and information criteria”Proceeding of the 2”dInternational Conference on Infomation Fusion, July 68,1999.
[3] Orsagh RF., Roemer MJ., et al “Development of Mettics for Mechanical Diagnostic Technique Qualification and V a l i d a t i o n ” , COMADEM Conference, Houston Tx, December 2000.
[4] Agosta, J. M, and Weiss, J. W., “Active Fusion for Diagnosis Guided by Mutual Information Measures”, Proceedmg of the 2nd International Conference on InformationFusion, July 68,1999

The net result of this feature performance and effectiveness study was an analytical and relatively unbiased evaluation of different feature diagnostic approaches. For this particular study, the Dempster-Shafer combination yielded the highest overall technical score and value because of its robustness and sensitivity;however, it was penalized for its complexity in the final effectiveness score. In other words, due to the fact that the Dempster-Shafer fusion process needs input from several different feature extraction algorithms makes it more complex to implement and therefore reduces its overall costlbenefitof implementation.

[SI Brooks, R R, and Iyengar, S. S, Multi-Sensor Fusion, Copyright 1998 by Rentice Hall, Inc., Upper Saddle River, New Jersey 07458

6-2987

[6] Roemer, M. J. and Kacprzynski, G.J., “Advanced Diagnostics and Prognostics for Gas Turbine Engine Risk Assessment,” Paper 2000-GT-30, ASME and IGTI Turbo Expo 2000, Munich, Germany, May 2000. [7] Roemer,M. J., and Ghiocel, D. M., “A ProbabilisticApproach to the Diagnosis of Gas Turbine Engine Faults” Paper 99-GT-363, ASME and IGTI Turbo Expo 1999, Indianapolis, Indiana, June 1999. [8] Roemer, M. J., and Atkinson, B., “Real-Time Engine Health Monitoring and Diagnostics for Gas Turbine Engines,” Paper 97-GT-30, ASME and IGTI Turbo EXPO1997, Orlando, Florida, June 1997. [9] Byingtoq C.S., Kozlows.ki, J.D., “Transitional D a t a for E s t i m a t i o n of Gearbox Remaining Life”, Proceedings of the 51“ Meeting of the MFPT, April 1997.
Dr. Michael J. Roemer is the Director of Engineering at Impact Technologies in Rochester, Ny and Adjunct Professor of Mechanical Engineering at the Rochester Institute of Technology. He was formerly a Vice President of Engineering at STI Technologiesprior to joining Impact Technologies. Mike has a Ph.D. in Mechanical Enginewing, MS. in Systems Engineering and B.S. in Electkal Engineering, all from the State University o f New York at Bu$alo. He has over 14 y e m acpen‘ence developing real-time, automated health management technologiesfor complex systems, including large steam and gas turbines, gas turbine engines, rotary/fied-wing aircrq? subsystems and naval propulsion systems. He has developed several diagnostic and prognostic capabilities for complex system utilizing probabilistic methodr that are directly linked to maintenance planning and system operation. He is the author or co-author of more than 50 technical pqers in these subject areas. He is current& the Chairman of the Machinery Failure Prevention Technology (MFPT) Society, a Divirion of the Vibration Institute, and Prognostics Lead for the sAE-E32 Engine Condition Monitoring Committee.

generation condition-based mmntenance systems, aircraft prognostics, and health management design for the Nay and USAF. Greg has publishedpapers and developed technologies in the area o f maintenance optimization, FMECA, Lge Cycle cost assessment, model-bared prognostim and data &ion technologks. Greg has his MS and BS in Mechanical Engineeringfrom Rochester Institute of Technology.

Gregory J. Kacptqvnski is a Project Manager at Impact Technologies with 5-y. exp’ence in the development and testing of diagnostic&ognostic systems for compressom, pumps, trmmksions, g a s and steam turbinas. He has been involved in developing real-time, intelligent health monitoring systems for gas turbine engines for on-wing and test cell applicah*om as well as for other air vehicle subsystems. He has managed mdtiple SBIRs dealing with next

6-2988


赞助商链接
相关文章:
更多相关标签: