Data processing: artificial intelligence – Knowledge processing system – Knowledge representation and reasoning technique
Reexamination Certificate
1998-12-02
2001-10-09
Powell, Mark (Department: 2122)
Data processing: artificial intelligence
Knowledge processing system
Knowledge representation and reasoning technique
C702S054000, C702S185000
Reexamination Certificate
active
06301572
ABSTRACT:
FIELD OF THE INVENTION
The invention pertains to systems and methods of vibration analysis. More particularly, the invention pertains to such systems and methods which incorporate fuzzy adaptive resonance theory neural networks for distinguishing between normal and abnormal vibrations.
BACKGROUND OF INVENTION
Gas turbines engines and power generation, such as the General Electric LM2500, are used in numerous marine applications and for power generation. This harsh environment demands early detection and analysis of impending turbine failure in order to prevent catastrophic failures which may endanger personnel as well as shipboard equipment.
The current “state-of-the-art” monitoring equipment incorporates the Wigner-Ville Distribution (WVD). The WVD monitors the once-per-rev vibration energy of the turbine components.
The once-per-rev energy occurs at the frequency corresponding to the rotational velocity of the turbine. This type of system is being used by the US and Foreign navies to monitor the vibration of the LM2500 gas turbine.
The existing system incorporates acceleration sensors that are mechanically connected to the turbine casing. These sensors provide an analog signals suitable for analysis. These signals are then fed into a data acquisition system to filter and digitize the signal.
Sensor output signals are converted to the frequency domain. In one aspect, the conversion can be implemented using Fast Fourier Transform-type (FFT) processing. The frequency domain representation can be processed by deleting those frequency components known to be of little or no interest.
In addition, the amplitudes of negative frequency components can be set to zero. Inverse FFT processing can then be carried out to produce a time domain signal having only frequencies of interest and no negative frequencies.
The technique of isolating sections of the frequency domain signal to create only a few frequency components in the signal allows the WVD to be applied without any smoothing to reduce cross-term energy. The cross-term energy was controlled to occur in spectral locations that were not used in the analysis. This technique can be used in other analysis efforts with the WVD.
The WVD thus provides a highly accurate measure of the turbine vibration amplitude at any instant. This has been used as an important feature in a condition-based monitoring application.
As the vibration level changes under identical operating conditions, the change can be recorded and used as a measure of turbine health. Steadily increasing vibration levels indicate deteriorating turbine health.
The output of the WVD can be used with a thresholding algorithm to detect excessive once-per-rev (1X) vibration. Thresholding of the 1X vibration is used in many gas turbine installations. This allowed detection of deteriorating turbine condition. Ideally, the deterioration would be detected at an early enough stage to prevent catastrophic damage, and to schedule maintenance activities.
Complex signals have been processed by neural networks. One known form of a neural net is the Fuzzy Adaptive Resonance Theory, Fuzzy (ART) neural network.
This is a neural network architecture developed by Stephen Grossberg and Gail A. Carpenter of the Department of Cognitive and Neural Systems at Boston University. The network uses a resonance concept that involves comparisons of new inputs to information that is already learned.
If the new input is sufficiently close to the old information, then resonance occurs and the network will adapt to learn the new information. The first implementations of ART operated only on binary (0 or 1) data. The theory has been extended to accommodate analog input values of the range zero to one, using Fuzzy ART. The “Fuzzy” prefix implies that the input numbers are analog; no fuzzy logic connotations need be ascribed to the numbers. The system learns to recognize analog values within subsets of its total input space. The information that is presented to the network occupies a section of the input space corresponding to the amplitude of the information.
Fuzzy ART is called an unsupervised neural network because the training set is presented to the network without information as to the desired classification. Instead, the network forms an internal representation of the data presented to it.
While the known vibration analysis systems have generally been effective for their intended purpose, it would be desirable to improve the accuracy and extent of the analysis. In this regard, it would be desirable to be able to detect the presence of non-stationary frequency components. It would also be desirable to be able to eliminate frequency components not of interest.
It would also be desirable to be able to characterize, in some sense, a particular turbine and then monitor, over time the performance of that unit. In this fashion, deviations from expected performance should be detectable at an early enough stage to avoid the occurrence of catastrophic failures. Detected deteriorating performance can trigger unit maintenance.
SUMMARY OF THE INVENTION
A system and method for monitoring dynamic performance of an operating turbine utilizing at least one vibration sensor coupled to the turbine incorporate a pattern recognition subsystem. The output signals from the sensor can be transformed to a frequency domain using FFT-type processing. These frequency components can be used as inputs to a neural network.
The solutions to many data analysis problems have been shown to be non-algorithmic in nature. That is, they cannot be described or predicted through application of repetitive numerical manipulation or analysis.
An inability to analyze a signal can be caused by the presence of noise or minor variations in signal or sensor data. These kinds of analysis have been shown to potentially benefit from the application of systems that can detect patterns in a signal or signals. One class of such a system is the neural network. This is a system which can be trained to “recognize” certain characteristics or trends in a signal.
In a preferred embodiment, the neural network of the present system and method is the Fuzzy Adaptive Resonance Theory (Fuzzy ART) neural network. The Fuzzy ART is a type of adaptive resonance theory neural network that can examine new input data, and decide if its already-learned prototypes sufficiently match the input (resonate) and if so then the new input is learned.
This network architecture is very good at novelty detection and can perform quick learning of new localized data without destroying the other stored information. This property eases the implementation of an on-line monitoring system, because retraining can occur in near real-time, as opposed to other neural networks that must have lengthy retraining to incorporate new information.
A Fuzzy ART system as incorporated herein includes two major components, an attentional subsystem and an orienting subsystem. The attentional subsystem activates the system in response to the input vector. The orienting subsystem finds the correct internal representation of the new information.
Each input vector presented to the network enters through the Input Layer. The F
0
input layer preprocesses the input layer vector, a, extending the representation of the input vector to allow the network to represent ranges of input vectors in a single neuron, as opposed to storing just a single vector.
The F
1
activity layer determines the amount of activity present in the different neurons when an input vector is presented. In the F
1
layer, the input vector is compared with the stored prototype information, using fuzzy arithmetic, to determine how close the new vector is to the stored prototype.
The F
2
category layer retains the prototypes that are checked for resonance with the input vector. The F
1
and F
2
layers contain multiple neurons, but the F
0
layer contains one neuron. In
FIG. 11
, the activity layer weights are shown as W
j
and the prototype weights are shown as w
j
.
In operation, the new input vector is compared, in the activity layer, F
1
to the
Lockheed Martin Corporation
Powell Mark
Rockey Milnamow & Katz Ltd.
Starks Wilbert
LandOfFree
Neural network based analysis system for vibration analysis... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Neural network based analysis system for vibration analysis..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Neural network based analysis system for vibration analysis... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2612138