| Summary: | 碩士 === 國立雲林科技大學 === 工業工程與管理研究所 === 87 === This research is to develop an on line fault diagnosis model based upon the neural network clusters. The algorithm of its neural network is developed from the probabilistic neural network for making diagnosis of unsteady dynamic systems.
There are many systems operating in an unsteady environment. The operator of these systems must often adjust the parameters to fit the environment's change, and that will make the system to keep in an unsteady state.
Most of the current diagnostic systems are divided into two different models. One is the steady model and the other is the unsteady model.
The steady model diagnoses systems with their instantaneous parameter values. It assumes that the system will get to some steady states when it is operating in some particular situations. And the diagnostic system will diagnose them with their parameter combinations. But because it does not take into account the system’s past behavior in diagnostic process, it can't deal with the diagnostic problems whose answer must depend on the system past operating process.
The unsteady model uses the time window of signals to help identifying the system's dynamic state. These time window data are used to train the neural network before using them to predict the system's variation in different states. While diagnosing, it maps the system's actual parameters with the neural network predictions then outputs the diagnostic result. But because signals vary continuously and consists of many combinations in the unsteady dynamic system, it is difficult to get enough samples for training the neural network. And if the training samples were not enough to describe the system whole behavior, the neural network could make the wrong predictions.
For resolving above problems, this research uses the probabilistic neural network cluster to develop a case-base diagnostic model. This model consists of two operating steps. At the first step, it identifies the system signals’ time window pattern and gives them a number. At the second step, it analyzes these time window pattern number combinations to output the diagnostic result. According to the result of experiment, this system will be able to diagnose the unsteady dynamic systems by observing their signal behaviors and it tends to not make the wrong diagnosis even in lack of training sample.
|
|---|
