| Summary: | 碩士 === 國立成功大學 === 製造資訊與系統研究所碩博士班 === 98 === Signals detected from accelerate and acoustic sensors should be processed to extract features by using such as time-domain, frequency-domain, and wavelet transform to diagnose machine faults. Effective features decide diagnostic performance. However, causing physics of fault modes, feature extraction and selection is a diagnostic analysis challenge.
This work proposed a heuristic dimension reduction procedure to derive limited feature dimensions with maximum diagnostic accuracy in fewer experiments. In this procedure, higher diagnostic effect is selected from features which possess similar meta-attributes, e.g., time duration, frequency bandwidth, and source. Moreover, selection of original signal and the agreeing wavelet band is supported to reduce analyzing time for deriving the agreeing feature sets of the following classification methods.
A machining bearing diagnostics case which included single and mixture fault modes of inner race, outer race, and roller is presented. It shows twenty-eight major features extracted from vibration data can be reduced to nine features in six experiments by using the proposed procedure. After training diagnostic models by using features, comparison of two conditions mixture failure modes classification accuracy of back-propagation neural network is improved from 90.78% to 92.23%; meanwhile, probabilistic neural network is improved from 86.67% to 87.21%.
|
|---|
