| Summary: | 博士 === 國立成功大學 === 航空太空工程學系碩博士班 === 91 === Gaspath analysis holds a central position in the engine condition monitoring (ECM) and fault diagnostics (FD) technique. However, popularization of this approach has been impeded when practical enforcements were tried in both civil and military sectors. Artificial neural network (ANN) arises as a new approach which avoids the fundamental difficulties associated with the classical model-based methods. The objective of the present work is to develop a reliable ANN-based diagnostic system that can be enforced in the practical applications. Back-propagation, feedforward neural nets are employed for constructing the engine diagnostic networks. Noise-contained training and testing data are generated using an influence coefficient matrix and the data scatters. The results indicate that for situations where sensor scatters are comparable to those of the normal engine operation, the success rates for both 4-input and 8-input ANN diagnoses achieve high scores which satisfy the requirement. The success rate of 4-input diagnosis is almost as good as that of the 8-input. Although the ANN-based method possesses certain capability in resisting the influence of input noise, the success rate of fault diagnosis still depends mainly on the quality of the measurements obtained. A high success rate of diagnosis can only be guaranteed when a correct set of measurement deltas is available. Thus, a design of a preprocessor that can perform sensor data validation is of paramount importance. To this end, the present work proposes a genetic auto-associative neural network algorithm that can perform off-line sensor data validation simultaneously for noise-filtering and bias detection and correction. Neural network-based sensor validation procedure usually suffers from the slow convergence in network training. In addition, the trained network often fails to provide an accurate accommodation when bias error is detected. To remedy these problems, the Levenberg-Marquardt (LM) algorithm is adopted to speed up the network training and a novel two-step approach is proposed for bias accommodation problems. The first step is the construction of a noise-filtering and a self-mapping auto-associative neural network based on the LMBP algorithm. It is shown that the noise can be greatly filtered by the noise-filtering auto-associative neural network. The second step uses an optimization procedure built on top of these noise-filtering and self-mapping nets to perform bias detection and correction. Non-gradient genetic algorithm search is employed as the optimization method. It is shown in the present work that effective sensor data validation can be achieved for noise-filtering, bias correction, and missing sensor data replacement incurred in the gaspath analysis. This newly developed algorithm can also serve as an intelligent trend detector. True performance delta and trend change can be identified with no delay to assure a timely and high-quality engine fault diagnosis.
|
|---|
