外文文献及翻译--应用高阶神经网络对旋转机器的多种故障进行诊断(编辑修改稿)

外文文献及翻译--应用高阶神经网络对旋转机器的多种故障进行诊断(编辑修改稿)内容摘要：

efficient diagnostic results and is useful for realtime condition monitoring and online diagnosis of rotating machines. Keywords: Ellipsoidal unit works。 Hierarchical diagnostic works。 Rotating machinery。 Simul taneous diagnosis of multiple faults。 ural work Introduction In recent years, more and more attention has been given to the theory, method and strategy of fault diagnosis for largescale rotating machines, and quantitative methods have been proposed, such as statistical pattern classification methods,system identificationbased parameter model methods,and so on. These methods always need plicate models and extensive calculation. The Artificial Neural Network(ANN)technique offers potential for fault diagnosis due to its parallel processing, associative memory, selforganising, selflearning and very strong nonlinear mapping abilities. In particular, the ability of neural works to deal with highdimension pattern classification and nonlinear pattern classification problems[3,4]are important in fault diagnosis. The authors have published numerous papers on this subject. As a pattern classification method, the standard feedforward neural work–the MultiLayer Perceptron(MLP)–divides decision space with hyperplanes and the decision regions so formed are always unbounded, which can lead to undesirable extrapolation problems. To overe this limitation highorder neural works using ellipsoidal units have been proposed, which can form bounded decision regions with hyperellipsoids, and are more useful for fault diagnosis applications. In this paper the ellipsoidal unit work is described, the method for initialising hyperellipsoids, and the training algorithm is described. A Hierarchical Diagnostic Artificial Neural Network(HDANN)is proposed with respect to simultaneous diagnosis of multiple faults for rotating machines, and test results are given and discussed. Network Architecture The ellipsoidal unit defined by Eq. (3)is capable of approximating the Gaussian distribution. For the symmetric distribution, one unit can achieve a satisfactory approximation, and the work consists of two layers of linear input units and ellipsoidal output units. Each hidden unit is connected to only one output unit and each output unit has a dedicated set of hidden units for some distribution. Considering the distribution characteristics of feature vectors abstracted from raw data from rotating machines, the work shown in is chosen in this paper. In this work each input unit is connected to each hidden unit with two weights, which determine the centre coordinate Fig1 work Architecture Applications of HDANN in。