Incipient Fault Detection in Rotating Machines Using a Neural Network

本文的主要目的是开发一种利用人工神经网络检测旋转机械早期故障的新方法。由于中型感应电动机的广泛应用和经济原因，本项目采用中型感应电动机作为旋转机械的原型。这个概念可以很容易地从感应电动机推广到其他旋转机械。基于异步电动机的稳态性能，设计了一种用于异步电动机早期故障检测的人工神经网络。但是，当电机偶尔受到干扰时，检测方案将实时应用，在这种情况下，电机将不会始终处于稳态。对故障检测器的不适当输入将在故障检测器上产生假警报。一个人工神经网络将被开发，以过滤掉瞬态测量，但保留稳态测量，从而提供正确的测量故障检测器。所提出的工作的预期意义将是开发一种用于旋转机械的在线早期故障检测器。该检测器由两部分组成：(1)测量扰动滤波人工神经网络和(2)早期故障检测人工神经网络。该检测器将能够检测电机常见的早期故障，例如匝间绝缘故障和轴承磨损，并且对电机中的干扰以及测量噪声具有鲁棒性。将发展理论以确保网络的性能，包括不同运动操作条件下的训练（学习）和召回方案。将干扰和噪声滤波神经网络的性能与传统的噪声滤波方案进行比较

The main objective of the proposed work is to develop a new approach to detection of incipient faults in rotating machines by using artificial neural networks. Medium size induction motors are used as prototypes for rotating machines in this project due to their wide application and also for economic reasons. The concept can be easily generalized from induction motors to other rotating machines. The design of an artificial neural network to detect incipient faults of induction motors is based on the steady-state performance of the motor. However, the detection scheme will be applied in real time when the motor experience occasional disturbances, and in this case, the motor will not always be in steady-state. Inappropriate inputs to the fault detector will yield false alarms at the fault dector. An artificial neural network will be developed to filter out the transient measurements but retain the steady-state measurements and thus feed the correct measurement to the fault detector. The expected significance of the proposed work will be development of an on-line incipient fault detector for rotating machines. The detector is composed of two parts: (1) a measurement disturbance filter artificial neural network, and (2) an incipient fault detection artificial neural network. The detector will be able to detect the common incipient faults of the motor, such as turn-to-turn insulation failure and bearing wear, and will be robust to disturbances in the motor as well as to measurement noise. Theory will be developed to ensure the performance of the networks, including training (learning) and recalling schemes under different motor operating conditions. The performance of the disturbance and noise filtering neural network will be compared to the conventional noise filtering schemes