Intelligent Condition Monitoring of Rotating Machinery

Abstract

This study investigated the use of pattern recognition techniques in intelligent diagnosis of rotating machinery. Existing literature on machine fault diagnosis suggested many approaches to machine diagnosis: notable among them are pattern recognition technique, data mining and Hidden Markov Modelling. A method using MLP neural network classifier for pattern recognition of fault in a centrifugal pump-rig has been developed. The study was primarily experimental and involved the simulation of six types of faults on a centrifugal pump, one at a time. These were bearing failure, seal-ring wear, misalignment, unbalance on impeller, cavitation, and unbalance on coupling. Data were collected using a portable data acquisition system: SKF Microlog. Data were collected when the pump was in no-fault condition. Each fault was trained on a separate neural network, giving a total of six types of networks with different number of inputs and only one output. The results obtained from the simulation work confirmed previous studies that pattern recognition technique is effective in recognising and classifying machine faults. Using the seal-ring wear as an example, the distribution of weight vectors showed low weight values distributed around zero. This is a sign of a healthy network. The distribution was also slightly skewed to the left, indicating the presence of large weight values, and subsequently, the network may have slightly over-fitted the data. The error associated with a decision made by the network was evaluated. After 240 epochs, an average error of 0.004070 was obtained. The validation set error obtained was 0.0%.