Vibration signature analysis using artificial neural networks

Abstract

Vibration signatures are useful predictors of incipient equipment failures in mechanical systems such as reciprocating engines, generators, and turbines. Properly interpreted, they can prevent expensive and potentially dangerous failures. Also, these signatures can be used to indicate the need for maintenance, allowing scheduled versus costly and inconvenient unscheduled downtime. Analyzing and classifying these signatures has proven difficult. Often, they are the result of multiple inputs from highly complex nonlinear coupled systems. With this kind of data, traditional classification methods often produce inaccurate and misleading results. Attempts have been made to use expert systems to perform the classification. This approach presupposes the availability of an expert with the ability to accurately perform the classification, as well as the willingness to undergo the lengthy process of rule development working with a knowledge engineer. This method has failed in many cases. Often the expert is simply unable to explain how he performs the classification; his actual methods may be largely the result of subjective decisions, not reducible to the precise rules needed for an expert system. Artificial neural networks have proven ideal under such circumstances. Learning by example, they automatically develop arbitrarily complex nonlinear relationships directly from experimental data; hence, there is no need for humans to identify the rules. Given an adequate training set they exhibit generalization, correctly classifying signatures despite noise and the inevitable variations in the system. This paper describes research done on artificial neural network classification of actual vibration signatures from large rotating machines. A group of these signatures and their correct classifications were collected from operating machines. These were then used as a training set for two types of artificial neural network classifiers; the first using backpropagation, the second using a probabilistic neural network. Once trained, both networks were then tested for accuracy on previously unseen signatures. Relative classification accuracy was then determined between the two networks, a conventional Baysian classifier, and a human expert.