NEURAL-BASED REMOTE DIAGNOSTICS OF MANUFACTURING MACHINERY

摘要

As industrial competition intensifies, and more emphasis is placed on increased productivity and cost reduction, development of means to automatically monitor and assess the health of plant equipment becomes more important. This paper presents results for our research in feature extraction, neural network technology, and advanced networking techniques for remote health assessment and failure diagnostics of laboratory and industrial machinery. A laboratory scale coupled motor and generator suspension system and an industrial 1500-ton ammonia liquefier system are used for testing purposes of this research. Welch's method is used to form a low variance Power Spectral Density estimate from the instrumentation signals, and feature input vectors are formed. Conditioning is performed on the input vectors to improve differentiation of the features. Finally, the neural network architecture used for real-time fault detection is found based on the number of input features and the number of separating classes. Successful real-time recognition of four failure modes was accomplished for the motor and generator suspension system by extracting 11 features from a 0 to 120 Hz frequency range. Pattern recognition became more robust as training feature variance increased and real-time feature variance decreased. For field implementation, the 1500-ton motor and compressor were sufficiently instrumented to provide a rich array of vibration information. Sensor information and fault diagnostics were displayed on a real-time basis via the World Wide Web.