Automated valve condition classification of a reciprocating compressor with seeded faults: experimentation and validation of classification strategy

• Published in: Smart materials and structures Vol. 18; no. 9; pp. 095020 - 095020 (19)
• Main Authors: Lin, Yih-Hwang, Liu, Huai-Sheng, Wu, Chung-Yung
• Format: Journal Article
• Language: English
• Published: Bristol IOP Publishing 01.09.2009; Institute of Physics
• Subjects: Applied sciences; Exact sciences and technology; Industrial metrology. Testing; Mechanical engineering. Machine design; Pump and compressors (turbocompressors, fans, etc.); Automatic classification; Seat; Fourier transformation; Processor; Time frequency domain method; Fastener; Modeling; Stationary system; Disk valve; Breakdown; Valve; Reciprocating movement machine; Monitoring; Bolted joint; Human; Reciprocating compressor; Electric breakdown; Leakproofness; Cracked plate; Vibration; Probabilistic approach; Non stationary condition; Bolt; Maintenance; Neural network; Experimental study; Spring; Data reduction
• ISSN: 0964-1726; 1361-665X
• DOI: 10.1088/0964-1726/18/9/095020

This paper deals with automatic valve condition classification of a reciprocating processor with seeded faults. The seeded faults are considered based on observation of valve faults in practice. They include the misplacement of valve and spring plates, incorrect tightness of the bolts for valve cover or valve seat, softening of the spring plate, and cracked or broken spring plate or valve plate. The seeded faults represent various stages of machine health condition and it is crucial to be able to correctly classify the conditions so that preventative maintenance can be performed before catastrophic breakdown of the compressor occurs. Considering the non-stationary characteristics of the system, time-frequency analysis techniques are applied to obtain the vibration spectrum as time develops. A data reduction algorithm is subsequently employed to extract the fault features from the formidable amount of time-frequency data and finally the probabilistic neural network is utilized to automate the classification process without the intervention of human experts. This study shows that the use of modification indices, as opposed to the original indices, greatly reduces the classification error, from about 80% down to about 20% misclassification for the 15 fault cases. Correct condition classification can be further enhanced if the use of similar fault cases is avoided. It is shown that 6.67% classification error is achievable when using the short-time Fourier transform and the mean variation method for the case of seven seeded faults with 10 training samples used. A stunning 100% correct classification can even be realized when the neural network is well trained with 30 training samples being used.