Study On Machine Fault Feature Extraction And Early Identification Based On Inner Product Transform

In this paper, the faults feature extraction and early identification of keyfans and pumps, gas turbines, reciprocating compressors and wind turbines arestudied. Theoretical and application researches are made.Theoretical research is made following the relation between inner productand machine fault diagnosis which was summarized by the former researchers.Two main points are presented based on Riesz representation theorem andDuhamel integral. First, if and only if linear transform is adopted by faultfeature extraction, the feature extraction process can be taken as inner product.Second, the response of linear machine system excited by fault forces is theinner product of the forces and the system's physical response function.Furthermore, by giving the physical meaning to frames and dual framesrelation proposed in signal processing field, this relation can be taken as ageneral machine fault feature extraction model termed inner product transform (IPT) model. Different concrete forms of this model are shown in thesituations of signal compression, impulsive vibration fault feature extractionand indicator diagram feature extraction of reciprocating compressors.The problem of huge amount of data and expensive bandwidth isencountered in the monitoring of key pumps on oil platforms and windturbines. Hence, vibration acceleration signal compression is researched.Compression approach based on wavelet transform and sparse representationis presented. This approach is a special form of the IPT model. Experimentaland application studies reveal that relatively optimal results are obtained bysym8wavelet and Matching Pursuit. Without loss of fault features,80%ofdata is compressed.The same mechanism—impulsive vibration of gearbox fault of windturbines and rolling bearing faults of fans, pumps, and double-rotor gasturbines is clarified and used in this work. Two envelope demodulationmethods based on optimal high-pass filter and optimal Antisymmetric RealLaplace Wavelet (ARLW) filter are proposed. These two methods share asame form of IPT model. More effective results of feature extraction andidentification of early or weak impulsive fault are obtained by experimentaland application study.Indicator diagram feature extraction based on Curvelet transform which isalso a special form of IPT model is proposed in the last part of this work. AndSVM is used to reciprocating compressors fault recognition. More clear features can be extracted by Curvelet transform than wavelet transform in theexperimental study. Combination of Curvelet transform and SVM can be usedto construct fault diagnosis expert system.