Study On Principle Of Nonstationary Signal Feature Extraction And Its Application In Fault Diagnosis Of Reciprocating Compressor

Vibration is the most important symbolization of the compressor operation condition. The vibration signals of reciprocating compressors are non-stationary. How to analyze effectively the non-stationary signals has become one of important problems for monitoring and diagnosis. In this dissertation, fault diagnosis techniques based on the principles of nonstationary signal feature extraction were studied, and they were applied in reciprocating compressors. The thesis is devoted to the following researches:1)High Resolution Pursuit(HRP) is a new adaptive approximation technique of signal. HRP is an enhanced version of the matching pursuit algorithm. The matching pursuit algorithm and the basis pursuit algorithm do not always efficiently yield representations, which are sparse and physically interpretable. HRP overcomes these shortcomings of the both algorithms; a new more locally sensitive similarity measure is proposed, and it can represent signals more precisely. In this dissertation, based on HRP, a model of fault diagnosis was established with multiresolution analysis of wavelet, and the B-spline wavelet was chosen as basis function. The model was applied to the fault diagnosis of the valves in reciprocating compressors. This model reveals clearly not only the time location of the impulsive components but also their frequency range.2)For the nonstationary signal process, Wigner-Ville distribution(WVD) function has many useful properties, but there is some cross-talk. In this dissertation an enhanced version of the Wigner-Ville distribution function (EWDF) was reported. This enhanced WDF is defined as the convolution integral of the original WDFs of the input to the system and the system impulse response function. The new version of WDF developed in this work actually improves the resolutions and furthermore minimizes the oscillating and zero-crossing phenomena. The EWDF is applied to analysis the vibration and noise signals of reciprocating compressors. Vibration analysis is widely used in machinery diagnostics, but acoustic signal is rarely used. It is the bottleneck of noise signal application in fault diagnosis how to separate effectively the fault noise from its background. In this dissertation, the EWDF was used to extract the features of background, and then the fault noise was separated from its background. On the other hand, the EWDF was used to analyze energy distributions of vibration signals in time-frequency zone, which were used to determine the operation condition of the compressor. The results suggest that acoustic signals are very effective for the early detection of faults and may provide a powerful tool to indicate the various types of proceeding faults in reciprocating compressors.3) Fault prognose is considered as the reasoning how long can a machine or its component work when the critical condition has been affirmed. Modern industry is concerned about extending the lifetime of its critical processes and maintaining them only when necessary. The significant aspects of these trends include the ability to diagnose impending failures, prognose the residual useful lifetime of the process and to schedule the maintenance operations so that uptime may be maximized. Prognosis is probably the key point of the three issues leading to condition-based maintenance. This dissertation attempts to address this challenging problem with radial basis function neural networks(RBFNN). A model of fault prognosis was built by means of wavelet basis function neural network(WBFNN). Gaussian radial basis functions and Mexican hat wavelet frames were used as the scaling functions. The centers of the basis functions are calculated by use of a dyadic expansion scheme and a k-means clustering algorithm. An example is presented in which a WBFNN successfully prognoses a crack between the piston and cylinder in a defective reciprocating compressor.